{ "pdf_info": [ { "preproc_blocks": [ { "type": "title", "bbox": [ 95, 67, 500, 100 ], "lines": [ { "bbox": [ 96, 67, 499, 86 ], "spans": [ { "bbox": [ 96, 67, 499, 86 ], "score": 1.0, "content": "SpeechGPT: Empowering Large Language Models with Intrinsic", "type": "text" } ], "index": 0 }, { "bbox": [ 181, 84, 414, 100 ], "spans": [ { "bbox": [ 181, 84, 414, 100 ], "score": 1.0, "content": "Cross-Modal Conversational Abilities", "type": "text" } ], "index": 1 } ], "index": 0.5 }, { "type": "text", "bbox": [ 95, 105, 504, 190 ], "lines": [ { "bbox": [ 134, 105, 466, 123 ], "spans": [ { "bbox": [ 134, 105, 466, 123 ], "score": 1.0, "content": "Dong Zhang, Shimin Li, Xin Zhang, Jun Zhan, Pengyu Wang,", "type": "text" } ], "index": 2 }, { "bbox": [ 232, 120, 372, 135 ], "spans": [ { "bbox": [ 232, 120, 372, 135 ], "score": 1.0, "content": "Yaqian Zhou∗, Xipeng Qiu∗", "type": "text" } ], "index": 3 }, { "bbox": [ 183, 134, 413, 150 ], "spans": [ { "bbox": [ 183, 134, 413, 150 ], "score": 1.0, "content": "School of Computer Science, Fudan University", "type": "text" } ], "index": 4 }, { "bbox": [ 101, 148, 496, 164 ], "spans": [ { "bbox": [ 101, 148, 496, 164 ], "score": 1.0, "content": "Shanghai Key Laboratory of Intelligent Information Processing, Fudan University", "type": "text" } ], "index": 5 }, { "bbox": [ 94, 163, 505, 177 ], "spans": [ { "bbox": [ 94, 163, 505, 177 ], "score": 1.0, "content": "{dongzhang22,xin_zhang22,jzhan22,pywang22}@m.fudan.edu.cn", "type": "text" } ], "index": 6 }, { "bbox": [ 163, 177, 437, 191 ], "spans": [ { "bbox": [ 163, 177, 437, 191 ], "score": 1.0, "content": "{smli20,zhouyaqian,xpqiu}@fudan.edu.cn", "type": "text" } ], "index": 7 } ], "index": 4.5 }, { "type": "title", "bbox": [ 157, 212, 202, 226 ], "lines": [ { "bbox": [ 155, 211, 205, 228 ], "spans": [ { "bbox": [ 155, 211, 205, 228 ], "score": 1.0, "content": "Abstract", "type": "text" } ], "index": 8 } ], "index": 8 }, { "type": "text", "bbox": [ 87, 238, 273, 559 ], "lines": [ { "bbox": [ 86, 236, 275, 250 ], "spans": [ { "bbox": [ 86, 236, 275, 250 ], "score": 1.0, "content": "Multi-modal large language models are re-", "type": "text" } ], "index": 9 }, { "bbox": [ 86, 249, 275, 261 ], "spans": [ { "bbox": [ 86, 249, 275, 261 ], "score": 1.0, "content": "garded as a crucial step towards Artificial Gen-", "type": "text" } ], "index": 10 }, { "bbox": [ 85, 259, 275, 275 ], "spans": [ { "bbox": [ 85, 259, 275, 275 ], "score": 1.0, "content": "eral Intelligence (AGI) and have garnered sig-", "type": "text" } ], "index": 11 }, { "bbox": [ 86, 272, 275, 285 ], "spans": [ { "bbox": [ 86, 272, 275, 285 ], "score": 1.0, "content": "nificant interest with the emergence of Chat-", "type": "text" } ], "index": 12 }, { "bbox": [ 86, 284, 275, 297 ], "spans": [ { "bbox": [ 86, 284, 275, 297 ], "score": 1.0, "content": "GPT. However, current speech-language mod-", "type": "text" } ], "index": 13 }, { "bbox": [ 86, 296, 275, 309 ], "spans": [ { "bbox": [ 86, 296, 275, 309 ], "score": 1.0, "content": "els typically adopt the cascade paradigm, pre-", "type": "text" } ], "index": 14 }, { "bbox": [ 86, 308, 273, 321 ], "spans": [ { "bbox": [ 86, 308, 273, 321 ], "score": 1.0, "content": "venting inter-modal knowledge transfer. In", "type": "text" } ], "index": 15 }, { "bbox": [ 86, 320, 273, 334 ], "spans": [ { "bbox": [ 86, 320, 273, 334 ], "score": 1.0, "content": "this paper, we propose SpeechGPT, a large", "type": "text" } ], "index": 16 }, { "bbox": [ 85, 332, 273, 344 ], "spans": [ { "bbox": [ 85, 332, 273, 344 ], "score": 1.0, "content": "language model with intrinsic cross-modal", "type": "text" } ], "index": 17 }, { "bbox": [ 85, 343, 274, 358 ], "spans": [ { "bbox": [ 85, 343, 274, 358 ], "score": 1.0, "content": "conversational abilities, capable of perceiving", "type": "text" } ], "index": 18 }, { "bbox": [ 86, 356, 273, 368 ], "spans": [ { "bbox": [ 86, 356, 273, 368 ], "score": 1.0, "content": "and generating multi-modal content. With", "type": "text" } ], "index": 19 }, { "bbox": [ 86, 368, 273, 381 ], "spans": [ { "bbox": [ 86, 368, 273, 381 ], "score": 1.0, "content": "discrete speech representations, we construct", "type": "text" } ], "index": 20 }, { "bbox": [ 86, 380, 274, 393 ], "spans": [ { "bbox": [ 86, 380, 274, 393 ], "score": 1.0, "content": "SpeechInstruct, the first large-scale cross-", "type": "text" } ], "index": 21 }, { "bbox": [ 86, 392, 274, 404 ], "spans": [ { "bbox": [ 86, 392, 274, 404 ], "score": 1.0, "content": "modal speech instruction dataset. Addition-", "type": "text" } ], "index": 22 }, { "bbox": [ 86, 404, 275, 417 ], "spans": [ { "bbox": [ 86, 404, 275, 417 ], "score": 1.0, "content": "ally, we employ a three-stage training strat-", "type": "text" } ], "index": 23 }, { "bbox": [ 85, 416, 275, 430 ], "spans": [ { "bbox": [ 85, 416, 275, 430 ], "score": 1.0, "content": "egy that includes modality-adaptation pre-", "type": "text" } ], "index": 24 }, { "bbox": [ 85, 426, 275, 442 ], "spans": [ { "bbox": [ 85, 426, 275, 442 ], "score": 1.0, "content": "training, cross-modal instruction fine-tuning,", "type": "text" } ], "index": 25 }, { "bbox": [ 85, 439, 275, 453 ], "spans": [ { "bbox": [ 85, 439, 275, 453 ], "score": 1.0, "content": "and chain-of-modality instruction fine-tuning.", "type": "text" } ], "index": 26 }, { "bbox": [ 86, 452, 273, 463 ], "spans": [ { "bbox": [ 86, 452, 273, 463 ], "score": 1.0, "content": "The experimental results demonstrate that", "type": "text" } ], "index": 27 }, { "bbox": [ 86, 463, 273, 477 ], "spans": [ { "bbox": [ 86, 463, 273, 477 ], "score": 1.0, "content": "SpeechGPT has an impressive capacity to", "type": "text" } ], "index": 28 }, { "bbox": [ 86, 475, 273, 488 ], "spans": [ { "bbox": [ 86, 475, 273, 488 ], "score": 1.0, "content": "follow cross-modal human instructions and", "type": "text" } ], "index": 29 }, { "bbox": [ 86, 488, 272, 500 ], "spans": [ { "bbox": [ 86, 488, 272, 500 ], "score": 1.0, "content": "highlight the potential of handling multiple", "type": "text" } ], "index": 30 }, { "bbox": [ 86, 500, 273, 511 ], "spans": [ { "bbox": [ 86, 500, 273, 511 ], "score": 1.0, "content": "modalities with one model. Code and models", "type": "text" } ], "index": 31 }, { "bbox": [ 86, 512, 272, 524 ], "spans": [ { "bbox": [ 86, 512, 272, 524 ], "score": 1.0, "content": "are available in https://github.com/", "type": "text" } ], "index": 32 }, { "bbox": [ 87, 524, 273, 536 ], "spans": [ { "bbox": [ 87, 524, 206, 536 ], "score": 1.0, "content": "0nutation/SpeechGPT.", "type": "text" }, { "bbox": [ 220, 524, 273, 535 ], "score": 1.0, "content": "Demos are", "type": "text" } ], "index": 33 }, { "bbox": [ 86, 536, 275, 549 ], "spans": [ { "bbox": [ 86, 536, 275, 549 ], "score": 1.0, "content": "shown in https://0nutation.github.", "type": "text" } ], "index": 34 }, { "bbox": [ 86, 548, 229, 560 ], "spans": [ { "bbox": [ 86, 548, 229, 560 ], "score": 1.0, "content": "io/SpeechGPT.github.io/.", "type": "text" } ], "index": 35 } ], "index": 22 }, { "type": "title", "bbox": [ 70, 571, 154, 585 ], "lines": [ { "bbox": [ 67, 569, 155, 586 ], "spans": [ { "bbox": [ 67, 569, 155, 586 ], "score": 1.0, "content": "1 Introduction", "type": "text" } ], "index": 36 } ], "index": 36 }, { "type": "text", "bbox": [ 69, 594, 290, 754 ], "lines": [ { "bbox": [ 69, 594, 290, 606 ], "spans": [ { "bbox": [ 69, 594, 290, 606 ], "score": 1.0, "content": "Large language models (OpenAI, 2023; Touvron", "type": "text" } ], "index": 37 }, { "bbox": [ 69, 607, 291, 620 ], "spans": [ { "bbox": [ 69, 607, 291, 620 ], "score": 1.0, "content": "et al., 2023) have performed astonishingly on vari-", "type": "text" } ], "index": 38 }, { "bbox": [ 68, 621, 291, 633 ], "spans": [ { "bbox": [ 68, 621, 291, 633 ], "score": 1.0, "content": "ous natural language processing tasks. Meanwhile,", "type": "text" } ], "index": 39 }, { "bbox": [ 68, 633, 292, 647 ], "spans": [ { "bbox": [ 68, 633, 292, 647 ], "score": 1.0, "content": "multi-modal large language models, such as GPT-", "type": "text" } ], "index": 40 }, { "bbox": [ 68, 647, 290, 660 ], "spans": [ { "bbox": [ 68, 647, 290, 660 ], "score": 1.0, "content": "4, PALM-E (Driess et al., 2023), and LLaVA (Liu", "type": "text" } ], "index": 41 }, { "bbox": [ 69, 661, 290, 673 ], "spans": [ { "bbox": [ 69, 661, 290, 673 ], "score": 1.0, "content": "et al., 2023), have explored the ability of LLMs to", "type": "text" } ], "index": 42 }, { "bbox": [ 69, 674, 290, 687 ], "spans": [ { "bbox": [ 69, 674, 290, 687 ], "score": 1.0, "content": "understand multi-modal information. However, a", "type": "text" } ], "index": 43 }, { "bbox": [ 69, 689, 290, 701 ], "spans": [ { "bbox": [ 69, 689, 290, 701 ], "score": 1.0, "content": "significant gap exists between current LLMs and", "type": "text" } ], "index": 44 }, { "bbox": [ 69, 702, 292, 714 ], "spans": [ { "bbox": [ 69, 702, 292, 714 ], "score": 1.0, "content": "general artificial intelligence (AGI). First, most cur-", "type": "text" } ], "index": 45 }, { "bbox": [ 68, 716, 291, 728 ], "spans": [ { "bbox": [ 68, 716, 291, 728 ], "score": 1.0, "content": "rent LLMs can only perceive and understand multi-", "type": "text" } ], "index": 46 }, { "bbox": [ 69, 729, 290, 742 ], "spans": [ { "bbox": [ 69, 729, 290, 742 ], "score": 1.0, "content": "modal content but cannot spontaneously generate", "type": "text" } ], "index": 47 }, { "bbox": [ 69, 742, 290, 755 ], "spans": [ { "bbox": [ 69, 742, 290, 755 ], "score": 1.0, "content": "multi-modal content. Second, continuous signals", "type": "text" } ], "index": 48 } ], "index": 42.5 }, { "type": "image", "bbox": [ 306, 212, 523, 328 ], "blocks": [ { "type": "image_body", "bbox": [ 306, 212, 523, 328 ], "group_id": 0, "lines": [ { "bbox": [ 306, 212, 523, 328 ], "spans": [ { "bbox": [ 306, 212, 523, 328 ], "score": 0.97, "type": "image", "image_path": "2914927effd25a1062b348dbbc7abea459d3b55a988bd889ae878e885b349122.jpg" } ] } ], "index": 52.5, "virtual_lines": [ { "bbox": [ 306, 212, 523, 226.5 ], "spans": [], "index": 49 }, { "bbox": [ 306, 226.5, 523, 241.0 ], "spans": [], "index": 50 }, { "bbox": [ 306, 241.0, 523, 255.5 ], "spans": [], "index": 51 }, { "bbox": [ 306, 255.5, 523, 270.0 ], "spans": [], "index": 52 }, { "bbox": [ 306, 270.0, 523, 284.5 ], "spans": [], "index": 53 }, { "bbox": [ 306, 284.5, 523, 299.0 ], "spans": [], "index": 54 }, { "bbox": [ 306, 299.0, 523, 313.5 ], "spans": [], "index": 55 }, { "bbox": [ 306, 313.5, 523, 328.0 ], "spans": [], "index": 56 } ] }, { "type": "image_caption", "bbox": [ 304, 340, 524, 365 ], "group_id": 0, "lines": [ { "bbox": [ 304, 340, 525, 353 ], "spans": [ { "bbox": [ 304, 340, 525, 353 ], "score": 1.0, "content": "Figure 1: SpeechGPT’s capabilities to tackle multiple", "type": "text" } ], "index": 57 }, { "bbox": [ 304, 353, 381, 365 ], "spans": [ { "bbox": [ 304, 353, 381, 365 ], "score": 1.0, "content": "cross-modal tasks.", "type": "text" } ], "index": 58 } ], "index": 57.5 } ], "index": 55.0 }, { "type": "text", "bbox": [ 304, 372, 524, 397 ], "lines": [ { "bbox": [ 304, 372, 525, 385 ], "spans": [ { "bbox": [ 304, 372, 525, 385 ], "score": 1.0, "content": "like images and speech cannot be adapted directly", "type": "text" } ], "index": 59 }, { "bbox": [ 304, 385, 469, 399 ], "spans": [ { "bbox": [ 304, 385, 469, 399 ], "score": 1.0, "content": "to LLMs that receive discrete tokens.", "type": "text" } ], "index": 60 } ], "index": 59.5 }, { "type": "text", "bbox": [ 305, 405, 525, 619 ], "lines": [ { "bbox": [ 315, 403, 525, 417 ], "spans": [ { "bbox": [ 315, 403, 525, 417 ], "score": 1.0, "content": "The current speech-language model mainly", "type": "text" } ], "index": 61 }, { "bbox": [ 305, 417, 522, 430 ], "spans": [ { "bbox": [ 305, 417, 522, 430 ], "score": 1.0, "content": "adopts a cascading paradigm (Huang et al., 2023a)", "type": "text" } ], "index": 62 }, { "bbox": [ 304, 431, 525, 443 ], "spans": [ { "bbox": [ 304, 431, 525, 443 ], "score": 1.0, "content": "i.e., the LLM is connected with an automatic", "type": "text" } ], "index": 63 }, { "bbox": [ 304, 446, 526, 457 ], "spans": [ { "bbox": [ 304, 446, 526, 457 ], "score": 1.0, "content": "speech recognition (ASR) model or a text-to-", "type": "text" } ], "index": 64 }, { "bbox": [ 304, 458, 526, 470 ], "spans": [ { "bbox": [ 304, 458, 526, 470 ], "score": 1.0, "content": "speech (TTS) model in tandem, or the LLM is", "type": "text" } ], "index": 65 }, { "bbox": [ 304, 472, 525, 484 ], "spans": [ { "bbox": [ 304, 472, 525, 484 ], "score": 1.0, "content": "employed as a control hub, with several speech", "type": "text" } ], "index": 66 }, { "bbox": [ 303, 486, 526, 498 ], "spans": [ { "bbox": [ 303, 486, 526, 498 ], "score": 1.0, "content": "processing models (Cheng et al., 2023a,b,c) are", "type": "text" } ], "index": 67 }, { "bbox": [ 304, 498, 525, 511 ], "spans": [ { "bbox": [ 304, 498, 525, 511 ], "score": 1.0, "content": "integrated to cover multiple audio or speech", "type": "text" } ], "index": 68 }, { "bbox": [ 304, 513, 525, 525 ], "spans": [ { "bbox": [ 304, 513, 525, 525 ], "score": 1.0, "content": "tasks (Huang et al., 2023a; Shen et al., 2023). Some", "type": "text" } ], "index": 69 }, { "bbox": [ 303, 526, 527, 539 ], "spans": [ { "bbox": [ 303, 526, 527, 539 ], "score": 1.0, "content": "prior work on generative spoken language mod-", "type": "text" } ], "index": 70 }, { "bbox": [ 304, 540, 526, 552 ], "spans": [ { "bbox": [ 304, 540, 526, 552 ], "score": 1.0, "content": "els involves encoding the speech signal into a dis-", "type": "text" } ], "index": 71 }, { "bbox": [ 304, 554, 527, 565 ], "spans": [ { "bbox": [ 304, 554, 527, 565 ], "score": 1.0, "content": "crete representation (Baevski et al., 2020; Hsu et al.,", "type": "text" } ], "index": 72 }, { "bbox": [ 304, 566, 525, 579 ], "spans": [ { "bbox": [ 304, 566, 525, 579 ], "score": 1.0, "content": "2021; Zhang et al., 2023a) and modeling it with", "type": "text" } ], "index": 73 }, { "bbox": [ 304, 580, 525, 592 ], "spans": [ { "bbox": [ 304, 580, 525, 592 ], "score": 1.0, "content": "language models (Lakhotia et al., 2021; Borsos", "type": "text" } ], "index": 74 }, { "bbox": [ 304, 593, 526, 606 ], "spans": [ { "bbox": [ 304, 593, 526, 606 ], "score": 1.0, "content": "et al., 2022; Zhang et al., 2023d; Wang et al., 2023;", "type": "text" } ], "index": 75 }, { "bbox": [ 304, 607, 397, 621 ], "spans": [ { "bbox": [ 304, 607, 397, 621 ], "score": 1.0, "content": "Zhang et al., 2023c).", "type": "text" } ], "index": 76 } ], "index": 68.5 }, { "type": "text", "bbox": [ 305, 626, 525, 773 ], "lines": [ { "bbox": [ 315, 624, 526, 641 ], "spans": [ { "bbox": [ 315, 624, 526, 641 ], "score": 1.0, "content": "While capable of perceiving and generating", "type": "text" } ], "index": 77 }, { "bbox": [ 304, 639, 525, 652 ], "spans": [ { "bbox": [ 304, 639, 525, 652 ], "score": 1.0, "content": "speech, the existing cascaded methods or spoken", "type": "text" } ], "index": 78 }, { "bbox": [ 304, 653, 527, 665 ], "spans": [ { "bbox": [ 304, 653, 527, 665 ], "score": 1.0, "content": "language models still have several limitations. First,", "type": "text" } ], "index": 79 }, { "bbox": [ 304, 666, 526, 680 ], "spans": [ { "bbox": [ 304, 666, 526, 680 ], "score": 1.0, "content": "the LLM in the cascaded model only functions as", "type": "text" } ], "index": 80 }, { "bbox": [ 304, 681, 526, 693 ], "spans": [ { "bbox": [ 304, 681, 526, 693 ], "score": 1.0, "content": "a content generator. Since the representations of", "type": "text" } ], "index": 81 }, { "bbox": [ 304, 694, 527, 705 ], "spans": [ { "bbox": [ 304, 694, 527, 705 ], "score": 1.0, "content": "speech and text are not aligned, the LLM’s knowl-", "type": "text" } ], "index": 82 }, { "bbox": [ 305, 707, 527, 720 ], "spans": [ { "bbox": [ 305, 707, 527, 720 ], "score": 1.0, "content": "edge cannot be transferred to the speech modality.", "type": "text" } ], "index": 83 }, { "bbox": [ 304, 720, 527, 734 ], "spans": [ { "bbox": [ 304, 720, 527, 734 ], "score": 1.0, "content": "Second, the cascade approach (Shen et al., 2023;", "type": "text" } ], "index": 84 }, { "bbox": [ 304, 735, 527, 747 ], "spans": [ { "bbox": [ 304, 735, 527, 747 ], "score": 1.0, "content": "Huang et al., 2023a) suffers from the loss of par-", "type": "text" } ], "index": 85 }, { "bbox": [ 305, 748, 527, 761 ], "spans": [ { "bbox": [ 305, 748, 527, 761 ], "score": 1.0, "content": "alinguistic signals such as emotion and prosody.", "type": "text" } ], "index": 86 }, { "bbox": [ 303, 760, 527, 776 ], "spans": [ { "bbox": [ 303, 760, 527, 776 ], "score": 1.0, "content": "Third, existing spoken language models (Wang", "type": "text" } ], "index": 87 } ], "index": 82 } ], "page_idx": 0, "page_size": [ 595, 841 ], "discarded_blocks": [ { "type": "discarded", "bbox": [ 86, 763, 169, 773 ], "lines": [ { "bbox": [ 85, 761, 172, 775 ], "spans": [ { "bbox": [ 85, 761, 172, 775 ], "score": 1.0, "content": "∗Corresponding author", "type": "text" } ] } ] } ], "para_blocks": [ { "type": "title", "bbox": [ 95, 67, 500, 100 ], "lines": [ { "bbox": [ 96, 67, 499, 86 ], "spans": [ { "bbox": [ 96, 67, 499, 86 ], "score": 1.0, "content": "SpeechGPT: Empowering Large Language Models with Intrinsic", "type": "text" } ], "index": 0 }, { "bbox": [ 181, 84, 414, 100 ], "spans": [ { "bbox": [ 181, 84, 414, 100 ], "score": 1.0, "content": "Cross-Modal Conversational Abilities", "type": "text" } ], "index": 1 } ], "index": 0.5 }, { "type": "list", "bbox": [ 95, 105, 504, 190 ], "lines": [ { "bbox": [ 134, 105, 466, 123 ], "spans": [ { "bbox": [ 134, 105, 466, 123 ], "score": 1.0, "content": "Dong Zhang, Shimin Li, Xin Zhang, Jun Zhan, Pengyu Wang,", "type": "text" } ], "index": 2, "is_list_end_line": true }, { "bbox": [ 232, 120, 372, 135 ], "spans": [ { "bbox": [ 232, 120, 372, 135 ], "score": 1.0, "content": "Yaqian Zhou∗, Xipeng Qiu∗", "type": "text" } ], "index": 3, "is_list_end_line": true }, { "bbox": [ 183, 134, 413, 150 ], "spans": [ { "bbox": [ 183, 134, 413, 150 ], "score": 1.0, "content": "School of Computer Science, Fudan University", "type": "text" } ], "index": 4, "is_list_end_line": true }, { "bbox": [ 101, 148, 496, 164 ], "spans": [ { "bbox": [ 101, 148, 496, 164 ], "score": 1.0, "content": "Shanghai Key Laboratory of Intelligent Information Processing, Fudan University", "type": "text" } ], "index": 5, "is_list_start_line": true }, { "bbox": [ 94, 163, 505, 177 ], "spans": [ { "bbox": [ 94, 163, 505, 177 ], "score": 1.0, "content": "{dongzhang22,xin_zhang22,jzhan22,pywang22}@m.fudan.edu.cn", "type": "text" } ], "index": 6, "is_list_start_line": true }, { "bbox": [ 163, 177, 437, 191 ], "spans": [ { "bbox": [ 163, 177, 437, 191 ], "score": 1.0, "content": "{smli20,zhouyaqian,xpqiu}@fudan.edu.cn", "type": "text" } ], "index": 7, "is_list_end_line": true } ], "index": 4.5, "bbox_fs": [ 94, 105, 505, 191 ] }, { "type": "title", "bbox": [ 157, 212, 202, 226 ], "lines": [ { "bbox": [ 155, 211, 205, 228 ], "spans": [ { "bbox": [ 155, 211, 205, 228 ], "score": 1.0, "content": "Abstract", "type": "text" } ], "index": 8 } ], "index": 8 }, { "type": "text", "bbox": [ 87, 238, 273, 559 ], "lines": [ { "bbox": [ 86, 236, 275, 250 ], "spans": [ { "bbox": [ 86, 236, 275, 250 ], "score": 1.0, "content": "Multi-modal large language models are re-", "type": "text" } ], "index": 9 }, { "bbox": [ 86, 249, 275, 261 ], "spans": [ { "bbox": [ 86, 249, 275, 261 ], "score": 1.0, "content": "garded as a crucial step towards Artificial Gen-", "type": "text" } ], "index": 10 }, { "bbox": [ 85, 259, 275, 275 ], "spans": [ { "bbox": [ 85, 259, 275, 275 ], "score": 1.0, "content": "eral Intelligence (AGI) and have garnered sig-", "type": "text" } ], "index": 11 }, { "bbox": [ 86, 272, 275, 285 ], "spans": [ { "bbox": [ 86, 272, 275, 285 ], "score": 1.0, "content": "nificant interest with the emergence of Chat-", "type": "text" } ], "index": 12 }, { "bbox": [ 86, 284, 275, 297 ], "spans": [ { "bbox": [ 86, 284, 275, 297 ], "score": 1.0, "content": "GPT. However, current speech-language mod-", "type": "text" } ], "index": 13 }, { "bbox": [ 86, 296, 275, 309 ], "spans": [ { "bbox": [ 86, 296, 275, 309 ], "score": 1.0, "content": "els typically adopt the cascade paradigm, pre-", "type": "text" } ], "index": 14 }, { "bbox": [ 86, 308, 273, 321 ], "spans": [ { "bbox": [ 86, 308, 273, 321 ], "score": 1.0, "content": "venting inter-modal knowledge transfer. In", "type": "text" } ], "index": 15 }, { "bbox": [ 86, 320, 273, 334 ], "spans": [ { "bbox": [ 86, 320, 273, 334 ], "score": 1.0, "content": "this paper, we propose SpeechGPT, a large", "type": "text" } ], "index": 16 }, { "bbox": [ 85, 332, 273, 344 ], "spans": [ { "bbox": [ 85, 332, 273, 344 ], "score": 1.0, "content": "language model with intrinsic cross-modal", "type": "text" } ], "index": 17 }, { "bbox": [ 85, 343, 274, 358 ], "spans": [ { "bbox": [ 85, 343, 274, 358 ], "score": 1.0, "content": "conversational abilities, capable of perceiving", "type": "text" } ], "index": 18 }, { "bbox": [ 86, 356, 273, 368 ], "spans": [ { "bbox": [ 86, 356, 273, 368 ], "score": 1.0, "content": "and generating multi-modal content. With", "type": "text" } ], "index": 19 }, { "bbox": [ 86, 368, 273, 381 ], "spans": [ { "bbox": [ 86, 368, 273, 381 ], "score": 1.0, "content": "discrete speech representations, we construct", "type": "text" } ], "index": 20 }, { "bbox": [ 86, 380, 274, 393 ], "spans": [ { "bbox": [ 86, 380, 274, 393 ], "score": 1.0, "content": "SpeechInstruct, the first large-scale cross-", "type": "text" } ], "index": 21 }, { "bbox": [ 86, 392, 274, 404 ], "spans": [ { "bbox": [ 86, 392, 274, 404 ], "score": 1.0, "content": "modal speech instruction dataset. Addition-", "type": "text" } ], "index": 22 }, { "bbox": [ 86, 404, 275, 417 ], "spans": [ { "bbox": [ 86, 404, 275, 417 ], "score": 1.0, "content": "ally, we employ a three-stage training strat-", "type": "text" } ], "index": 23 }, { "bbox": [ 85, 416, 275, 430 ], "spans": [ { "bbox": [ 85, 416, 275, 430 ], "score": 1.0, "content": "egy that includes modality-adaptation pre-", "type": "text" } ], "index": 24 }, { "bbox": [ 85, 426, 275, 442 ], "spans": [ { "bbox": [ 85, 426, 275, 442 ], "score": 1.0, "content": "training, cross-modal instruction fine-tuning,", "type": "text" } ], "index": 25 }, { "bbox": [ 85, 439, 275, 453 ], "spans": [ { "bbox": [ 85, 439, 275, 453 ], "score": 1.0, "content": "and chain-of-modality instruction fine-tuning.", "type": "text" } ], "index": 26 }, { "bbox": [ 86, 452, 273, 463 ], "spans": [ { "bbox": [ 86, 452, 273, 463 ], "score": 1.0, "content": "The experimental results demonstrate that", "type": "text" } ], "index": 27 }, { "bbox": [ 86, 463, 273, 477 ], "spans": [ { "bbox": [ 86, 463, 273, 477 ], "score": 1.0, "content": "SpeechGPT has an impressive capacity to", "type": "text" } ], "index": 28 }, { "bbox": [ 86, 475, 273, 488 ], "spans": [ { "bbox": [ 86, 475, 273, 488 ], "score": 1.0, "content": "follow cross-modal human instructions and", "type": "text" } ], "index": 29 }, { "bbox": [ 86, 488, 272, 500 ], "spans": [ { "bbox": [ 86, 488, 272, 500 ], "score": 1.0, "content": "highlight the potential of handling multiple", "type": "text" } ], "index": 30 }, { "bbox": [ 86, 500, 273, 511 ], "spans": [ { "bbox": [ 86, 500, 273, 511 ], "score": 1.0, "content": "modalities with one model. Code and models", "type": "text" } ], "index": 31 }, { "bbox": [ 86, 512, 272, 524 ], "spans": [ { "bbox": [ 86, 512, 272, 524 ], "score": 1.0, "content": "are available in https://github.com/", "type": "text" } ], "index": 32 }, { "bbox": [ 87, 524, 273, 536 ], "spans": [ { "bbox": [ 87, 524, 206, 536 ], "score": 1.0, "content": "0nutation/SpeechGPT.", "type": "text" }, { "bbox": [ 220, 524, 273, 535 ], "score": 1.0, "content": "Demos are", "type": "text" } ], "index": 33 }, { "bbox": [ 86, 536, 275, 549 ], "spans": [ { "bbox": [ 86, 536, 275, 549 ], "score": 1.0, "content": "shown in https://0nutation.github.", "type": "text" } ], "index": 34 }, { "bbox": [ 86, 548, 229, 560 ], "spans": [ { "bbox": [ 86, 548, 229, 560 ], "score": 1.0, "content": "io/SpeechGPT.github.io/.", "type": "text" } ], "index": 35 } ], "index": 22, "bbox_fs": [ 85, 236, 275, 560 ] }, { "type": "title", "bbox": [ 70, 571, 154, 585 ], "lines": [ { "bbox": [ 67, 569, 155, 586 ], "spans": [ { "bbox": [ 67, 569, 155, 586 ], "score": 1.0, "content": "1 Introduction", "type": "text" } ], "index": 36 } ], "index": 36 }, { "type": "text", "bbox": [ 69, 594, 290, 754 ], "lines": [ { "bbox": [ 69, 594, 290, 606 ], "spans": [ { "bbox": [ 69, 594, 290, 606 ], "score": 1.0, "content": "Large language models (OpenAI, 2023; Touvron", "type": "text" } ], "index": 37 }, { "bbox": [ 69, 607, 291, 620 ], "spans": [ { "bbox": [ 69, 607, 291, 620 ], "score": 1.0, "content": "et al., 2023) have performed astonishingly on vari-", "type": "text" } ], "index": 38 }, { "bbox": [ 68, 621, 291, 633 ], "spans": [ { "bbox": [ 68, 621, 291, 633 ], "score": 1.0, "content": "ous natural language processing tasks. Meanwhile,", "type": "text" } ], "index": 39 }, { "bbox": [ 68, 633, 292, 647 ], "spans": [ { "bbox": [ 68, 633, 292, 647 ], "score": 1.0, "content": "multi-modal large language models, such as GPT-", "type": "text" } ], "index": 40 }, { "bbox": [ 68, 647, 290, 660 ], "spans": [ { "bbox": [ 68, 647, 290, 660 ], "score": 1.0, "content": "4, PALM-E (Driess et al., 2023), and LLaVA (Liu", "type": "text" } ], "index": 41 }, { "bbox": [ 69, 661, 290, 673 ], "spans": [ { "bbox": [ 69, 661, 290, 673 ], "score": 1.0, "content": "et al., 2023), have explored the ability of LLMs to", "type": "text" } ], "index": 42 }, { "bbox": [ 69, 674, 290, 687 ], "spans": [ { "bbox": [ 69, 674, 290, 687 ], "score": 1.0, "content": "understand multi-modal information. However, a", "type": "text" } ], "index": 43 }, { "bbox": [ 69, 689, 290, 701 ], "spans": [ { "bbox": [ 69, 689, 290, 701 ], "score": 1.0, "content": "significant gap exists between current LLMs and", "type": "text" } ], "index": 44 }, { "bbox": [ 69, 702, 292, 714 ], "spans": [ { "bbox": [ 69, 702, 292, 714 ], "score": 1.0, "content": "general artificial intelligence (AGI). First, most cur-", "type": "text" } ], "index": 45 }, { "bbox": [ 68, 716, 291, 728 ], "spans": [ { "bbox": [ 68, 716, 291, 728 ], "score": 1.0, "content": "rent LLMs can only perceive and understand multi-", "type": "text" } ], "index": 46 }, { "bbox": [ 69, 729, 290, 742 ], "spans": [ { "bbox": [ 69, 729, 290, 742 ], "score": 1.0, "content": "modal content but cannot spontaneously generate", "type": "text" } ], "index": 47 }, { "bbox": [ 69, 742, 290, 755 ], "spans": [ { "bbox": [ 69, 742, 290, 755 ], "score": 1.0, "content": "multi-modal content. Second, continuous signals", "type": "text" } ], "index": 48 }, { "bbox": [ 304, 372, 525, 385 ], "spans": [ { "bbox": [ 304, 372, 525, 385 ], "score": 1.0, "content": "like images and speech cannot be adapted directly", "type": "text" } ], "index": 59 }, { "bbox": [ 304, 385, 469, 399 ], "spans": [ { "bbox": [ 304, 385, 469, 399 ], "score": 1.0, "content": "to LLMs that receive discrete tokens.", "type": "text" } ], "index": 60 } ], "index": 42.5, "bbox_fs": [ 68, 594, 292, 755 ] }, { "type": "image", "bbox": [ 306, 212, 523, 328 ], "blocks": [ { "type": "image_body", "bbox": [ 306, 212, 523, 328 ], "group_id": 0, "lines": [ { "bbox": [ 306, 212, 523, 328 ], "spans": [ { "bbox": [ 306, 212, 523, 328 ], "score": 0.97, "type": "image", "image_path": "2914927effd25a1062b348dbbc7abea459d3b55a988bd889ae878e885b349122.jpg" } ] } ], "index": 52.5, "virtual_lines": [ { "bbox": [ 306, 212, 523, 226.5 ], "spans": [], "index": 49 }, { "bbox": [ 306, 226.5, 523, 241.0 ], "spans": [], "index": 50 }, { "bbox": [ 306, 241.0, 523, 255.5 ], "spans": [], "index": 51 }, { "bbox": [ 306, 255.5, 523, 270.0 ], "spans": [], "index": 52 }, { "bbox": [ 306, 270.0, 523, 284.5 ], "spans": [], "index": 53 }, { "bbox": [ 306, 284.5, 523, 299.0 ], "spans": [], "index": 54 }, { "bbox": [ 306, 299.0, 523, 313.5 ], "spans": [], "index": 55 }, { "bbox": [ 306, 313.5, 523, 328.0 ], "spans": [], "index": 56 } ] }, { "type": "image_caption", "bbox": [ 304, 340, 524, 365 ], "group_id": 0, "lines": [ { "bbox": [ 304, 340, 525, 353 ], "spans": [ { "bbox": [ 304, 340, 525, 353 ], "score": 1.0, "content": "Figure 1: SpeechGPT’s capabilities to tackle multiple", "type": "text" } ], "index": 57 }, { "bbox": [ 304, 353, 381, 365 ], "spans": [ { "bbox": [ 304, 353, 381, 365 ], "score": 1.0, "content": "cross-modal tasks.", "type": "text" } ], "index": 58 } ], "index": 57.5 } ], "index": 55.0 }, { "type": "text", "bbox": [ 304, 372, 524, 397 ], "lines": [], "index": 59.5, "bbox_fs": [ 304, 372, 525, 399 ], "lines_deleted": true }, { "type": "text", "bbox": [ 305, 405, 525, 619 ], "lines": [ { "bbox": [ 315, 403, 525, 417 ], "spans": [ { "bbox": [ 315, 403, 525, 417 ], "score": 1.0, "content": "The current speech-language model mainly", "type": "text" } ], "index": 61 }, { "bbox": [ 305, 417, 522, 430 ], "spans": [ { "bbox": [ 305, 417, 522, 430 ], "score": 1.0, "content": "adopts a cascading paradigm (Huang et al., 2023a)", "type": "text" } ], "index": 62 }, { "bbox": [ 304, 431, 525, 443 ], "spans": [ { "bbox": [ 304, 431, 525, 443 ], "score": 1.0, "content": "i.e., the LLM is connected with an automatic", "type": "text" } ], "index": 63 }, { "bbox": [ 304, 446, 526, 457 ], "spans": [ { "bbox": [ 304, 446, 526, 457 ], "score": 1.0, "content": "speech recognition (ASR) model or a text-to-", "type": "text" } ], "index": 64 }, { "bbox": [ 304, 458, 526, 470 ], "spans": [ { "bbox": [ 304, 458, 526, 470 ], "score": 1.0, "content": "speech (TTS) model in tandem, or the LLM is", "type": "text" } ], "index": 65 }, { "bbox": [ 304, 472, 525, 484 ], "spans": [ { "bbox": [ 304, 472, 525, 484 ], "score": 1.0, "content": "employed as a control hub, with several speech", "type": "text" } ], "index": 66 }, { "bbox": [ 303, 486, 526, 498 ], "spans": [ { "bbox": [ 303, 486, 526, 498 ], "score": 1.0, "content": "processing models (Cheng et al., 2023a,b,c) are", "type": "text" } ], "index": 67 }, { "bbox": [ 304, 498, 525, 511 ], "spans": [ { "bbox": [ 304, 498, 525, 511 ], "score": 1.0, "content": "integrated to cover multiple audio or speech", "type": "text" } ], "index": 68 }, { "bbox": [ 304, 513, 525, 525 ], "spans": [ { "bbox": [ 304, 513, 525, 525 ], "score": 1.0, "content": "tasks (Huang et al., 2023a; Shen et al., 2023). Some", "type": "text" } ], "index": 69 }, { "bbox": [ 303, 526, 527, 539 ], "spans": [ { "bbox": [ 303, 526, 527, 539 ], "score": 1.0, "content": "prior work on generative spoken language mod-", "type": "text" } ], "index": 70 }, { "bbox": [ 304, 540, 526, 552 ], "spans": [ { "bbox": [ 304, 540, 526, 552 ], "score": 1.0, "content": "els involves encoding the speech signal into a dis-", "type": "text" } ], "index": 71 }, { "bbox": [ 304, 554, 527, 565 ], "spans": [ { "bbox": [ 304, 554, 527, 565 ], "score": 1.0, "content": "crete representation (Baevski et al., 2020; Hsu et al.,", "type": "text" } ], "index": 72 }, { "bbox": [ 304, 566, 525, 579 ], "spans": [ { "bbox": [ 304, 566, 525, 579 ], "score": 1.0, "content": "2021; Zhang et al., 2023a) and modeling it with", "type": "text" } ], "index": 73 }, { "bbox": [ 304, 580, 525, 592 ], "spans": [ { "bbox": [ 304, 580, 525, 592 ], "score": 1.0, "content": "language models (Lakhotia et al., 2021; Borsos", "type": "text" } ], "index": 74 }, { "bbox": [ 304, 593, 526, 606 ], "spans": [ { "bbox": [ 304, 593, 526, 606 ], "score": 1.0, "content": "et al., 2022; Zhang et al., 2023d; Wang et al., 2023;", "type": "text" } ], "index": 75 }, { "bbox": [ 304, 607, 397, 621 ], "spans": [ { "bbox": [ 304, 607, 397, 621 ], "score": 1.0, "content": "Zhang et al., 2023c).", "type": "text" } ], "index": 76 } ], "index": 68.5, "bbox_fs": [ 303, 403, 527, 621 ] }, { "type": "text", "bbox": [ 305, 626, 525, 773 ], "lines": [ { "bbox": [ 315, 624, 526, 641 ], "spans": [ { "bbox": [ 315, 624, 526, 641 ], "score": 1.0, "content": "While capable of perceiving and generating", "type": "text" } ], "index": 77 }, { "bbox": [ 304, 639, 525, 652 ], "spans": [ { "bbox": [ 304, 639, 525, 652 ], "score": 1.0, "content": "speech, the existing cascaded methods or spoken", "type": "text" } ], "index": 78 }, { "bbox": [ 304, 653, 527, 665 ], "spans": [ { "bbox": [ 304, 653, 527, 665 ], "score": 1.0, "content": "language models still have several limitations. First,", "type": "text" } ], "index": 79 }, { "bbox": [ 304, 666, 526, 680 ], "spans": [ { "bbox": [ 304, 666, 526, 680 ], "score": 1.0, "content": "the LLM in the cascaded model only functions as", "type": "text" } ], "index": 80 }, { "bbox": [ 304, 681, 526, 693 ], "spans": [ { "bbox": [ 304, 681, 526, 693 ], "score": 1.0, "content": "a content generator. Since the representations of", "type": "text" } ], "index": 81 }, { "bbox": [ 304, 694, 527, 705 ], "spans": [ { "bbox": [ 304, 694, 527, 705 ], "score": 1.0, "content": "speech and text are not aligned, the LLM’s knowl-", "type": "text" } ], "index": 82 }, { "bbox": [ 305, 707, 527, 720 ], "spans": [ { "bbox": [ 305, 707, 527, 720 ], "score": 1.0, "content": "edge cannot be transferred to the speech modality.", "type": "text" } ], "index": 83 }, { "bbox": [ 304, 720, 527, 734 ], "spans": [ { "bbox": [ 304, 720, 527, 734 ], "score": 1.0, "content": "Second, the cascade approach (Shen et al., 2023;", "type": "text" } ], "index": 84 }, { "bbox": [ 304, 735, 527, 747 ], "spans": [ { "bbox": [ 304, 735, 527, 747 ], "score": 1.0, "content": "Huang et al., 2023a) suffers from the loss of par-", "type": "text" } ], "index": 85 }, { "bbox": [ 305, 748, 527, 761 ], "spans": [ { "bbox": [ 305, 748, 527, 761 ], "score": 1.0, "content": "alinguistic signals such as emotion and prosody.", "type": "text" } ], "index": 86 }, { "bbox": [ 303, 760, 527, 776 ], "spans": [ { "bbox": [ 303, 760, 527, 776 ], "score": 1.0, "content": "Third, existing spoken language models (Wang", "type": "text" } ], "index": 87 }, { "bbox": [ 68, 70, 290, 85 ], "spans": [ { "bbox": [ 68, 70, 290, 85 ], "score": 1.0, "content": "et al., 2023; Zhang et al., 2023d) only synthesize", "type": "text", "cross_page": true } ], "index": 0 }, { "bbox": [ 69, 86, 292, 98 ], "spans": [ { "bbox": [ 69, 86, 292, 98 ], "score": 1.0, "content": "speech but fail to comprehend its semantic infor-", "type": "text", "cross_page": true } ], "index": 1 }, { "bbox": [ 69, 100, 292, 112 ], "spans": [ { "bbox": [ 69, 100, 292, 112 ], "score": 1.0, "content": "mation, preventing them from achieving true cross-", "type": "text", "cross_page": true } ], "index": 2 }, { "bbox": [ 69, 112, 218, 126 ], "spans": [ { "bbox": [ 69, 112, 218, 126 ], "score": 1.0, "content": "modal perception and generation.", "type": "text", "cross_page": true } ], "index": 3 } ], "index": 82, "bbox_fs": [ 303, 624, 527, 776 ] } ] }, { "preproc_blocks": [ { "type": "text", "bbox": [ 69, 72, 290, 125 ], "lines": [ { "bbox": [ 68, 70, 290, 85 ], "spans": [ { "bbox": [ 68, 70, 290, 85 ], "score": 1.0, "content": "et al., 2023; Zhang et al., 2023d) only synthesize", "type": "text" } ], "index": 0 }, { "bbox": [ 69, 86, 292, 98 ], "spans": [ { "bbox": [ 69, 86, 292, 98 ], "score": 1.0, "content": "speech but fail to comprehend its semantic infor-", "type": "text" } ], "index": 1 }, { "bbox": [ 69, 100, 292, 112 ], "spans": [ { "bbox": [ 69, 100, 292, 112 ], "score": 1.0, "content": "mation, preventing them from achieving true cross-", "type": "text" } ], "index": 2 }, { "bbox": [ 69, 112, 218, 126 ], "spans": [ { "bbox": [ 69, 112, 218, 126 ], "score": 1.0, "content": "modal perception and generation.", "type": "text" } ], "index": 3 } ], "index": 1.5 }, { "type": "text", "bbox": [ 70, 128, 290, 262 ], "lines": [ { "bbox": [ 79, 127, 290, 142 ], "spans": [ { "bbox": [ 79, 127, 290, 142 ], "score": 1.0, "content": "In this paper, we propose SpeechGPT, a large", "type": "text" } ], "index": 4 }, { "bbox": [ 68, 141, 292, 155 ], "spans": [ { "bbox": [ 68, 141, 292, 155 ], "score": 1.0, "content": "language model with intrinsic cross-modal conver-", "type": "text" } ], "index": 5 }, { "bbox": [ 69, 155, 291, 168 ], "spans": [ { "bbox": [ 69, 155, 291, 168 ], "score": 1.0, "content": "sational abilities, capable of perceiving and gen-", "type": "text" } ], "index": 6 }, { "bbox": [ 69, 169, 290, 182 ], "spans": [ { "bbox": [ 69, 169, 290, 182 ], "score": 1.0, "content": "erating multi-modal content. We perform speech", "type": "text" } ], "index": 7 }, { "bbox": [ 69, 182, 290, 195 ], "spans": [ { "bbox": [ 69, 182, 290, 195 ], "score": 1.0, "content": "discretization with a self-supervised trained speech", "type": "text" } ], "index": 8 }, { "bbox": [ 69, 196, 290, 208 ], "spans": [ { "bbox": [ 69, 196, 290, 208 ], "score": 1.0, "content": "model to unify the modality between speech and", "type": "text" } ], "index": 9 }, { "bbox": [ 69, 209, 290, 222 ], "spans": [ { "bbox": [ 69, 209, 290, 222 ], "score": 1.0, "content": "text. The discrete speech tokens are then expanded", "type": "text" } ], "index": 10 }, { "bbox": [ 68, 222, 291, 236 ], "spans": [ { "bbox": [ 68, 222, 291, 236 ], "score": 1.0, "content": "into the vocabulary of the LLM, thus endowing the", "type": "text" } ], "index": 11 }, { "bbox": [ 69, 236, 290, 249 ], "spans": [ { "bbox": [ 69, 236, 290, 249 ], "score": 1.0, "content": "model with an inherent competence to perceive and", "type": "text" } ], "index": 12 }, { "bbox": [ 69, 251, 159, 263 ], "spans": [ { "bbox": [ 69, 251, 159, 263 ], "score": 1.0, "content": "generate the speech.", "type": "text" } ], "index": 13 } ], "index": 8.5 }, { "type": "text", "bbox": [ 69, 265, 289, 467 ], "lines": [ { "bbox": [ 79, 263, 291, 279 ], "spans": [ { "bbox": [ 79, 263, 291, 279 ], "score": 1.0, "content": "To provide the model with the capacity to", "type": "text" } ], "index": 14 }, { "bbox": [ 69, 279, 290, 291 ], "spans": [ { "bbox": [ 69, 279, 290, 291 ], "score": 1.0, "content": "handle multi-modal instructions, we build the", "type": "text" } ], "index": 15 }, { "bbox": [ 69, 293, 290, 306 ], "spans": [ { "bbox": [ 69, 293, 290, 306 ], "score": 1.0, "content": "first speech-text cross-modal instruction-following", "type": "text" } ], "index": 16 }, { "bbox": [ 69, 306, 290, 318 ], "spans": [ { "bbox": [ 69, 306, 290, 318 ], "score": 1.0, "content": "dataset SpeechInstruct. Specifically, we discretize", "type": "text" } ], "index": 17 }, { "bbox": [ 69, 320, 290, 332 ], "spans": [ { "bbox": [ 69, 320, 290, 332 ], "score": 1.0, "content": "the speech to discrete units (Hsu et al., 2021) and", "type": "text" } ], "index": 18 }, { "bbox": [ 69, 333, 290, 346 ], "spans": [ { "bbox": [ 69, 333, 290, 346 ], "score": 1.0, "content": "construct the cross-modal unit-text pair based on", "type": "text" } ], "index": 19 }, { "bbox": [ 69, 347, 291, 359 ], "spans": [ { "bbox": [ 69, 347, 291, 359 ], "score": 1.0, "content": "the existing ASR dataset. Meanwhile, we construct", "type": "text" } ], "index": 20 }, { "bbox": [ 69, 360, 291, 372 ], "spans": [ { "bbox": [ 69, 360, 291, 372 ], "score": 1.0, "content": "hundreds of instructions for diverse tasks with GPT-", "type": "text" } ], "index": 21 }, { "bbox": [ 68, 374, 290, 385 ], "spans": [ { "bbox": [ 68, 374, 290, 385 ], "score": 1.0, "content": "4 to simulate actual user instructions as illustrated", "type": "text" } ], "index": 22 }, { "bbox": [ 68, 387, 290, 399 ], "spans": [ { "bbox": [ 68, 387, 290, 399 ], "score": 1.0, "content": "in Appendix B. In addition, to further enhance the", "type": "text" } ], "index": 23 }, { "bbox": [ 68, 400, 290, 414 ], "spans": [ { "bbox": [ 68, 400, 290, 414 ], "score": 1.0, "content": "model’s cross-modal capability, we designed the", "type": "text" } ], "index": 24 }, { "bbox": [ 69, 414, 290, 426 ], "spans": [ { "bbox": [ 69, 414, 290, 426 ], "score": 1.0, "content": "Chain-of-Modality instruction data, i.e., the model", "type": "text" } ], "index": 25 }, { "bbox": [ 68, 428, 289, 439 ], "spans": [ { "bbox": [ 68, 428, 289, 439 ], "score": 1.0, "content": "receives the speech command, thinks about the", "type": "text" } ], "index": 26 }, { "bbox": [ 68, 442, 290, 455 ], "spans": [ { "bbox": [ 68, 442, 290, 455 ], "score": 1.0, "content": "process in text, and then outputs the response in", "type": "text" } ], "index": 27 }, { "bbox": [ 68, 456, 104, 467 ], "spans": [ { "bbox": [ 68, 456, 104, 467 ], "score": 1.0, "content": "speech.", "type": "text" } ], "index": 28 } ], "index": 21 }, { "type": "text", "bbox": [ 70, 470, 290, 618 ], "lines": [ { "bbox": [ 79, 470, 290, 483 ], "spans": [ { "bbox": [ 79, 470, 290, 483 ], "score": 1.0, "content": "For better cross-modal transfer and efficient", "type": "text" } ], "index": 29 }, { "bbox": [ 68, 483, 291, 497 ], "spans": [ { "bbox": [ 68, 483, 291, 497 ], "score": 1.0, "content": "training, SpeechGPT undergoes a three-stage train-", "type": "text" } ], "index": 30 }, { "bbox": [ 69, 498, 291, 510 ], "spans": [ { "bbox": [ 69, 498, 291, 510 ], "score": 1.0, "content": "ing process: modality-adaptation pre-training,", "type": "text" } ], "index": 31 }, { "bbox": [ 68, 511, 291, 523 ], "spans": [ { "bbox": [ 68, 511, 291, 523 ], "score": 1.0, "content": "cross-modal instruction fine-tuning, and chain-of-", "type": "text" } ], "index": 32 }, { "bbox": [ 69, 525, 291, 537 ], "spans": [ { "bbox": [ 69, 525, 291, 537 ], "score": 1.0, "content": "modality instruction fine-tuning. The first stage en-", "type": "text" } ], "index": 33 }, { "bbox": [ 69, 538, 290, 551 ], "spans": [ { "bbox": [ 69, 538, 290, 551 ], "score": 1.0, "content": "ables speech comprehension for SpeechGPT with", "type": "text" } ], "index": 34 }, { "bbox": [ 68, 551, 292, 565 ], "spans": [ { "bbox": [ 68, 551, 292, 565 ], "score": 1.0, "content": "the discrete speech unit continuation task. The sec-", "type": "text" } ], "index": 35 }, { "bbox": [ 69, 565, 290, 578 ], "spans": [ { "bbox": [ 69, 565, 290, 578 ], "score": 1.0, "content": "ond stage employs the SpeechInstruct to improve", "type": "text" } ], "index": 36 }, { "bbox": [ 69, 579, 290, 591 ], "spans": [ { "bbox": [ 69, 579, 290, 591 ], "score": 1.0, "content": "the model’s cross-modal capabilities. The third", "type": "text" } ], "index": 37 }, { "bbox": [ 69, 593, 291, 605 ], "spans": [ { "bbox": [ 69, 593, 291, 605 ], "score": 1.0, "content": "stage utilizes parameter-efficient LoRA (Hu et al.,", "type": "text" } ], "index": 38 }, { "bbox": [ 69, 605, 286, 619 ], "spans": [ { "bbox": [ 69, 605, 286, 619 ], "score": 1.0, "content": "2021) fine-tuning for further modality alignment.", "type": "text" } ], "index": 39 } ], "index": 34 }, { "type": "text", "bbox": [ 70, 621, 290, 728 ], "lines": [ { "bbox": [ 79, 621, 291, 634 ], "spans": [ { "bbox": [ 79, 621, 291, 634 ], "score": 1.0, "content": "To evaluate the effectiveness of SpeechGPT,", "type": "text" } ], "index": 40 }, { "bbox": [ 69, 635, 290, 647 ], "spans": [ { "bbox": [ 69, 635, 290, 647 ], "score": 1.0, "content": "we conduct a wide range of human evaluations", "type": "text" } ], "index": 41 }, { "bbox": [ 69, 648, 291, 661 ], "spans": [ { "bbox": [ 69, 648, 291, 661 ], "score": 1.0, "content": "and case analyses to estimate the performance of", "type": "text" } ], "index": 42 }, { "bbox": [ 69, 662, 292, 675 ], "spans": [ { "bbox": [ 69, 662, 292, 675 ], "score": 1.0, "content": "SpeechGPT on textual tasks, speech-text cross-", "type": "text" } ], "index": 43 }, { "bbox": [ 69, 676, 291, 688 ], "spans": [ { "bbox": [ 69, 676, 291, 688 ], "score": 1.0, "content": "modal tasks, and spoken dialogue tasks. The re-", "type": "text" } ], "index": 44 }, { "bbox": [ 69, 688, 290, 703 ], "spans": [ { "bbox": [ 69, 688, 290, 703 ], "score": 1.0, "content": "sults demonstrate that SpeechGPT exhibits a strong", "type": "text" } ], "index": 45 }, { "bbox": [ 69, 703, 290, 715 ], "spans": [ { "bbox": [ 69, 703, 290, 715 ], "score": 1.0, "content": "ability for unimodal and cross-modal instruction", "type": "text" } ], "index": 46 }, { "bbox": [ 69, 715, 141, 729 ], "spans": [ { "bbox": [ 69, 715, 141, 729 ], "score": 1.0, "content": "following tasks.", "type": "text" } ], "index": 47 } ], "index": 43.5 }, { "type": "text", "bbox": [ 79, 731, 258, 744 ], "lines": [ { "bbox": [ 79, 729, 259, 746 ], "spans": [ { "bbox": [ 79, 729, 259, 746 ], "score": 1.0, "content": "Our contributions include the following:", "type": "text" } ], "index": 48 } ], "index": 48 }, { "type": "text", "bbox": [ 71, 747, 290, 773 ], "lines": [ { "bbox": [ 69, 745, 290, 762 ], "spans": [ { "bbox": [ 69, 745, 290, 762 ], "score": 1.0, "content": "• We build the first multi-modal large language", "type": "text" } ], "index": 49 }, { "bbox": [ 79, 761, 291, 774 ], "spans": [ { "bbox": [ 79, 761, 291, 774 ], "score": 1.0, "content": "model that can perceive and generate multi-", "type": "text" } ], "index": 50 } ], "index": 49.5 }, { "type": "text", "bbox": [ 306, 72, 526, 208 ], "lines": [ { "bbox": [ 314, 72, 386, 84 ], "spans": [ { "bbox": [ 314, 72, 386, 84 ], "score": 1.0, "content": "modal contents.", "type": "text" } ], "index": 51 }, { "bbox": [ 305, 85, 526, 100 ], "spans": [ { "bbox": [ 305, 85, 526, 100 ], "score": 1.0, "content": "• We construct and release SpeechInstruct, the first", "type": "text" } ], "index": 52 }, { "bbox": [ 314, 100, 527, 113 ], "spans": [ { "bbox": [ 314, 100, 527, 113 ], "score": 1.0, "content": "large-scale speech-text cross-modal instruction-", "type": "text" } ], "index": 53 }, { "bbox": [ 314, 113, 396, 127 ], "spans": [ { "bbox": [ 314, 113, 396, 127 ], "score": 1.0, "content": "following dataset.", "type": "text" } ], "index": 54 }, { "bbox": [ 305, 126, 525, 142 ], "spans": [ { "bbox": [ 305, 126, 525, 142 ], "score": 1.0, "content": "• We build the first spoken dialogue LLM with", "type": "text" } ], "index": 55 }, { "bbox": [ 315, 142, 525, 154 ], "spans": [ { "bbox": [ 315, 142, 525, 154 ], "score": 1.0, "content": "strong human instruction following ability and", "type": "text" } ], "index": 56 }, { "bbox": [ 313, 154, 421, 169 ], "spans": [ { "bbox": [ 313, 154, 421, 169 ], "score": 1.0, "content": "spoken dialogue ability.", "type": "text" } ], "index": 57 }, { "bbox": [ 305, 169, 526, 183 ], "spans": [ { "bbox": [ 305, 169, 526, 183 ], "score": 1.0, "content": "• We show great potential to incorporate other", "type": "text" } ], "index": 58 }, { "bbox": [ 313, 182, 527, 197 ], "spans": [ { "bbox": [ 313, 182, 527, 197 ], "score": 1.0, "content": "modalities into LLMs through discrete represen-", "type": "text" } ], "index": 59 }, { "bbox": [ 314, 197, 349, 209 ], "spans": [ { "bbox": [ 314, 197, 349, 209 ], "score": 1.0, "content": "tations.", "type": "text" } ], "index": 60 } ], "index": 55.5 }, { "type": "title", "bbox": [ 305, 222, 395, 235 ], "lines": [ { "bbox": [ 303, 221, 397, 237 ], "spans": [ { "bbox": [ 303, 221, 397, 237 ], "score": 1.0, "content": "2 Related Work", "type": "text" } ], "index": 61 } ], "index": 61 }, { "type": "text", "bbox": [ 304, 246, 525, 569 ], "lines": [ { "bbox": [ 304, 245, 525, 258 ], "spans": [ { "bbox": [ 304, 245, 525, 258 ], "score": 1.0, "content": "Multi-modal Large Language Model Current", "type": "text" } ], "index": 62 }, { "bbox": [ 304, 259, 527, 272 ], "spans": [ { "bbox": [ 304, 259, 527, 272 ], "score": 1.0, "content": "multi-modal LLMs predominantly focus on the vi-", "type": "text" } ], "index": 63 }, { "bbox": [ 304, 272, 525, 286 ], "spans": [ { "bbox": [ 304, 272, 525, 286 ], "score": 1.0, "content": "sual domain, feeding continuous representations", "type": "text" } ], "index": 64 }, { "bbox": [ 304, 286, 525, 299 ], "spans": [ { "bbox": [ 304, 286, 525, 299 ], "score": 1.0, "content": "obtained from pre-trained visual encoders into", "type": "text" } ], "index": 65 }, { "bbox": [ 303, 299, 528, 313 ], "spans": [ { "bbox": [ 303, 299, 528, 313 ], "score": 1.0, "content": "LLMs, facilitating full-parameter or parameter-", "type": "text" } ], "index": 66 }, { "bbox": [ 304, 312, 527, 327 ], "spans": [ { "bbox": [ 304, 312, 527, 327 ], "score": 1.0, "content": "efficient training on visual-language data (Ope-", "type": "text" } ], "index": 67 }, { "bbox": [ 303, 326, 527, 340 ], "spans": [ { "bbox": [ 303, 326, 527, 340 ], "score": 1.0, "content": "nAI, 2023; Huang et al., 2023b; Zhang et al.,", "type": "text" } ], "index": 68 }, { "bbox": [ 304, 339, 525, 353 ], "spans": [ { "bbox": [ 304, 339, 525, 353 ], "score": 1.0, "content": "2023b). Palm-E (Driess et al., 2023) integrates", "type": "text" } ], "index": 69 }, { "bbox": [ 304, 354, 525, 366 ], "spans": [ { "bbox": [ 304, 354, 525, 366 ], "score": 1.0, "content": "the 540B PaLM (Chowdhery et al., 2022) and 22B", "type": "text" } ], "index": 70 }, { "bbox": [ 304, 367, 525, 380 ], "spans": [ { "bbox": [ 304, 367, 525, 380 ], "score": 1.0, "content": "Vision Transformer (Dosovitskiy et al., 2021) into", "type": "text" } ], "index": 71 }, { "bbox": [ 304, 381, 525, 394 ], "spans": [ { "bbox": [ 304, 381, 525, 394 ], "score": 1.0, "content": "the largest vision-language model. LLaVA (Liu", "type": "text" } ], "index": 72 }, { "bbox": [ 303, 394, 525, 407 ], "spans": [ { "bbox": [ 303, 394, 525, 407 ], "score": 1.0, "content": "et al., 2023) leverages pre-trained CLIP (Radford", "type": "text" } ], "index": 73 }, { "bbox": [ 304, 408, 525, 420 ], "spans": [ { "bbox": [ 304, 408, 525, 420 ], "score": 1.0, "content": "et al., 2021) visual encoder and LLaMA (Touvron", "type": "text" } ], "index": 74 }, { "bbox": [ 304, 421, 527, 434 ], "spans": [ { "bbox": [ 304, 421, 527, 434 ], "score": 1.0, "content": "et al., 2023) and conduct instruct tuning on GPT4-", "type": "text" } ], "index": 75 }, { "bbox": [ 304, 434, 525, 448 ], "spans": [ { "bbox": [ 304, 434, 525, 448 ], "score": 1.0, "content": "assisted visual instruction data. X-LLM (Chen", "type": "text" } ], "index": 76 }, { "bbox": [ 304, 448, 527, 462 ], "spans": [ { "bbox": [ 304, 448, 527, 462 ], "score": 1.0, "content": "et al., 2023) converts multi-modalities into repre-", "type": "text" } ], "index": 77 }, { "bbox": [ 304, 462, 525, 475 ], "spans": [ { "bbox": [ 304, 462, 525, 475 ], "score": 1.0, "content": "sentations with X2L interfaces as the inputs of the", "type": "text" } ], "index": 78 }, { "bbox": [ 304, 476, 526, 489 ], "spans": [ { "bbox": [ 304, 476, 526, 489 ], "score": 1.0, "content": "large language model. However, such structures", "type": "text" } ], "index": 79 }, { "bbox": [ 304, 489, 527, 502 ], "spans": [ { "bbox": [ 304, 489, 527, 502 ], "score": 1.0, "content": "only enable LLMs to process multi-modal input,", "type": "text" } ], "index": 80 }, { "bbox": [ 304, 503, 527, 516 ], "spans": [ { "bbox": [ 304, 503, 527, 516 ], "score": 1.0, "content": "without ability to generate multi-modal output. Di-", "type": "text" } ], "index": 81 }, { "bbox": [ 304, 517, 527, 530 ], "spans": [ { "bbox": [ 304, 517, 527, 530 ], "score": 1.0, "content": "verging from prior studies, our approach empha-", "type": "text" } ], "index": 82 }, { "bbox": [ 304, 530, 527, 542 ], "spans": [ { "bbox": [ 304, 530, 527, 542 ], "score": 1.0, "content": "sizes the development of a speech-centric multi-", "type": "text" } ], "index": 83 }, { "bbox": [ 304, 543, 525, 557 ], "spans": [ { "bbox": [ 304, 543, 525, 557 ], "score": 1.0, "content": "modal LLM, endowing it with the proficiency to", "type": "text" } ], "index": 84 }, { "bbox": [ 304, 557, 523, 570 ], "spans": [ { "bbox": [ 304, 557, 523, 570 ], "score": 1.0, "content": "accommodate both multi-modal input and output.", "type": "text" } ], "index": 85 } ], "index": 73.5 }, { "type": "text", "bbox": [ 305, 572, 526, 773 ], "lines": [ { "bbox": [ 305, 572, 525, 584 ], "spans": [ { "bbox": [ 305, 572, 525, 584 ], "score": 1.0, "content": "Generative Spoken Language Model Discrete", "type": "text" } ], "index": 86 }, { "bbox": [ 304, 585, 527, 599 ], "spans": [ { "bbox": [ 304, 585, 527, 599 ], "score": 1.0, "content": "self-supervised representation based spoken gen-", "type": "text" } ], "index": 87 }, { "bbox": [ 304, 598, 527, 612 ], "spans": [ { "bbox": [ 304, 598, 527, 612 ], "score": 1.0, "content": "erative language modeling is making remark-", "type": "text" } ], "index": 88 }, { "bbox": [ 304, 612, 527, 625 ], "spans": [ { "bbox": [ 304, 612, 527, 625 ], "score": 1.0, "content": "able progress on large-scale speech dataset train-", "type": "text" } ], "index": 89 }, { "bbox": [ 304, 626, 526, 639 ], "spans": [ { "bbox": [ 304, 626, 526, 639 ], "score": 1.0, "content": "ing (Nguyen et al., 2022). AudioLM (Borsos et al.,", "type": "text" } ], "index": 90 }, { "bbox": [ 304, 639, 525, 651 ], "spans": [ { "bbox": [ 304, 639, 525, 651 ], "score": 1.0, "content": "2022) proposes to model speech based on audio", "type": "text" } ], "index": 91 }, { "bbox": [ 304, 653, 525, 665 ], "spans": [ { "bbox": [ 304, 653, 525, 665 ], "score": 1.0, "content": "codecs together with semantic codes, which can", "type": "text" } ], "index": 92 }, { "bbox": [ 304, 666, 527, 680 ], "spans": [ { "bbox": [ 304, 666, 527, 680 ], "score": 1.0, "content": "synthesize speech in a textlesss setting. VALL-", "type": "text" } ], "index": 93 }, { "bbox": [ 303, 679, 526, 694 ], "spans": [ { "bbox": [ 303, 679, 526, 694 ], "score": 1.0, "content": "E (Wang et al., 2023) builds a generative spoken", "type": "text" } ], "index": 94 }, { "bbox": [ 304, 694, 527, 706 ], "spans": [ { "bbox": [ 304, 694, 527, 706 ], "score": 1.0, "content": "language model on audio codecs and treat Text-", "type": "text" } ], "index": 95 }, { "bbox": [ 304, 708, 527, 720 ], "spans": [ { "bbox": [ 304, 708, 527, 720 ], "score": 1.0, "content": "to-Speech as a conditional generation task. How-", "type": "text" } ], "index": 96 }, { "bbox": [ 304, 721, 526, 733 ], "spans": [ { "bbox": [ 304, 721, 526, 733 ], "score": 1.0, "content": "ever, these models are designed for a specific task", "type": "text" } ], "index": 97 }, { "bbox": [ 304, 734, 526, 747 ], "spans": [ { "bbox": [ 304, 734, 526, 747 ], "score": 1.0, "content": "and failed to benefit from LLMs. SpeechGPT", "type": "text" } ], "index": 98 }, { "bbox": [ 304, 748, 527, 760 ], "spans": [ { "bbox": [ 304, 748, 527, 760 ], "score": 1.0, "content": "is built upon the foundation of LLM and trans-", "type": "text" } ], "index": 99 }, { "bbox": [ 304, 761, 527, 775 ], "spans": [ { "bbox": [ 304, 761, 527, 775 ], "score": 1.0, "content": "fers LLM’s knowledge to speech modality, con-", "type": "text" } ], "index": 100 } ], "index": 93 } ], "page_idx": 1, "page_size": [ 595, 841 ], "discarded_blocks": [], "para_blocks": [ { "type": "text", "bbox": [ 69, 72, 290, 125 ], "lines": [], "index": 1.5, "bbox_fs": [ 68, 70, 292, 126 ], "lines_deleted": true }, { "type": "text", "bbox": [ 70, 128, 290, 262 ], "lines": [ { "bbox": [ 79, 127, 290, 142 ], "spans": [ { "bbox": [ 79, 127, 290, 142 ], "score": 1.0, "content": "In this paper, we propose SpeechGPT, a large", "type": "text" } ], "index": 4 }, { "bbox": [ 68, 141, 292, 155 ], "spans": [ { "bbox": [ 68, 141, 292, 155 ], "score": 1.0, "content": "language model with intrinsic cross-modal conver-", "type": "text" } ], "index": 5 }, { "bbox": [ 69, 155, 291, 168 ], "spans": [ { "bbox": [ 69, 155, 291, 168 ], "score": 1.0, "content": "sational abilities, capable of perceiving and gen-", "type": "text" } ], "index": 6 }, { "bbox": [ 69, 169, 290, 182 ], "spans": [ { "bbox": [ 69, 169, 290, 182 ], "score": 1.0, "content": "erating multi-modal content. We perform speech", "type": "text" } ], "index": 7 }, { "bbox": [ 69, 182, 290, 195 ], "spans": [ { "bbox": [ 69, 182, 290, 195 ], "score": 1.0, "content": "discretization with a self-supervised trained speech", "type": "text" } ], "index": 8 }, { "bbox": [ 69, 196, 290, 208 ], "spans": [ { "bbox": [ 69, 196, 290, 208 ], "score": 1.0, "content": "model to unify the modality between speech and", "type": "text" } ], "index": 9 }, { "bbox": [ 69, 209, 290, 222 ], "spans": [ { "bbox": [ 69, 209, 290, 222 ], "score": 1.0, "content": "text. The discrete speech tokens are then expanded", "type": "text" } ], "index": 10 }, { "bbox": [ 68, 222, 291, 236 ], "spans": [ { "bbox": [ 68, 222, 291, 236 ], "score": 1.0, "content": "into the vocabulary of the LLM, thus endowing the", "type": "text" } ], "index": 11 }, { "bbox": [ 69, 236, 290, 249 ], "spans": [ { "bbox": [ 69, 236, 290, 249 ], "score": 1.0, "content": "model with an inherent competence to perceive and", "type": "text" } ], "index": 12 }, { "bbox": [ 69, 251, 159, 263 ], "spans": [ { "bbox": [ 69, 251, 159, 263 ], "score": 1.0, "content": "generate the speech.", "type": "text" } ], "index": 13 } ], "index": 8.5, "bbox_fs": [ 68, 127, 292, 263 ] }, { "type": "text", "bbox": [ 69, 265, 289, 467 ], "lines": [ { "bbox": [ 79, 263, 291, 279 ], "spans": [ { "bbox": [ 79, 263, 291, 279 ], "score": 1.0, "content": "To provide the model with the capacity to", "type": "text" } ], "index": 14 }, { "bbox": [ 69, 279, 290, 291 ], "spans": [ { "bbox": [ 69, 279, 290, 291 ], "score": 1.0, "content": "handle multi-modal instructions, we build the", "type": "text" } ], "index": 15 }, { "bbox": [ 69, 293, 290, 306 ], "spans": [ { "bbox": [ 69, 293, 290, 306 ], "score": 1.0, "content": "first speech-text cross-modal instruction-following", "type": "text" } ], "index": 16 }, { "bbox": [ 69, 306, 290, 318 ], "spans": [ { "bbox": [ 69, 306, 290, 318 ], "score": 1.0, "content": "dataset SpeechInstruct. Specifically, we discretize", "type": "text" } ], "index": 17 }, { "bbox": [ 69, 320, 290, 332 ], "spans": [ { "bbox": [ 69, 320, 290, 332 ], "score": 1.0, "content": "the speech to discrete units (Hsu et al., 2021) and", "type": "text" } ], "index": 18 }, { "bbox": [ 69, 333, 290, 346 ], "spans": [ { "bbox": [ 69, 333, 290, 346 ], "score": 1.0, "content": "construct the cross-modal unit-text pair based on", "type": "text" } ], "index": 19 }, { "bbox": [ 69, 347, 291, 359 ], "spans": [ { "bbox": [ 69, 347, 291, 359 ], "score": 1.0, "content": "the existing ASR dataset. Meanwhile, we construct", "type": "text" } ], "index": 20 }, { "bbox": [ 69, 360, 291, 372 ], "spans": [ { "bbox": [ 69, 360, 291, 372 ], "score": 1.0, "content": "hundreds of instructions for diverse tasks with GPT-", "type": "text" } ], "index": 21 }, { "bbox": [ 68, 374, 290, 385 ], "spans": [ { "bbox": [ 68, 374, 290, 385 ], "score": 1.0, "content": "4 to simulate actual user instructions as illustrated", "type": "text" } ], "index": 22 }, { "bbox": [ 68, 387, 290, 399 ], "spans": [ { "bbox": [ 68, 387, 290, 399 ], "score": 1.0, "content": "in Appendix B. In addition, to further enhance the", "type": "text" } ], "index": 23 }, { "bbox": [ 68, 400, 290, 414 ], "spans": [ { "bbox": [ 68, 400, 290, 414 ], "score": 1.0, "content": "model’s cross-modal capability, we designed the", "type": "text" } ], "index": 24 }, { "bbox": [ 69, 414, 290, 426 ], "spans": [ { "bbox": [ 69, 414, 290, 426 ], "score": 1.0, "content": "Chain-of-Modality instruction data, i.e., the model", "type": "text" } ], "index": 25 }, { "bbox": [ 68, 428, 289, 439 ], "spans": [ { "bbox": [ 68, 428, 289, 439 ], "score": 1.0, "content": "receives the speech command, thinks about the", "type": "text" } ], "index": 26 }, { "bbox": [ 68, 442, 290, 455 ], "spans": [ { "bbox": [ 68, 442, 290, 455 ], "score": 1.0, "content": "process in text, and then outputs the response in", "type": "text" } ], "index": 27 }, { "bbox": [ 68, 456, 104, 467 ], "spans": [ { "bbox": [ 68, 456, 104, 467 ], "score": 1.0, "content": "speech.", "type": "text" } ], "index": 28 } ], "index": 21, "bbox_fs": [ 68, 263, 291, 467 ] }, { "type": "text", "bbox": [ 70, 470, 290, 618 ], "lines": [ { "bbox": [ 79, 470, 290, 483 ], "spans": [ { "bbox": [ 79, 470, 290, 483 ], "score": 1.0, "content": "For better cross-modal transfer and efficient", "type": "text" } ], "index": 29 }, { "bbox": [ 68, 483, 291, 497 ], "spans": [ { "bbox": [ 68, 483, 291, 497 ], "score": 1.0, "content": "training, SpeechGPT undergoes a three-stage train-", "type": "text" } ], "index": 30 }, { "bbox": [ 69, 498, 291, 510 ], "spans": [ { "bbox": [ 69, 498, 291, 510 ], "score": 1.0, "content": "ing process: modality-adaptation pre-training,", "type": "text" } ], "index": 31 }, { "bbox": [ 68, 511, 291, 523 ], "spans": [ { "bbox": [ 68, 511, 291, 523 ], "score": 1.0, "content": "cross-modal instruction fine-tuning, and chain-of-", "type": "text" } ], "index": 32 }, { "bbox": [ 69, 525, 291, 537 ], "spans": [ { "bbox": [ 69, 525, 291, 537 ], "score": 1.0, "content": "modality instruction fine-tuning. The first stage en-", "type": "text" } ], "index": 33 }, { "bbox": [ 69, 538, 290, 551 ], "spans": [ { "bbox": [ 69, 538, 290, 551 ], "score": 1.0, "content": "ables speech comprehension for SpeechGPT with", "type": "text" } ], "index": 34 }, { "bbox": [ 68, 551, 292, 565 ], "spans": [ { "bbox": [ 68, 551, 292, 565 ], "score": 1.0, "content": "the discrete speech unit continuation task. The sec-", "type": "text" } ], "index": 35 }, { "bbox": [ 69, 565, 290, 578 ], "spans": [ { "bbox": [ 69, 565, 290, 578 ], "score": 1.0, "content": "ond stage employs the SpeechInstruct to improve", "type": "text" } ], "index": 36 }, { "bbox": [ 69, 579, 290, 591 ], "spans": [ { "bbox": [ 69, 579, 290, 591 ], "score": 1.0, "content": "the model’s cross-modal capabilities. The third", "type": "text" } ], "index": 37 }, { "bbox": [ 69, 593, 291, 605 ], "spans": [ { "bbox": [ 69, 593, 291, 605 ], "score": 1.0, "content": "stage utilizes parameter-efficient LoRA (Hu et al.,", "type": "text" } ], "index": 38 }, { "bbox": [ 69, 605, 286, 619 ], "spans": [ { "bbox": [ 69, 605, 286, 619 ], "score": 1.0, "content": "2021) fine-tuning for further modality alignment.", "type": "text" } ], "index": 39 } ], "index": 34, "bbox_fs": [ 68, 470, 292, 619 ] }, { "type": "text", "bbox": [ 70, 621, 290, 728 ], "lines": [ { "bbox": [ 79, 621, 291, 634 ], "spans": [ { "bbox": [ 79, 621, 291, 634 ], "score": 1.0, "content": "To evaluate the effectiveness of SpeechGPT,", "type": "text" } ], "index": 40 }, { "bbox": [ 69, 635, 290, 647 ], "spans": [ { "bbox": [ 69, 635, 290, 647 ], "score": 1.0, "content": "we conduct a wide range of human evaluations", "type": "text" } ], "index": 41 }, { "bbox": [ 69, 648, 291, 661 ], "spans": [ { "bbox": [ 69, 648, 291, 661 ], "score": 1.0, "content": "and case analyses to estimate the performance of", "type": "text" } ], "index": 42 }, { "bbox": [ 69, 662, 292, 675 ], "spans": [ { "bbox": [ 69, 662, 292, 675 ], "score": 1.0, "content": "SpeechGPT on textual tasks, speech-text cross-", "type": "text" } ], "index": 43 }, { "bbox": [ 69, 676, 291, 688 ], "spans": [ { "bbox": [ 69, 676, 291, 688 ], "score": 1.0, "content": "modal tasks, and spoken dialogue tasks. The re-", "type": "text" } ], "index": 44 }, { "bbox": [ 69, 688, 290, 703 ], "spans": [ { "bbox": [ 69, 688, 290, 703 ], "score": 1.0, "content": "sults demonstrate that SpeechGPT exhibits a strong", "type": "text" } ], "index": 45 }, { "bbox": [ 69, 703, 290, 715 ], "spans": [ { "bbox": [ 69, 703, 290, 715 ], "score": 1.0, "content": "ability for unimodal and cross-modal instruction", "type": "text" } ], "index": 46 }, { "bbox": [ 69, 715, 141, 729 ], "spans": [ { "bbox": [ 69, 715, 141, 729 ], "score": 1.0, "content": "following tasks.", "type": "text" } ], "index": 47 } ], "index": 43.5, "bbox_fs": [ 69, 621, 292, 729 ] }, { "type": "text", "bbox": [ 79, 731, 258, 744 ], "lines": [ { "bbox": [ 79, 729, 259, 746 ], "spans": [ { "bbox": [ 79, 729, 259, 746 ], "score": 1.0, "content": "Our contributions include the following:", "type": "text" } ], "index": 48 } ], "index": 48, "bbox_fs": [ 79, 729, 259, 746 ] }, { "type": "text", "bbox": [ 71, 747, 290, 773 ], "lines": [ { "bbox": [ 69, 745, 290, 762 ], "spans": [ { "bbox": [ 69, 745, 290, 762 ], "score": 1.0, "content": "• We build the first multi-modal large language", "type": "text" } ], "index": 49 }, { "bbox": [ 79, 761, 291, 774 ], "spans": [ { "bbox": [ 79, 761, 291, 774 ], "score": 1.0, "content": "model that can perceive and generate multi-", "type": "text" } ], "index": 50 } ], "index": 49.5, "bbox_fs": [ 69, 745, 291, 774 ] }, { "type": "list", "bbox": [ 306, 72, 526, 208 ], "lines": [ { "bbox": [ 314, 72, 386, 84 ], "spans": [ { "bbox": [ 314, 72, 386, 84 ], "score": 1.0, "content": "modal contents.", "type": "text" } ], "index": 51, "is_list_end_line": true }, { "bbox": [ 305, 85, 526, 100 ], "spans": [ { "bbox": [ 305, 85, 526, 100 ], "score": 1.0, "content": "• We construct and release SpeechInstruct, the first", "type": "text" } ], "index": 52, "is_list_start_line": true }, { "bbox": [ 314, 100, 527, 113 ], "spans": [ { "bbox": [ 314, 100, 527, 113 ], "score": 1.0, "content": "large-scale speech-text cross-modal instruction-", "type": "text" } ], "index": 53 }, { "bbox": [ 314, 113, 396, 127 ], "spans": [ { "bbox": [ 314, 113, 396, 127 ], "score": 1.0, "content": "following dataset.", "type": "text" } ], "index": 54, "is_list_end_line": true }, { "bbox": [ 305, 126, 525, 142 ], "spans": [ { "bbox": [ 305, 126, 525, 142 ], "score": 1.0, "content": "• We build the first spoken dialogue LLM with", "type": "text" } ], "index": 55, "is_list_start_line": true }, { "bbox": [ 315, 142, 525, 154 ], "spans": [ { "bbox": [ 315, 142, 525, 154 ], "score": 1.0, "content": "strong human instruction following ability and", "type": "text" } ], "index": 56 }, { "bbox": [ 313, 154, 421, 169 ], "spans": [ { "bbox": [ 313, 154, 421, 169 ], "score": 1.0, "content": "spoken dialogue ability.", "type": "text" } ], "index": 57, "is_list_end_line": true }, { "bbox": [ 305, 169, 526, 183 ], "spans": [ { "bbox": [ 305, 169, 526, 183 ], "score": 1.0, "content": "• We show great potential to incorporate other", "type": "text" } ], "index": 58, "is_list_start_line": true }, { "bbox": [ 313, 182, 527, 197 ], "spans": [ { "bbox": [ 313, 182, 527, 197 ], "score": 1.0, "content": "modalities into LLMs through discrete represen-", "type": "text" } ], "index": 59 }, { "bbox": [ 314, 197, 349, 209 ], "spans": [ { "bbox": [ 314, 197, 349, 209 ], "score": 1.0, "content": "tations.", "type": "text" } ], "index": 60, "is_list_end_line": true } ], "index": 55.5, "bbox_fs": [ 305, 72, 527, 209 ] }, { "type": "title", "bbox": [ 305, 222, 395, 235 ], "lines": [ { "bbox": [ 303, 221, 397, 237 ], "spans": [ { "bbox": [ 303, 221, 397, 237 ], "score": 1.0, "content": "2 Related Work", "type": "text" } ], "index": 61 } ], "index": 61 }, { "type": "text", "bbox": [ 304, 246, 525, 569 ], "lines": [ { "bbox": [ 304, 245, 525, 258 ], "spans": [ { "bbox": [ 304, 245, 525, 258 ], "score": 1.0, "content": "Multi-modal Large Language Model Current", "type": "text" } ], "index": 62 }, { "bbox": [ 304, 259, 527, 272 ], "spans": [ { "bbox": [ 304, 259, 527, 272 ], "score": 1.0, "content": "multi-modal LLMs predominantly focus on the vi-", "type": "text" } ], "index": 63 }, { "bbox": [ 304, 272, 525, 286 ], "spans": [ { "bbox": [ 304, 272, 525, 286 ], "score": 1.0, "content": "sual domain, feeding continuous representations", "type": "text" } ], "index": 64 }, { "bbox": [ 304, 286, 525, 299 ], "spans": [ { "bbox": [ 304, 286, 525, 299 ], "score": 1.0, "content": "obtained from pre-trained visual encoders into", "type": "text" } ], "index": 65 }, { "bbox": [ 303, 299, 528, 313 ], "spans": [ { "bbox": [ 303, 299, 528, 313 ], "score": 1.0, "content": "LLMs, facilitating full-parameter or parameter-", "type": "text" } ], "index": 66 }, { "bbox": [ 304, 312, 527, 327 ], "spans": [ { "bbox": [ 304, 312, 527, 327 ], "score": 1.0, "content": "efficient training on visual-language data (Ope-", "type": "text" } ], "index": 67 }, { "bbox": [ 303, 326, 527, 340 ], "spans": [ { "bbox": [ 303, 326, 527, 340 ], "score": 1.0, "content": "nAI, 2023; Huang et al., 2023b; Zhang et al.,", "type": "text" } ], "index": 68 }, { "bbox": [ 304, 339, 525, 353 ], "spans": [ { "bbox": [ 304, 339, 525, 353 ], "score": 1.0, "content": "2023b). Palm-E (Driess et al., 2023) integrates", "type": "text" } ], "index": 69 }, { "bbox": [ 304, 354, 525, 366 ], "spans": [ { "bbox": [ 304, 354, 525, 366 ], "score": 1.0, "content": "the 540B PaLM (Chowdhery et al., 2022) and 22B", "type": "text" } ], "index": 70 }, { "bbox": [ 304, 367, 525, 380 ], "spans": [ { "bbox": [ 304, 367, 525, 380 ], "score": 1.0, "content": "Vision Transformer (Dosovitskiy et al., 2021) into", "type": "text" } ], "index": 71 }, { "bbox": [ 304, 381, 525, 394 ], "spans": [ { "bbox": [ 304, 381, 525, 394 ], "score": 1.0, "content": "the largest vision-language model. LLaVA (Liu", "type": "text" } ], "index": 72 }, { "bbox": [ 303, 394, 525, 407 ], "spans": [ { "bbox": [ 303, 394, 525, 407 ], "score": 1.0, "content": "et al., 2023) leverages pre-trained CLIP (Radford", "type": "text" } ], "index": 73 }, { "bbox": [ 304, 408, 525, 420 ], "spans": [ { "bbox": [ 304, 408, 525, 420 ], "score": 1.0, "content": "et al., 2021) visual encoder and LLaMA (Touvron", "type": "text" } ], "index": 74 }, { "bbox": [ 304, 421, 527, 434 ], "spans": [ { "bbox": [ 304, 421, 527, 434 ], "score": 1.0, "content": "et al., 2023) and conduct instruct tuning on GPT4-", "type": "text" } ], "index": 75 }, { "bbox": [ 304, 434, 525, 448 ], "spans": [ { "bbox": [ 304, 434, 525, 448 ], "score": 1.0, "content": "assisted visual instruction data. X-LLM (Chen", "type": "text" } ], "index": 76 }, { "bbox": [ 304, 448, 527, 462 ], "spans": [ { "bbox": [ 304, 448, 527, 462 ], "score": 1.0, "content": "et al., 2023) converts multi-modalities into repre-", "type": "text" } ], "index": 77 }, { "bbox": [ 304, 462, 525, 475 ], "spans": [ { "bbox": [ 304, 462, 525, 475 ], "score": 1.0, "content": "sentations with X2L interfaces as the inputs of the", "type": "text" } ], "index": 78 }, { "bbox": [ 304, 476, 526, 489 ], "spans": [ { "bbox": [ 304, 476, 526, 489 ], "score": 1.0, "content": "large language model. However, such structures", "type": "text" } ], "index": 79 }, { "bbox": [ 304, 489, 527, 502 ], "spans": [ { "bbox": [ 304, 489, 527, 502 ], "score": 1.0, "content": "only enable LLMs to process multi-modal input,", "type": "text" } ], "index": 80 }, { "bbox": [ 304, 503, 527, 516 ], "spans": [ { "bbox": [ 304, 503, 527, 516 ], "score": 1.0, "content": "without ability to generate multi-modal output. Di-", "type": "text" } ], "index": 81 }, { "bbox": [ 304, 517, 527, 530 ], "spans": [ { "bbox": [ 304, 517, 527, 530 ], "score": 1.0, "content": "verging from prior studies, our approach empha-", "type": "text" } ], "index": 82 }, { "bbox": [ 304, 530, 527, 542 ], "spans": [ { "bbox": [ 304, 530, 527, 542 ], "score": 1.0, "content": "sizes the development of a speech-centric multi-", "type": "text" } ], "index": 83 }, { "bbox": [ 304, 543, 525, 557 ], "spans": [ { "bbox": [ 304, 543, 525, 557 ], "score": 1.0, "content": "modal LLM, endowing it with the proficiency to", "type": "text" } ], "index": 84 }, { "bbox": [ 304, 557, 523, 570 ], "spans": [ { "bbox": [ 304, 557, 523, 570 ], "score": 1.0, "content": "accommodate both multi-modal input and output.", "type": "text" } ], "index": 85 } ], "index": 73.5, "bbox_fs": [ 303, 245, 528, 570 ] }, { "type": "text", "bbox": [ 305, 572, 526, 773 ], "lines": [ { "bbox": [ 305, 572, 525, 584 ], "spans": [ { "bbox": [ 305, 572, 525, 584 ], "score": 1.0, "content": "Generative Spoken Language Model Discrete", "type": "text" } ], "index": 86 }, { "bbox": [ 304, 585, 527, 599 ], "spans": [ { "bbox": [ 304, 585, 527, 599 ], "score": 1.0, "content": "self-supervised representation based spoken gen-", "type": "text" } ], "index": 87 }, { "bbox": [ 304, 598, 527, 612 ], "spans": [ { "bbox": [ 304, 598, 527, 612 ], "score": 1.0, "content": "erative language modeling is making remark-", "type": "text" } ], "index": 88 }, { "bbox": [ 304, 612, 527, 625 ], "spans": [ { "bbox": [ 304, 612, 527, 625 ], "score": 1.0, "content": "able progress on large-scale speech dataset train-", "type": "text" } ], "index": 89 }, { "bbox": [ 304, 626, 526, 639 ], "spans": [ { "bbox": [ 304, 626, 526, 639 ], "score": 1.0, "content": "ing (Nguyen et al., 2022). AudioLM (Borsos et al.,", "type": "text" } ], "index": 90 }, { "bbox": [ 304, 639, 525, 651 ], "spans": [ { "bbox": [ 304, 639, 525, 651 ], "score": 1.0, "content": "2022) proposes to model speech based on audio", "type": "text" } ], "index": 91 }, { "bbox": [ 304, 653, 525, 665 ], "spans": [ { "bbox": [ 304, 653, 525, 665 ], "score": 1.0, "content": "codecs together with semantic codes, which can", "type": "text" } ], "index": 92 }, { "bbox": [ 304, 666, 527, 680 ], "spans": [ { "bbox": [ 304, 666, 527, 680 ], "score": 1.0, "content": "synthesize speech in a textlesss setting. VALL-", "type": "text" } ], "index": 93 }, { "bbox": [ 303, 679, 526, 694 ], "spans": [ { "bbox": [ 303, 679, 526, 694 ], "score": 1.0, "content": "E (Wang et al., 2023) builds a generative spoken", "type": "text" } ], "index": 94 }, { "bbox": [ 304, 694, 527, 706 ], "spans": [ { "bbox": [ 304, 694, 527, 706 ], "score": 1.0, "content": "language model on audio codecs and treat Text-", "type": "text" } ], "index": 95 }, { "bbox": [ 304, 708, 527, 720 ], "spans": [ { "bbox": [ 304, 708, 527, 720 ], "score": 1.0, "content": "to-Speech as a conditional generation task. How-", "type": "text" } ], "index": 96 }, { "bbox": [ 304, 721, 526, 733 ], "spans": [ { "bbox": [ 304, 721, 526, 733 ], "score": 1.0, "content": "ever, these models are designed for a specific task", "type": "text" } ], "index": 97 }, { "bbox": [ 304, 734, 526, 747 ], "spans": [ { "bbox": [ 304, 734, 526, 747 ], "score": 1.0, "content": "and failed to benefit from LLMs. SpeechGPT", "type": "text" } ], "index": 98 }, { "bbox": [ 304, 748, 527, 760 ], "spans": [ { "bbox": [ 304, 748, 527, 760 ], "score": 1.0, "content": "is built upon the foundation of LLM and trans-", "type": "text" } ], "index": 99 }, { "bbox": [ 304, 761, 527, 775 ], "spans": [ { "bbox": [ 304, 761, 527, 775 ], "score": 1.0, "content": "fers LLM’s knowledge to speech modality, con-", "type": "text" } ], "index": 100 }, { "bbox": [ 68, 71, 290, 86 ], "spans": [ { "bbox": [ 68, 71, 290, 86 ], "score": 1.0, "content": "sequently obtaining better task generalization and", "type": "text", "cross_page": true } ], "index": 0 }, { "bbox": [ 68, 83, 229, 100 ], "spans": [ { "bbox": [ 68, 83, 229, 100 ], "score": 1.0, "content": "human-instruction following ability.", "type": "text", "cross_page": true } ], "index": 1 } ], "index": 93, "bbox_fs": [ 303, 572, 527, 775 ] } ] }, { "preproc_blocks": [ { "type": "text", "bbox": [ 69, 72, 289, 97 ], "lines": [ { "bbox": [ 68, 71, 290, 86 ], "spans": [ { "bbox": [ 68, 71, 290, 86 ], "score": 1.0, "content": "sequently obtaining better task generalization and", "type": "text" } ], "index": 0 }, { "bbox": [ 68, 83, 229, 100 ], "spans": [ { "bbox": [ 68, 83, 229, 100 ], "score": 1.0, "content": "human-instruction following ability.", "type": "text" } ], "index": 1 } ], "index": 0.5 }, { "type": "text", "bbox": [ 69, 99, 290, 369 ], "lines": [ { "bbox": [ 68, 97, 291, 114 ], "spans": [ { "bbox": [ 68, 97, 291, 114 ], "score": 1.0, "content": "Speech-Enabled LLM Interaction Following", "type": "text" } ], "index": 2 }, { "bbox": [ 69, 113, 290, 126 ], "spans": [ { "bbox": [ 69, 113, 290, 126 ], "score": 1.0, "content": "the emergence of ChatGPT, several studies have", "type": "text" } ], "index": 3 }, { "bbox": [ 69, 127, 289, 139 ], "spans": [ { "bbox": [ 69, 127, 289, 139 ], "score": 1.0, "content": "concentrated on the integration of expert speech", "type": "text" } ], "index": 4 }, { "bbox": [ 69, 140, 291, 153 ], "spans": [ { "bbox": [ 69, 140, 291, 153 ], "score": 1.0, "content": "models with LLMs to enable direct speech interac-", "type": "text" } ], "index": 5 }, { "bbox": [ 69, 154, 290, 166 ], "spans": [ { "bbox": [ 69, 154, 290, 166 ], "score": 1.0, "content": "tion with LLMs. HuggingGPT (Shen et al., 2023)", "type": "text" } ], "index": 6 }, { "bbox": [ 69, 167, 291, 179 ], "spans": [ { "bbox": [ 69, 167, 291, 179 ], "score": 1.0, "content": "facilitates task decomposition of human instruc-", "type": "text" } ], "index": 7 }, { "bbox": [ 69, 180, 292, 193 ], "spans": [ { "bbox": [ 69, 180, 292, 193 ], "score": 1.0, "content": "tions by LLMs and allows the invocation of mod-", "type": "text" } ], "index": 8 }, { "bbox": [ 68, 194, 292, 208 ], "spans": [ { "bbox": [ 68, 194, 292, 208 ], "score": 1.0, "content": "els from Huggingface to accomplish specific tasks,", "type": "text" } ], "index": 9 }, { "bbox": [ 68, 208, 291, 222 ], "spans": [ { "bbox": [ 68, 208, 291, 222 ], "score": 1.0, "content": "encompassing a range of automatic speech recog-", "type": "text" } ], "index": 10 }, { "bbox": [ 69, 222, 291, 234 ], "spans": [ { "bbox": [ 69, 222, 291, 234 ], "score": 1.0, "content": "nition (ASR) and text-to-speech models. Audio-", "type": "text" } ], "index": 11 }, { "bbox": [ 69, 235, 291, 248 ], "spans": [ { "bbox": [ 69, 235, 291, 248 ], "score": 1.0, "content": "GPT (Huang et al., 2023a) leverages a variety of", "type": "text" } ], "index": 12 }, { "bbox": [ 69, 248, 290, 261 ], "spans": [ { "bbox": [ 69, 248, 290, 261 ], "score": 1.0, "content": "audio foundation models to process complex audio", "type": "text" } ], "index": 13 }, { "bbox": [ 69, 262, 290, 275 ], "spans": [ { "bbox": [ 69, 262, 290, 275 ], "score": 1.0, "content": "information and connect LLMs with input/output", "type": "text" } ], "index": 14 }, { "bbox": [ 68, 275, 292, 289 ], "spans": [ { "bbox": [ 68, 275, 292, 289 ], "score": 1.0, "content": "interface (ASR, TTS) for speech conversations.", "type": "text" } ], "index": 15 }, { "bbox": [ 69, 289, 291, 302 ], "spans": [ { "bbox": [ 69, 289, 291, 302 ], "score": 1.0, "content": "However, these models exhibit increased complex-", "type": "text" } ], "index": 16 }, { "bbox": [ 69, 302, 290, 316 ], "spans": [ { "bbox": [ 69, 302, 290, 316 ], "score": 1.0, "content": "ity, demand extensive resources, and are prone to", "type": "text" } ], "index": 17 }, { "bbox": [ 69, 316, 290, 329 ], "spans": [ { "bbox": [ 69, 316, 290, 329 ], "score": 1.0, "content": "the unavoidable error accumulation problems. Our", "type": "text" } ], "index": 18 }, { "bbox": [ 68, 330, 291, 343 ], "spans": [ { "bbox": [ 68, 330, 291, 343 ], "score": 1.0, "content": "approach enables speech interaction with LLMs", "type": "text" } ], "index": 19 }, { "bbox": [ 69, 343, 291, 356 ], "spans": [ { "bbox": [ 69, 343, 291, 356 ], "score": 1.0, "content": "without relying on ASR or TTS systems, circum-", "type": "text" } ], "index": 20 }, { "bbox": [ 69, 357, 243, 369 ], "spans": [ { "bbox": [ 69, 357, 243, 369 ], "score": 1.0, "content": "venting the aforementioned drawbacks.", "type": "text" } ], "index": 21 } ], "index": 11.5 }, { "type": "title", "bbox": [ 70, 380, 236, 394 ], "lines": [ { "bbox": [ 67, 380, 237, 397 ], "spans": [ { "bbox": [ 67, 380, 237, 397 ], "score": 1.0, "content": "3 SpeechInstruct Construction", "type": "text" } ], "index": 22 } ], "index": 22 }, { "type": "text", "bbox": [ 69, 403, 290, 511 ], "lines": [ { "bbox": [ 69, 403, 290, 416 ], "spans": [ { "bbox": [ 69, 403, 290, 416 ], "score": 1.0, "content": "Due to the limitations in publicly available speech", "type": "text" } ], "index": 23 }, { "bbox": [ 69, 417, 291, 431 ], "spans": [ { "bbox": [ 69, 417, 291, 431 ], "score": 1.0, "content": "data and the lack of variety of speech-text tasks,", "type": "text" } ], "index": 24 }, { "bbox": [ 69, 431, 291, 443 ], "spans": [ { "bbox": [ 69, 431, 291, 443 ], "score": 1.0, "content": "we construct SpeechInstruct, a speech-text cross-", "type": "text" } ], "index": 25 }, { "bbox": [ 69, 445, 290, 456 ], "spans": [ { "bbox": [ 69, 445, 290, 456 ], "score": 1.0, "content": "modal instruction-following dataset. This dataset", "type": "text" } ], "index": 26 }, { "bbox": [ 69, 459, 291, 470 ], "spans": [ { "bbox": [ 69, 459, 291, 470 ], "score": 1.0, "content": "consists of two parts, the first part is called Cross-", "type": "text" } ], "index": 27 }, { "bbox": [ 68, 471, 290, 484 ], "spans": [ { "bbox": [ 68, 471, 290, 484 ], "score": 1.0, "content": "Modal Instruction, and the second part is called", "type": "text" } ], "index": 28 }, { "bbox": [ 69, 485, 290, 497 ], "spans": [ { "bbox": [ 69, 485, 290, 497 ], "score": 1.0, "content": "Chain-of-Modality Instruction. The construction", "type": "text" } ], "index": 29 }, { "bbox": [ 68, 498, 292, 512 ], "spans": [ { "bbox": [ 68, 498, 292, 512 ], "score": 1.0, "content": "process of SpeechInstruct is illustrated in Figure 2.", "type": "text" } ], "index": 30 } ], "index": 26.5 }, { "type": "title", "bbox": [ 70, 522, 210, 534 ], "lines": [ { "bbox": [ 68, 520, 211, 537 ], "spans": [ { "bbox": [ 68, 520, 211, 537 ], "score": 1.0, "content": "3.1 Cross-modal Instruction", "type": "text" } ], "index": 31 } ], "index": 31 }, { "type": "text", "bbox": [ 70, 540, 290, 661 ], "lines": [ { "bbox": [ 68, 539, 290, 553 ], "spans": [ { "bbox": [ 68, 539, 290, 553 ], "score": 1.0, "content": "Data Collection We collect several large-scale", "type": "text" } ], "index": 32 }, { "bbox": [ 68, 554, 290, 567 ], "spans": [ { "bbox": [ 68, 554, 290, 567 ], "score": 1.0, "content": "English ASR datasets to construct Cross-Modal", "type": "text" } ], "index": 33 }, { "bbox": [ 68, 567, 291, 581 ], "spans": [ { "bbox": [ 68, 567, 291, 581 ], "score": 1.0, "content": "Instruction, including Gigaspeech (Chen et al.,", "type": "text" } ], "index": 34 }, { "bbox": [ 69, 581, 290, 593 ], "spans": [ { "bbox": [ 69, 581, 290, 593 ], "score": 1.0, "content": "2021), Common Voice (Ardila et al., 2020), and", "type": "text" } ], "index": 35 }, { "bbox": [ 68, 594, 290, 608 ], "spans": [ { "bbox": [ 68, 594, 290, 608 ], "score": 1.0, "content": "LibriSpeech (Panayotov et al., 2015). We employ", "type": "text" } ], "index": 36 }, { "bbox": [ 69, 608, 290, 621 ], "spans": [ { "bbox": [ 69, 608, 290, 621 ], "score": 1.0, "content": "mHuBERT1 as the speech tokenizer to discretize", "type": "text" } ], "index": 37 }, { "bbox": [ 69, 622, 291, 635 ], "spans": [ { "bbox": [ 69, 622, 291, 635 ], "score": 1.0, "content": "speech data into discrete units and remove the repet-", "type": "text" } ], "index": 38 }, { "bbox": [ 69, 635, 292, 648 ], "spans": [ { "bbox": [ 69, 635, 292, 648 ], "score": 1.0, "content": "itive units of adjacent frames to get reduced units.", "type": "text" } ], "index": 39 }, { "bbox": [ 69, 649, 292, 661 ], "spans": [ { "bbox": [ 69, 649, 292, 661 ], "score": 1.0, "content": "Ultimately, we obtain 9 million unit-text data pairs.", "type": "text" } ], "index": 40 } ], "index": 36 }, { "type": "text", "bbox": [ 69, 663, 290, 743 ], "lines": [ { "bbox": [ 69, 663, 290, 676 ], "spans": [ { "bbox": [ 69, 663, 290, 676 ], "score": 1.0, "content": "Task Description Generation We generate ASR", "type": "text" } ], "index": 41 }, { "bbox": [ 69, 676, 290, 689 ], "spans": [ { "bbox": [ 69, 676, 290, 689 ], "score": 1.0, "content": "and TTS task descriptions that are compatible with", "type": "text" } ], "index": 42 }, { "bbox": [ 68, 690, 291, 702 ], "spans": [ { "bbox": [ 68, 690, 291, 702 ], "score": 1.0, "content": "speech-text data pairs. Unlike the Self-Instruct", "type": "text" } ], "index": 43 }, { "bbox": [ 68, 703, 291, 717 ], "spans": [ { "bbox": [ 68, 703, 291, 717 ], "score": 1.0, "content": "method (Wang et al., 2022), we generate descrip-", "type": "text" } ], "index": 44 }, { "bbox": [ 69, 717, 291, 730 ], "spans": [ { "bbox": [ 69, 717, 291, 730 ], "score": 1.0, "content": "tions through a zero-shot approach. Specifically,", "type": "text" } ], "index": 45 }, { "bbox": [ 68, 730, 291, 744 ], "spans": [ { "bbox": [ 68, 730, 291, 744 ], "score": 1.0, "content": "we directly input the prompts shown in Appendix A", "type": "text" } ], "index": 46 } ], "index": 43.5 }, { "type": "text", "bbox": [ 304, 72, 525, 124 ], "lines": [ { "bbox": [ 303, 70, 528, 85 ], "spans": [ { "bbox": [ 303, 70, 528, 85 ], "score": 1.0, "content": "into OpenAI GPT-4 to generate task descriptions.", "type": "text" } ], "index": 47 }, { "bbox": [ 304, 85, 527, 98 ], "spans": [ { "bbox": [ 304, 85, 527, 98 ], "score": 1.0, "content": "Our generation method yields 100 instructions for", "type": "text" } ], "index": 48 }, { "bbox": [ 304, 99, 526, 112 ], "spans": [ { "bbox": [ 304, 99, 526, 112 ], "score": 1.0, "content": "each task and some examples are shown in Ap-", "type": "text" } ], "index": 49 }, { "bbox": [ 303, 113, 350, 125 ], "spans": [ { "bbox": [ 303, 113, 350, 125 ], "score": 1.0, "content": "pendix B.", "type": "text" } ], "index": 50 } ], "index": 48.5 }, { "type": "text", "bbox": [ 305, 127, 525, 289 ], "lines": [ { "bbox": [ 304, 126, 527, 140 ], "spans": [ { "bbox": [ 304, 126, 527, 140 ], "score": 1.0, "content": "Instruction Formatting For a discrete unit se-", "type": "text" } ], "index": 51 }, { "bbox": [ 303, 140, 527, 154 ], "spans": [ { "bbox": [ 303, 140, 339, 154 ], "score": 1.0, "content": "quence", "type": "text" }, { "bbox": [ 340, 140, 350, 151 ], "score": 0.72, "content": "U", "type": "inline_equation" }, { "bbox": [ 350, 140, 495, 154 ], "score": 1.0, "content": "and its associated transcription", "type": "text" }, { "bbox": [ 495, 140, 505, 151 ], "score": 0.67, "content": "T", "type": "inline_equation" }, { "bbox": [ 505, 140, 527, 154 ], "score": 1.0, "content": ", we", "type": "text" } ], "index": 52 }, { "bbox": [ 304, 154, 527, 166 ], "spans": [ { "bbox": [ 304, 154, 527, 166 ], "score": 1.0, "content": "determine whether it will be used for construct-", "type": "text" } ], "index": 53 }, { "bbox": [ 304, 168, 526, 180 ], "spans": [ { "bbox": [ 304, 168, 526, 180 ], "score": 1.0, "content": "ing an ASR task or a TTS task based on the", "type": "text" } ], "index": 54 }, { "bbox": [ 303, 181, 526, 194 ], "spans": [ { "bbox": [ 303, 181, 356, 194 ], "score": 1.0, "content": "probability", "type": "text" }, { "bbox": [ 357, 183, 363, 194 ], "score": 0.7, "content": "p", "type": "inline_equation" }, { "bbox": [ 364, 181, 526, 194 ], "score": 1.0, "content": ". Subsequently, we randomly select", "type": "text" } ], "index": 55 }, { "bbox": [ 303, 194, 527, 208 ], "spans": [ { "bbox": [ 303, 194, 366, 208 ], "score": 1.0, "content": "a description", "type": "text" }, { "bbox": [ 366, 195, 377, 205 ], "score": 0.74, "content": "D", "type": "inline_equation" }, { "bbox": [ 378, 194, 527, 208 ], "score": 1.0, "content": "from the corresponding task de-", "type": "text" } ], "index": 56 }, { "bbox": [ 304, 209, 527, 221 ], "spans": [ { "bbox": [ 304, 209, 527, 221 ], "score": 1.0, "content": "scription. This results in a triplet consisting of", "type": "text" } ], "index": 57 }, { "bbox": [ 304, 221, 525, 234 ], "spans": [ { "bbox": [ 304, 221, 525, 234 ], "score": 1.0, "content": "the task description, discrete unit sequence, and", "type": "text" } ], "index": 58 }, { "bbox": [ 304, 234, 526, 249 ], "spans": [ { "bbox": [ 304, 234, 423, 249 ], "score": 1.0, "content": "transcription, denoted as", "type": "text" }, { "bbox": [ 423, 235, 466, 248 ], "score": 0.93, "content": "( D , U , T )", "type": "inline_equation" }, { "bbox": [ 467, 234, 526, 249 ], "score": 1.0, "content": ". Following", "type": "text" } ], "index": 59 }, { "bbox": [ 304, 249, 526, 262 ], "spans": [ { "bbox": [ 304, 249, 526, 262 ], "score": 1.0, "content": "this, the triplet is assembled into an instruction", "type": "text" } ], "index": 60 }, { "bbox": [ 303, 262, 527, 276 ], "spans": [ { "bbox": [ 303, 262, 437, 276 ], "score": 1.0, "content": "using the template: [Human]:", "type": "text" }, { "bbox": [ 437, 262, 458, 275 ], "score": 0.9, "content": "\\{ D \\}", "type": "inline_equation" }, { "bbox": [ 458, 262, 527, 276 ], "score": 1.0, "content": ". This is input:", "type": "text" } ], "index": 61 }, { "bbox": [ 305, 275, 478, 290 ], "spans": [ { "bbox": [ 305, 275, 325, 289 ], "score": 0.78, "content": "\\{ U \\}", "type": "inline_equation" }, { "bbox": [ 325, 276, 424, 290 ], "score": 1.0, "content": ".[SpeechGPT]:", "type": "text" }, { "bbox": [ 425, 276, 473, 289 ], "score": 0.79, "content": "\\{ T \\} { \\ < } { \\bf e } { \\bf 0 } { \\bf s } > .", "type": "inline_equation" }, { "bbox": [ 473, 276, 478, 290 ], "score": 1.0, "content": "..", "type": "text" } ], "index": 62 } ], "index": 56.5 }, { "type": "title", "bbox": [ 306, 300, 473, 313 ], "lines": [ { "bbox": [ 302, 298, 474, 317 ], "spans": [ { "bbox": [ 302, 298, 474, 317 ], "score": 1.0, "content": "3.2 Chain-of-Modality Instruction", "type": "text" } ], "index": 63 } ], "index": 63 }, { "type": "text", "bbox": [ 305, 320, 525, 534 ], "lines": [ { "bbox": [ 304, 320, 526, 331 ], "spans": [ { "bbox": [ 304, 320, 526, 331 ], "score": 1.0, "content": "Speech Instruction Generation Due to the lack", "type": "text" } ], "index": 64 }, { "bbox": [ 304, 333, 525, 345 ], "spans": [ { "bbox": [ 304, 333, 525, 345 ], "score": 1.0, "content": "of instruction data with speech input and speech", "type": "text" } ], "index": 65 }, { "bbox": [ 304, 348, 526, 359 ], "spans": [ { "bbox": [ 304, 348, 526, 359 ], "score": 1.0, "content": "output, we trained a text-to-unit generator to con-", "type": "text" } ], "index": 66 }, { "bbox": [ 304, 361, 524, 372 ], "spans": [ { "bbox": [ 304, 361, 524, 372 ], "score": 1.0, "content": "vert text instruction data into speech instruction", "type": "text" } ], "index": 67 }, { "bbox": [ 304, 373, 525, 387 ], "spans": [ { "bbox": [ 304, 373, 525, 387 ], "score": 1.0, "content": "data. Specifically, the text-to-unit generator adopts", "type": "text" } ], "index": 68 }, { "bbox": [ 304, 387, 525, 399 ], "spans": [ { "bbox": [ 304, 387, 525, 399 ], "score": 1.0, "content": "a Transformer encoder-decoder architecture. We", "type": "text" } ], "index": 69 }, { "bbox": [ 304, 400, 527, 414 ], "spans": [ { "bbox": [ 304, 400, 527, 414 ], "score": 1.0, "content": "trained it on LibriSpeech unit-text pairs in Cross-", "type": "text" } ], "index": 70 }, { "bbox": [ 304, 414, 525, 427 ], "spans": [ { "bbox": [ 304, 414, 525, 427 ], "score": 1.0, "content": "modal Instruction. We select 37,969 samples from", "type": "text" } ], "index": 71 }, { "bbox": [ 304, 426, 526, 441 ], "spans": [ { "bbox": [ 304, 426, 526, 441 ], "score": 1.0, "content": "the moss-002-sft-data dataset 2 whose response", "type": "text" } ], "index": 72 }, { "bbox": [ 304, 442, 526, 454 ], "spans": [ { "bbox": [ 304, 442, 526, 454 ], "score": 1.0, "content": "length is shorter than 35 words. And we convert", "type": "text" } ], "index": 73 }, { "bbox": [ 304, 455, 526, 468 ], "spans": [ { "bbox": [ 304, 455, 526, 468 ], "score": 1.0, "content": "both their instructions and responses into unit se-", "type": "text" } ], "index": 74 }, { "bbox": [ 304, 469, 526, 482 ], "spans": [ { "bbox": [ 304, 469, 526, 482 ], "score": 1.0, "content": "quences through the text-to-unit generator. As", "type": "text" } ], "index": 75 }, { "bbox": [ 303, 481, 527, 495 ], "spans": [ { "bbox": [ 303, 481, 527, 495 ], "score": 1.0, "content": "a result, we obtained 37,969 quadruplets com-", "type": "text" } ], "index": 76 }, { "bbox": [ 304, 496, 527, 508 ], "spans": [ { "bbox": [ 304, 496, 527, 508 ], "score": 1.0, "content": "posed of speech instructions, text instructions,", "type": "text" } ], "index": 77 }, { "bbox": [ 304, 509, 526, 523 ], "spans": [ { "bbox": [ 304, 509, 526, 523 ], "score": 1.0, "content": "text responses, and speech responses, denoted as", "type": "text" } ], "index": 78 }, { "bbox": [ 304, 523, 473, 535 ], "spans": [ { "bbox": [ 304, 523, 473, 535 ], "score": 1.0, "content": "(SpeechI, T extI, T extR, SpeechR).", "type": "text" } ], "index": 79 } ], "index": 71.5 }, { "type": "text", "bbox": [ 305, 537, 525, 644 ], "lines": [ { "bbox": [ 304, 537, 526, 550 ], "spans": [ { "bbox": [ 304, 537, 526, 550 ], "score": 1.0, "content": "Instruction Formatting Using the above quadru-", "type": "text" } ], "index": 80 }, { "bbox": [ 304, 550, 525, 563 ], "spans": [ { "bbox": [ 304, 550, 525, 563 ], "score": 1.0, "content": "plets, we could construct chain-of-thought style", "type": "text" } ], "index": 81 }, { "bbox": [ 304, 564, 525, 577 ], "spans": [ { "bbox": [ 304, 564, 525, 577 ], "score": 1.0, "content": "instructions for four input-output formats, namely", "type": "text" } ], "index": 82 }, { "bbox": [ 304, 577, 525, 590 ], "spans": [ { "bbox": [ 304, 577, 525, 590 ], "score": 1.0, "content": "Speech Instruction-Speech Response, Speech", "type": "text" } ], "index": 83 }, { "bbox": [ 304, 590, 526, 605 ], "spans": [ { "bbox": [ 304, 590, 526, 605 ], "score": 1.0, "content": "Instruction-Text Response, Text Instruction-Speech", "type": "text" } ], "index": 84 }, { "bbox": [ 304, 605, 527, 618 ], "spans": [ { "bbox": [ 304, 605, 527, 618 ], "score": 1.0, "content": "Response, and Text Instruction-Text Response.", "type": "text" } ], "index": 85 }, { "bbox": [ 304, 618, 527, 631 ], "spans": [ { "bbox": [ 304, 618, 527, 631 ], "score": 1.0, "content": "Their corresponding templates can be found in Ap-", "type": "text" } ], "index": 86 }, { "bbox": [ 303, 632, 349, 644 ], "spans": [ { "bbox": [ 303, 632, 349, 644 ], "score": 1.0, "content": "pendix C.", "type": "text" } ], "index": 87 } ], "index": 83.5 }, { "type": "title", "bbox": [ 306, 656, 471, 669 ], "lines": [ { "bbox": [ 303, 655, 474, 670 ], "spans": [ { "bbox": [ 303, 655, 474, 670 ], "score": 1.0, "content": "3.3 SpeechInstruct Evaluation Set", "type": "text" } ], "index": 88 } ], "index": 88 }, { "type": "text", "bbox": [ 305, 675, 525, 742 ], "lines": [ { "bbox": [ 303, 674, 525, 688 ], "spans": [ { "bbox": [ 303, 674, 525, 688 ], "score": 1.0, "content": "We constructed cross-modal dialogue datasets", "type": "text" } ], "index": 89 }, { "bbox": [ 304, 689, 526, 701 ], "spans": [ { "bbox": [ 304, 689, 526, 701 ], "score": 1.0, "content": "under different scenarios to evaluate whether", "type": "text" } ], "index": 90 }, { "bbox": [ 304, 702, 527, 714 ], "spans": [ { "bbox": [ 304, 702, 527, 714 ], "score": 1.0, "content": "SpeechGPT could take on various roles. Specif-", "type": "text" } ], "index": 91 }, { "bbox": [ 304, 716, 527, 729 ], "spans": [ { "bbox": [ 304, 716, 527, 729 ], "score": 1.0, "content": "ically, these included a talking encyclopedia, per-", "type": "text" } ], "index": 92 }, { "bbox": [ 304, 729, 527, 743 ], "spans": [ { "bbox": [ 304, 729, 527, 743 ], "score": 1.0, "content": "sonal assistant, chat partner, poet, psychologist,", "type": "text" } ], "index": 93 } ], "index": 91 } ], "page_idx": 2, "page_size": [ 595, 841 ], "discarded_blocks": [ { "type": "discarded", "bbox": [ 69, 753, 275, 773 ], "lines": [ { "bbox": [ 82, 751, 254, 764 ], "spans": [ { "bbox": [ 82, 751, 254, 764 ], "score": 1.0, "content": "1https://dl.fbaipublicfiles.com/", "type": "text" } ] }, { "bbox": [ 68, 762, 276, 774 ], "spans": [ { "bbox": [ 68, 762, 276, 774 ], "score": 1.0, "content": "hubert/mhubert_base_vp_en_es_fr_it3.pt", "type": "text" } ] } ] }, { "type": "discarded", "bbox": [ 304, 753, 521, 772 ], "lines": [ { "bbox": [ 316, 750, 522, 765 ], "spans": [ { "bbox": [ 316, 750, 522, 765 ], "score": 1.0, "content": "2https://huggingface.co/datasets/fnlp/", "type": "text" } ] }, { "bbox": [ 304, 763, 400, 773 ], "spans": [ { "bbox": [ 304, 763, 400, 773 ], "score": 1.0, "content": "moss-002-sft-data", "type": "text" } ] } ] } ], "para_blocks": [ { "type": "text", "bbox": [ 69, 72, 289, 97 ], "lines": [], "index": 0.5, "bbox_fs": [ 68, 71, 290, 100 ], "lines_deleted": true }, { "type": "text", "bbox": [ 69, 99, 290, 369 ], "lines": [ { "bbox": [ 68, 97, 291, 114 ], "spans": [ { "bbox": [ 68, 97, 291, 114 ], "score": 1.0, "content": "Speech-Enabled LLM Interaction Following", "type": "text" } ], "index": 2 }, { "bbox": [ 69, 113, 290, 126 ], "spans": [ { "bbox": [ 69, 113, 290, 126 ], "score": 1.0, "content": "the emergence of ChatGPT, several studies have", "type": "text" } ], "index": 3 }, { "bbox": [ 69, 127, 289, 139 ], "spans": [ { "bbox": [ 69, 127, 289, 139 ], "score": 1.0, "content": "concentrated on the integration of expert speech", "type": "text" } ], "index": 4 }, { "bbox": [ 69, 140, 291, 153 ], "spans": [ { "bbox": [ 69, 140, 291, 153 ], "score": 1.0, "content": "models with LLMs to enable direct speech interac-", "type": "text" } ], "index": 5 }, { "bbox": [ 69, 154, 290, 166 ], "spans": [ { "bbox": [ 69, 154, 290, 166 ], "score": 1.0, "content": "tion with LLMs. HuggingGPT (Shen et al., 2023)", "type": "text" } ], "index": 6 }, { "bbox": [ 69, 167, 291, 179 ], "spans": [ { "bbox": [ 69, 167, 291, 179 ], "score": 1.0, "content": "facilitates task decomposition of human instruc-", "type": "text" } ], "index": 7 }, { "bbox": [ 69, 180, 292, 193 ], "spans": [ { "bbox": [ 69, 180, 292, 193 ], "score": 1.0, "content": "tions by LLMs and allows the invocation of mod-", "type": "text" } ], "index": 8 }, { "bbox": [ 68, 194, 292, 208 ], "spans": [ { "bbox": [ 68, 194, 292, 208 ], "score": 1.0, "content": "els from Huggingface to accomplish specific tasks,", "type": "text" } ], "index": 9 }, { "bbox": [ 68, 208, 291, 222 ], "spans": [ { "bbox": [ 68, 208, 291, 222 ], "score": 1.0, "content": "encompassing a range of automatic speech recog-", "type": "text" } ], "index": 10 }, { "bbox": [ 69, 222, 291, 234 ], "spans": [ { "bbox": [ 69, 222, 291, 234 ], "score": 1.0, "content": "nition (ASR) and text-to-speech models. Audio-", "type": "text" } ], "index": 11 }, { "bbox": [ 69, 235, 291, 248 ], "spans": [ { "bbox": [ 69, 235, 291, 248 ], "score": 1.0, "content": "GPT (Huang et al., 2023a) leverages a variety of", "type": "text" } ], "index": 12 }, { "bbox": [ 69, 248, 290, 261 ], "spans": [ { "bbox": [ 69, 248, 290, 261 ], "score": 1.0, "content": "audio foundation models to process complex audio", "type": "text" } ], "index": 13 }, { "bbox": [ 69, 262, 290, 275 ], "spans": [ { "bbox": [ 69, 262, 290, 275 ], "score": 1.0, "content": "information and connect LLMs with input/output", "type": "text" } ], "index": 14 }, { "bbox": [ 68, 275, 292, 289 ], "spans": [ { "bbox": [ 68, 275, 292, 289 ], "score": 1.0, "content": "interface (ASR, TTS) for speech conversations.", "type": "text" } ], "index": 15 }, { "bbox": [ 69, 289, 291, 302 ], "spans": [ { "bbox": [ 69, 289, 291, 302 ], "score": 1.0, "content": "However, these models exhibit increased complex-", "type": "text" } ], "index": 16 }, { "bbox": [ 69, 302, 290, 316 ], "spans": [ { "bbox": [ 69, 302, 290, 316 ], "score": 1.0, "content": "ity, demand extensive resources, and are prone to", "type": "text" } ], "index": 17 }, { "bbox": [ 69, 316, 290, 329 ], "spans": [ { "bbox": [ 69, 316, 290, 329 ], "score": 1.0, "content": "the unavoidable error accumulation problems. Our", "type": "text" } ], "index": 18 }, { "bbox": [ 68, 330, 291, 343 ], "spans": [ { "bbox": [ 68, 330, 291, 343 ], "score": 1.0, "content": "approach enables speech interaction with LLMs", "type": "text" } ], "index": 19 }, { "bbox": [ 69, 343, 291, 356 ], "spans": [ { "bbox": [ 69, 343, 291, 356 ], "score": 1.0, "content": "without relying on ASR or TTS systems, circum-", "type": "text" } ], "index": 20 }, { "bbox": [ 69, 357, 243, 369 ], "spans": [ { "bbox": [ 69, 357, 243, 369 ], "score": 1.0, "content": "venting the aforementioned drawbacks.", "type": "text" } ], "index": 21 } ], "index": 11.5, "bbox_fs": [ 68, 97, 292, 369 ] }, { "type": "title", "bbox": [ 70, 380, 236, 394 ], "lines": [ { "bbox": [ 67, 380, 237, 397 ], "spans": [ { "bbox": [ 67, 380, 237, 397 ], "score": 1.0, "content": "3 SpeechInstruct Construction", "type": "text" } ], "index": 22 } ], "index": 22 }, { "type": "text", "bbox": [ 69, 403, 290, 511 ], "lines": [ { "bbox": [ 69, 403, 290, 416 ], "spans": [ { "bbox": [ 69, 403, 290, 416 ], "score": 1.0, "content": "Due to the limitations in publicly available speech", "type": "text" } ], "index": 23 }, { "bbox": [ 69, 417, 291, 431 ], "spans": [ { "bbox": [ 69, 417, 291, 431 ], "score": 1.0, "content": "data and the lack of variety of speech-text tasks,", "type": "text" } ], "index": 24 }, { "bbox": [ 69, 431, 291, 443 ], "spans": [ { "bbox": [ 69, 431, 291, 443 ], "score": 1.0, "content": "we construct SpeechInstruct, a speech-text cross-", "type": "text" } ], "index": 25 }, { "bbox": [ 69, 445, 290, 456 ], "spans": [ { "bbox": [ 69, 445, 290, 456 ], "score": 1.0, "content": "modal instruction-following dataset. This dataset", "type": "text" } ], "index": 26 }, { "bbox": [ 69, 459, 291, 470 ], "spans": [ { "bbox": [ 69, 459, 291, 470 ], "score": 1.0, "content": "consists of two parts, the first part is called Cross-", "type": "text" } ], "index": 27 }, { "bbox": [ 68, 471, 290, 484 ], "spans": [ { "bbox": [ 68, 471, 290, 484 ], "score": 1.0, "content": "Modal Instruction, and the second part is called", "type": "text" } ], "index": 28 }, { "bbox": [ 69, 485, 290, 497 ], "spans": [ { "bbox": [ 69, 485, 290, 497 ], "score": 1.0, "content": "Chain-of-Modality Instruction. The construction", "type": "text" } ], "index": 29 }, { "bbox": [ 68, 498, 292, 512 ], "spans": [ { "bbox": [ 68, 498, 292, 512 ], "score": 1.0, "content": "process of SpeechInstruct is illustrated in Figure 2.", "type": "text" } ], "index": 30 } ], "index": 26.5, "bbox_fs": [ 68, 403, 292, 512 ] }, { "type": "title", "bbox": [ 70, 522, 210, 534 ], "lines": [ { "bbox": [ 68, 520, 211, 537 ], "spans": [ { "bbox": [ 68, 520, 211, 537 ], "score": 1.0, "content": "3.1 Cross-modal Instruction", "type": "text" } ], "index": 31 } ], "index": 31 }, { "type": "text", "bbox": [ 70, 540, 290, 661 ], "lines": [ { "bbox": [ 68, 539, 290, 553 ], "spans": [ { "bbox": [ 68, 539, 290, 553 ], "score": 1.0, "content": "Data Collection We collect several large-scale", "type": "text" } ], "index": 32 }, { "bbox": [ 68, 554, 290, 567 ], "spans": [ { "bbox": [ 68, 554, 290, 567 ], "score": 1.0, "content": "English ASR datasets to construct Cross-Modal", "type": "text" } ], "index": 33 }, { "bbox": [ 68, 567, 291, 581 ], "spans": [ { "bbox": [ 68, 567, 291, 581 ], "score": 1.0, "content": "Instruction, including Gigaspeech (Chen et al.,", "type": "text" } ], "index": 34 }, { "bbox": [ 69, 581, 290, 593 ], "spans": [ { "bbox": [ 69, 581, 290, 593 ], "score": 1.0, "content": "2021), Common Voice (Ardila et al., 2020), and", "type": "text" } ], "index": 35 }, { "bbox": [ 68, 594, 290, 608 ], "spans": [ { "bbox": [ 68, 594, 290, 608 ], "score": 1.0, "content": "LibriSpeech (Panayotov et al., 2015). We employ", "type": "text" } ], "index": 36 }, { "bbox": [ 69, 608, 290, 621 ], "spans": [ { "bbox": [ 69, 608, 290, 621 ], "score": 1.0, "content": "mHuBERT1 as the speech tokenizer to discretize", "type": "text" } ], "index": 37 }, { "bbox": [ 69, 622, 291, 635 ], "spans": [ { "bbox": [ 69, 622, 291, 635 ], "score": 1.0, "content": "speech data into discrete units and remove the repet-", "type": "text" } ], "index": 38 }, { "bbox": [ 69, 635, 292, 648 ], "spans": [ { "bbox": [ 69, 635, 292, 648 ], "score": 1.0, "content": "itive units of adjacent frames to get reduced units.", "type": "text" } ], "index": 39 }, { "bbox": [ 69, 649, 292, 661 ], "spans": [ { "bbox": [ 69, 649, 292, 661 ], "score": 1.0, "content": "Ultimately, we obtain 9 million unit-text data pairs.", "type": "text" } ], "index": 40 } ], "index": 36, "bbox_fs": [ 68, 539, 292, 661 ] }, { "type": "text", "bbox": [ 69, 663, 290, 743 ], "lines": [ { "bbox": [ 69, 663, 290, 676 ], "spans": [ { "bbox": [ 69, 663, 290, 676 ], "score": 1.0, "content": "Task Description Generation We generate ASR", "type": "text" } ], "index": 41 }, { "bbox": [ 69, 676, 290, 689 ], "spans": [ { "bbox": [ 69, 676, 290, 689 ], "score": 1.0, "content": "and TTS task descriptions that are compatible with", "type": "text" } ], "index": 42 }, { "bbox": [ 68, 690, 291, 702 ], "spans": [ { "bbox": [ 68, 690, 291, 702 ], "score": 1.0, "content": "speech-text data pairs. Unlike the Self-Instruct", "type": "text" } ], "index": 43 }, { "bbox": [ 68, 703, 291, 717 ], "spans": [ { "bbox": [ 68, 703, 291, 717 ], "score": 1.0, "content": "method (Wang et al., 2022), we generate descrip-", "type": "text" } ], "index": 44 }, { "bbox": [ 69, 717, 291, 730 ], "spans": [ { "bbox": [ 69, 717, 291, 730 ], "score": 1.0, "content": "tions through a zero-shot approach. Specifically,", "type": "text" } ], "index": 45 }, { "bbox": [ 68, 730, 291, 744 ], "spans": [ { "bbox": [ 68, 730, 291, 744 ], "score": 1.0, "content": "we directly input the prompts shown in Appendix A", "type": "text" } ], "index": 46 }, { "bbox": [ 303, 70, 528, 85 ], "spans": [ { "bbox": [ 303, 70, 528, 85 ], "score": 1.0, "content": "into OpenAI GPT-4 to generate task descriptions.", "type": "text" } ], "index": 47 }, { "bbox": [ 304, 85, 527, 98 ], "spans": [ { "bbox": [ 304, 85, 527, 98 ], "score": 1.0, "content": "Our generation method yields 100 instructions for", "type": "text" } ], "index": 48 }, { "bbox": [ 304, 99, 526, 112 ], "spans": [ { "bbox": [ 304, 99, 526, 112 ], "score": 1.0, "content": "each task and some examples are shown in Ap-", "type": "text" } ], "index": 49 }, { "bbox": [ 303, 113, 350, 125 ], "spans": [ { "bbox": [ 303, 113, 350, 125 ], "score": 1.0, "content": "pendix B.", "type": "text" } ], "index": 50 } ], "index": 43.5, "bbox_fs": [ 68, 663, 291, 744 ] }, { "type": "text", "bbox": [ 304, 72, 525, 124 ], "lines": [], "index": 48.5, "bbox_fs": [ 303, 70, 528, 125 ], "lines_deleted": true }, { "type": "text", "bbox": [ 305, 127, 525, 289 ], "lines": [ { "bbox": [ 304, 126, 527, 140 ], "spans": [ { "bbox": [ 304, 126, 527, 140 ], "score": 1.0, "content": "Instruction Formatting For a discrete unit se-", "type": "text" } ], "index": 51 }, { "bbox": [ 303, 140, 527, 154 ], "spans": [ { "bbox": [ 303, 140, 339, 154 ], "score": 1.0, "content": "quence", "type": "text" }, { "bbox": [ 340, 140, 350, 151 ], "score": 0.72, "content": "U", "type": "inline_equation" }, { "bbox": [ 350, 140, 495, 154 ], "score": 1.0, "content": "and its associated transcription", "type": "text" }, { "bbox": [ 495, 140, 505, 151 ], "score": 0.67, "content": "T", "type": "inline_equation" }, { "bbox": [ 505, 140, 527, 154 ], "score": 1.0, "content": ", we", "type": "text" } ], "index": 52 }, { "bbox": [ 304, 154, 527, 166 ], "spans": [ { "bbox": [ 304, 154, 527, 166 ], "score": 1.0, "content": "determine whether it will be used for construct-", "type": "text" } ], "index": 53 }, { "bbox": [ 304, 168, 526, 180 ], "spans": [ { "bbox": [ 304, 168, 526, 180 ], "score": 1.0, "content": "ing an ASR task or a TTS task based on the", "type": "text" } ], "index": 54 }, { "bbox": [ 303, 181, 526, 194 ], "spans": [ { "bbox": [ 303, 181, 356, 194 ], "score": 1.0, "content": "probability", "type": "text" }, { "bbox": [ 357, 183, 363, 194 ], "score": 0.7, "content": "p", "type": "inline_equation" }, { "bbox": [ 364, 181, 526, 194 ], "score": 1.0, "content": ". Subsequently, we randomly select", "type": "text" } ], "index": 55 }, { "bbox": [ 303, 194, 527, 208 ], "spans": [ { "bbox": [ 303, 194, 366, 208 ], "score": 1.0, "content": "a description", "type": "text" }, { "bbox": [ 366, 195, 377, 205 ], "score": 0.74, "content": "D", "type": "inline_equation" }, { "bbox": [ 378, 194, 527, 208 ], "score": 1.0, "content": "from the corresponding task de-", "type": "text" } ], "index": 56 }, { "bbox": [ 304, 209, 527, 221 ], "spans": [ { "bbox": [ 304, 209, 527, 221 ], "score": 1.0, "content": "scription. This results in a triplet consisting of", "type": "text" } ], "index": 57 }, { "bbox": [ 304, 221, 525, 234 ], "spans": [ { "bbox": [ 304, 221, 525, 234 ], "score": 1.0, "content": "the task description, discrete unit sequence, and", "type": "text" } ], "index": 58 }, { "bbox": [ 304, 234, 526, 249 ], "spans": [ { "bbox": [ 304, 234, 423, 249 ], "score": 1.0, "content": "transcription, denoted as", "type": "text" }, { "bbox": [ 423, 235, 466, 248 ], "score": 0.93, "content": "( D , U , T )", "type": "inline_equation" }, { "bbox": [ 467, 234, 526, 249 ], "score": 1.0, "content": ". Following", "type": "text" } ], "index": 59 }, { "bbox": [ 304, 249, 526, 262 ], "spans": [ { "bbox": [ 304, 249, 526, 262 ], "score": 1.0, "content": "this, the triplet is assembled into an instruction", "type": "text" } ], "index": 60 }, { "bbox": [ 303, 262, 527, 276 ], "spans": [ { "bbox": [ 303, 262, 437, 276 ], "score": 1.0, "content": "using the template: [Human]:", "type": "text" }, { "bbox": [ 437, 262, 458, 275 ], "score": 0.9, "content": "\\{ D \\}", "type": "inline_equation" }, { "bbox": [ 458, 262, 527, 276 ], "score": 1.0, "content": ". This is input:", "type": "text" } ], "index": 61 }, { "bbox": [ 305, 275, 478, 290 ], "spans": [ { "bbox": [ 305, 275, 325, 289 ], "score": 0.78, "content": "\\{ U \\}", "type": "inline_equation" }, { "bbox": [ 325, 276, 424, 290 ], "score": 1.0, "content": ".[SpeechGPT]:", "type": "text" }, { "bbox": [ 425, 276, 473, 289 ], "score": 0.79, "content": "\\{ T \\} { \\ < } { \\bf e } { \\bf 0 } { \\bf s } > .", "type": "inline_equation" }, { "bbox": [ 473, 276, 478, 290 ], "score": 1.0, "content": "..", "type": "text" } ], "index": 62 } ], "index": 56.5, "bbox_fs": [ 303, 126, 527, 290 ] }, { "type": "title", "bbox": [ 306, 300, 473, 313 ], "lines": [ { "bbox": [ 302, 298, 474, 317 ], "spans": [ { "bbox": [ 302, 298, 474, 317 ], "score": 1.0, "content": "3.2 Chain-of-Modality Instruction", "type": "text" } ], "index": 63 } ], "index": 63 }, { "type": "text", "bbox": [ 305, 320, 525, 534 ], "lines": [ { "bbox": [ 304, 320, 526, 331 ], "spans": [ { "bbox": [ 304, 320, 526, 331 ], "score": 1.0, "content": "Speech Instruction Generation Due to the lack", "type": "text" } ], "index": 64 }, { "bbox": [ 304, 333, 525, 345 ], "spans": [ { "bbox": [ 304, 333, 525, 345 ], "score": 1.0, "content": "of instruction data with speech input and speech", "type": "text" } ], "index": 65 }, { "bbox": [ 304, 348, 526, 359 ], "spans": [ { "bbox": [ 304, 348, 526, 359 ], "score": 1.0, "content": "output, we trained a text-to-unit generator to con-", "type": "text" } ], "index": 66 }, { "bbox": [ 304, 361, 524, 372 ], "spans": [ { "bbox": [ 304, 361, 524, 372 ], "score": 1.0, "content": "vert text instruction data into speech instruction", "type": "text" } ], "index": 67 }, { "bbox": [ 304, 373, 525, 387 ], "spans": [ { "bbox": [ 304, 373, 525, 387 ], "score": 1.0, "content": "data. Specifically, the text-to-unit generator adopts", "type": "text" } ], "index": 68 }, { "bbox": [ 304, 387, 525, 399 ], "spans": [ { "bbox": [ 304, 387, 525, 399 ], "score": 1.0, "content": "a Transformer encoder-decoder architecture. We", "type": "text" } ], "index": 69 }, { "bbox": [ 304, 400, 527, 414 ], "spans": [ { "bbox": [ 304, 400, 527, 414 ], "score": 1.0, "content": "trained it on LibriSpeech unit-text pairs in Cross-", "type": "text" } ], "index": 70 }, { "bbox": [ 304, 414, 525, 427 ], "spans": [ { "bbox": [ 304, 414, 525, 427 ], "score": 1.0, "content": "modal Instruction. We select 37,969 samples from", "type": "text" } ], "index": 71 }, { "bbox": [ 304, 426, 526, 441 ], "spans": [ { "bbox": [ 304, 426, 526, 441 ], "score": 1.0, "content": "the moss-002-sft-data dataset 2 whose response", "type": "text" } ], "index": 72 }, { "bbox": [ 304, 442, 526, 454 ], "spans": [ { "bbox": [ 304, 442, 526, 454 ], "score": 1.0, "content": "length is shorter than 35 words. And we convert", "type": "text" } ], "index": 73 }, { "bbox": [ 304, 455, 526, 468 ], "spans": [ { "bbox": [ 304, 455, 526, 468 ], "score": 1.0, "content": "both their instructions and responses into unit se-", "type": "text" } ], "index": 74 }, { "bbox": [ 304, 469, 526, 482 ], "spans": [ { "bbox": [ 304, 469, 526, 482 ], "score": 1.0, "content": "quences through the text-to-unit generator. As", "type": "text" } ], "index": 75 }, { "bbox": [ 303, 481, 527, 495 ], "spans": [ { "bbox": [ 303, 481, 527, 495 ], "score": 1.0, "content": "a result, we obtained 37,969 quadruplets com-", "type": "text" } ], "index": 76 }, { "bbox": [ 304, 496, 527, 508 ], "spans": [ { "bbox": [ 304, 496, 527, 508 ], "score": 1.0, "content": "posed of speech instructions, text instructions,", "type": "text" } ], "index": 77 }, { "bbox": [ 304, 509, 526, 523 ], "spans": [ { "bbox": [ 304, 509, 526, 523 ], "score": 1.0, "content": "text responses, and speech responses, denoted as", "type": "text" } ], "index": 78 }, { "bbox": [ 304, 523, 473, 535 ], "spans": [ { "bbox": [ 304, 523, 473, 535 ], "score": 1.0, "content": "(SpeechI, T extI, T extR, SpeechR).", "type": "text" } ], "index": 79 } ], "index": 71.5, "bbox_fs": [ 303, 320, 527, 535 ] }, { "type": "text", "bbox": [ 305, 537, 525, 644 ], "lines": [ { "bbox": [ 304, 537, 526, 550 ], "spans": [ { "bbox": [ 304, 537, 526, 550 ], "score": 1.0, "content": "Instruction Formatting Using the above quadru-", "type": "text" } ], "index": 80 }, { "bbox": [ 304, 550, 525, 563 ], "spans": [ { "bbox": [ 304, 550, 525, 563 ], "score": 1.0, "content": "plets, we could construct chain-of-thought style", "type": "text" } ], "index": 81 }, { "bbox": [ 304, 564, 525, 577 ], "spans": [ { "bbox": [ 304, 564, 525, 577 ], "score": 1.0, "content": "instructions for four input-output formats, namely", "type": "text" } ], "index": 82 }, { "bbox": [ 304, 577, 525, 590 ], "spans": [ { "bbox": [ 304, 577, 525, 590 ], "score": 1.0, "content": "Speech Instruction-Speech Response, Speech", "type": "text" } ], "index": 83 }, { "bbox": [ 304, 590, 526, 605 ], "spans": [ { "bbox": [ 304, 590, 526, 605 ], "score": 1.0, "content": "Instruction-Text Response, Text Instruction-Speech", "type": "text" } ], "index": 84 }, { "bbox": [ 304, 605, 527, 618 ], "spans": [ { "bbox": [ 304, 605, 527, 618 ], "score": 1.0, "content": "Response, and Text Instruction-Text Response.", "type": "text" } ], "index": 85 }, { "bbox": [ 304, 618, 527, 631 ], "spans": [ { "bbox": [ 304, 618, 527, 631 ], "score": 1.0, "content": "Their corresponding templates can be found in Ap-", "type": "text" } ], "index": 86 }, { "bbox": [ 303, 632, 349, 644 ], "spans": [ { "bbox": [ 303, 632, 349, 644 ], "score": 1.0, "content": "pendix C.", "type": "text" } ], "index": 87 } ], "index": 83.5, "bbox_fs": [ 303, 537, 527, 644 ] }, { "type": "title", "bbox": [ 306, 656, 471, 669 ], "lines": [ { "bbox": [ 303, 655, 474, 670 ], "spans": [ { "bbox": [ 303, 655, 474, 670 ], "score": 1.0, "content": "3.3 SpeechInstruct Evaluation Set", "type": "text" } ], "index": 88 } ], "index": 88 }, { "type": "text", "bbox": [ 305, 675, 525, 742 ], "lines": [ { "bbox": [ 303, 674, 525, 688 ], "spans": [ { "bbox": [ 303, 674, 525, 688 ], "score": 1.0, "content": "We constructed cross-modal dialogue datasets", "type": "text" } ], "index": 89 }, { "bbox": [ 304, 689, 526, 701 ], "spans": [ { "bbox": [ 304, 689, 526, 701 ], "score": 1.0, "content": "under different scenarios to evaluate whether", "type": "text" } ], "index": 90 }, { "bbox": [ 304, 702, 527, 714 ], "spans": [ { "bbox": [ 304, 702, 527, 714 ], "score": 1.0, "content": "SpeechGPT could take on various roles. Specif-", "type": "text" } ], "index": 91 }, { "bbox": [ 304, 716, 527, 729 ], "spans": [ { "bbox": [ 304, 716, 527, 729 ], "score": 1.0, "content": "ically, these included a talking encyclopedia, per-", "type": "text" } ], "index": 92 }, { "bbox": [ 304, 729, 527, 743 ], "spans": [ { "bbox": [ 304, 729, 527, 743 ], "score": 1.0, "content": "sonal assistant, chat partner, poet, psychologist,", "type": "text" } ], "index": 93 }, { "bbox": [ 69, 308, 291, 322 ], "spans": [ { "bbox": [ 69, 308, 291, 322 ], "score": 1.0, "content": "and educational assistant. For each role, we pro-", "type": "text", "cross_page": true } ], "index": 6 }, { "bbox": [ 69, 322, 290, 336 ], "spans": [ { "bbox": [ 69, 322, 290, 336 ], "score": 1.0, "content": "vide 10 manually authored instruction-response", "type": "text", "cross_page": true } ], "index": 7 }, { "bbox": [ 68, 336, 291, 349 ], "spans": [ { "bbox": [ 68, 336, 291, 349 ], "score": 1.0, "content": "pairs written by ourselves. We use a pre-trained", "type": "text", "cross_page": true } ], "index": 8 }, { "bbox": [ 68, 349, 291, 362 ], "spans": [ { "bbox": [ 68, 349, 291, 362 ], "score": 1.0, "content": "text-to-speech model 3 to convert the text into cor-", "type": "text", "cross_page": true } ], "index": 9 }, { "bbox": [ 68, 363, 291, 376 ], "spans": [ { "bbox": [ 68, 363, 291, 376 ], "score": 1.0, "content": "responding speech. We then employ mHuBERT", "type": "text", "cross_page": true } ], "index": 10 }, { "bbox": [ 69, 377, 291, 389 ], "spans": [ { "bbox": [ 69, 377, 291, 389 ], "score": 1.0, "content": "to discretize speech data into discrete units as de-", "type": "text", "cross_page": true } ], "index": 11 }, { "bbox": [ 68, 389, 292, 403 ], "spans": [ { "bbox": [ 68, 389, 292, 403 ], "score": 1.0, "content": "scribed in Section 3.1. Ultimately, for each role,", "type": "text", "cross_page": true } ], "index": 12 }, { "bbox": [ 69, 404, 290, 416 ], "spans": [ { "bbox": [ 69, 404, 290, 416 ], "score": 1.0, "content": "we obtained 10 quadruplets composed of speech", "type": "text", "cross_page": true } ], "index": 13 }, { "bbox": [ 69, 418, 290, 430 ], "spans": [ { "bbox": [ 69, 418, 290, 430 ], "score": 1.0, "content": "instructions, text instructions, text responses, and", "type": "text", "cross_page": true } ], "index": 14 }, { "bbox": [ 69, 431, 150, 445 ], "spans": [ { "bbox": [ 69, 431, 150, 445 ], "score": 1.0, "content": "speech responses.", "type": "text", "cross_page": true } ], "index": 15 } ], "index": 91, "bbox_fs": [ 303, 674, 527, 743 ] } ] }, { "preproc_blocks": [ { "type": "image", "bbox": [ 72, 69, 526, 241 ], "blocks": [ { "type": "image_body", "bbox": [ 72, 69, 526, 241 ], "group_id": 0, "lines": [ { "bbox": [ 72, 69, 526, 241 ], "spans": [ { "bbox": [ 72, 69, 526, 241 ], "score": 0.973, "type": "image", "image_path": "18e296fb0af1200b4362fce127e53612728fab4197c4d19c9740c246b9e519ce.jpg" } ] } ], "index": 1, "virtual_lines": [ { "bbox": [ 72, 69, 526, 126.33333333333334 ], "spans": [], "index": 0 }, { "bbox": [ 72, 126.33333333333334, 526, 183.66666666666669 ], "spans": [], "index": 1 }, { "bbox": [ 72, 183.66666666666669, 526, 241.00000000000003 ], "spans": [], "index": 2 } ] }, { "type": "image_caption", "bbox": [ 70, 250, 527, 286 ], "group_id": 0, "lines": [ { "bbox": [ 69, 250, 526, 264 ], "spans": [ { "bbox": [ 69, 250, 526, 264 ], "score": 1.0, "content": "Figure 2: Left: An overview of SpeechInstruct construction process. The SpeechInstruct dataset consists of", "type": "text" } ], "index": 3 }, { "bbox": [ 69, 263, 527, 275 ], "spans": [ { "bbox": [ 69, 263, 527, 275 ], "score": 1.0, "content": "two parts: Cross-modal Instruction data and Chain-of-Modality Instruction data. T emplate1 is shown in 3.1.", "type": "text" } ], "index": 4 }, { "bbox": [ 68, 273, 438, 288 ], "spans": [ { "bbox": [ 68, 273, 438, 288 ], "score": 1.0, "content": "T emplate2 is shown in Appendix C. Right: An illustration of SpeechGPT model structure.", "type": "text" } ], "index": 5 } ], "index": 4 } ], "index": 2.5 }, { "type": "text", "bbox": [ 70, 308, 290, 443 ], "lines": [ { "bbox": [ 69, 308, 291, 322 ], "spans": [ { "bbox": [ 69, 308, 291, 322 ], "score": 1.0, "content": "and educational assistant. For each role, we pro-", "type": "text" } ], "index": 6 }, { "bbox": [ 69, 322, 290, 336 ], "spans": [ { "bbox": [ 69, 322, 290, 336 ], "score": 1.0, "content": "vide 10 manually authored instruction-response", "type": "text" } ], "index": 7 }, { "bbox": [ 68, 336, 291, 349 ], "spans": [ { "bbox": [ 68, 336, 291, 349 ], "score": 1.0, "content": "pairs written by ourselves. We use a pre-trained", "type": "text" } ], "index": 8 }, { "bbox": [ 68, 349, 291, 362 ], "spans": [ { "bbox": [ 68, 349, 291, 362 ], "score": 1.0, "content": "text-to-speech model 3 to convert the text into cor-", "type": "text" } ], "index": 9 }, { "bbox": [ 68, 363, 291, 376 ], "spans": [ { "bbox": [ 68, 363, 291, 376 ], "score": 1.0, "content": "responding speech. We then employ mHuBERT", "type": "text" } ], "index": 10 }, { "bbox": [ 69, 377, 291, 389 ], "spans": [ { "bbox": [ 69, 377, 291, 389 ], "score": 1.0, "content": "to discretize speech data into discrete units as de-", "type": "text" } ], "index": 11 }, { "bbox": [ 68, 389, 292, 403 ], "spans": [ { "bbox": [ 68, 389, 292, 403 ], "score": 1.0, "content": "scribed in Section 3.1. Ultimately, for each role,", "type": "text" } ], "index": 12 }, { "bbox": [ 69, 404, 290, 416 ], "spans": [ { "bbox": [ 69, 404, 290, 416 ], "score": 1.0, "content": "we obtained 10 quadruplets composed of speech", "type": "text" } ], "index": 13 }, { "bbox": [ 69, 418, 290, 430 ], "spans": [ { "bbox": [ 69, 418, 290, 430 ], "score": 1.0, "content": "instructions, text instructions, text responses, and", "type": "text" } ], "index": 14 }, { "bbox": [ 69, 431, 150, 445 ], "spans": [ { "bbox": [ 69, 431, 150, 445 ], "score": 1.0, "content": "speech responses.", "type": "text" } ], "index": 15 } ], "index": 10.5 }, { "type": "title", "bbox": [ 70, 455, 149, 469 ], "lines": [ { "bbox": [ 67, 454, 150, 471 ], "spans": [ { "bbox": [ 67, 454, 150, 471 ], "score": 1.0, "content": "4 SpeechGPT", "type": "text" } ], "index": 16 } ], "index": 16 }, { "type": "title", "bbox": [ 70, 479, 172, 492 ], "lines": [ { "bbox": [ 67, 477, 174, 494 ], "spans": [ { "bbox": [ 67, 477, 174, 494 ], "score": 1.0, "content": "4.1 Model Structure", "type": "text" } ], "index": 17 } ], "index": 17 }, { "type": "text", "bbox": [ 70, 498, 290, 592 ], "lines": [ { "bbox": [ 69, 498, 291, 511 ], "spans": [ { "bbox": [ 69, 498, 291, 511 ], "score": 1.0, "content": "A unified framework is designed to provide archi-", "type": "text" } ], "index": 18 }, { "bbox": [ 68, 512, 292, 524 ], "spans": [ { "bbox": [ 68, 512, 292, 524 ], "score": 1.0, "content": "tecture compatibility across different modalities.", "type": "text" } ], "index": 19 }, { "bbox": [ 69, 526, 291, 538 ], "spans": [ { "bbox": [ 69, 526, 291, 538 ], "score": 1.0, "content": "As shown in Figure 2, our model consists of three", "type": "text" } ], "index": 20 }, { "bbox": [ 69, 540, 290, 552 ], "spans": [ { "bbox": [ 69, 540, 290, 552 ], "score": 1.0, "content": "main components: discrete unit extractor, large", "type": "text" } ], "index": 21 }, { "bbox": [ 69, 553, 291, 564 ], "spans": [ { "bbox": [ 69, 553, 291, 564 ], "score": 1.0, "content": "language modal and unit vocoder. Under this ar-", "type": "text" } ], "index": 22 }, { "bbox": [ 68, 565, 290, 579 ], "spans": [ { "bbox": [ 68, 565, 290, 579 ], "score": 1.0, "content": "chitecture, LLM can perceive multi-modal inputs", "type": "text" } ], "index": 23 }, { "bbox": [ 70, 580, 221, 592 ], "spans": [ { "bbox": [ 70, 580, 221, 592 ], "score": 1.0, "content": "and generate multi-modal outputs.", "type": "text" } ], "index": 24 } ], "index": 21 }, { "type": "text", "bbox": [ 69, 594, 290, 741 ], "lines": [ { "bbox": [ 69, 593, 292, 606 ], "spans": [ { "bbox": [ 69, 593, 292, 606 ], "score": 1.0, "content": "Discrete Unit Extractor The discrete unit ex-", "type": "text" } ], "index": 25 }, { "bbox": [ 69, 606, 291, 620 ], "spans": [ { "bbox": [ 69, 606, 291, 620 ], "score": 1.0, "content": "tractor utilizes the Hidden-unit BERT (HuBERT)", "type": "text" } ], "index": 26 }, { "bbox": [ 69, 621, 290, 633 ], "spans": [ { "bbox": [ 69, 621, 290, 633 ], "score": 1.0, "content": "model (Hsu et al., 2021) to transform continuous", "type": "text" } ], "index": 27 }, { "bbox": [ 69, 635, 291, 648 ], "spans": [ { "bbox": [ 69, 635, 291, 648 ], "score": 1.0, "content": "speech signals into a sequence of discrete units, .", "type": "text" } ], "index": 28 }, { "bbox": [ 69, 648, 290, 660 ], "spans": [ { "bbox": [ 69, 648, 290, 660 ], "score": 1.0, "content": "HuBERT is a self-supervised model that learns", "type": "text" } ], "index": 29 }, { "bbox": [ 68, 662, 290, 674 ], "spans": [ { "bbox": [ 68, 662, 290, 674 ], "score": 1.0, "content": "by predicting discrete labels for masked audio", "type": "text" } ], "index": 30 }, { "bbox": [ 69, 676, 290, 688 ], "spans": [ { "bbox": [ 69, 676, 155, 688 ], "score": 1.0, "content": "segments based on", "type": "text" }, { "bbox": [ 155, 676, 163, 686 ], "score": 0.44, "content": "\\mathbf { k }", "type": "inline_equation" }, { "bbox": [ 163, 676, 290, 688 ], "score": 1.0, "content": "-means clustering applied to", "type": "text" } ], "index": 31 }, { "bbox": [ 69, 689, 292, 701 ], "spans": [ { "bbox": [ 69, 689, 292, 701 ], "score": 1.0, "content": "the model’s intermediate representations. It fea-", "type": "text" } ], "index": 32 }, { "bbox": [ 69, 703, 290, 715 ], "spans": [ { "bbox": [ 69, 703, 290, 715 ], "score": 1.0, "content": "tures a combination of 1-D convolutional layers", "type": "text" } ], "index": 33 }, { "bbox": [ 69, 716, 290, 728 ], "spans": [ { "bbox": [ 69, 716, 290, 728 ], "score": 1.0, "content": "and a Transformer encoder to encode speech into", "type": "text" } ], "index": 34 }, { "bbox": [ 69, 730, 291, 741 ], "spans": [ { "bbox": [ 69, 730, 291, 741 ], "score": 1.0, "content": "continuous intermediate representations, with a k-", "type": "text" } ], "index": 35 } ], "index": 30 }, { "type": "text", "bbox": [ 305, 308, 526, 402 ], "lines": [ { "bbox": [ 303, 308, 527, 322 ], "spans": [ { "bbox": [ 303, 308, 527, 322 ], "score": 1.0, "content": "means model further converting these representa-", "type": "text" } ], "index": 36 }, { "bbox": [ 304, 322, 527, 334 ], "spans": [ { "bbox": [ 304, 322, 527, 334 ], "score": 1.0, "content": "tions into a sequence of cluster indices. Subse-", "type": "text" } ], "index": 37 }, { "bbox": [ 304, 336, 527, 348 ], "spans": [ { "bbox": [ 304, 336, 527, 348 ], "score": 1.0, "content": "quently, adjacent duplicate indices are removed,", "type": "text" } ], "index": 38 }, { "bbox": [ 303, 349, 527, 362 ], "spans": [ { "bbox": [ 303, 349, 527, 362 ], "score": 1.0, "content": "resulting in a discrete units sequence represented", "type": "text" } ], "index": 39 }, { "bbox": [ 303, 362, 527, 376 ], "spans": [ { "bbox": [ 303, 362, 317, 376 ], "score": 1.0, "content": "as", "type": "text" }, { "bbox": [ 318, 362, 419, 376 ], "score": 0.9, "content": "U = ( u _ { 1 } , u _ { 2 } , . . . , u _ { T } )", "type": "inline_equation" }, { "bbox": [ 420, 362, 425, 376 ], "score": 1.0, "content": ",", "type": "text" }, { "bbox": [ 425, 362, 523, 375 ], "score": 0.89, "content": "u _ { i } \\in { 0 , 1 , . . . , K - 1 }", "type": "inline_equation" }, { "bbox": [ 523, 362, 527, 376 ], "score": 1.0, "content": ",", "type": "text" } ], "index": 40 }, { "bbox": [ 304, 376, 527, 389 ], "spans": [ { "bbox": [ 304, 376, 358, 388 ], "score": 0.91, "content": "\\forall 1 \\leq i \\leq T", "type": "inline_equation" }, { "bbox": [ 358, 376, 385, 389 ], "score": 1.0, "content": ", with", "type": "text" }, { "bbox": [ 385, 376, 396, 387 ], "score": 0.77, "content": "K", "type": "inline_equation" }, { "bbox": [ 397, 376, 527, 389 ], "score": 1.0, "content": "denoting the total number of", "type": "text" } ], "index": 41 }, { "bbox": [ 304, 389, 344, 402 ], "spans": [ { "bbox": [ 304, 389, 344, 402 ], "score": 1.0, "content": "clusters.", "type": "text" } ], "index": 42 } ], "index": 39 }, { "type": "text", "bbox": [ 305, 420, 525, 553 ], "lines": [ { "bbox": [ 303, 418, 526, 434 ], "spans": [ { "bbox": [ 303, 418, 526, 434 ], "score": 1.0, "content": "Large Language Model We employ the Meta", "type": "text" } ], "index": 43 }, { "bbox": [ 304, 433, 526, 446 ], "spans": [ { "bbox": [ 304, 433, 526, 446 ], "score": 1.0, "content": "AI LLaMA (Touvron et al., 2023) model as our", "type": "text" } ], "index": 44 }, { "bbox": [ 303, 448, 527, 460 ], "spans": [ { "bbox": [ 303, 448, 527, 460 ], "score": 1.0, "content": "Large Language Model. LLaMA comprises an em-", "type": "text" } ], "index": 45 }, { "bbox": [ 304, 460, 525, 473 ], "spans": [ { "bbox": [ 304, 460, 525, 473 ], "score": 1.0, "content": "bedding layer, multiple transformer blocks, and", "type": "text" } ], "index": 46 }, { "bbox": [ 304, 473, 527, 487 ], "spans": [ { "bbox": [ 304, 473, 527, 487 ], "score": 1.0, "content": "an LM head layer. The total number of parame-", "type": "text" } ], "index": 47 }, { "bbox": [ 303, 487, 526, 501 ], "spans": [ { "bbox": [ 303, 487, 526, 501 ], "score": 1.0, "content": "ters in LLaMA ranges from 7B to 65B. Drawing", "type": "text" } ], "index": 48 }, { "bbox": [ 304, 501, 525, 513 ], "spans": [ { "bbox": [ 304, 501, 525, 513 ], "score": 1.0, "content": "from an extensive training dataset of 1.0 trillion", "type": "text" } ], "index": 49 }, { "bbox": [ 304, 515, 527, 528 ], "spans": [ { "bbox": [ 304, 515, 527, 528 ], "score": 1.0, "content": "tokens, LLaMA demonstrates competitive perfor-", "type": "text" } ], "index": 50 }, { "bbox": [ 303, 528, 525, 542 ], "spans": [ { "bbox": [ 303, 528, 525, 542 ], "score": 1.0, "content": "mance compared to the substantially larger 175B", "type": "text" } ], "index": 51 }, { "bbox": [ 305, 542, 482, 554 ], "spans": [ { "bbox": [ 305, 542, 482, 554 ], "score": 1.0, "content": "GPT-3 across various NLP benchmarks.", "type": "text" } ], "index": 52 } ], "index": 47.5 }, { "type": "text", "bbox": [ 305, 572, 525, 773 ], "lines": [ { "bbox": [ 304, 570, 525, 585 ], "spans": [ { "bbox": [ 304, 570, 525, 585 ], "score": 1.0, "content": "Unit Vocoder Due to limition of single speaker", "type": "text" } ], "index": 53 }, { "bbox": [ 304, 585, 526, 598 ], "spans": [ { "bbox": [ 304, 585, 526, 598 ], "score": 1.0, "content": "unit vocoder in (Polyak et al., 2021), we train a", "type": "text" } ], "index": 54 }, { "bbox": [ 304, 599, 525, 611 ], "spans": [ { "bbox": [ 304, 599, 525, 611 ], "score": 1.0, "content": "multi-speaker unit HiFi-GAN to decode the speech", "type": "text" } ], "index": 55 }, { "bbox": [ 304, 612, 527, 625 ], "spans": [ { "bbox": [ 304, 612, 527, 625 ], "score": 1.0, "content": "signal from the discrete representation. The HiFi-", "type": "text" } ], "index": 56 }, { "bbox": [ 304, 626, 525, 638 ], "spans": [ { "bbox": [ 304, 626, 493, 638 ], "score": 1.0, "content": "GAN architecture consists of a generator", "type": "text" }, { "bbox": [ 494, 626, 505, 637 ], "score": 0.49, "content": "\\mathbf { G }", "type": "inline_equation" }, { "bbox": [ 506, 626, 525, 638 ], "score": 1.0, "content": "and", "type": "text" } ], "index": 57 }, { "bbox": [ 304, 639, 526, 653 ], "spans": [ { "bbox": [ 304, 639, 526, 653 ], "score": 1.0, "content": "multiple discriminators D. The generator uses", "type": "text" } ], "index": 58 }, { "bbox": [ 304, 652, 527, 666 ], "spans": [ { "bbox": [ 304, 652, 527, 666 ], "score": 1.0, "content": "look-up tables (LUT) to embed discrete representa-", "type": "text" } ], "index": 59 }, { "bbox": [ 303, 666, 527, 680 ], "spans": [ { "bbox": [ 303, 666, 527, 680 ], "score": 1.0, "content": "tions and the embedding sequences are up-sampled", "type": "text" } ], "index": 60 }, { "bbox": [ 304, 681, 527, 693 ], "spans": [ { "bbox": [ 304, 681, 527, 693 ], "score": 1.0, "content": "by a series of blocks composed of transposed con-", "type": "text" } ], "index": 61 }, { "bbox": [ 304, 693, 527, 706 ], "spans": [ { "bbox": [ 304, 693, 527, 706 ], "score": 1.0, "content": "volution and a residual block with dilated layers.", "type": "text" } ], "index": 62 }, { "bbox": [ 304, 708, 526, 720 ], "spans": [ { "bbox": [ 304, 708, 526, 720 ], "score": 1.0, "content": "The speaker embedding is concatenated to each", "type": "text" } ], "index": 63 }, { "bbox": [ 304, 720, 527, 733 ], "spans": [ { "bbox": [ 304, 720, 527, 733 ], "score": 1.0, "content": "frame in the up-sampled sequence. The discrimi-", "type": "text" } ], "index": 64 }, { "bbox": [ 304, 735, 526, 747 ], "spans": [ { "bbox": [ 304, 735, 526, 747 ], "score": 1.0, "content": "nator features a Multi-Period Discriminator (MPD)", "type": "text" } ], "index": 65 }, { "bbox": [ 304, 748, 525, 760 ], "spans": [ { "bbox": [ 304, 748, 525, 760 ], "score": 1.0, "content": "and a Multi-Scale Discriminator (MSD), which", "type": "text" } ], "index": 66 }, { "bbox": [ 304, 762, 527, 774 ], "spans": [ { "bbox": [ 304, 762, 527, 774 ], "score": 1.0, "content": "have the same architecture as (Polyak et al., 2021).", "type": "text" } ], "index": 67 } ], "index": 60 } ], "page_idx": 3, "page_size": [ 595, 841 ], "discarded_blocks": [ { "type": "discarded", "bbox": [ 70, 752, 259, 773 ], "lines": [ { "bbox": [ 81, 750, 259, 765 ], "spans": [ { "bbox": [ 81, 750, 259, 765 ], "score": 1.0, "content": "3https://huggingface.co/facebook/", "type": "text" } ] }, { "bbox": [ 69, 762, 196, 774 ], "spans": [ { "bbox": [ 69, 762, 196, 774 ], "score": 1.0, "content": "fastspeech2-en-ljspeech", "type": "text" } ] } ] } ], "para_blocks": [ { "type": "image", "bbox": [ 72, 69, 526, 241 ], "blocks": [ { "type": "image_body", "bbox": [ 72, 69, 526, 241 ], "group_id": 0, "lines": [ { "bbox": [ 72, 69, 526, 241 ], "spans": [ { "bbox": [ 72, 69, 526, 241 ], "score": 0.973, "type": "image", "image_path": "18e296fb0af1200b4362fce127e53612728fab4197c4d19c9740c246b9e519ce.jpg" } ] } ], "index": 1, "virtual_lines": [ { "bbox": [ 72, 69, 526, 126.33333333333334 ], "spans": [], "index": 0 }, { "bbox": [ 72, 126.33333333333334, 526, 183.66666666666669 ], "spans": [], "index": 1 }, { "bbox": [ 72, 183.66666666666669, 526, 241.00000000000003 ], "spans": [], "index": 2 } ] }, { "type": "image_caption", "bbox": [ 70, 250, 527, 286 ], "group_id": 0, "lines": [ { "bbox": [ 69, 250, 526, 264 ], "spans": [ { "bbox": [ 69, 250, 526, 264 ], "score": 1.0, "content": "Figure 2: Left: An overview of SpeechInstruct construction process. The SpeechInstruct dataset consists of", "type": "text" } ], "index": 3 }, { "bbox": [ 69, 263, 527, 275 ], "spans": [ { "bbox": [ 69, 263, 527, 275 ], "score": 1.0, "content": "two parts: Cross-modal Instruction data and Chain-of-Modality Instruction data. T emplate1 is shown in 3.1.", "type": "text" } ], "index": 4 }, { "bbox": [ 68, 273, 438, 288 ], "spans": [ { "bbox": [ 68, 273, 438, 288 ], "score": 1.0, "content": "T emplate2 is shown in Appendix C. Right: An illustration of SpeechGPT model structure.", "type": "text" } ], "index": 5 } ], "index": 4 } ], "index": 2.5 }, { "type": "text", "bbox": [ 70, 308, 290, 443 ], "lines": [], "index": 10.5, "bbox_fs": [ 68, 308, 292, 445 ], "lines_deleted": true }, { "type": "title", "bbox": [ 70, 455, 149, 469 ], "lines": [ { "bbox": [ 67, 454, 150, 471 ], "spans": [ { "bbox": [ 67, 454, 150, 471 ], "score": 1.0, "content": "4 SpeechGPT", "type": "text" } ], "index": 16 } ], "index": 16 }, { "type": "title", "bbox": [ 70, 479, 172, 492 ], "lines": [ { "bbox": [ 67, 477, 174, 494 ], "spans": [ { "bbox": [ 67, 477, 174, 494 ], "score": 1.0, "content": "4.1 Model Structure", "type": "text" } ], "index": 17 } ], "index": 17 }, { "type": "text", "bbox": [ 70, 498, 290, 592 ], "lines": [ { "bbox": [ 69, 498, 291, 511 ], "spans": [ { "bbox": [ 69, 498, 291, 511 ], "score": 1.0, "content": "A unified framework is designed to provide archi-", "type": "text" } ], "index": 18 }, { "bbox": [ 68, 512, 292, 524 ], "spans": [ { "bbox": [ 68, 512, 292, 524 ], "score": 1.0, "content": "tecture compatibility across different modalities.", "type": "text" } ], "index": 19 }, { "bbox": [ 69, 526, 291, 538 ], "spans": [ { "bbox": [ 69, 526, 291, 538 ], "score": 1.0, "content": "As shown in Figure 2, our model consists of three", "type": "text" } ], "index": 20 }, { "bbox": [ 69, 540, 290, 552 ], "spans": [ { "bbox": [ 69, 540, 290, 552 ], "score": 1.0, "content": "main components: discrete unit extractor, large", "type": "text" } ], "index": 21 }, { "bbox": [ 69, 553, 291, 564 ], "spans": [ { "bbox": [ 69, 553, 291, 564 ], "score": 1.0, "content": "language modal and unit vocoder. Under this ar-", "type": "text" } ], "index": 22 }, { "bbox": [ 68, 565, 290, 579 ], "spans": [ { "bbox": [ 68, 565, 290, 579 ], "score": 1.0, "content": "chitecture, LLM can perceive multi-modal inputs", "type": "text" } ], "index": 23 }, { "bbox": [ 70, 580, 221, 592 ], "spans": [ { "bbox": [ 70, 580, 221, 592 ], "score": 1.0, "content": "and generate multi-modal outputs.", "type": "text" } ], "index": 24 } ], "index": 21, "bbox_fs": [ 68, 498, 292, 592 ] }, { "type": "text", "bbox": [ 69, 594, 290, 741 ], "lines": [ { "bbox": [ 69, 593, 292, 606 ], "spans": [ { "bbox": [ 69, 593, 292, 606 ], "score": 1.0, "content": "Discrete Unit Extractor The discrete unit ex-", "type": "text" } ], "index": 25 }, { "bbox": [ 69, 606, 291, 620 ], "spans": [ { "bbox": [ 69, 606, 291, 620 ], "score": 1.0, "content": "tractor utilizes the Hidden-unit BERT (HuBERT)", "type": "text" } ], "index": 26 }, { "bbox": [ 69, 621, 290, 633 ], "spans": [ { "bbox": [ 69, 621, 290, 633 ], "score": 1.0, "content": "model (Hsu et al., 2021) to transform continuous", "type": "text" } ], "index": 27 }, { "bbox": [ 69, 635, 291, 648 ], "spans": [ { "bbox": [ 69, 635, 291, 648 ], "score": 1.0, "content": "speech signals into a sequence of discrete units, .", "type": "text" } ], "index": 28 }, { "bbox": [ 69, 648, 290, 660 ], "spans": [ { "bbox": [ 69, 648, 290, 660 ], "score": 1.0, "content": "HuBERT is a self-supervised model that learns", "type": "text" } ], "index": 29 }, { "bbox": [ 68, 662, 290, 674 ], "spans": [ { "bbox": [ 68, 662, 290, 674 ], "score": 1.0, "content": "by predicting discrete labels for masked audio", "type": "text" } ], "index": 30 }, { "bbox": [ 69, 676, 290, 688 ], "spans": [ { "bbox": [ 69, 676, 155, 688 ], "score": 1.0, "content": "segments based on", "type": "text" }, { "bbox": [ 155, 676, 163, 686 ], "score": 0.44, "content": "\\mathbf { k }", "type": "inline_equation" }, { "bbox": [ 163, 676, 290, 688 ], "score": 1.0, "content": "-means clustering applied to", "type": "text" } ], "index": 31 }, { "bbox": [ 69, 689, 292, 701 ], "spans": [ { "bbox": [ 69, 689, 292, 701 ], "score": 1.0, "content": "the model’s intermediate representations. It fea-", "type": "text" } ], "index": 32 }, { "bbox": [ 69, 703, 290, 715 ], "spans": [ { "bbox": [ 69, 703, 290, 715 ], "score": 1.0, "content": "tures a combination of 1-D convolutional layers", "type": "text" } ], "index": 33 }, { "bbox": [ 69, 716, 290, 728 ], "spans": [ { "bbox": [ 69, 716, 290, 728 ], "score": 1.0, "content": "and a Transformer encoder to encode speech into", "type": "text" } ], "index": 34 }, { "bbox": [ 69, 730, 291, 741 ], "spans": [ { "bbox": [ 69, 730, 291, 741 ], "score": 1.0, "content": "continuous intermediate representations, with a k-", "type": "text" } ], "index": 35 }, { "bbox": [ 303, 308, 527, 322 ], "spans": [ { "bbox": [ 303, 308, 527, 322 ], "score": 1.0, "content": "means model further converting these representa-", "type": "text" } ], "index": 36 }, { "bbox": [ 304, 322, 527, 334 ], "spans": [ { "bbox": [ 304, 322, 527, 334 ], "score": 1.0, "content": "tions into a sequence of cluster indices. Subse-", "type": "text" } ], "index": 37 }, { "bbox": [ 304, 336, 527, 348 ], "spans": [ { "bbox": [ 304, 336, 527, 348 ], "score": 1.0, "content": "quently, adjacent duplicate indices are removed,", "type": "text" } ], "index": 38 }, { "bbox": [ 303, 349, 527, 362 ], "spans": [ { "bbox": [ 303, 349, 527, 362 ], "score": 1.0, "content": "resulting in a discrete units sequence represented", "type": "text" } ], "index": 39 }, { "bbox": [ 303, 362, 527, 376 ], "spans": [ { "bbox": [ 303, 362, 317, 376 ], "score": 1.0, "content": "as", "type": "text" }, { "bbox": [ 318, 362, 419, 376 ], "score": 0.9, "content": "U = ( u _ { 1 } , u _ { 2 } , . . . , u _ { T } )", "type": "inline_equation" }, { "bbox": [ 420, 362, 425, 376 ], "score": 1.0, "content": ",", "type": "text" }, { "bbox": [ 425, 362, 523, 375 ], "score": 0.89, "content": "u _ { i } \\in { 0 , 1 , . . . , K - 1 }", "type": "inline_equation" }, { "bbox": [ 523, 362, 527, 376 ], "score": 1.0, "content": ",", "type": "text" } ], "index": 40 }, { "bbox": [ 304, 376, 527, 389 ], "spans": [ { "bbox": [ 304, 376, 358, 388 ], "score": 0.91, "content": "\\forall 1 \\leq i \\leq T", "type": "inline_equation" }, { "bbox": [ 358, 376, 385, 389 ], "score": 1.0, "content": ", with", "type": "text" }, { "bbox": [ 385, 376, 396, 387 ], "score": 0.77, "content": "K", "type": "inline_equation" }, { "bbox": [ 397, 376, 527, 389 ], "score": 1.0, "content": "denoting the total number of", "type": "text" } ], "index": 41 }, { "bbox": [ 304, 389, 344, 402 ], "spans": [ { "bbox": [ 304, 389, 344, 402 ], "score": 1.0, "content": "clusters.", "type": "text" } ], "index": 42 } ], "index": 30, "bbox_fs": [ 68, 593, 292, 741 ] }, { "type": "text", "bbox": [ 305, 308, 526, 402 ], "lines": [], "index": 39, "bbox_fs": [ 303, 308, 527, 402 ], "lines_deleted": true }, { "type": "text", "bbox": [ 305, 420, 525, 553 ], "lines": [ { "bbox": [ 303, 418, 526, 434 ], "spans": [ { "bbox": [ 303, 418, 526, 434 ], "score": 1.0, "content": "Large Language Model We employ the Meta", "type": "text" } ], "index": 43 }, { "bbox": [ 304, 433, 526, 446 ], "spans": [ { "bbox": [ 304, 433, 526, 446 ], "score": 1.0, "content": "AI LLaMA (Touvron et al., 2023) model as our", "type": "text" } ], "index": 44 }, { "bbox": [ 303, 448, 527, 460 ], "spans": [ { "bbox": [ 303, 448, 527, 460 ], "score": 1.0, "content": "Large Language Model. LLaMA comprises an em-", "type": "text" } ], "index": 45 }, { "bbox": [ 304, 460, 525, 473 ], "spans": [ { "bbox": [ 304, 460, 525, 473 ], "score": 1.0, "content": "bedding layer, multiple transformer blocks, and", "type": "text" } ], "index": 46 }, { "bbox": [ 304, 473, 527, 487 ], "spans": [ { "bbox": [ 304, 473, 527, 487 ], "score": 1.0, "content": "an LM head layer. The total number of parame-", "type": "text" } ], "index": 47 }, { "bbox": [ 303, 487, 526, 501 ], "spans": [ { "bbox": [ 303, 487, 526, 501 ], "score": 1.0, "content": "ters in LLaMA ranges from 7B to 65B. Drawing", "type": "text" } ], "index": 48 }, { "bbox": [ 304, 501, 525, 513 ], "spans": [ { "bbox": [ 304, 501, 525, 513 ], "score": 1.0, "content": "from an extensive training dataset of 1.0 trillion", "type": "text" } ], "index": 49 }, { "bbox": [ 304, 515, 527, 528 ], "spans": [ { "bbox": [ 304, 515, 527, 528 ], "score": 1.0, "content": "tokens, LLaMA demonstrates competitive perfor-", "type": "text" } ], "index": 50 }, { "bbox": [ 303, 528, 525, 542 ], "spans": [ { "bbox": [ 303, 528, 525, 542 ], "score": 1.0, "content": "mance compared to the substantially larger 175B", "type": "text" } ], "index": 51 }, { "bbox": [ 305, 542, 482, 554 ], "spans": [ { "bbox": [ 305, 542, 482, 554 ], "score": 1.0, "content": "GPT-3 across various NLP benchmarks.", "type": "text" } ], "index": 52 } ], "index": 47.5, "bbox_fs": [ 303, 418, 527, 554 ] }, { "type": "text", "bbox": [ 305, 572, 525, 773 ], "lines": [ { "bbox": [ 304, 570, 525, 585 ], "spans": [ { "bbox": [ 304, 570, 525, 585 ], "score": 1.0, "content": "Unit Vocoder Due to limition of single speaker", "type": "text" } ], "index": 53 }, { "bbox": [ 304, 585, 526, 598 ], "spans": [ { "bbox": [ 304, 585, 526, 598 ], "score": 1.0, "content": "unit vocoder in (Polyak et al., 2021), we train a", "type": "text" } ], "index": 54 }, { "bbox": [ 304, 599, 525, 611 ], "spans": [ { "bbox": [ 304, 599, 525, 611 ], "score": 1.0, "content": "multi-speaker unit HiFi-GAN to decode the speech", "type": "text" } ], "index": 55 }, { "bbox": [ 304, 612, 527, 625 ], "spans": [ { "bbox": [ 304, 612, 527, 625 ], "score": 1.0, "content": "signal from the discrete representation. The HiFi-", "type": "text" } ], "index": 56 }, { "bbox": [ 304, 626, 525, 638 ], "spans": [ { "bbox": [ 304, 626, 493, 638 ], "score": 1.0, "content": "GAN architecture consists of a generator", "type": "text" }, { "bbox": [ 494, 626, 505, 637 ], "score": 0.49, "content": "\\mathbf { G }", "type": "inline_equation" }, { "bbox": [ 506, 626, 525, 638 ], "score": 1.0, "content": "and", "type": "text" } ], "index": 57 }, { "bbox": [ 304, 639, 526, 653 ], "spans": [ { "bbox": [ 304, 639, 526, 653 ], "score": 1.0, "content": "multiple discriminators D. The generator uses", "type": "text" } ], "index": 58 }, { "bbox": [ 304, 652, 527, 666 ], "spans": [ { "bbox": [ 304, 652, 527, 666 ], "score": 1.0, "content": "look-up tables (LUT) to embed discrete representa-", "type": "text" } ], "index": 59 }, { "bbox": [ 303, 666, 527, 680 ], "spans": [ { "bbox": [ 303, 666, 527, 680 ], "score": 1.0, "content": "tions and the embedding sequences are up-sampled", "type": "text" } ], "index": 60 }, { "bbox": [ 304, 681, 527, 693 ], "spans": [ { "bbox": [ 304, 681, 527, 693 ], "score": 1.0, "content": "by a series of blocks composed of transposed con-", "type": "text" } ], "index": 61 }, { "bbox": [ 304, 693, 527, 706 ], "spans": [ { "bbox": [ 304, 693, 527, 706 ], "score": 1.0, "content": "volution and a residual block with dilated layers.", "type": "text" } ], "index": 62 }, { "bbox": [ 304, 708, 526, 720 ], "spans": [ { "bbox": [ 304, 708, 526, 720 ], "score": 1.0, "content": "The speaker embedding is concatenated to each", "type": "text" } ], "index": 63 }, { "bbox": [ 304, 720, 527, 733 ], "spans": [ { "bbox": [ 304, 720, 527, 733 ], "score": 1.0, "content": "frame in the up-sampled sequence. The discrimi-", "type": "text" } ], "index": 64 }, { "bbox": [ 304, 735, 526, 747 ], "spans": [ { "bbox": [ 304, 735, 526, 747 ], "score": 1.0, "content": "nator features a Multi-Period Discriminator (MPD)", "type": "text" } ], "index": 65 }, { "bbox": [ 304, 748, 525, 760 ], "spans": [ { "bbox": [ 304, 748, 525, 760 ], "score": 1.0, "content": "and a Multi-Scale Discriminator (MSD), which", "type": "text" } ], "index": 66 }, { "bbox": [ 304, 762, 527, 774 ], "spans": [ { "bbox": [ 304, 762, 527, 774 ], "score": 1.0, "content": "have the same architecture as (Polyak et al., 2021).", "type": "text" } ], "index": 67 } ], "index": 60, "bbox_fs": [ 303, 570, 527, 774 ] } ] }, { "preproc_blocks": [ { "type": "title", "bbox": [ 70, 71, 136, 84 ], "lines": [ { "bbox": [ 67, 68, 138, 88 ], "spans": [ { "bbox": [ 67, 68, 138, 88 ], "score": 1.0, "content": "4.2 Training", "type": "text" } ], "index": 0 } ], "index": 0 }, { "type": "text", "bbox": [ 69, 90, 290, 199 ], "lines": [ { "bbox": [ 68, 91, 290, 104 ], "spans": [ { "bbox": [ 68, 91, 290, 104 ], "score": 1.0, "content": "To incorporate speech discrete representation into", "type": "text" } ], "index": 1 }, { "bbox": [ 68, 104, 290, 118 ], "spans": [ { "bbox": [ 68, 104, 290, 118 ], "score": 1.0, "content": "LLM, we expand the vocabulary and corresponding", "type": "text" } ], "index": 2 }, { "bbox": [ 69, 118, 291, 131 ], "spans": [ { "bbox": [ 69, 118, 291, 131 ], "score": 1.0, "content": "embedding matrix first. We divide the training pro-", "type": "text" } ], "index": 3 }, { "bbox": [ 68, 130, 292, 145 ], "spans": [ { "bbox": [ 68, 130, 292, 145 ], "score": 1.0, "content": "cess into three stages. The first stage is Modality-", "type": "text" } ], "index": 4 }, { "bbox": [ 69, 146, 291, 158 ], "spans": [ { "bbox": [ 69, 146, 291, 158 ], "score": 1.0, "content": "Adaptation Pre-training on unpaired speech data.", "type": "text" } ], "index": 5 }, { "bbox": [ 69, 159, 291, 171 ], "spans": [ { "bbox": [ 69, 159, 291, 171 ], "score": 1.0, "content": "The second stage is Cross-modal Instruction Fine-", "type": "text" } ], "index": 6 }, { "bbox": [ 68, 172, 291, 185 ], "spans": [ { "bbox": [ 68, 172, 291, 185 ], "score": 1.0, "content": "Tuning. The third stage is Chain-of-Modality In-", "type": "text" } ], "index": 7 }, { "bbox": [ 69, 185, 170, 200 ], "spans": [ { "bbox": [ 69, 185, 170, 200 ], "score": 1.0, "content": "struction Fine-Tuning.", "type": "text" } ], "index": 8 } ], "index": 4.5 }, { "type": "text", "bbox": [ 69, 200, 290, 293 ], "lines": [ { "bbox": [ 69, 200, 291, 213 ], "spans": [ { "bbox": [ 69, 200, 291, 213 ], "score": 1.0, "content": "Expanding Vocabulary Given original LLM vo-", "type": "text" } ], "index": 9 }, { "bbox": [ 69, 213, 290, 227 ], "spans": [ { "bbox": [ 69, 214, 109, 226 ], "score": 1.0, "content": "cabulary", "type": "text" }, { "bbox": [ 109, 214, 119, 225 ], "score": 0.69, "content": "V", "type": "inline_equation" }, { "bbox": [ 119, 214, 151, 226 ], "score": 1.0, "content": "of size", "type": "text" }, { "bbox": [ 151, 213, 167, 227 ], "score": 0.91, "content": "| V |", "type": "inline_equation" }, { "bbox": [ 168, 214, 290, 226 ], "score": 1.0, "content": ", to integrate speech discrete", "type": "text" } ], "index": 10 }, { "bbox": [ 68, 227, 291, 240 ], "spans": [ { "bbox": [ 68, 227, 291, 240 ], "score": 1.0, "content": "representations into LLM, we expand the vocab-", "type": "text" } ], "index": 11 }, { "bbox": [ 68, 240, 291, 253 ], "spans": [ { "bbox": [ 68, 240, 261, 253 ], "score": 1.0, "content": "ulary with an additional set of unit tokens", "type": "text" }, { "bbox": [ 261, 240, 273, 252 ], "score": 0.86, "content": "V ^ { \\prime }", "type": "inline_equation" }, { "bbox": [ 274, 240, 291, 253 ], "score": 1.0, "content": ", of", "type": "text" } ], "index": 12 }, { "bbox": [ 68, 253, 291, 268 ], "spans": [ { "bbox": [ 68, 253, 90, 268 ], "score": 1.0, "content": "size", "type": "text" }, { "bbox": [ 90, 254, 136, 267 ], "score": 0.93, "content": "| V ^ { \\prime } | = K", "type": "inline_equation" }, { "bbox": [ 137, 253, 263, 268 ], "score": 1.0, "content": ". The expanded vocabulary", "type": "text" }, { "bbox": [ 263, 254, 278, 266 ], "score": 0.87, "content": "V ^ { \\prime \\prime }", "type": "inline_equation" }, { "bbox": [ 279, 253, 291, 268 ], "score": 1.0, "content": "is", "type": "text" } ], "index": 13 }, { "bbox": [ 68, 267, 291, 281 ], "spans": [ { "bbox": [ 68, 267, 225, 281 ], "score": 1.0, "content": "the union of the original vocabulary", "type": "text" }, { "bbox": [ 225, 268, 235, 279 ], "score": 0.78, "content": "V", "type": "inline_equation" }, { "bbox": [ 236, 267, 291, 281 ], "score": 1.0, "content": "and the new", "type": "text" } ], "index": 14 }, { "bbox": [ 68, 281, 115, 293 ], "spans": [ { "bbox": [ 68, 281, 98, 293 ], "score": 1.0, "content": "words", "type": "text" }, { "bbox": [ 99, 281, 111, 292 ], "score": 0.88, "content": "V ^ { \\prime }", "type": "inline_equation" }, { "bbox": [ 111, 281, 115, 293 ], "score": 1.0, "content": ":", "type": "text" } ], "index": 15 } ], "index": 12 }, { "type": "interline_equation", "bbox": [ 147, 309, 212, 324 ], "lines": [ { "bbox": [ 147, 309, 212, 324 ], "spans": [ { "bbox": [ 147, 309, 212, 324 ], "score": 0.85, "content": "V ^ { \\prime \\prime } = V \\cup V ^ { \\prime }", "type": "interline_equation", "image_path": "ecf3e6ed2da4cca7d61a363613f9a1609345ce3c1d02ce37899c0cd76172bc7f.jpg" } ] } ], "index": 16, "virtual_lines": [ { "bbox": [ 147, 309, 212, 324 ], "spans": [], "index": 16 } ] }, { "type": "text", "bbox": [ 69, 334, 290, 429 ], "lines": [ { "bbox": [ 68, 333, 291, 349 ], "spans": [ { "bbox": [ 68, 333, 291, 349 ], "score": 1.0, "content": "We denote the original word embedding matrix as", "type": "text" } ], "index": 17 }, { "bbox": [ 69, 345, 291, 362 ], "spans": [ { "bbox": [ 69, 347, 127, 360 ], "score": 0.92, "content": "E \\in \\mathbb { R } ^ { | V | \\times d }", "type": "inline_equation" }, { "bbox": [ 128, 345, 163, 362 ], "score": 1.0, "content": ", where", "type": "text" }, { "bbox": [ 164, 349, 171, 360 ], "score": 0.79, "content": "d", "type": "inline_equation" }, { "bbox": [ 172, 345, 291, 362 ], "score": 1.0, "content": "is the dimension of word", "type": "text" } ], "index": 18 }, { "bbox": [ 69, 362, 291, 375 ], "spans": [ { "bbox": [ 69, 362, 291, 375 ], "score": 1.0, "content": "embeddings. To accommodate the expanded vo-", "type": "text" } ], "index": 19 }, { "bbox": [ 68, 375, 291, 389 ], "spans": [ { "bbox": [ 68, 375, 291, 389 ], "score": 1.0, "content": "cabulary, we need to create a randomly initialized", "type": "text" } ], "index": 20 }, { "bbox": [ 67, 386, 293, 403 ], "spans": [ { "bbox": [ 67, 386, 181, 403 ], "score": 1.0, "content": "word embedding matrix", "type": "text" }, { "bbox": [ 181, 387, 246, 401 ], "score": 0.93, "content": "E ^ { \\prime } \\in \\mathbb { R } ^ { | V ^ { \\prime \\prime } | \\times d }", "type": "inline_equation" }, { "bbox": [ 246, 386, 293, 403 ], "score": 1.0, "content": ". We pre-", "type": "text" } ], "index": 21 }, { "bbox": [ 69, 403, 290, 416 ], "spans": [ { "bbox": [ 69, 403, 290, 416 ], "score": 1.0, "content": "serve the original word embeddings by copying the", "type": "text" } ], "index": 22 }, { "bbox": [ 69, 415, 240, 429 ], "spans": [ { "bbox": [ 69, 415, 111, 428 ], "score": 1.0, "content": "values of", "type": "text" }, { "bbox": [ 111, 416, 121, 427 ], "score": 0.85, "content": "E", "type": "inline_equation" }, { "bbox": [ 122, 415, 169, 428 ], "score": 1.0, "content": "to the first", "type": "text" }, { "bbox": [ 170, 416, 186, 429 ], "score": 0.9, "content": "| V |", "type": "inline_equation" }, { "bbox": [ 186, 415, 222, 428 ], "score": 1.0, "content": "rows of", "type": "text" }, { "bbox": [ 223, 416, 235, 427 ], "score": 0.89, "content": "E ^ { \\prime }", "type": "inline_equation" }, { "bbox": [ 236, 415, 240, 428 ], "score": 1.0, "content": ":", "type": "text" } ], "index": 23 } ], "index": 20 }, { "type": "interline_equation", "bbox": [ 139, 444, 220, 460 ], "lines": [ { "bbox": [ 139, 444, 220, 460 ], "spans": [ { "bbox": [ 139, 444, 220, 460 ], "score": 0.9, "content": "E ^ { \\prime } [ 0 : | V | , : ] = E", "type": "interline_equation", "image_path": "2a5ff459674a5e6a9098d06007bb122a5347290ea0bb0eb37dcdfe681b740ee4.jpg" } ] } ], "index": 24, "virtual_lines": [ { "bbox": [ 139, 444, 220, 460 ], "spans": [], "index": 24 } ] }, { "type": "text", "bbox": [ 69, 469, 290, 510 ], "lines": [ { "bbox": [ 69, 470, 290, 483 ], "spans": [ { "bbox": [ 69, 470, 290, 483 ], "score": 1.0, "content": "Finally, we replace the original vocabulary and", "type": "text" } ], "index": 25 }, { "bbox": [ 68, 483, 290, 497 ], "spans": [ { "bbox": [ 68, 483, 290, 497 ], "score": 1.0, "content": "word embedding matrix with the new vocabulary", "type": "text" } ], "index": 26 }, { "bbox": [ 69, 496, 245, 511 ], "spans": [ { "bbox": [ 69, 497, 84, 508 ], "score": 0.85, "content": "V ^ { \\prime \\prime }", "type": "inline_equation" }, { "bbox": [ 85, 496, 228, 511 ], "score": 1.0, "content": "and the word embedding matrix", "type": "text" }, { "bbox": [ 228, 497, 240, 508 ], "score": 0.87, "content": "E ^ { \\prime }", "type": "inline_equation" }, { "bbox": [ 241, 496, 245, 511 ], "score": 1.0, "content": ".", "type": "text" } ], "index": 27 } ], "index": 26 }, { "type": "text", "bbox": [ 69, 511, 290, 619 ], "lines": [ { "bbox": [ 69, 511, 290, 525 ], "spans": [ { "bbox": [ 69, 511, 290, 525 ], "score": 1.0, "content": "Stage 1: Modality-Adaptation Pre-training To", "type": "text" } ], "index": 28 }, { "bbox": [ 68, 524, 290, 539 ], "spans": [ { "bbox": [ 68, 524, 290, 539 ], "score": 1.0, "content": "enable LLM to handle discrete units modality, we", "type": "text" } ], "index": 29 }, { "bbox": [ 69, 539, 290, 551 ], "spans": [ { "bbox": [ 69, 539, 290, 551 ], "score": 1.0, "content": "utilize an unlabeled speech corpus to train LLM in", "type": "text" } ], "index": 30 }, { "bbox": [ 68, 552, 290, 565 ], "spans": [ { "bbox": [ 68, 552, 290, 565 ], "score": 1.0, "content": "a next-token prediction task. This approach aligns", "type": "text" } ], "index": 31 }, { "bbox": [ 69, 566, 290, 579 ], "spans": [ { "bbox": [ 69, 566, 290, 579 ], "score": 1.0, "content": "with the text pre-training objective of LLM. Given", "type": "text" } ], "index": 32 }, { "bbox": [ 68, 578, 290, 594 ], "spans": [ { "bbox": [ 68, 578, 183, 594 ], "score": 1.0, "content": "unlabeled speech corpus", "type": "text" }, { "bbox": [ 183, 579, 194, 590 ], "score": 0.8, "content": "C", "type": "inline_equation" }, { "bbox": [ 194, 578, 290, 594 ], "score": 1.0, "content": "consisting of speech", "type": "text" } ], "index": 33 }, { "bbox": [ 69, 591, 292, 607 ], "spans": [ { "bbox": [ 69, 592, 140, 605 ], "score": 0.92, "content": "U _ { 1 } , U _ { 2 } , \\dots , U _ { m }", "type": "inline_equation" }, { "bbox": [ 140, 591, 232, 607 ], "score": 1.0, "content": "and LLM denoted as", "type": "text" }, { "bbox": [ 232, 593, 246, 605 ], "score": 0.88, "content": "L _ { 1 }", "type": "inline_equation" }, { "bbox": [ 246, 591, 292, 607 ], "score": 1.0, "content": ", the nega-", "type": "text" } ], "index": 34 }, { "bbox": [ 69, 606, 269, 619 ], "spans": [ { "bbox": [ 69, 606, 269, 619 ], "score": 1.0, "content": "tive log-likelihood loss can be formulated as:", "type": "text" } ], "index": 35 } ], "index": 31.5 }, { "type": "interline_equation", "bbox": [ 81, 642, 266, 681 ], "lines": [ { "bbox": [ 81, 642, 266, 681 ], "spans": [ { "bbox": [ 81, 642, 266, 681 ], "score": 0.93, "content": "\\mathcal { L } ( L | C ) = - \\sum _ { j = 1 } ^ { m } \\sum _ { i = 1 } ^ { n _ { j } } \\log P ( u _ { i , j } | u _ { < i , j } ; L )", "type": "interline_equation", "image_path": "5c24d3200a6c1c1a8ab469f2b58bd7a939e61229813d75742509df2d4030eccd.jpg" } ] } ], "index": 36.5, "virtual_lines": [ { "bbox": [ 81, 642, 266, 661.5 ], "spans": [], "index": 36 }, { "bbox": [ 81, 661.5, 266, 681.0 ], "spans": [], "index": 37 } ] }, { "type": "text", "bbox": [ 69, 692, 290, 746 ], "lines": [ { "bbox": [ 68, 691, 289, 707 ], "spans": [ { "bbox": [ 68, 691, 98, 707 ], "score": 1.0, "content": "where", "type": "text" }, { "bbox": [ 99, 695, 110, 704 ], "score": 0.78, "content": "m", "type": "inline_equation" }, { "bbox": [ 110, 691, 262, 707 ], "score": 1.0, "content": "is the number of speech in dataset", "type": "text" }, { "bbox": [ 263, 693, 272, 704 ], "score": 0.7, "content": "C", "type": "inline_equation" }, { "bbox": [ 273, 691, 276, 707 ], "score": 1.0, "content": ",", "type": "text" }, { "bbox": [ 277, 694, 289, 706 ], "score": 0.76, "content": "n _ { j }", "type": "inline_equation" } ], "index": 38 }, { "bbox": [ 68, 705, 291, 720 ], "spans": [ { "bbox": [ 68, 705, 274, 720 ], "score": 1.0, "content": "is the number of discrete unit token in speech", "type": "text" }, { "bbox": [ 274, 707, 287, 720 ], "score": 0.87, "content": "U _ { j }", "type": "inline_equation" }, { "bbox": [ 288, 705, 291, 720 ], "score": 1.0, "content": ",", "type": "text" } ], "index": 39 }, { "bbox": [ 68, 719, 290, 734 ], "spans": [ { "bbox": [ 68, 719, 88, 733 ], "score": 1.0, "content": "and", "type": "text" }, { "bbox": [ 89, 721, 106, 734 ], "score": 0.88, "content": "u _ { i , j }", "type": "inline_equation" }, { "bbox": [ 106, 719, 271, 733 ], "score": 1.0, "content": "represents the i-th unit token in the", "type": "text" }, { "bbox": [ 271, 721, 277, 732 ], "score": 0.3, "content": "\\mathrm { j } \\cdot", "type": "inline_equation" }, { "bbox": [ 277, 719, 290, 733 ], "score": 1.0, "content": "-th", "type": "text" } ], "index": 40 }, { "bbox": [ 67, 733, 105, 747 ], "spans": [ { "bbox": [ 67, 733, 105, 747 ], "score": 1.0, "content": "speech.", "type": "text" } ], "index": 41 } ], "index": 39.5 }, { "type": "text", "bbox": [ 69, 747, 290, 774 ], "lines": [ { "bbox": [ 69, 746, 292, 761 ], "spans": [ { "bbox": [ 69, 746, 292, 761 ], "score": 1.0, "content": "Stage 2: Cross-modal Instruction Fine-", "type": "text" } ], "index": 42 }, { "bbox": [ 69, 760, 291, 775 ], "spans": [ { "bbox": [ 69, 760, 291, 775 ], "score": 1.0, "content": "Tuning In this stage, we align speech and text", "type": "text" } ], "index": 43 } ], "index": 42.5 }, { "type": "text", "bbox": [ 304, 71, 526, 138 ], "lines": [ { "bbox": [ 304, 71, 527, 84 ], "spans": [ { "bbox": [ 304, 71, 527, 84 ], "score": 1.0, "content": "modalities utilizing paired data. We mix Cross-", "type": "text" } ], "index": 44 }, { "bbox": [ 304, 85, 527, 97 ], "spans": [ { "bbox": [ 304, 85, 527, 97 ], "score": 1.0, "content": "modal Instruction in SpeechInstruct with moss-002-", "type": "text" } ], "index": 45 }, { "bbox": [ 304, 99, 526, 111 ], "spans": [ { "bbox": [ 304, 99, 446, 111 ], "score": 1.0, "content": "sft dataset to derive mix dataset", "type": "text" }, { "bbox": [ 447, 99, 454, 110 ], "score": 0.69, "content": "I", "type": "inline_equation" }, { "bbox": [ 454, 99, 526, 111 ], "score": 1.0, "content": ", which consists", "type": "text" } ], "index": 46 }, { "bbox": [ 304, 112, 525, 125 ], "spans": [ { "bbox": [ 304, 113, 353, 125 ], "score": 1.0, "content": "of samples", "type": "text" }, { "bbox": [ 354, 112, 417, 125 ], "score": 0.93, "content": "T _ { 1 } , T _ { 2 } , \\dots , T _ { x }", "type": "inline_equation" }, { "bbox": [ 418, 113, 525, 125 ], "score": 1.0, "content": ". We fine-tune the model", "type": "text" } ], "index": 47 }, { "bbox": [ 304, 126, 464, 139 ], "spans": [ { "bbox": [ 304, 126, 313, 137 ], "score": 0.74, "content": "L", "type": "inline_equation" }, { "bbox": [ 314, 126, 453, 139 ], "score": 1.0, "content": "obtained from the first stage on", "type": "text" }, { "bbox": [ 454, 127, 461, 137 ], "score": 0.75, "content": "I", "type": "inline_equation" }, { "bbox": [ 461, 126, 464, 139 ], "score": 1.0, "content": ".", "type": "text" } ], "index": 48 } ], "index": 46 }, { "type": "text", "bbox": [ 305, 140, 526, 219 ], "lines": [ { "bbox": [ 303, 138, 527, 155 ], "spans": [ { "bbox": [ 303, 138, 369, 155 ], "score": 1.0, "content": "Each sample", "type": "text" }, { "bbox": [ 369, 140, 381, 154 ], "score": 0.88, "content": "T _ { j }", "type": "inline_equation" }, { "bbox": [ 382, 138, 450, 155 ], "score": 1.0, "content": "consisting of", "type": "text" }, { "bbox": [ 450, 140, 511, 154 ], "score": 0.92, "content": "t _ { 1 } , t _ { 2 } , \\ldots , t _ { n _ { j } }", "type": "inline_equation" }, { "bbox": [ 512, 138, 527, 155 ], "score": 1.0, "content": "is", "type": "text" } ], "index": 49 }, { "bbox": [ 304, 154, 525, 166 ], "spans": [ { "bbox": [ 304, 154, 525, 166 ], "score": 1.0, "content": "formed by concatenating a prefix and a text. The", "type": "text" } ], "index": 50 }, { "bbox": [ 303, 166, 527, 181 ], "spans": [ { "bbox": [ 303, 166, 527, 181 ], "score": 1.0, "content": "training objective is to minimize the negative log-", "type": "text" } ], "index": 51 }, { "bbox": [ 303, 180, 526, 194 ], "spans": [ { "bbox": [ 303, 180, 526, 194 ], "score": 1.0, "content": "likelihood and the loss calculation only considers", "type": "text" } ], "index": 52 }, { "bbox": [ 304, 194, 527, 207 ], "spans": [ { "bbox": [ 304, 194, 527, 207 ], "score": 1.0, "content": "the text part, ignoring the prefix, which can be for-", "type": "text" } ], "index": 53 }, { "bbox": [ 304, 207, 349, 220 ], "spans": [ { "bbox": [ 304, 207, 349, 220 ], "score": 1.0, "content": "mated as:", "type": "text" } ], "index": 54 } ], "index": 51.5 }, { "type": "interline_equation", "bbox": [ 313, 240, 504, 281 ], "lines": [ { "bbox": [ 313, 240, 504, 281 ], "spans": [ { "bbox": [ 313, 240, 504, 281 ], "score": 0.93, "content": "\\mathcal { L } ( L | I ) = - \\sum _ { j = 1 } ^ { x } \\sum _ { i = p _ { j } + 1 } ^ { y _ { j } } \\log P ( t _ { i , j } | t _ { < i , j } ; L )", "type": "interline_equation", "image_path": "736c4a0fae6a175438d4a57d6bf4d60b0599e47380be66ebaf270706a95705b4.jpg" } ] } ], "index": 55.5, "virtual_lines": [ { "bbox": [ 313, 240, 504, 260.5 ], "spans": [], "index": 55 }, { "bbox": [ 313, 260.5, 504, 281.0 ], "spans": [], "index": 56 } ] }, { "type": "text", "bbox": [ 304, 290, 525, 344 ], "lines": [ { "bbox": [ 303, 288, 527, 305 ], "spans": [ { "bbox": [ 303, 288, 333, 305 ], "score": 1.0, "content": "where", "type": "text" }, { "bbox": [ 333, 293, 340, 301 ], "score": 0.77, "content": "x", "type": "inline_equation" }, { "bbox": [ 341, 288, 492, 305 ], "score": 1.0, "content": "is the number of samples in corpus", "type": "text" }, { "bbox": [ 493, 291, 500, 301 ], "score": 0.38, "content": "I", "type": "inline_equation" }, { "bbox": [ 500, 288, 503, 305 ], "score": 1.0, "content": ",", "type": "text" }, { "bbox": [ 504, 291, 514, 304 ], "score": 0.61, "content": "y _ { j }", "type": "inline_equation" }, { "bbox": [ 515, 288, 527, 305 ], "score": 1.0, "content": "is", "type": "text" } ], "index": 57 }, { "bbox": [ 304, 303, 525, 318 ], "spans": [ { "bbox": [ 304, 303, 469, 318 ], "score": 1.0, "content": "the total number of tokens in sample", "type": "text" }, { "bbox": [ 470, 304, 497, 317 ], "score": 0.28, "content": "T _ { j } , p _ { j }", "type": "inline_equation" }, { "bbox": [ 498, 303, 525, 318 ], "score": 1.0, "content": "is the", "type": "text" } ], "index": 58 }, { "bbox": [ 303, 316, 524, 332 ], "spans": [ { "bbox": [ 303, 316, 474, 332 ], "score": 1.0, "content": "number of tokens in the prefix part of", "type": "text" }, { "bbox": [ 474, 318, 486, 331 ], "score": 0.89, "content": "T _ { j }", "type": "inline_equation" }, { "bbox": [ 486, 316, 509, 332 ], "score": 1.0, "content": ", and", "type": "text" }, { "bbox": [ 509, 318, 524, 331 ], "score": 0.89, "content": "t _ { i , j }", "type": "inline_equation" } ], "index": 59 }, { "bbox": [ 304, 331, 438, 345 ], "spans": [ { "bbox": [ 304, 331, 421, 345 ], "score": 1.0, "content": "represents the i-th word in", "type": "text" }, { "bbox": [ 422, 331, 434, 345 ], "score": 0.88, "content": "T _ { j }", "type": "inline_equation" }, { "bbox": [ 434, 331, 438, 345 ], "score": 1.0, "content": ".", "type": "text" } ], "index": 60 } ], "index": 58.5 }, { "type": "text", "bbox": [ 304, 345, 526, 464 ], "lines": [ { "bbox": [ 304, 343, 527, 359 ], "spans": [ { "bbox": [ 304, 343, 527, 359 ], "score": 1.0, "content": "Stage 3: Chain-of-Modality Instruction Fine-", "type": "text" } ], "index": 61 }, { "bbox": [ 304, 358, 527, 372 ], "spans": [ { "bbox": [ 304, 358, 527, 372 ], "score": 1.0, "content": "Tuning After obtaining the model in stage 2,", "type": "text" } ], "index": 62 }, { "bbox": [ 303, 371, 527, 386 ], "spans": [ { "bbox": [ 303, 371, 527, 386 ], "score": 1.0, "content": "we utilizes parameter-efficient Low-Rank Adap-", "type": "text" } ], "index": 63 }, { "bbox": [ 304, 385, 526, 398 ], "spans": [ { "bbox": [ 304, 385, 526, 398 ], "score": 1.0, "content": "tation (LoRA) (Hu et al., 2021) to fine-tune it on", "type": "text" } ], "index": 64 }, { "bbox": [ 304, 399, 527, 412 ], "spans": [ { "bbox": [ 304, 399, 527, 412 ], "score": 1.0, "content": "Chain-of-Modality Instruction in SpeechInstruct.", "type": "text" } ], "index": 65 }, { "bbox": [ 304, 412, 525, 426 ], "spans": [ { "bbox": [ 304, 412, 525, 426 ], "score": 1.0, "content": "We add LoRA weights (adapters) to the attention", "type": "text" } ], "index": 66 }, { "bbox": [ 304, 425, 526, 439 ], "spans": [ { "bbox": [ 304, 425, 526, 439 ], "score": 1.0, "content": "mechanisms and train the newly added LoRA pa-", "type": "text" } ], "index": 67 }, { "bbox": [ 303, 439, 525, 453 ], "spans": [ { "bbox": [ 303, 439, 525, 453 ], "score": 1.0, "content": "rameters. We adopt the same loss function as stage", "type": "text" } ], "index": 68 }, { "bbox": [ 304, 453, 316, 466 ], "spans": [ { "bbox": [ 304, 453, 316, 466 ], "score": 1.0, "content": "2.", "type": "text" } ], "index": 69 } ], "index": 65 }, { "type": "title", "bbox": [ 305, 477, 389, 491 ], "lines": [ { "bbox": [ 302, 474, 392, 494 ], "spans": [ { "bbox": [ 302, 474, 392, 494 ], "score": 1.0, "content": "5 Experiments", "type": "text" } ], "index": 70 } ], "index": 70 }, { "type": "title", "bbox": [ 305, 498, 428, 512 ], "lines": [ { "bbox": [ 303, 497, 429, 515 ], "spans": [ { "bbox": [ 303, 497, 429, 515 ], "score": 1.0, "content": "5.1 Experimental Setups", "type": "text" } ], "index": 71 } ], "index": 71 }, { "type": "text", "bbox": [ 304, 517, 525, 665 ], "lines": [ { "bbox": [ 304, 515, 525, 531 ], "spans": [ { "bbox": [ 304, 515, 525, 531 ], "score": 1.0, "content": "Datasets For modality-adaption pre-training, we", "type": "text" } ], "index": 72 }, { "bbox": [ 304, 530, 525, 543 ], "spans": [ { "bbox": [ 304, 530, 525, 543 ], "score": 1.0, "content": "use LibriLight (Kahn et al., 2020) which contains", "type": "text" } ], "index": 73 }, { "bbox": [ 304, 544, 527, 557 ], "spans": [ { "bbox": [ 304, 544, 527, 557 ], "score": 1.0, "content": "60K hours of unlabelled English audiobook speech.", "type": "text" } ], "index": 74 }, { "bbox": [ 304, 558, 526, 571 ], "spans": [ { "bbox": [ 304, 558, 526, 571 ], "score": 1.0, "content": "For cross-modal instruction fine-tuning stage, we", "type": "text" } ], "index": 75 }, { "bbox": [ 304, 572, 525, 584 ], "spans": [ { "bbox": [ 304, 572, 525, 584 ], "score": 1.0, "content": "use Gigaspeech (Chen et al., 2021), Common", "type": "text" } ], "index": 76 }, { "bbox": [ 304, 583, 527, 599 ], "spans": [ { "bbox": [ 304, 583, 527, 599 ], "score": 1.0, "content": "voice (Ardila et al., 2020) and LibriSpeech (Panay-", "type": "text" } ], "index": 77 }, { "bbox": [ 304, 598, 526, 610 ], "spans": [ { "bbox": [ 304, 598, 526, 610 ], "score": 1.0, "content": "otov et al., 2015) dataset and moss-002-sft-data", "type": "text" } ], "index": 78 }, { "bbox": [ 304, 612, 526, 624 ], "spans": [ { "bbox": [ 304, 612, 526, 624 ], "score": 1.0, "content": "dataset, which is illustrated in detail in 3.1. For", "type": "text" } ], "index": 79 }, { "bbox": [ 304, 625, 526, 639 ], "spans": [ { "bbox": [ 304, 625, 526, 639 ], "score": 1.0, "content": "chain-of-modality instruction fine-tuning stage, we", "type": "text" } ], "index": 80 }, { "bbox": [ 304, 639, 525, 650 ], "spans": [ { "bbox": [ 304, 639, 525, 650 ], "score": 1.0, "content": "use moss-002-sft-data dataset, which is illustrated", "type": "text" } ], "index": 81 }, { "bbox": [ 304, 653, 372, 664 ], "spans": [ { "bbox": [ 304, 653, 372, 664 ], "score": 1.0, "content": "in detail in 3.2.", "type": "text" } ], "index": 82 } ], "index": 77 }, { "type": "text", "bbox": [ 304, 666, 525, 774 ], "lines": [ { "bbox": [ 305, 666, 526, 679 ], "spans": [ { "bbox": [ 305, 666, 526, 679 ], "score": 1.0, "content": "Configuration We employ LLaMA-13B (Touvron", "type": "text" } ], "index": 83 }, { "bbox": [ 304, 680, 525, 692 ], "spans": [ { "bbox": [ 304, 680, 525, 692 ], "score": 1.0, "content": "et al., 2023) as our backbone model for a trade-off", "type": "text" } ], "index": 84 }, { "bbox": [ 304, 693, 525, 706 ], "spans": [ { "bbox": [ 304, 693, 525, 706 ], "score": 1.0, "content": "between performance and computational resources", "type": "text" } ], "index": 85 }, { "bbox": [ 304, 707, 525, 719 ], "spans": [ { "bbox": [ 304, 707, 525, 719 ], "score": 1.0, "content": "available. For stage 1, we use 96 A100 GPUs and", "type": "text" } ], "index": 86 }, { "bbox": [ 304, 720, 525, 734 ], "spans": [ { "bbox": [ 304, 720, 525, 734 ], "score": 1.0, "content": "train for 900 steps with batch size 768. For stage", "type": "text" } ], "index": 87 }, { "bbox": [ 304, 734, 525, 747 ], "spans": [ { "bbox": [ 304, 734, 525, 747 ], "score": 1.0, "content": "2, we use 96 A100 GPUs and train for 2100 steps", "type": "text" } ], "index": 88 }, { "bbox": [ 304, 747, 525, 760 ], "spans": [ { "bbox": [ 304, 747, 525, 760 ], "score": 1.0, "content": "with batch size 1536. For stage 3, we use 8 A100", "type": "text" } ], "index": 89 }, { "bbox": [ 305, 761, 527, 773 ], "spans": [ { "bbox": [ 305, 761, 527, 773 ], "score": 1.0, "content": "GPUs and train for 4200 steps with batch size 128.", "type": "text" } ], "index": 90 } ], "index": 86.5 } ], "page_idx": 4, "page_size": [ 595, 841 ], "discarded_blocks": [], "para_blocks": [ { "type": "title", "bbox": [ 70, 71, 136, 84 ], "lines": [ { "bbox": [ 67, 68, 138, 88 ], "spans": [ { "bbox": [ 67, 68, 138, 88 ], "score": 1.0, "content": "4.2 Training", "type": "text" } ], "index": 0 } ], "index": 0 }, { "type": "text", "bbox": [ 69, 90, 290, 199 ], "lines": [ { "bbox": [ 68, 91, 290, 104 ], "spans": [ { "bbox": [ 68, 91, 290, 104 ], "score": 1.0, "content": "To incorporate speech discrete representation into", "type": "text" } ], "index": 1 }, { "bbox": [ 68, 104, 290, 118 ], "spans": [ { "bbox": [ 68, 104, 290, 118 ], "score": 1.0, "content": "LLM, we expand the vocabulary and corresponding", "type": "text" } ], "index": 2 }, { "bbox": [ 69, 118, 291, 131 ], "spans": [ { "bbox": [ 69, 118, 291, 131 ], "score": 1.0, "content": "embedding matrix first. We divide the training pro-", "type": "text" } ], "index": 3 }, { "bbox": [ 68, 130, 292, 145 ], "spans": [ { "bbox": [ 68, 130, 292, 145 ], "score": 1.0, "content": "cess into three stages. The first stage is Modality-", "type": "text" } ], "index": 4 }, { "bbox": [ 69, 146, 291, 158 ], "spans": [ { "bbox": [ 69, 146, 291, 158 ], "score": 1.0, "content": "Adaptation Pre-training on unpaired speech data.", "type": "text" } ], "index": 5 }, { "bbox": [ 69, 159, 291, 171 ], "spans": [ { "bbox": [ 69, 159, 291, 171 ], "score": 1.0, "content": "The second stage is Cross-modal Instruction Fine-", "type": "text" } ], "index": 6 }, { "bbox": [ 68, 172, 291, 185 ], "spans": [ { "bbox": [ 68, 172, 291, 185 ], "score": 1.0, "content": "Tuning. The third stage is Chain-of-Modality In-", "type": "text" } ], "index": 7 }, { "bbox": [ 69, 185, 170, 200 ], "spans": [ { "bbox": [ 69, 185, 170, 200 ], "score": 1.0, "content": "struction Fine-Tuning.", "type": "text" } ], "index": 8 } ], "index": 4.5, "bbox_fs": [ 68, 91, 292, 200 ] }, { "type": "text", "bbox": [ 69, 200, 290, 293 ], "lines": [ { "bbox": [ 69, 200, 291, 213 ], "spans": [ { "bbox": [ 69, 200, 291, 213 ], "score": 1.0, "content": "Expanding Vocabulary Given original LLM vo-", "type": "text" } ], "index": 9 }, { "bbox": [ 69, 213, 290, 227 ], "spans": [ { "bbox": [ 69, 214, 109, 226 ], "score": 1.0, "content": "cabulary", "type": "text" }, { "bbox": [ 109, 214, 119, 225 ], "score": 0.69, "content": "V", "type": "inline_equation" }, { "bbox": [ 119, 214, 151, 226 ], "score": 1.0, "content": "of size", "type": "text" }, { "bbox": [ 151, 213, 167, 227 ], "score": 0.91, "content": "| V |", "type": "inline_equation" }, { "bbox": [ 168, 214, 290, 226 ], "score": 1.0, "content": ", to integrate speech discrete", "type": "text" } ], "index": 10 }, { "bbox": [ 68, 227, 291, 240 ], "spans": [ { "bbox": [ 68, 227, 291, 240 ], "score": 1.0, "content": "representations into LLM, we expand the vocab-", "type": "text" } ], "index": 11 }, { "bbox": [ 68, 240, 291, 253 ], "spans": [ { "bbox": [ 68, 240, 261, 253 ], "score": 1.0, "content": "ulary with an additional set of unit tokens", "type": "text" }, { "bbox": [ 261, 240, 273, 252 ], "score": 0.86, "content": "V ^ { \\prime }", "type": "inline_equation" }, { "bbox": [ 274, 240, 291, 253 ], "score": 1.0, "content": ", of", "type": "text" } ], "index": 12 }, { "bbox": [ 68, 253, 291, 268 ], "spans": [ { "bbox": [ 68, 253, 90, 268 ], "score": 1.0, "content": "size", "type": "text" }, { "bbox": [ 90, 254, 136, 267 ], "score": 0.93, "content": "| V ^ { \\prime } | = K", "type": "inline_equation" }, { "bbox": [ 137, 253, 263, 268 ], "score": 1.0, "content": ". The expanded vocabulary", "type": "text" }, { "bbox": [ 263, 254, 278, 266 ], "score": 0.87, "content": "V ^ { \\prime \\prime }", "type": "inline_equation" }, { "bbox": [ 279, 253, 291, 268 ], "score": 1.0, "content": "is", "type": "text" } ], "index": 13 }, { "bbox": [ 68, 267, 291, 281 ], "spans": [ { "bbox": [ 68, 267, 225, 281 ], "score": 1.0, "content": "the union of the original vocabulary", "type": "text" }, { "bbox": [ 225, 268, 235, 279 ], "score": 0.78, "content": "V", "type": "inline_equation" }, { "bbox": [ 236, 267, 291, 281 ], "score": 1.0, "content": "and the new", "type": "text" } ], "index": 14 }, { "bbox": [ 68, 281, 115, 293 ], "spans": [ { "bbox": [ 68, 281, 98, 293 ], "score": 1.0, "content": "words", "type": "text" }, { "bbox": [ 99, 281, 111, 292 ], "score": 0.88, "content": "V ^ { \\prime }", "type": "inline_equation" }, { "bbox": [ 111, 281, 115, 293 ], "score": 1.0, "content": ":", "type": "text" } ], "index": 15 } ], "index": 12, "bbox_fs": [ 68, 200, 291, 293 ] }, { "type": "interline_equation", "bbox": [ 147, 309, 212, 324 ], "lines": [ { "bbox": [ 147, 309, 212, 324 ], "spans": [ { "bbox": [ 147, 309, 212, 324 ], "score": 0.85, "content": "V ^ { \\prime \\prime } = V \\cup V ^ { \\prime }", "type": "interline_equation", "image_path": "ecf3e6ed2da4cca7d61a363613f9a1609345ce3c1d02ce37899c0cd76172bc7f.jpg" } ] } ], "index": 16, "virtual_lines": [ { "bbox": [ 147, 309, 212, 324 ], "spans": [], "index": 16 } ] }, { "type": "text", "bbox": [ 69, 334, 290, 429 ], "lines": [ { "bbox": [ 68, 333, 291, 349 ], "spans": [ { "bbox": [ 68, 333, 291, 349 ], "score": 1.0, "content": "We denote the original word embedding matrix as", "type": "text" } ], "index": 17 }, { "bbox": [ 69, 345, 291, 362 ], "spans": [ { "bbox": [ 69, 347, 127, 360 ], "score": 0.92, "content": "E \\in \\mathbb { R } ^ { | V | \\times d }", "type": "inline_equation" }, { "bbox": [ 128, 345, 163, 362 ], "score": 1.0, "content": ", where", "type": "text" }, { "bbox": [ 164, 349, 171, 360 ], "score": 0.79, "content": "d", "type": "inline_equation" }, { "bbox": [ 172, 345, 291, 362 ], "score": 1.0, "content": "is the dimension of word", "type": "text" } ], "index": 18 }, { "bbox": [ 69, 362, 291, 375 ], "spans": [ { "bbox": [ 69, 362, 291, 375 ], "score": 1.0, "content": "embeddings. To accommodate the expanded vo-", "type": "text" } ], "index": 19 }, { "bbox": [ 68, 375, 291, 389 ], "spans": [ { "bbox": [ 68, 375, 291, 389 ], "score": 1.0, "content": "cabulary, we need to create a randomly initialized", "type": "text" } ], "index": 20 }, { "bbox": [ 67, 386, 293, 403 ], "spans": [ { "bbox": [ 67, 386, 181, 403 ], "score": 1.0, "content": "word embedding matrix", "type": "text" }, { "bbox": [ 181, 387, 246, 401 ], "score": 0.93, "content": "E ^ { \\prime } \\in \\mathbb { R } ^ { | V ^ { \\prime \\prime } | \\times d }", "type": "inline_equation" }, { "bbox": [ 246, 386, 293, 403 ], "score": 1.0, "content": ". We pre-", "type": "text" } ], "index": 21 }, { "bbox": [ 69, 403, 290, 416 ], "spans": [ { "bbox": [ 69, 403, 290, 416 ], "score": 1.0, "content": "serve the original word embeddings by copying the", "type": "text" } ], "index": 22 }, { "bbox": [ 69, 415, 240, 429 ], "spans": [ { "bbox": [ 69, 415, 111, 428 ], "score": 1.0, "content": "values of", "type": "text" }, { "bbox": [ 111, 416, 121, 427 ], "score": 0.85, "content": "E", "type": "inline_equation" }, { "bbox": [ 122, 415, 169, 428 ], "score": 1.0, "content": "to the first", "type": "text" }, { "bbox": [ 170, 416, 186, 429 ], "score": 0.9, "content": "| V |", "type": "inline_equation" }, { "bbox": [ 186, 415, 222, 428 ], "score": 1.0, "content": "rows of", "type": "text" }, { "bbox": [ 223, 416, 235, 427 ], "score": 0.89, "content": "E ^ { \\prime }", "type": "inline_equation" }, { "bbox": [ 236, 415, 240, 428 ], "score": 1.0, "content": ":", "type": "text" } ], "index": 23 } ], "index": 20, "bbox_fs": [ 67, 333, 293, 429 ] }, { "type": "interline_equation", "bbox": [ 139, 444, 220, 460 ], "lines": [ { "bbox": [ 139, 444, 220, 460 ], "spans": [ { "bbox": [ 139, 444, 220, 460 ], "score": 0.9, "content": "E ^ { \\prime } [ 0 : | V | , : ] = E", "type": "interline_equation", "image_path": "2a5ff459674a5e6a9098d06007bb122a5347290ea0bb0eb37dcdfe681b740ee4.jpg" } ] } ], "index": 24, "virtual_lines": [ { "bbox": [ 139, 444, 220, 460 ], "spans": [], "index": 24 } ] }, { "type": "text", "bbox": [ 69, 469, 290, 510 ], "lines": [ { "bbox": [ 69, 470, 290, 483 ], "spans": [ { "bbox": [ 69, 470, 290, 483 ], "score": 1.0, "content": "Finally, we replace the original vocabulary and", "type": "text" } ], "index": 25 }, { "bbox": [ 68, 483, 290, 497 ], "spans": [ { "bbox": [ 68, 483, 290, 497 ], "score": 1.0, "content": "word embedding matrix with the new vocabulary", "type": "text" } ], "index": 26 }, { "bbox": [ 69, 496, 245, 511 ], "spans": [ { "bbox": [ 69, 497, 84, 508 ], "score": 0.85, "content": "V ^ { \\prime \\prime }", "type": "inline_equation" }, { "bbox": [ 85, 496, 228, 511 ], "score": 1.0, "content": "and the word embedding matrix", "type": "text" }, { "bbox": [ 228, 497, 240, 508 ], "score": 0.87, "content": "E ^ { \\prime }", "type": "inline_equation" }, { "bbox": [ 241, 496, 245, 511 ], "score": 1.0, "content": ".", "type": "text" } ], "index": 27 } ], "index": 26, "bbox_fs": [ 68, 470, 290, 511 ] }, { "type": "text", "bbox": [ 69, 511, 290, 619 ], "lines": [ { "bbox": [ 69, 511, 290, 525 ], "spans": [ { "bbox": [ 69, 511, 290, 525 ], "score": 1.0, "content": "Stage 1: Modality-Adaptation Pre-training To", "type": "text" } ], "index": 28 }, { "bbox": [ 68, 524, 290, 539 ], "spans": [ { "bbox": [ 68, 524, 290, 539 ], "score": 1.0, "content": "enable LLM to handle discrete units modality, we", "type": "text" } ], "index": 29 }, { "bbox": [ 69, 539, 290, 551 ], "spans": [ { "bbox": [ 69, 539, 290, 551 ], "score": 1.0, "content": "utilize an unlabeled speech corpus to train LLM in", "type": "text" } ], "index": 30 }, { "bbox": [ 68, 552, 290, 565 ], "spans": [ { "bbox": [ 68, 552, 290, 565 ], "score": 1.0, "content": "a next-token prediction task. This approach aligns", "type": "text" } ], "index": 31 }, { "bbox": [ 69, 566, 290, 579 ], "spans": [ { "bbox": [ 69, 566, 290, 579 ], "score": 1.0, "content": "with the text pre-training objective of LLM. Given", "type": "text" } ], "index": 32 }, { "bbox": [ 68, 578, 290, 594 ], "spans": [ { "bbox": [ 68, 578, 183, 594 ], "score": 1.0, "content": "unlabeled speech corpus", "type": "text" }, { "bbox": [ 183, 579, 194, 590 ], "score": 0.8, "content": "C", "type": "inline_equation" }, { "bbox": [ 194, 578, 290, 594 ], "score": 1.0, "content": "consisting of speech", "type": "text" } ], "index": 33 }, { "bbox": [ 69, 591, 292, 607 ], "spans": [ { "bbox": [ 69, 592, 140, 605 ], "score": 0.92, "content": "U _ { 1 } , U _ { 2 } , \\dots , U _ { m }", "type": "inline_equation" }, { "bbox": [ 140, 591, 232, 607 ], "score": 1.0, "content": "and LLM denoted as", "type": "text" }, { "bbox": [ 232, 593, 246, 605 ], "score": 0.88, "content": "L _ { 1 }", "type": "inline_equation" }, { "bbox": [ 246, 591, 292, 607 ], "score": 1.0, "content": ", the nega-", "type": "text" } ], "index": 34 }, { "bbox": [ 69, 606, 269, 619 ], "spans": [ { "bbox": [ 69, 606, 269, 619 ], "score": 1.0, "content": "tive log-likelihood loss can be formulated as:", "type": "text" } ], "index": 35 } ], "index": 31.5, "bbox_fs": [ 68, 511, 292, 619 ] }, { "type": "interline_equation", "bbox": [ 81, 642, 266, 681 ], "lines": [ { "bbox": [ 81, 642, 266, 681 ], "spans": [ { "bbox": [ 81, 642, 266, 681 ], "score": 0.93, "content": "\\mathcal { L } ( L | C ) = - \\sum _ { j = 1 } ^ { m } \\sum _ { i = 1 } ^ { n _ { j } } \\log P ( u _ { i , j } | u _ { < i , j } ; L )", "type": "interline_equation", "image_path": "5c24d3200a6c1c1a8ab469f2b58bd7a939e61229813d75742509df2d4030eccd.jpg" } ] } ], "index": 36.5, "virtual_lines": [ { "bbox": [ 81, 642, 266, 661.5 ], "spans": [], "index": 36 }, { "bbox": [ 81, 661.5, 266, 681.0 ], "spans": [], "index": 37 } ] }, { "type": "text", "bbox": [ 69, 692, 290, 746 ], "lines": [ { "bbox": [ 68, 691, 289, 707 ], "spans": [ { "bbox": [ 68, 691, 98, 707 ], "score": 1.0, "content": "where", "type": "text" }, { "bbox": [ 99, 695, 110, 704 ], "score": 0.78, "content": "m", "type": "inline_equation" }, { "bbox": [ 110, 691, 262, 707 ], "score": 1.0, "content": "is the number of speech in dataset", "type": "text" }, { "bbox": [ 263, 693, 272, 704 ], "score": 0.7, "content": "C", "type": "inline_equation" }, { "bbox": [ 273, 691, 276, 707 ], "score": 1.0, "content": ",", "type": "text" }, { "bbox": [ 277, 694, 289, 706 ], "score": 0.76, "content": "n _ { j }", "type": "inline_equation" } ], "index": 38 }, { "bbox": [ 68, 705, 291, 720 ], "spans": [ { "bbox": [ 68, 705, 274, 720 ], "score": 1.0, "content": "is the number of discrete unit token in speech", "type": "text" }, { "bbox": [ 274, 707, 287, 720 ], "score": 0.87, "content": "U _ { j }", "type": "inline_equation" }, { "bbox": [ 288, 705, 291, 720 ], "score": 1.0, "content": ",", "type": "text" } ], "index": 39 }, { "bbox": [ 68, 719, 290, 734 ], "spans": [ { "bbox": [ 68, 719, 88, 733 ], "score": 1.0, "content": "and", "type": "text" }, { "bbox": [ 89, 721, 106, 734 ], "score": 0.88, "content": "u _ { i , j }", "type": "inline_equation" }, { "bbox": [ 106, 719, 271, 733 ], "score": 1.0, "content": "represents the i-th unit token in the", "type": "text" }, { "bbox": [ 271, 721, 277, 732 ], "score": 0.3, "content": "\\mathrm { j } \\cdot", "type": "inline_equation" }, { "bbox": [ 277, 719, 290, 733 ], "score": 1.0, "content": "-th", "type": "text" } ], "index": 40 }, { "bbox": [ 67, 733, 105, 747 ], "spans": [ { "bbox": [ 67, 733, 105, 747 ], "score": 1.0, "content": "speech.", "type": "text" } ], "index": 41 } ], "index": 39.5, "bbox_fs": [ 67, 691, 291, 747 ] }, { "type": "text", "bbox": [ 69, 747, 290, 774 ], "lines": [ { "bbox": [ 69, 746, 292, 761 ], "spans": [ { "bbox": [ 69, 746, 292, 761 ], "score": 1.0, "content": "Stage 2: Cross-modal Instruction Fine-", "type": "text" } ], "index": 42 }, { "bbox": [ 69, 760, 291, 775 ], "spans": [ { "bbox": [ 69, 760, 291, 775 ], "score": 1.0, "content": "Tuning In this stage, we align speech and text", "type": "text" } ], "index": 43 }, { "bbox": [ 304, 71, 527, 84 ], "spans": [ { "bbox": [ 304, 71, 527, 84 ], "score": 1.0, "content": "modalities utilizing paired data. We mix Cross-", "type": "text" } ], "index": 44 }, { "bbox": [ 304, 85, 527, 97 ], "spans": [ { "bbox": [ 304, 85, 527, 97 ], "score": 1.0, "content": "modal Instruction in SpeechInstruct with moss-002-", "type": "text" } ], "index": 45 }, { "bbox": [ 304, 99, 526, 111 ], "spans": [ { "bbox": [ 304, 99, 446, 111 ], "score": 1.0, "content": "sft dataset to derive mix dataset", "type": "text" }, { "bbox": [ 447, 99, 454, 110 ], "score": 0.69, "content": "I", "type": "inline_equation" }, { "bbox": [ 454, 99, 526, 111 ], "score": 1.0, "content": ", which consists", "type": "text" } ], "index": 46 }, { "bbox": [ 304, 112, 525, 125 ], "spans": [ { "bbox": [ 304, 113, 353, 125 ], "score": 1.0, "content": "of samples", "type": "text" }, { "bbox": [ 354, 112, 417, 125 ], "score": 0.93, "content": "T _ { 1 } , T _ { 2 } , \\dots , T _ { x }", "type": "inline_equation" }, { "bbox": [ 418, 113, 525, 125 ], "score": 1.0, "content": ". We fine-tune the model", "type": "text" } ], "index": 47 }, { "bbox": [ 304, 126, 464, 139 ], "spans": [ { "bbox": [ 304, 126, 313, 137 ], "score": 0.74, "content": "L", "type": "inline_equation" }, { "bbox": [ 314, 126, 453, 139 ], "score": 1.0, "content": "obtained from the first stage on", "type": "text" }, { "bbox": [ 454, 127, 461, 137 ], "score": 0.75, "content": "I", "type": "inline_equation" }, { "bbox": [ 461, 126, 464, 139 ], "score": 1.0, "content": ".", "type": "text" } ], "index": 48 } ], "index": 42.5, "bbox_fs": [ 69, 746, 292, 775 ] }, { "type": "text", "bbox": [ 304, 71, 526, 138 ], "lines": [], "index": 46, "bbox_fs": [ 304, 71, 527, 139 ], "lines_deleted": true }, { "type": "text", "bbox": [ 305, 140, 526, 219 ], "lines": [ { "bbox": [ 303, 138, 527, 155 ], "spans": [ { "bbox": [ 303, 138, 369, 155 ], "score": 1.0, "content": "Each sample", "type": "text" }, { "bbox": [ 369, 140, 381, 154 ], "score": 0.88, "content": "T _ { j }", "type": "inline_equation" }, { "bbox": [ 382, 138, 450, 155 ], "score": 1.0, "content": "consisting of", "type": "text" }, { "bbox": [ 450, 140, 511, 154 ], "score": 0.92, "content": "t _ { 1 } , t _ { 2 } , \\ldots , t _ { n _ { j } }", "type": "inline_equation" }, { "bbox": [ 512, 138, 527, 155 ], "score": 1.0, "content": "is", "type": "text" } ], "index": 49 }, { "bbox": [ 304, 154, 525, 166 ], "spans": [ { "bbox": [ 304, 154, 525, 166 ], "score": 1.0, "content": "formed by concatenating a prefix and a text. The", "type": "text" } ], "index": 50 }, { "bbox": [ 303, 166, 527, 181 ], "spans": [ { "bbox": [ 303, 166, 527, 181 ], "score": 1.0, "content": "training objective is to minimize the negative log-", "type": "text" } ], "index": 51 }, { "bbox": [ 303, 180, 526, 194 ], "spans": [ { "bbox": [ 303, 180, 526, 194 ], "score": 1.0, "content": "likelihood and the loss calculation only considers", "type": "text" } ], "index": 52 }, { "bbox": [ 304, 194, 527, 207 ], "spans": [ { "bbox": [ 304, 194, 527, 207 ], "score": 1.0, "content": "the text part, ignoring the prefix, which can be for-", "type": "text" } ], "index": 53 }, { "bbox": [ 304, 207, 349, 220 ], "spans": [ { "bbox": [ 304, 207, 349, 220 ], "score": 1.0, "content": "mated as:", "type": "text" } ], "index": 54 } ], "index": 51.5, "bbox_fs": [ 303, 138, 527, 220 ] }, { "type": "interline_equation", "bbox": [ 313, 240, 504, 281 ], "lines": [ { "bbox": [ 313, 240, 504, 281 ], "spans": [ { "bbox": [ 313, 240, 504, 281 ], "score": 0.93, "content": "\\mathcal { L } ( L | I ) = - \\sum _ { j = 1 } ^ { x } \\sum _ { i = p _ { j } + 1 } ^ { y _ { j } } \\log P ( t _ { i , j } | t _ { < i , j } ; L )", "type": "interline_equation", "image_path": "736c4a0fae6a175438d4a57d6bf4d60b0599e47380be66ebaf270706a95705b4.jpg" } ] } ], "index": 55.5, "virtual_lines": [ { "bbox": [ 313, 240, 504, 260.5 ], "spans": [], "index": 55 }, { "bbox": [ 313, 260.5, 504, 281.0 ], "spans": [], "index": 56 } ] }, { "type": "text", "bbox": [ 304, 290, 525, 344 ], "lines": [ { "bbox": [ 303, 288, 527, 305 ], "spans": [ { "bbox": [ 303, 288, 333, 305 ], "score": 1.0, "content": "where", "type": "text" }, { "bbox": [ 333, 293, 340, 301 ], "score": 0.77, "content": "x", "type": "inline_equation" }, { "bbox": [ 341, 288, 492, 305 ], "score": 1.0, "content": "is the number of samples in corpus", "type": "text" }, { "bbox": [ 493, 291, 500, 301 ], "score": 0.38, "content": "I", "type": "inline_equation" }, { "bbox": [ 500, 288, 503, 305 ], "score": 1.0, "content": ",", "type": "text" }, { "bbox": [ 504, 291, 514, 304 ], "score": 0.61, "content": "y _ { j }", "type": "inline_equation" }, { "bbox": [ 515, 288, 527, 305 ], "score": 1.0, "content": "is", "type": "text" } ], "index": 57 }, { "bbox": [ 304, 303, 525, 318 ], "spans": [ { "bbox": [ 304, 303, 469, 318 ], "score": 1.0, "content": "the total number of tokens in sample", "type": "text" }, { "bbox": [ 470, 304, 497, 317 ], "score": 0.28, "content": "T _ { j } , p _ { j }", "type": "inline_equation" }, { "bbox": [ 498, 303, 525, 318 ], "score": 1.0, "content": "is the", "type": "text" } ], "index": 58 }, { "bbox": [ 303, 316, 524, 332 ], "spans": [ { "bbox": [ 303, 316, 474, 332 ], "score": 1.0, "content": "number of tokens in the prefix part of", "type": "text" }, { "bbox": [ 474, 318, 486, 331 ], "score": 0.89, "content": "T _ { j }", "type": "inline_equation" }, { "bbox": [ 486, 316, 509, 332 ], "score": 1.0, "content": ", and", "type": "text" }, { "bbox": [ 509, 318, 524, 331 ], "score": 0.89, "content": "t _ { i , j }", "type": "inline_equation" } ], "index": 59 }, { "bbox": [ 304, 331, 438, 345 ], "spans": [ { "bbox": [ 304, 331, 421, 345 ], "score": 1.0, "content": "represents the i-th word in", "type": "text" }, { "bbox": [ 422, 331, 434, 345 ], "score": 0.88, "content": "T _ { j }", "type": "inline_equation" }, { "bbox": [ 434, 331, 438, 345 ], "score": 1.0, "content": ".", "type": "text" } ], "index": 60 } ], "index": 58.5, "bbox_fs": [ 303, 288, 527, 345 ] }, { "type": "text", "bbox": [ 304, 345, 526, 464 ], "lines": [ { "bbox": [ 304, 343, 527, 359 ], "spans": [ { "bbox": [ 304, 343, 527, 359 ], "score": 1.0, "content": "Stage 3: Chain-of-Modality Instruction Fine-", "type": "text" } ], "index": 61 }, { "bbox": [ 304, 358, 527, 372 ], "spans": [ { "bbox": [ 304, 358, 527, 372 ], "score": 1.0, "content": "Tuning After obtaining the model in stage 2,", "type": "text" } ], "index": 62 }, { "bbox": [ 303, 371, 527, 386 ], "spans": [ { "bbox": [ 303, 371, 527, 386 ], "score": 1.0, "content": "we utilizes parameter-efficient Low-Rank Adap-", "type": "text" } ], "index": 63 }, { "bbox": [ 304, 385, 526, 398 ], "spans": [ { "bbox": [ 304, 385, 526, 398 ], "score": 1.0, "content": "tation (LoRA) (Hu et al., 2021) to fine-tune it on", "type": "text" } ], "index": 64 }, { "bbox": [ 304, 399, 527, 412 ], "spans": [ { "bbox": [ 304, 399, 527, 412 ], "score": 1.0, "content": "Chain-of-Modality Instruction in SpeechInstruct.", "type": "text" } ], "index": 65 }, { "bbox": [ 304, 412, 525, 426 ], "spans": [ { "bbox": [ 304, 412, 525, 426 ], "score": 1.0, "content": "We add LoRA weights (adapters) to the attention", "type": "text" } ], "index": 66 }, { "bbox": [ 304, 425, 526, 439 ], "spans": [ { "bbox": [ 304, 425, 526, 439 ], "score": 1.0, "content": "mechanisms and train the newly added LoRA pa-", "type": "text" } ], "index": 67 }, { "bbox": [ 303, 439, 525, 453 ], "spans": [ { "bbox": [ 303, 439, 525, 453 ], "score": 1.0, "content": "rameters. We adopt the same loss function as stage", "type": "text" } ], "index": 68 }, { "bbox": [ 304, 453, 316, 466 ], "spans": [ { "bbox": [ 304, 453, 316, 466 ], "score": 1.0, "content": "2.", "type": "text" } ], "index": 69 } ], "index": 65, "bbox_fs": [ 303, 343, 527, 466 ] }, { "type": "title", "bbox": [ 305, 477, 389, 491 ], "lines": [ { "bbox": [ 302, 474, 392, 494 ], "spans": [ { "bbox": [ 302, 474, 392, 494 ], "score": 1.0, "content": "5 Experiments", "type": "text" } ], "index": 70 } ], "index": 70 }, { "type": "title", "bbox": [ 305, 498, 428, 512 ], "lines": [ { "bbox": [ 303, 497, 429, 515 ], "spans": [ { "bbox": [ 303, 497, 429, 515 ], "score": 1.0, "content": "5.1 Experimental Setups", "type": "text" } ], "index": 71 } ], "index": 71 }, { "type": "text", "bbox": [ 304, 517, 525, 665 ], "lines": [ { "bbox": [ 304, 515, 525, 531 ], "spans": [ { "bbox": [ 304, 515, 525, 531 ], "score": 1.0, "content": "Datasets For modality-adaption pre-training, we", "type": "text" } ], "index": 72 }, { "bbox": [ 304, 530, 525, 543 ], "spans": [ { "bbox": [ 304, 530, 525, 543 ], "score": 1.0, "content": "use LibriLight (Kahn et al., 2020) which contains", "type": "text" } ], "index": 73 }, { "bbox": [ 304, 544, 527, 557 ], "spans": [ { "bbox": [ 304, 544, 527, 557 ], "score": 1.0, "content": "60K hours of unlabelled English audiobook speech.", "type": "text" } ], "index": 74 }, { "bbox": [ 304, 558, 526, 571 ], "spans": [ { "bbox": [ 304, 558, 526, 571 ], "score": 1.0, "content": "For cross-modal instruction fine-tuning stage, we", "type": "text" } ], "index": 75 }, { "bbox": [ 304, 572, 525, 584 ], "spans": [ { "bbox": [ 304, 572, 525, 584 ], "score": 1.0, "content": "use Gigaspeech (Chen et al., 2021), Common", "type": "text" } ], "index": 76 }, { "bbox": [ 304, 583, 527, 599 ], "spans": [ { "bbox": [ 304, 583, 527, 599 ], "score": 1.0, "content": "voice (Ardila et al., 2020) and LibriSpeech (Panay-", "type": "text" } ], "index": 77 }, { "bbox": [ 304, 598, 526, 610 ], "spans": [ { "bbox": [ 304, 598, 526, 610 ], "score": 1.0, "content": "otov et al., 2015) dataset and moss-002-sft-data", "type": "text" } ], "index": 78 }, { "bbox": [ 304, 612, 526, 624 ], "spans": [ { "bbox": [ 304, 612, 526, 624 ], "score": 1.0, "content": "dataset, which is illustrated in detail in 3.1. For", "type": "text" } ], "index": 79 }, { "bbox": [ 304, 625, 526, 639 ], "spans": [ { "bbox": [ 304, 625, 526, 639 ], "score": 1.0, "content": "chain-of-modality instruction fine-tuning stage, we", "type": "text" } ], "index": 80 }, { "bbox": [ 304, 639, 525, 650 ], "spans": [ { "bbox": [ 304, 639, 525, 650 ], "score": 1.0, "content": "use moss-002-sft-data dataset, which is illustrated", "type": "text" } ], "index": 81 }, { "bbox": [ 304, 653, 372, 664 ], "spans": [ { "bbox": [ 304, 653, 372, 664 ], "score": 1.0, "content": "in detail in 3.2.", "type": "text" } ], "index": 82 } ], "index": 77, "bbox_fs": [ 304, 515, 527, 664 ] }, { "type": "text", "bbox": [ 304, 666, 525, 774 ], "lines": [ { "bbox": [ 305, 666, 526, 679 ], "spans": [ { "bbox": [ 305, 666, 526, 679 ], "score": 1.0, "content": "Configuration We employ LLaMA-13B (Touvron", "type": "text" } ], "index": 83 }, { "bbox": [ 304, 680, 525, 692 ], "spans": [ { "bbox": [ 304, 680, 525, 692 ], "score": 1.0, "content": "et al., 2023) as our backbone model for a trade-off", "type": "text" } ], "index": 84 }, { "bbox": [ 304, 693, 525, 706 ], "spans": [ { "bbox": [ 304, 693, 525, 706 ], "score": 1.0, "content": "between performance and computational resources", "type": "text" } ], "index": 85 }, { "bbox": [ 304, 707, 525, 719 ], "spans": [ { "bbox": [ 304, 707, 525, 719 ], "score": 1.0, "content": "available. For stage 1, we use 96 A100 GPUs and", "type": "text" } ], "index": 86 }, { "bbox": [ 304, 720, 525, 734 ], "spans": [ { "bbox": [ 304, 720, 525, 734 ], "score": 1.0, "content": "train for 900 steps with batch size 768. For stage", "type": "text" } ], "index": 87 }, { "bbox": [ 304, 734, 525, 747 ], "spans": [ { "bbox": [ 304, 734, 525, 747 ], "score": 1.0, "content": "2, we use 96 A100 GPUs and train for 2100 steps", "type": "text" } ], "index": 88 }, { "bbox": [ 304, 747, 525, 760 ], "spans": [ { "bbox": [ 304, 747, 525, 760 ], "score": 1.0, "content": "with batch size 1536. For stage 3, we use 8 A100", "type": "text" } ], "index": 89 }, { "bbox": [ 305, 761, 527, 773 ], "spans": [ { "bbox": [ 305, 761, 527, 773 ], "score": 1.0, "content": "GPUs and train for 4200 steps with batch size 128.", "type": "text" } ], "index": 90 } ], "index": 86.5, "bbox_fs": [ 304, 666, 527, 773 ] } ] }, { "preproc_blocks": [ { "type": "text", "bbox": [ 69, 72, 289, 139 ], "lines": [ { "bbox": [ 69, 72, 290, 85 ], "spans": [ { "bbox": [ 69, 72, 290, 85 ], "score": 1.0, "content": "Details about training hyperparameters are shown", "type": "text" } ], "index": 0 }, { "bbox": [ 68, 85, 291, 99 ], "spans": [ { "bbox": [ 68, 85, 291, 99 ], "score": 1.0, "content": "in Appendix D. For decoding, we set the maximum", "type": "text" } ], "index": 1 }, { "bbox": [ 68, 99, 291, 113 ], "spans": [ { "bbox": [ 68, 99, 291, 113 ], "score": 1.0, "content": "sequence length to 2048 and set the temperature to", "type": "text" } ], "index": 2 }, { "bbox": [ 68, 111, 291, 126 ], "spans": [ { "bbox": [ 68, 111, 146, 126 ], "score": 1.0, "content": "0.8. We use Top-", "type": "text" }, { "bbox": [ 146, 113, 154, 123 ], "score": 0.79, "content": "k", "type": "inline_equation" }, { "bbox": [ 154, 111, 221, 126 ], "score": 1.0, "content": "sampling with", "type": "text" }, { "bbox": [ 222, 112, 247, 124 ], "score": 0.86, "content": "k { = } 6 0", "type": "inline_equation" }, { "bbox": [ 247, 111, 291, 126 ], "score": 1.0, "content": ". We also", "type": "text" } ], "index": 3 }, { "bbox": [ 68, 126, 210, 139 ], "spans": [ { "bbox": [ 68, 126, 107, 139 ], "score": 1.0, "content": "use Top-", "type": "text" }, { "bbox": [ 107, 128, 114, 138 ], "score": 0.78, "content": "p", "type": "inline_equation" }, { "bbox": [ 114, 126, 179, 139 ], "score": 1.0, "content": "sampling with", "type": "text" }, { "bbox": [ 180, 126, 207, 139 ], "score": 0.81, "content": "\\mathrm { p { = } } 0 . 8", "type": "inline_equation" }, { "bbox": [ 207, 126, 210, 139 ], "score": 1.0, "content": ".", "type": "text" } ], "index": 4 } ], "index": 2 }, { "type": "title", "bbox": [ 70, 150, 138, 163 ], "lines": [ { "bbox": [ 67, 148, 140, 165 ], "spans": [ { "bbox": [ 67, 148, 140, 165 ], "score": 1.0, "content": "5.2 Baselines", "type": "text" } ], "index": 5 } ], "index": 5 }, { "type": "text", "bbox": [ 70, 169, 290, 303 ], "lines": [ { "bbox": [ 68, 168, 292, 182 ], "spans": [ { "bbox": [ 68, 168, 292, 182 ], "score": 1.0, "content": "We establish two cascaded cross-modal conversa-", "type": "text" } ], "index": 6 }, { "bbox": [ 69, 183, 291, 195 ], "spans": [ { "bbox": [ 69, 183, 291, 195 ], "score": 1.0, "content": "tional systems as our baselines. The first model, re-", "type": "text" } ], "index": 7 }, { "bbox": [ 68, 196, 292, 209 ], "spans": [ { "bbox": [ 68, 196, 292, 209 ], "score": 1.0, "content": "ferred to as Speech-Alpaca-13B, consists of an off-", "type": "text" } ], "index": 8 }, { "bbox": [ 69, 210, 291, 223 ], "spans": [ { "bbox": [ 69, 210, 291, 223 ], "score": 1.0, "content": "the-shell ASR system 4, Alpaca 13B (Taori et al.,", "type": "text" } ], "index": 9 }, { "bbox": [ 69, 222, 291, 237 ], "spans": [ { "bbox": [ 69, 222, 291, 237 ], "score": 1.0, "content": "2023) as well as a pre-trained TTS system 5. The", "type": "text" } ], "index": 10 }, { "bbox": [ 69, 237, 291, 249 ], "spans": [ { "bbox": [ 69, 237, 291, 249 ], "score": 1.0, "content": "second model, named Speech-LLaMA-MOSS-002,", "type": "text" } ], "index": 11 }, { "bbox": [ 68, 249, 292, 265 ], "spans": [ { "bbox": [ 68, 249, 292, 265 ], "score": 1.0, "content": "incorporates the same ASR and TTS system, along", "type": "text" } ], "index": 12 }, { "bbox": [ 68, 263, 292, 278 ], "spans": [ { "bbox": [ 68, 263, 292, 278 ], "score": 1.0, "content": "with a large language model obtained by perform-", "type": "text" } ], "index": 13 }, { "bbox": [ 69, 278, 290, 291 ], "spans": [ { "bbox": [ 69, 278, 290, 291 ], "score": 1.0, "content": "ing supervised fine-tuning on LLaMA-13B using", "type": "text" } ], "index": 14 }, { "bbox": [ 68, 290, 236, 305 ], "spans": [ { "bbox": [ 68, 290, 236, 305 ], "score": 1.0, "content": "MOSS-sft-002 as the training dataset.", "type": "text" } ], "index": 15 } ], "index": 10.5 }, { "type": "title", "bbox": [ 70, 316, 146, 328 ], "lines": [ { "bbox": [ 67, 313, 148, 330 ], "spans": [ { "bbox": [ 67, 313, 148, 330 ], "score": 1.0, "content": "5.3 Evaluation", "type": "text" } ], "index": 16 } ], "index": 16 }, { "type": "text", "bbox": [ 70, 334, 290, 415 ], "lines": [ { "bbox": [ 69, 332, 290, 349 ], "spans": [ { "bbox": [ 69, 332, 290, 349 ], "score": 1.0, "content": "We evaluate the cross-modal instruction-following", "type": "text" } ], "index": 17 }, { "bbox": [ 69, 348, 291, 361 ], "spans": [ { "bbox": [ 69, 348, 291, 361 ], "score": 1.0, "content": "capabilities of SpeechGPT across four tasks:", "type": "text" } ], "index": 18 }, { "bbox": [ 68, 362, 291, 375 ], "spans": [ { "bbox": [ 68, 362, 291, 375 ], "score": 1.0, "content": "speech-to-speech instruction-following (S2SIF),", "type": "text" } ], "index": 19 }, { "bbox": [ 68, 376, 292, 389 ], "spans": [ { "bbox": [ 68, 376, 292, 389 ], "score": 1.0, "content": "speech-to-text instruction-following (S2TIF), text-", "type": "text" } ], "index": 20 }, { "bbox": [ 68, 389, 292, 402 ], "spans": [ { "bbox": [ 68, 389, 292, 402 ], "score": 1.0, "content": "to-speech instruction-following (T2SIF), and text-", "type": "text" } ], "index": 21 }, { "bbox": [ 69, 403, 235, 416 ], "spans": [ { "bbox": [ 69, 403, 235, 416 ], "score": 1.0, "content": "to-text instruction-following (T2TIF).", "type": "text" } ], "index": 22 } ], "index": 19.5 }, { "type": "text", "bbox": [ 70, 416, 290, 564 ], "lines": [ { "bbox": [ 69, 416, 290, 429 ], "spans": [ { "bbox": [ 69, 416, 290, 429 ], "score": 1.0, "content": "Data We randomly select 40 samples from the", "type": "text" } ], "index": 23 }, { "bbox": [ 68, 429, 290, 443 ], "spans": [ { "bbox": [ 68, 429, 290, 443 ], "score": 1.0, "content": "AlpacaEval dataset 6 and use the pre-trained TTS", "type": "text" } ], "index": 24 }, { "bbox": [ 69, 444, 291, 455 ], "spans": [ { "bbox": [ 69, 444, 291, 455 ], "score": 1.0, "content": "model in Section 3.3 to convert the text into cor-", "type": "text" } ], "index": 25 }, { "bbox": [ 68, 458, 290, 469 ], "spans": [ { "bbox": [ 68, 458, 290, 469 ], "score": 1.0, "content": "responding speech. We then employ mHuBERT", "type": "text" } ], "index": 26 }, { "bbox": [ 69, 471, 291, 483 ], "spans": [ { "bbox": [ 69, 471, 291, 483 ], "score": 1.0, "content": "to discretize speech data into discrete units as de-", "type": "text" } ], "index": 27 }, { "bbox": [ 69, 484, 290, 496 ], "spans": [ { "bbox": [ 69, 484, 290, 496 ], "score": 1.0, "content": "scribed in Section 3.1. These are combined with", "type": "text" } ], "index": 28 }, { "bbox": [ 69, 498, 291, 510 ], "spans": [ { "bbox": [ 69, 498, 291, 510 ], "score": 1.0, "content": "the SpeechInstruct Evaluation Set to constitute our", "type": "text" } ], "index": 29 }, { "bbox": [ 68, 511, 290, 524 ], "spans": [ { "bbox": [ 68, 511, 290, 524 ], "score": 1.0, "content": "test set, which contains 100 samples. Each sample", "type": "text" } ], "index": 30 }, { "bbox": [ 68, 525, 291, 538 ], "spans": [ { "bbox": [ 68, 525, 291, 538 ], "score": 1.0, "content": "is a quadruplet composed of a speech instruction,", "type": "text" } ], "index": 31 }, { "bbox": [ 68, 538, 292, 552 ], "spans": [ { "bbox": [ 68, 538, 292, 552 ], "score": 1.0, "content": "text instruction, text response, and speech response.", "type": "text" } ], "index": 32 }, { "bbox": [ 69, 551, 213, 565 ], "spans": [ { "bbox": [ 69, 551, 213, 565 ], "score": 1.0, "content": "We denote them as ground truth.", "type": "text" } ], "index": 33 } ], "index": 28 }, { "type": "text", "bbox": [ 69, 566, 290, 700 ], "lines": [ { "bbox": [ 69, 565, 291, 578 ], "spans": [ { "bbox": [ 69, 565, 291, 578 ], "score": 1.0, "content": "ChatGPT Score We utilize ChatGPT (GPT-", "type": "text" } ], "index": 34 }, { "bbox": [ 69, 580, 291, 592 ], "spans": [ { "bbox": [ 69, 580, 291, 592 ], "score": 1.0, "content": "3.5-turbo) to assess the cross-modal instruction-", "type": "text" } ], "index": 35 }, { "bbox": [ 69, 593, 290, 605 ], "spans": [ { "bbox": [ 69, 593, 290, 605 ], "score": 1.0, "content": "following performance. For tasks that include", "type": "text" } ], "index": 36 }, { "bbox": [ 68, 607, 290, 619 ], "spans": [ { "bbox": [ 68, 607, 290, 619 ], "score": 1.0, "content": "speech, we leveraged the pre-trained ASR model", "type": "text" } ], "index": 37 }, { "bbox": [ 69, 620, 291, 632 ], "spans": [ { "bbox": [ 69, 620, 291, 632 ], "score": 1.0, "content": "in section 5.2 to transform the speech into its corre-", "type": "text" } ], "index": 38 }, { "bbox": [ 68, 634, 290, 646 ], "spans": [ { "bbox": [ 68, 634, 290, 646 ], "score": 1.0, "content": "sponding text, which is subsequently submitted for", "type": "text" } ], "index": 39 }, { "bbox": [ 68, 647, 290, 660 ], "spans": [ { "bbox": [ 68, 647, 290, 660 ], "score": 1.0, "content": "evaluation. Inspired from (Zhou et al., 2023), we", "type": "text" } ], "index": 40 }, { "bbox": [ 69, 660, 290, 673 ], "spans": [ { "bbox": [ 69, 660, 290, 673 ], "score": 1.0, "content": "feed the prompt in appendix F to ChatGPT to score", "type": "text" } ], "index": 41 }, { "bbox": [ 69, 675, 290, 688 ], "spans": [ { "bbox": [ 69, 675, 290, 688 ], "score": 1.0, "content": "the model’s outputs based on response quality, with", "type": "text" } ], "index": 42 }, { "bbox": [ 68, 689, 189, 700 ], "spans": [ { "bbox": [ 68, 689, 189, 700 ], "score": 1.0, "content": "scores ranging from 1 to 5.", "type": "text" } ], "index": 43 } ], "index": 38.5 }, { "type": "text", "bbox": [ 305, 72, 525, 354 ], "lines": [ { "bbox": [ 304, 70, 527, 85 ], "spans": [ { "bbox": [ 304, 70, 527, 85 ], "score": 1.0, "content": "Human Opinion Score Following (Nguyen et al.,", "type": "text" } ], "index": 48 }, { "bbox": [ 304, 84, 526, 99 ], "spans": [ { "bbox": [ 304, 84, 526, 99 ], "score": 1.0, "content": "2022), we calculate the human opinion score of the", "type": "text" } ], "index": 49 }, { "bbox": [ 304, 100, 525, 112 ], "spans": [ { "bbox": [ 304, 100, 525, 112 ], "score": 1.0, "content": "generated examples through crowdsourcing. These", "type": "text" } ], "index": 50 }, { "bbox": [ 304, 113, 527, 125 ], "spans": [ { "bbox": [ 304, 113, 527, 125 ], "score": 1.0, "content": "opinions are based on two dimensions: the con-", "type": "text" } ], "index": 51 }, { "bbox": [ 304, 127, 525, 138 ], "spans": [ { "bbox": [ 304, 127, 525, 138 ], "score": 1.0, "content": "tent mean opinion score (CMOS) for content and", "type": "text" } ], "index": 52 }, { "bbox": [ 304, 140, 525, 153 ], "spans": [ { "bbox": [ 304, 140, 525, 153 ], "score": 1.0, "content": "meaningfulness quality, and the naturalness mean", "type": "text" } ], "index": 53 }, { "bbox": [ 304, 153, 526, 165 ], "spans": [ { "bbox": [ 304, 153, 526, 165 ], "score": 1.0, "content": "opinion score (NMOS) for speech naturalness and", "type": "text" } ], "index": 54 }, { "bbox": [ 304, 167, 527, 180 ], "spans": [ { "bbox": [ 304, 167, 527, 180 ], "score": 1.0, "content": "fluency. For CMOS, we ask participants to fo-", "type": "text" } ], "index": 55 }, { "bbox": [ 304, 181, 525, 193 ], "spans": [ { "bbox": [ 304, 181, 525, 193 ], "score": 1.0, "content": "cus on the correctness of the content in speech or", "type": "text" } ], "index": 56 }, { "bbox": [ 304, 194, 525, 207 ], "spans": [ { "bbox": [ 304, 194, 525, 207 ], "score": 1.0, "content": "text, without paying attention to the quality of the", "type": "text" } ], "index": 57 }, { "bbox": [ 304, 208, 527, 221 ], "spans": [ { "bbox": [ 304, 208, 527, 221 ], "score": 1.0, "content": "speech. For NMOS, we direct participants to fo-", "type": "text" } ], "index": 58 }, { "bbox": [ 303, 221, 527, 233 ], "spans": [ { "bbox": [ 303, 221, 527, 233 ], "score": 1.0, "content": "cus on the quality, smoothness, and naturalness of", "type": "text" } ], "index": 59 }, { "bbox": [ 304, 235, 527, 247 ], "spans": [ { "bbox": [ 304, 235, 527, 247 ], "score": 1.0, "content": "the speech, without considering its content. We in-", "type": "text" } ], "index": 60 }, { "bbox": [ 304, 248, 526, 262 ], "spans": [ { "bbox": [ 304, 248, 526, 262 ], "score": 1.0, "content": "vited five volunteers to perform the evaluation, and", "type": "text" } ], "index": 61 }, { "bbox": [ 304, 262, 526, 274 ], "spans": [ { "bbox": [ 304, 262, 526, 274 ], "score": 1.0, "content": "asked them to rate within a range of 1-5, where 1", "type": "text" } ], "index": 62 }, { "bbox": [ 304, 276, 526, 288 ], "spans": [ { "bbox": [ 304, 276, 526, 288 ], "score": 1.0, "content": "represents the worst and 5 represents the best. For", "type": "text" } ], "index": 63 }, { "bbox": [ 303, 289, 527, 302 ], "spans": [ { "bbox": [ 303, 289, 527, 302 ], "score": 1.0, "content": "speech-to-speech instruction-following and text-", "type": "text" } ], "index": 64 }, { "bbox": [ 303, 302, 527, 316 ], "spans": [ { "bbox": [ 303, 302, 527, 316 ], "score": 1.0, "content": "to-speech instruction-following tasks, we calcu-", "type": "text" } ], "index": 65 }, { "bbox": [ 304, 316, 526, 329 ], "spans": [ { "bbox": [ 304, 316, 526, 329 ], "score": 1.0, "content": "late both CMOS and NMOS. For speech-to-text", "type": "text" } ], "index": 66 }, { "bbox": [ 304, 330, 527, 342 ], "spans": [ { "bbox": [ 304, 330, 527, 342 ], "score": 1.0, "content": "instruction-following and text-to-text instruction-", "type": "text" } ], "index": 67 }, { "bbox": [ 304, 343, 470, 355 ], "spans": [ { "bbox": [ 304, 343, 470, 355 ], "score": 1.0, "content": "following tasks, we calculate CMOS.", "type": "text" } ], "index": 68 } ], "index": 58 }, { "type": "title", "bbox": [ 305, 371, 392, 384 ], "lines": [ { "bbox": [ 303, 369, 394, 387 ], "spans": [ { "bbox": [ 303, 369, 394, 387 ], "score": 1.0, "content": "5.4 Main Results", "type": "text" } ], "index": 69 } ], "index": 69 }, { "type": "text", "bbox": [ 305, 394, 525, 717 ], "lines": [ { "bbox": [ 305, 393, 527, 406 ], "spans": [ { "bbox": [ 305, 393, 527, 406 ], "score": 1.0, "content": "Content As shown in Table 1, taking into ac-", "type": "text" } ], "index": 70 }, { "bbox": [ 304, 407, 526, 419 ], "spans": [ { "bbox": [ 304, 407, 526, 419 ], "score": 1.0, "content": "count the comprehensive evaluation of ChatGPT", "type": "text" } ], "index": 71 }, { "bbox": [ 304, 421, 527, 433 ], "spans": [ { "bbox": [ 304, 421, 527, 433 ], "score": 1.0, "content": "Score and CMOS, SpeechGPT demonstrates su-", "type": "text" } ], "index": 72 }, { "bbox": [ 303, 434, 526, 446 ], "spans": [ { "bbox": [ 303, 434, 526, 446 ], "score": 1.0, "content": "perior performance on speech instructions (S2SIF", "type": "text" } ], "index": 73 }, { "bbox": [ 303, 447, 527, 461 ], "spans": [ { "bbox": [ 303, 447, 527, 461 ], "score": 1.0, "content": "and S2TIF) compared to the two baseline systems.", "type": "text" } ], "index": 74 }, { "bbox": [ 303, 461, 525, 474 ], "spans": [ { "bbox": [ 303, 461, 525, 474 ], "score": 1.0, "content": "This indicates that SpeechGPT outperforms the", "type": "text" } ], "index": 75 }, { "bbox": [ 304, 474, 525, 488 ], "spans": [ { "bbox": [ 304, 474, 525, 488 ], "score": 1.0, "content": "ASR model in the cascaded system when it comes", "type": "text" } ], "index": 76 }, { "bbox": [ 304, 489, 527, 501 ], "spans": [ { "bbox": [ 304, 489, 527, 501 ], "score": 1.0, "content": "to understanding speech content. From the per-", "type": "text" } ], "index": 77 }, { "bbox": [ 304, 501, 527, 515 ], "spans": [ { "bbox": [ 304, 501, 527, 515 ], "score": 1.0, "content": "spective of CMOS, SpeechGPT achieves perfor-", "type": "text" } ], "index": 78 }, { "bbox": [ 304, 515, 526, 528 ], "spans": [ { "bbox": [ 304, 515, 526, 528 ], "score": 1.0, "content": "mance similar to the baseline systems on T2SIF", "type": "text" } ], "index": 79 }, { "bbox": [ 304, 529, 525, 542 ], "spans": [ { "bbox": [ 304, 529, 525, 542 ], "score": 1.0, "content": "and T2TIF tasks, indicating that SpeechGPT still", "type": "text" } ], "index": 80 }, { "bbox": [ 303, 543, 525, 555 ], "spans": [ { "bbox": [ 303, 543, 525, 555 ], "score": 1.0, "content": "possesses commendable text and speech generation", "type": "text" } ], "index": 81 }, { "bbox": [ 304, 556, 525, 568 ], "spans": [ { "bbox": [ 304, 556, 525, 568 ], "score": 1.0, "content": "capabilities. In S2SIF and T2SIF tasks, ChatGPT", "type": "text" } ], "index": 82 }, { "bbox": [ 304, 569, 525, 583 ], "spans": [ { "bbox": [ 304, 569, 525, 583 ], "score": 1.0, "content": "Score and CMOS values exhibit ambiguity in the", "type": "text" } ], "index": 83 }, { "bbox": [ 304, 583, 525, 596 ], "spans": [ { "bbox": [ 304, 583, 525, 596 ], "score": 1.0, "content": "ground truth and baseline systems. This can be", "type": "text" } ], "index": 84 }, { "bbox": [ 304, 596, 525, 610 ], "spans": [ { "bbox": [ 304, 596, 525, 610 ], "score": 1.0, "content": "attributed to speech responses being synthesized by", "type": "text" } ], "index": 85 }, { "bbox": [ 303, 610, 526, 624 ], "spans": [ { "bbox": [ 303, 610, 526, 624 ], "score": 1.0, "content": "TTS system, which can have errors in pauses be-", "type": "text" } ], "index": 86 }, { "bbox": [ 303, 623, 525, 637 ], "spans": [ { "bbox": [ 303, 623, 525, 637 ], "score": 1.0, "content": "tween sentences. This introduces significant errors", "type": "text" } ], "index": 87 }, { "bbox": [ 304, 637, 525, 650 ], "spans": [ { "bbox": [ 304, 637, 525, 650 ], "score": 1.0, "content": "for longer responses, leading to incorrect text after", "type": "text" } ], "index": 88 }, { "bbox": [ 304, 651, 527, 664 ], "spans": [ { "bbox": [ 304, 651, 527, 664 ], "score": 1.0, "content": "being processed by the ASR system, thereby reduc-", "type": "text" } ], "index": 89 }, { "bbox": [ 304, 664, 527, 677 ], "spans": [ { "bbox": [ 304, 664, 527, 677 ], "score": 1.0, "content": "ing the ChatGPT score. However, humans can un-", "type": "text" } ], "index": 90 }, { "bbox": [ 304, 677, 525, 691 ], "spans": [ { "bbox": [ 304, 677, 525, 691 ], "score": 1.0, "content": "derstand the content of such speech, so the CMOS", "type": "text" } ], "index": 91 }, { "bbox": [ 304, 692, 526, 704 ], "spans": [ { "bbox": [ 304, 692, 526, 704 ], "score": 1.0, "content": "score is normal. Cases of cross-modal instruction-", "type": "text" } ], "index": 92 }, { "bbox": [ 304, 705, 478, 718 ], "spans": [ { "bbox": [ 304, 705, 478, 718 ], "score": 1.0, "content": "following can be found in Appendix G.", "type": "text" } ], "index": 93 } ], "index": 81.5 }, { "type": "text", "bbox": [ 304, 720, 525, 774 ], "lines": [ { "bbox": [ 304, 720, 526, 733 ], "spans": [ { "bbox": [ 304, 720, 526, 733 ], "score": 1.0, "content": "Speech Quality As shown in Table 1, SpeechGPT", "type": "text" } ], "index": 94 }, { "bbox": [ 304, 735, 527, 747 ], "spans": [ { "bbox": [ 304, 735, 527, 747 ], "score": 1.0, "content": "exhibits significantly higher NMOS values com-", "type": "text" } ], "index": 95 }, { "bbox": [ 304, 748, 526, 760 ], "spans": [ { "bbox": [ 304, 748, 526, 760 ], "score": 1.0, "content": "pared to the baseline systems. This indicates that", "type": "text" } ], "index": 96 }, { "bbox": [ 304, 761, 527, 775 ], "spans": [ { "bbox": [ 304, 761, 527, 775 ], "score": 1.0, "content": "the speech responses generated by SpeechGPT out-", "type": "text" } ], "index": 97 } ], "index": 95.5 } ], "page_idx": 5, "page_size": [ 595, 841 ], "discarded_blocks": [ { "type": "discarded", "bbox": [ 69, 710, 281, 772 ], "lines": [ { "bbox": [ 81, 709, 259, 723 ], "spans": [ { "bbox": [ 81, 709, 259, 723 ], "score": 1.0, "content": "4https://huggingface.co/facebook/", "type": "text" } ] }, { "bbox": [ 69, 721, 230, 731 ], "spans": [ { "bbox": [ 69, 721, 230, 731 ], "score": 1.0, "content": "wav2vec2-large-960h-lv60-self", "type": "text" } ] }, { "bbox": [ 81, 730, 259, 744 ], "spans": [ { "bbox": [ 81, 730, 259, 744 ], "score": 1.0, "content": "5https://huggingface.co/facebook/", "type": "text" } ] }, { "bbox": [ 69, 741, 196, 753 ], "spans": [ { "bbox": [ 69, 741, 196, 753 ], "score": 1.0, "content": "fastspeech2-en-ljspeech", "type": "text" } ] }, { "bbox": [ 81, 751, 282, 765 ], "spans": [ { "bbox": [ 81, 751, 282, 765 ], "score": 1.0, "content": "6https://github.com/tatsu-lab/alpaca_", "type": "text" } ] }, { "bbox": [ 68, 763, 93, 772 ], "spans": [ { "bbox": [ 68, 763, 93, 772 ], "score": 1.0, "content": "eval", "type": "text" } ] } ] } ], "para_blocks": [ { "type": "text", "bbox": [ 69, 72, 289, 139 ], "lines": [ { "bbox": [ 69, 72, 290, 85 ], "spans": [ { "bbox": [ 69, 72, 290, 85 ], "score": 1.0, "content": "Details about training hyperparameters are shown", "type": "text" } ], "index": 0 }, { "bbox": [ 68, 85, 291, 99 ], "spans": [ { "bbox": [ 68, 85, 291, 99 ], "score": 1.0, "content": "in Appendix D. For decoding, we set the maximum", "type": "text" } ], "index": 1 }, { "bbox": [ 68, 99, 291, 113 ], "spans": [ { "bbox": [ 68, 99, 291, 113 ], "score": 1.0, "content": "sequence length to 2048 and set the temperature to", "type": "text" } ], "index": 2 }, { "bbox": [ 68, 111, 291, 126 ], "spans": [ { "bbox": [ 68, 111, 146, 126 ], "score": 1.0, "content": "0.8. We use Top-", "type": "text" }, { "bbox": [ 146, 113, 154, 123 ], "score": 0.79, "content": "k", "type": "inline_equation" }, { "bbox": [ 154, 111, 221, 126 ], "score": 1.0, "content": "sampling with", "type": "text" }, { "bbox": [ 222, 112, 247, 124 ], "score": 0.86, "content": "k { = } 6 0", "type": "inline_equation" }, { "bbox": [ 247, 111, 291, 126 ], "score": 1.0, "content": ". We also", "type": "text" } ], "index": 3 }, { "bbox": [ 68, 126, 210, 139 ], "spans": [ { "bbox": [ 68, 126, 107, 139 ], "score": 1.0, "content": "use Top-", "type": "text" }, { "bbox": [ 107, 128, 114, 138 ], "score": 0.78, "content": "p", "type": "inline_equation" }, { "bbox": [ 114, 126, 179, 139 ], "score": 1.0, "content": "sampling with", "type": "text" }, { "bbox": [ 180, 126, 207, 139 ], "score": 0.81, "content": "\\mathrm { p { = } } 0 . 8", "type": "inline_equation" }, { "bbox": [ 207, 126, 210, 139 ], "score": 1.0, "content": ".", "type": "text" } ], "index": 4 } ], "index": 2, "bbox_fs": [ 68, 72, 291, 139 ] }, { "type": "title", "bbox": [ 70, 150, 138, 163 ], "lines": [ { "bbox": [ 67, 148, 140, 165 ], "spans": [ { "bbox": [ 67, 148, 140, 165 ], "score": 1.0, "content": "5.2 Baselines", "type": "text" } ], "index": 5 } ], "index": 5 }, { "type": "text", "bbox": [ 70, 169, 290, 303 ], "lines": [ { "bbox": [ 68, 168, 292, 182 ], "spans": [ { "bbox": [ 68, 168, 292, 182 ], "score": 1.0, "content": "We establish two cascaded cross-modal conversa-", "type": "text" } ], "index": 6 }, { "bbox": [ 69, 183, 291, 195 ], "spans": [ { "bbox": [ 69, 183, 291, 195 ], "score": 1.0, "content": "tional systems as our baselines. The first model, re-", "type": "text" } ], "index": 7 }, { "bbox": [ 68, 196, 292, 209 ], "spans": [ { "bbox": [ 68, 196, 292, 209 ], "score": 1.0, "content": "ferred to as Speech-Alpaca-13B, consists of an off-", "type": "text" } ], "index": 8 }, { "bbox": [ 69, 210, 291, 223 ], "spans": [ { "bbox": [ 69, 210, 291, 223 ], "score": 1.0, "content": "the-shell ASR system 4, Alpaca 13B (Taori et al.,", "type": "text" } ], "index": 9 }, { "bbox": [ 69, 222, 291, 237 ], "spans": [ { "bbox": [ 69, 222, 291, 237 ], "score": 1.0, "content": "2023) as well as a pre-trained TTS system 5. The", "type": "text" } ], "index": 10 }, { "bbox": [ 69, 237, 291, 249 ], "spans": [ { "bbox": [ 69, 237, 291, 249 ], "score": 1.0, "content": "second model, named Speech-LLaMA-MOSS-002,", "type": "text" } ], "index": 11 }, { "bbox": [ 68, 249, 292, 265 ], "spans": [ { "bbox": [ 68, 249, 292, 265 ], "score": 1.0, "content": "incorporates the same ASR and TTS system, along", "type": "text" } ], "index": 12 }, { "bbox": [ 68, 263, 292, 278 ], "spans": [ { "bbox": [ 68, 263, 292, 278 ], "score": 1.0, "content": "with a large language model obtained by perform-", "type": "text" } ], "index": 13 }, { "bbox": [ 69, 278, 290, 291 ], "spans": [ { "bbox": [ 69, 278, 290, 291 ], "score": 1.0, "content": "ing supervised fine-tuning on LLaMA-13B using", "type": "text" } ], "index": 14 }, { "bbox": [ 68, 290, 236, 305 ], "spans": [ { "bbox": [ 68, 290, 236, 305 ], "score": 1.0, "content": "MOSS-sft-002 as the training dataset.", "type": "text" } ], "index": 15 } ], "index": 10.5, "bbox_fs": [ 68, 168, 292, 305 ] }, { "type": "title", "bbox": [ 70, 316, 146, 328 ], "lines": [ { "bbox": [ 67, 313, 148, 330 ], "spans": [ { "bbox": [ 67, 313, 148, 330 ], "score": 1.0, "content": "5.3 Evaluation", "type": "text" } ], "index": 16 } ], "index": 16 }, { "type": "text", "bbox": [ 70, 334, 290, 415 ], "lines": [ { "bbox": [ 69, 332, 290, 349 ], "spans": [ { "bbox": [ 69, 332, 290, 349 ], "score": 1.0, "content": "We evaluate the cross-modal instruction-following", "type": "text" } ], "index": 17 }, { "bbox": [ 69, 348, 291, 361 ], "spans": [ { "bbox": [ 69, 348, 291, 361 ], "score": 1.0, "content": "capabilities of SpeechGPT across four tasks:", "type": "text" } ], "index": 18 }, { "bbox": [ 68, 362, 291, 375 ], "spans": [ { "bbox": [ 68, 362, 291, 375 ], "score": 1.0, "content": "speech-to-speech instruction-following (S2SIF),", "type": "text" } ], "index": 19 }, { "bbox": [ 68, 376, 292, 389 ], "spans": [ { "bbox": [ 68, 376, 292, 389 ], "score": 1.0, "content": "speech-to-text instruction-following (S2TIF), text-", "type": "text" } ], "index": 20 }, { "bbox": [ 68, 389, 292, 402 ], "spans": [ { "bbox": [ 68, 389, 292, 402 ], "score": 1.0, "content": "to-speech instruction-following (T2SIF), and text-", "type": "text" } ], "index": 21 }, { "bbox": [ 69, 403, 235, 416 ], "spans": [ { "bbox": [ 69, 403, 235, 416 ], "score": 1.0, "content": "to-text instruction-following (T2TIF).", "type": "text" } ], "index": 22 } ], "index": 19.5, "bbox_fs": [ 68, 332, 292, 416 ] }, { "type": "text", "bbox": [ 70, 416, 290, 564 ], "lines": [ { "bbox": [ 69, 416, 290, 429 ], "spans": [ { "bbox": [ 69, 416, 290, 429 ], "score": 1.0, "content": "Data We randomly select 40 samples from the", "type": "text" } ], "index": 23 }, { "bbox": [ 68, 429, 290, 443 ], "spans": [ { "bbox": [ 68, 429, 290, 443 ], "score": 1.0, "content": "AlpacaEval dataset 6 and use the pre-trained TTS", "type": "text" } ], "index": 24 }, { "bbox": [ 69, 444, 291, 455 ], "spans": [ { "bbox": [ 69, 444, 291, 455 ], "score": 1.0, "content": "model in Section 3.3 to convert the text into cor-", "type": "text" } ], "index": 25 }, { "bbox": [ 68, 458, 290, 469 ], "spans": [ { "bbox": [ 68, 458, 290, 469 ], "score": 1.0, "content": "responding speech. We then employ mHuBERT", "type": "text" } ], "index": 26 }, { "bbox": [ 69, 471, 291, 483 ], "spans": [ { "bbox": [ 69, 471, 291, 483 ], "score": 1.0, "content": "to discretize speech data into discrete units as de-", "type": "text" } ], "index": 27 }, { "bbox": [ 69, 484, 290, 496 ], "spans": [ { "bbox": [ 69, 484, 290, 496 ], "score": 1.0, "content": "scribed in Section 3.1. These are combined with", "type": "text" } ], "index": 28 }, { "bbox": [ 69, 498, 291, 510 ], "spans": [ { "bbox": [ 69, 498, 291, 510 ], "score": 1.0, "content": "the SpeechInstruct Evaluation Set to constitute our", "type": "text" } ], "index": 29 }, { "bbox": [ 68, 511, 290, 524 ], "spans": [ { "bbox": [ 68, 511, 290, 524 ], "score": 1.0, "content": "test set, which contains 100 samples. Each sample", "type": "text" } ], "index": 30 }, { "bbox": [ 68, 525, 291, 538 ], "spans": [ { "bbox": [ 68, 525, 291, 538 ], "score": 1.0, "content": "is a quadruplet composed of a speech instruction,", "type": "text" } ], "index": 31 }, { "bbox": [ 68, 538, 292, 552 ], "spans": [ { "bbox": [ 68, 538, 292, 552 ], "score": 1.0, "content": "text instruction, text response, and speech response.", "type": "text" } ], "index": 32 }, { "bbox": [ 69, 551, 213, 565 ], "spans": [ { "bbox": [ 69, 551, 213, 565 ], "score": 1.0, "content": "We denote them as ground truth.", "type": "text" } ], "index": 33 } ], "index": 28, "bbox_fs": [ 68, 416, 292, 565 ] }, { "type": "text", "bbox": [ 69, 566, 290, 700 ], "lines": [ { "bbox": [ 69, 565, 291, 578 ], "spans": [ { "bbox": [ 69, 565, 291, 578 ], "score": 1.0, "content": "ChatGPT Score We utilize ChatGPT (GPT-", "type": "text" } ], "index": 34 }, { "bbox": [ 69, 580, 291, 592 ], "spans": [ { "bbox": [ 69, 580, 291, 592 ], "score": 1.0, "content": "3.5-turbo) to assess the cross-modal instruction-", "type": "text" } ], "index": 35 }, { "bbox": [ 69, 593, 290, 605 ], "spans": [ { "bbox": [ 69, 593, 290, 605 ], "score": 1.0, "content": "following performance. For tasks that include", "type": "text" } ], "index": 36 }, { "bbox": [ 68, 607, 290, 619 ], "spans": [ { "bbox": [ 68, 607, 290, 619 ], "score": 1.0, "content": "speech, we leveraged the pre-trained ASR model", "type": "text" } ], "index": 37 }, { "bbox": [ 69, 620, 291, 632 ], "spans": [ { "bbox": [ 69, 620, 291, 632 ], "score": 1.0, "content": "in section 5.2 to transform the speech into its corre-", "type": "text" } ], "index": 38 }, { "bbox": [ 68, 634, 290, 646 ], "spans": [ { "bbox": [ 68, 634, 290, 646 ], "score": 1.0, "content": "sponding text, which is subsequently submitted for", "type": "text" } ], "index": 39 }, { "bbox": [ 68, 647, 290, 660 ], "spans": [ { "bbox": [ 68, 647, 290, 660 ], "score": 1.0, "content": "evaluation. Inspired from (Zhou et al., 2023), we", "type": "text" } ], "index": 40 }, { "bbox": [ 69, 660, 290, 673 ], "spans": [ { "bbox": [ 69, 660, 290, 673 ], "score": 1.0, "content": "feed the prompt in appendix F to ChatGPT to score", "type": "text" } ], "index": 41 }, { "bbox": [ 69, 675, 290, 688 ], "spans": [ { "bbox": [ 69, 675, 290, 688 ], "score": 1.0, "content": "the model’s outputs based on response quality, with", "type": "text" } ], "index": 42 }, { "bbox": [ 68, 689, 189, 700 ], "spans": [ { "bbox": [ 68, 689, 189, 700 ], "score": 1.0, "content": "scores ranging from 1 to 5.", "type": "text" } ], "index": 43 } ], "index": 38.5, "bbox_fs": [ 68, 565, 291, 700 ] }, { "type": "text", "bbox": [ 305, 72, 525, 354 ], "lines": [ { "bbox": [ 304, 70, 527, 85 ], "spans": [ { "bbox": [ 304, 70, 527, 85 ], "score": 1.0, "content": "Human Opinion Score Following (Nguyen et al.,", "type": "text" } ], "index": 48 }, { "bbox": [ 304, 84, 526, 99 ], "spans": [ { "bbox": [ 304, 84, 526, 99 ], "score": 1.0, "content": "2022), we calculate the human opinion score of the", "type": "text" } ], "index": 49 }, { "bbox": [ 304, 100, 525, 112 ], "spans": [ { "bbox": [ 304, 100, 525, 112 ], "score": 1.0, "content": "generated examples through crowdsourcing. These", "type": "text" } ], "index": 50 }, { "bbox": [ 304, 113, 527, 125 ], "spans": [ { "bbox": [ 304, 113, 527, 125 ], "score": 1.0, "content": "opinions are based on two dimensions: the con-", "type": "text" } ], "index": 51 }, { "bbox": [ 304, 127, 525, 138 ], "spans": [ { "bbox": [ 304, 127, 525, 138 ], "score": 1.0, "content": "tent mean opinion score (CMOS) for content and", "type": "text" } ], "index": 52 }, { "bbox": [ 304, 140, 525, 153 ], "spans": [ { "bbox": [ 304, 140, 525, 153 ], "score": 1.0, "content": "meaningfulness quality, and the naturalness mean", "type": "text" } ], "index": 53 }, { "bbox": [ 304, 153, 526, 165 ], "spans": [ { "bbox": [ 304, 153, 526, 165 ], "score": 1.0, "content": "opinion score (NMOS) for speech naturalness and", "type": "text" } ], "index": 54 }, { "bbox": [ 304, 167, 527, 180 ], "spans": [ { "bbox": [ 304, 167, 527, 180 ], "score": 1.0, "content": "fluency. For CMOS, we ask participants to fo-", "type": "text" } ], "index": 55 }, { "bbox": [ 304, 181, 525, 193 ], "spans": [ { "bbox": [ 304, 181, 525, 193 ], "score": 1.0, "content": "cus on the correctness of the content in speech or", "type": "text" } ], "index": 56 }, { "bbox": [ 304, 194, 525, 207 ], "spans": [ { "bbox": [ 304, 194, 525, 207 ], "score": 1.0, "content": "text, without paying attention to the quality of the", "type": "text" } ], "index": 57 }, { "bbox": [ 304, 208, 527, 221 ], "spans": [ { "bbox": [ 304, 208, 527, 221 ], "score": 1.0, "content": "speech. For NMOS, we direct participants to fo-", "type": "text" } ], "index": 58 }, { "bbox": [ 303, 221, 527, 233 ], "spans": [ { "bbox": [ 303, 221, 527, 233 ], "score": 1.0, "content": "cus on the quality, smoothness, and naturalness of", "type": "text" } ], "index": 59 }, { "bbox": [ 304, 235, 527, 247 ], "spans": [ { "bbox": [ 304, 235, 527, 247 ], "score": 1.0, "content": "the speech, without considering its content. We in-", "type": "text" } ], "index": 60 }, { "bbox": [ 304, 248, 526, 262 ], "spans": [ { "bbox": [ 304, 248, 526, 262 ], "score": 1.0, "content": "vited five volunteers to perform the evaluation, and", "type": "text" } ], "index": 61 }, { "bbox": [ 304, 262, 526, 274 ], "spans": [ { "bbox": [ 304, 262, 526, 274 ], "score": 1.0, "content": "asked them to rate within a range of 1-5, where 1", "type": "text" } ], "index": 62 }, { "bbox": [ 304, 276, 526, 288 ], "spans": [ { "bbox": [ 304, 276, 526, 288 ], "score": 1.0, "content": "represents the worst and 5 represents the best. For", "type": "text" } ], "index": 63 }, { "bbox": [ 303, 289, 527, 302 ], "spans": [ { "bbox": [ 303, 289, 527, 302 ], "score": 1.0, "content": "speech-to-speech instruction-following and text-", "type": "text" } ], "index": 64 }, { "bbox": [ 303, 302, 527, 316 ], "spans": [ { "bbox": [ 303, 302, 527, 316 ], "score": 1.0, "content": "to-speech instruction-following tasks, we calcu-", "type": "text" } ], "index": 65 }, { "bbox": [ 304, 316, 526, 329 ], "spans": [ { "bbox": [ 304, 316, 526, 329 ], "score": 1.0, "content": "late both CMOS and NMOS. For speech-to-text", "type": "text" } ], "index": 66 }, { "bbox": [ 304, 330, 527, 342 ], "spans": [ { "bbox": [ 304, 330, 527, 342 ], "score": 1.0, "content": "instruction-following and text-to-text instruction-", "type": "text" } ], "index": 67 }, { "bbox": [ 304, 343, 470, 355 ], "spans": [ { "bbox": [ 304, 343, 470, 355 ], "score": 1.0, "content": "following tasks, we calculate CMOS.", "type": "text" } ], "index": 68 } ], "index": 58, "bbox_fs": [ 303, 70, 527, 355 ] }, { "type": "title", "bbox": [ 305, 371, 392, 384 ], "lines": [ { "bbox": [ 303, 369, 394, 387 ], "spans": [ { "bbox": [ 303, 369, 394, 387 ], "score": 1.0, "content": "5.4 Main Results", "type": "text" } ], "index": 69 } ], "index": 69 }, { "type": "text", "bbox": [ 305, 394, 525, 717 ], "lines": [ { "bbox": [ 305, 393, 527, 406 ], "spans": [ { "bbox": [ 305, 393, 527, 406 ], "score": 1.0, "content": "Content As shown in Table 1, taking into ac-", "type": "text" } ], "index": 70 }, { "bbox": [ 304, 407, 526, 419 ], "spans": [ { "bbox": [ 304, 407, 526, 419 ], "score": 1.0, "content": "count the comprehensive evaluation of ChatGPT", "type": "text" } ], "index": 71 }, { "bbox": [ 304, 421, 527, 433 ], "spans": [ { "bbox": [ 304, 421, 527, 433 ], "score": 1.0, "content": "Score and CMOS, SpeechGPT demonstrates su-", "type": "text" } ], "index": 72 }, { "bbox": [ 303, 434, 526, 446 ], "spans": [ { "bbox": [ 303, 434, 526, 446 ], "score": 1.0, "content": "perior performance on speech instructions (S2SIF", "type": "text" } ], "index": 73 }, { "bbox": [ 303, 447, 527, 461 ], "spans": [ { "bbox": [ 303, 447, 527, 461 ], "score": 1.0, "content": "and S2TIF) compared to the two baseline systems.", "type": "text" } ], "index": 74 }, { "bbox": [ 303, 461, 525, 474 ], "spans": [ { "bbox": [ 303, 461, 525, 474 ], "score": 1.0, "content": "This indicates that SpeechGPT outperforms the", "type": "text" } ], "index": 75 }, { "bbox": [ 304, 474, 525, 488 ], "spans": [ { "bbox": [ 304, 474, 525, 488 ], "score": 1.0, "content": "ASR model in the cascaded system when it comes", "type": "text" } ], "index": 76 }, { "bbox": [ 304, 489, 527, 501 ], "spans": [ { "bbox": [ 304, 489, 527, 501 ], "score": 1.0, "content": "to understanding speech content. From the per-", "type": "text" } ], "index": 77 }, { "bbox": [ 304, 501, 527, 515 ], "spans": [ { "bbox": [ 304, 501, 527, 515 ], "score": 1.0, "content": "spective of CMOS, SpeechGPT achieves perfor-", "type": "text" } ], "index": 78 }, { "bbox": [ 304, 515, 526, 528 ], "spans": [ { "bbox": [ 304, 515, 526, 528 ], "score": 1.0, "content": "mance similar to the baseline systems on T2SIF", "type": "text" } ], "index": 79 }, { "bbox": [ 304, 529, 525, 542 ], "spans": [ { "bbox": [ 304, 529, 525, 542 ], "score": 1.0, "content": "and T2TIF tasks, indicating that SpeechGPT still", "type": "text" } ], "index": 80 }, { "bbox": [ 303, 543, 525, 555 ], "spans": [ { "bbox": [ 303, 543, 525, 555 ], "score": 1.0, "content": "possesses commendable text and speech generation", "type": "text" } ], "index": 81 }, { "bbox": [ 304, 556, 525, 568 ], "spans": [ { "bbox": [ 304, 556, 525, 568 ], "score": 1.0, "content": "capabilities. In S2SIF and T2SIF tasks, ChatGPT", "type": "text" } ], "index": 82 }, { "bbox": [ 304, 569, 525, 583 ], "spans": [ { "bbox": [ 304, 569, 525, 583 ], "score": 1.0, "content": "Score and CMOS values exhibit ambiguity in the", "type": "text" } ], "index": 83 }, { "bbox": [ 304, 583, 525, 596 ], "spans": [ { "bbox": [ 304, 583, 525, 596 ], "score": 1.0, "content": "ground truth and baseline systems. This can be", "type": "text" } ], "index": 84 }, { "bbox": [ 304, 596, 525, 610 ], "spans": [ { "bbox": [ 304, 596, 525, 610 ], "score": 1.0, "content": "attributed to speech responses being synthesized by", "type": "text" } ], "index": 85 }, { "bbox": [ 303, 610, 526, 624 ], "spans": [ { "bbox": [ 303, 610, 526, 624 ], "score": 1.0, "content": "TTS system, which can have errors in pauses be-", "type": "text" } ], "index": 86 }, { "bbox": [ 303, 623, 525, 637 ], "spans": [ { "bbox": [ 303, 623, 525, 637 ], "score": 1.0, "content": "tween sentences. This introduces significant errors", "type": "text" } ], "index": 87 }, { "bbox": [ 304, 637, 525, 650 ], "spans": [ { "bbox": [ 304, 637, 525, 650 ], "score": 1.0, "content": "for longer responses, leading to incorrect text after", "type": "text" } ], "index": 88 }, { "bbox": [ 304, 651, 527, 664 ], "spans": [ { "bbox": [ 304, 651, 527, 664 ], "score": 1.0, "content": "being processed by the ASR system, thereby reduc-", "type": "text" } ], "index": 89 }, { "bbox": [ 304, 664, 527, 677 ], "spans": [ { "bbox": [ 304, 664, 527, 677 ], "score": 1.0, "content": "ing the ChatGPT score. However, humans can un-", "type": "text" } ], "index": 90 }, { "bbox": [ 304, 677, 525, 691 ], "spans": [ { "bbox": [ 304, 677, 525, 691 ], "score": 1.0, "content": "derstand the content of such speech, so the CMOS", "type": "text" } ], "index": 91 }, { "bbox": [ 304, 692, 526, 704 ], "spans": [ { "bbox": [ 304, 692, 526, 704 ], "score": 1.0, "content": "score is normal. Cases of cross-modal instruction-", "type": "text" } ], "index": 92 }, { "bbox": [ 304, 705, 478, 718 ], "spans": [ { "bbox": [ 304, 705, 478, 718 ], "score": 1.0, "content": "following can be found in Appendix G.", "type": "text" } ], "index": 93 } ], "index": 81.5, "bbox_fs": [ 303, 393, 527, 718 ] }, { "type": "text", "bbox": [ 304, 720, 525, 774 ], "lines": [ { "bbox": [ 304, 720, 526, 733 ], "spans": [ { "bbox": [ 304, 720, 526, 733 ], "score": 1.0, "content": "Speech Quality As shown in Table 1, SpeechGPT", "type": "text" } ], "index": 94 }, { "bbox": [ 304, 735, 527, 747 ], "spans": [ { "bbox": [ 304, 735, 527, 747 ], "score": 1.0, "content": "exhibits significantly higher NMOS values com-", "type": "text" } ], "index": 95 }, { "bbox": [ 304, 748, 526, 760 ], "spans": [ { "bbox": [ 304, 748, 526, 760 ], "score": 1.0, "content": "pared to the baseline systems. This indicates that", "type": "text" } ], "index": 96 }, { "bbox": [ 304, 761, 527, 775 ], "spans": [ { "bbox": [ 304, 761, 527, 775 ], "score": 1.0, "content": "the speech responses generated by SpeechGPT out-", "type": "text" } ], "index": 97 } ], "index": 95.5, "bbox_fs": [ 304, 720, 527, 775 ] } ] }, { "preproc_blocks": [ { "type": "table", "bbox": [ 70, 69, 525, 177 ], "blocks": [ { "type": "table_body", "bbox": [ 70, 69, 525, 177 ], "group_id": 1, "lines": [ { "bbox": [ 70, 69, 525, 177 ], "spans": [ { "bbox": [ 70, 69, 525, 177 ], "score": 0.977, "html": "
MethodsChatGPT ScoreHuman Opinion Score
CMOSNMOS
S2SIFS2TIFT2SIFT2TIFS2SIFS2TIFT2SIFT2TIFS2SIFS2TIFT2SIFT2TIF
Ground Truth2.85*3.742.91*3.933.783.893.954.123.18-3.20-
Baselines: cascaded cross-modal conversational systems
Speech-Alpaca-13B2.743.312.713.833.393.423.713.753.123.131
Speech-LLaMA-MOSS-0022.873.503.233.823.383.443.743.833.143.111
SpeechGPT3.423.523.533.643.423.493.573.693.65-3.621
", "type": "table", "image_path": "9a7fcd22dd35afec89ad794b756a0b89d8a71276f648f06dff3208200eb58e3d.jpg" } ] } ], "index": 1, "virtual_lines": [ { "bbox": [ 70, 69, 525, 105.0 ], "spans": [], "index": 0 }, { "bbox": [ 70, 105.0, 525, 141.0 ], "spans": [], "index": 1 }, { "bbox": [ 70, 141.0, 525, 177.0 ], "spans": [], "index": 2 } ] } ], "index": 1 }, { "type": "text", "bbox": [ 69, 184, 527, 280 ], "lines": [ { "bbox": [ 68, 185, 527, 198 ], "spans": [ { "bbox": [ 68, 185, 527, 198 ], "score": 1.0, "content": "Table 1: Main Results of SpeechGPT. S2SIF refers to speech-to-speech instruction-following, S2TIF is speech-to-", "type": "text" } ], "index": 3 }, { "bbox": [ 68, 196, 526, 210 ], "spans": [ { "bbox": [ 68, 196, 526, 210 ], "score": 1.0, "content": "text instruction-following, T2SIF denotes text-to-speech instruction-following and T2TIF represents text-to-text", "type": "text" } ], "index": 4 }, { "bbox": [ 68, 208, 526, 222 ], "spans": [ { "bbox": [ 68, 208, 526, 222 ], "score": 1.0, "content": "instruction-following. ChatGPT score is obtained through ChatGPT evaluatation. CMOS refers to content mean", "type": "text" } ], "index": 5 }, { "bbox": [ 68, 221, 526, 234 ], "spans": [ { "bbox": [ 68, 221, 526, 234 ], "score": 1.0, "content": "opinion score. NMOS denotes naturalness mean opinion score. ∗: The low ChatGPT Score for speech responses in", "type": "text" } ], "index": 6 }, { "bbox": [ 69, 233, 527, 246 ], "spans": [ { "bbox": [ 69, 233, 527, 246 ], "score": 1.0, "content": "Ground Truth is due to them being synthesized by TTS system, which can have errors in pauses between sentences.", "type": "text" } ], "index": 7 }, { "bbox": [ 69, 245, 526, 258 ], "spans": [ { "bbox": [ 69, 245, 526, 258 ], "score": 1.0, "content": "This introduces significant errors for longer responses, leading to incorrect text after being processed by the ASR", "type": "text" } ], "index": 8 }, { "bbox": [ 68, 256, 526, 270 ], "spans": [ { "bbox": [ 68, 256, 526, 270 ], "score": 1.0, "content": "system, thereby reducing the score. However, humans can understand the content of such speech, so the CMOS", "type": "text" } ], "index": 9 }, { "bbox": [ 68, 268, 135, 281 ], "spans": [ { "bbox": [ 68, 268, 135, 281 ], "score": 1.0, "content": "score is normal.", "type": "text" } ], "index": 10 } ], "index": 6.5 }, { "type": "table", "bbox": [ 81, 294, 276, 368 ], "blocks": [ { "type": "table_body", "bbox": [ 81, 294, 276, 368 ], "group_id": 0, "lines": [ { "bbox": [ 81, 294, 276, 368 ], "spans": [ { "bbox": [ 81, 294, 276, 368 ], "score": 0.976, "html": "
TrainingInferenceChatGPT Score
StandardStandard2.15
StandardCoM2.12
CoMStandard2.35
CoMCoM3.42
", "type": "table", "image_path": "1d61b01892386d5a98c656e490472f5748ac72f5299b244e6050df23e1219fca.jpg" } ] } ], "index": 13, "virtual_lines": [ { "bbox": [ 81, 294, 276, 308.8 ], "spans": [], "index": 11 }, { "bbox": [ 81, 308.8, 276, 323.6 ], "spans": [], "index": 12 }, { "bbox": [ 81, 323.6, 276, 338.40000000000003 ], "spans": [], "index": 13 }, { "bbox": [ 81, 338.40000000000003, 276, 353.20000000000005 ], "spans": [], "index": 14 }, { "bbox": [ 81, 353.20000000000005, 276, 368.00000000000006 ], "spans": [], "index": 15 } ] }, { "type": "table_caption", "bbox": [ 69, 376, 290, 425 ], "group_id": 0, "lines": [ { "bbox": [ 68, 375, 290, 389 ], "spans": [ { "bbox": [ 68, 375, 290, 389 ], "score": 1.0, "content": "Table 2: ChatGPT Score on speech-to-speech", "type": "text" } ], "index": 16 }, { "bbox": [ 69, 388, 291, 399 ], "spans": [ { "bbox": [ 69, 388, 291, 399 ], "score": 1.0, "content": "instruction-following task. CoM refers to chain-of-", "type": "text" } ], "index": 17 }, { "bbox": [ 68, 400, 291, 413 ], "spans": [ { "bbox": [ 68, 400, 291, 413 ], "score": 1.0, "content": "modality prompting and Standard denotes standard", "type": "text" } ], "index": 18 }, { "bbox": [ 68, 412, 116, 426 ], "spans": [ { "bbox": [ 68, 412, 116, 426 ], "score": 1.0, "content": "prompting.", "type": "text" } ], "index": 19 } ], "index": 17.5 } ], "index": 15.25 }, { "type": "text", "bbox": [ 70, 431, 290, 471 ], "lines": [ { "bbox": [ 68, 430, 291, 445 ], "spans": [ { "bbox": [ 68, 430, 291, 445 ], "score": 1.0, "content": "perform the TTS system in the cascaded system in", "type": "text" } ], "index": 20 }, { "bbox": [ 68, 446, 291, 457 ], "spans": [ { "bbox": [ 68, 446, 291, 457 ], "score": 1.0, "content": "terms of audio quality and prosody. More detailed", "type": "text" } ], "index": 21 }, { "bbox": [ 69, 459, 291, 471 ], "spans": [ { "bbox": [ 69, 459, 291, 471 ], "score": 1.0, "content": "speech prosody analysis are located in Section ??.", "type": "text" } ], "index": 22 } ], "index": 21 }, { "type": "title", "bbox": [ 70, 486, 132, 499 ], "lines": [ { "bbox": [ 67, 484, 134, 502 ], "spans": [ { "bbox": [ 67, 484, 134, 502 ], "score": 1.0, "content": "6 Analysis", "type": "text" } ], "index": 23 } ], "index": 23 }, { "type": "title", "bbox": [ 70, 511, 272, 524 ], "lines": [ { "bbox": [ 67, 509, 274, 527 ], "spans": [ { "bbox": [ 67, 509, 274, 527 ], "score": 1.0, "content": "6.1 Chain-of-modality prompting matters", "type": "text" } ], "index": 24 } ], "index": 24 }, { "type": "text", "bbox": [ 69, 530, 290, 774 ], "lines": [ { "bbox": [ 68, 529, 291, 544 ], "spans": [ { "bbox": [ 68, 529, 291, 544 ], "score": 1.0, "content": "Table 2 shows ChatGPT Scores on speech-to-", "type": "text" } ], "index": 25 }, { "bbox": [ 69, 545, 292, 557 ], "spans": [ { "bbox": [ 69, 545, 292, 557 ], "score": 1.0, "content": "speech instruction-following task for models uti-", "type": "text" } ], "index": 26 }, { "bbox": [ 68, 557, 291, 572 ], "spans": [ { "bbox": [ 68, 557, 291, 572 ], "score": 1.0, "content": "lizing standard prompting and chain-of-modality", "type": "text" } ], "index": 27 }, { "bbox": [ 69, 572, 291, 585 ], "spans": [ { "bbox": [ 69, 572, 291, 585 ], "score": 1.0, "content": "prompting during training and inference stages re-", "type": "text" } ], "index": 28 }, { "bbox": [ 68, 585, 290, 599 ], "spans": [ { "bbox": [ 68, 585, 290, 599 ], "score": 1.0, "content": "spectively. Standard prompting refers to directly", "type": "text" } ], "index": 29 }, { "bbox": [ 69, 598, 291, 612 ], "spans": [ { "bbox": [ 69, 598, 291, 612 ], "score": 1.0, "content": "obtaining a speech response from a speech in-", "type": "text" } ], "index": 30 }, { "bbox": [ 68, 612, 292, 626 ], "spans": [ { "bbox": [ 68, 612, 292, 626 ], "score": 1.0, "content": "struction without transitioning through an inter-", "type": "text" } ], "index": 31 }, { "bbox": [ 69, 626, 290, 639 ], "spans": [ { "bbox": [ 69, 626, 290, 639 ], "score": 1.0, "content": "mediate text form. The template can be located", "type": "text" } ], "index": 32 }, { "bbox": [ 68, 639, 291, 653 ], "spans": [ { "bbox": [ 68, 639, 291, 653 ], "score": 1.0, "content": "in Appendix E. For standard prompting training,", "type": "text" } ], "index": 33 }, { "bbox": [ 68, 653, 291, 666 ], "spans": [ { "bbox": [ 68, 653, 291, 666 ], "score": 1.0, "content": "we use this template to construct training data.", "type": "text" } ], "index": 34 }, { "bbox": [ 68, 666, 291, 680 ], "spans": [ { "bbox": [ 68, 666, 291, 680 ], "score": 1.0, "content": "We discovered that if standard prompting is used,", "type": "text" } ], "index": 35 }, { "bbox": [ 69, 680, 291, 693 ], "spans": [ { "bbox": [ 69, 680, 291, 693 ], "score": 1.0, "content": "the performance is rather poor when either stan-", "type": "text" } ], "index": 36 }, { "bbox": [ 68, 693, 290, 708 ], "spans": [ { "bbox": [ 68, 693, 290, 708 ], "score": 1.0, "content": "dard prompting or chain-of-modality prompting is", "type": "text" } ], "index": 37 }, { "bbox": [ 68, 705, 291, 721 ], "spans": [ { "bbox": [ 68, 705, 291, 721 ], "score": 1.0, "content": "used for inference. If chain-of-modality prompt-", "type": "text" } ], "index": 38 }, { "bbox": [ 68, 720, 291, 733 ], "spans": [ { "bbox": [ 68, 720, 291, 733 ], "score": 1.0, "content": "ing is employed during training, ChatGPT Score", "type": "text" } ], "index": 39 }, { "bbox": [ 68, 734, 290, 747 ], "spans": [ { "bbox": [ 68, 734, 290, 747 ], "score": 1.0, "content": "sees an enhancement, and when the inference also", "type": "text" } ], "index": 40 }, { "bbox": [ 68, 747, 291, 761 ], "spans": [ { "bbox": [ 68, 747, 291, 761 ], "score": 1.0, "content": "applies chain-of-modality prompting, there is a", "type": "text" } ], "index": 41 }, { "bbox": [ 68, 761, 292, 774 ], "spans": [ { "bbox": [ 68, 761, 292, 774 ], "score": 1.0, "content": "huge improvement in performance. This indi-", "type": "text" } ], "index": 42 } ], "index": 33.5 }, { "type": "image", "bbox": [ 306, 309, 504, 459 ], "blocks": [ { "type": "image_body", "bbox": [ 306, 309, 504, 459 ], "group_id": 0, "lines": [ { "bbox": [ 306, 309, 504, 459 ], "spans": [ { "bbox": [ 306, 309, 504, 459 ], "score": 0.969, "type": "image", "image_path": "da7e7b4aab6aa78969c08f258ac74e2b540f2e61c6360fd974bc580739c6c728.jpg" } ] } ], "index": 48, "virtual_lines": [ { "bbox": [ 306, 309, 504, 322.6363636363636 ], "spans": [], "index": 43 }, { "bbox": [ 306, 322.6363636363636, 504, 336.27272727272725 ], "spans": [], "index": 44 }, { "bbox": [ 306, 336.27272727272725, 504, 349.9090909090909 ], "spans": [], "index": 45 }, { "bbox": [ 306, 349.9090909090909, 504, 363.5454545454545 ], "spans": [], "index": 46 }, { "bbox": [ 306, 363.5454545454545, 504, 377.18181818181813 ], "spans": [], "index": 47 }, { "bbox": [ 306, 377.18181818181813, 504, 390.81818181818176 ], "spans": [], "index": 48 }, { "bbox": [ 306, 390.81818181818176, 504, 404.4545454545454 ], "spans": [], "index": 49 }, { "bbox": [ 306, 404.4545454545454, 504, 418.090909090909 ], "spans": [], "index": 50 }, { "bbox": [ 306, 418.090909090909, 504, 431.72727272727263 ], "spans": [], "index": 51 }, { "bbox": [ 306, 431.72727272727263, 504, 445.36363636363626 ], "spans": [], "index": 52 }, { "bbox": [ 306, 445.36363636363626, 504, 458.9999999999999 ], "spans": [], "index": 53 } ] }, { "type": "image_caption", "bbox": [ 304, 468, 525, 528 ], "group_id": 0, "lines": [ { "bbox": [ 304, 469, 527, 480 ], "spans": [ { "bbox": [ 304, 469, 527, 480 ], "score": 1.0, "content": "Figure 3: ASR-PPL of speech continue task on 100 ut-", "type": "text" } ], "index": 54 }, { "bbox": [ 304, 481, 525, 492 ], "spans": [ { "bbox": [ 304, 481, 525, 492 ], "score": 1.0, "content": "terances from LibriSpeech test-clean set. From scratch", "type": "text" } ], "index": 55 }, { "bbox": [ 304, 492, 525, 504 ], "spans": [ { "bbox": [ 304, 492, 525, 504 ], "score": 1.0, "content": "refers to model pre-trained from randomly-initialized", "type": "text" } ], "index": 56 }, { "bbox": [ 303, 505, 525, 516 ], "spans": [ { "bbox": [ 303, 505, 525, 516 ], "score": 1.0, "content": "parameters. From LLaMA denotes model pre-trained", "type": "text" } ], "index": 57 }, { "bbox": [ 304, 517, 364, 528 ], "spans": [ { "bbox": [ 304, 517, 364, 528 ], "score": 1.0, "content": "from LLaMA.", "type": "text" } ], "index": 58 } ], "index": 56 } ], "index": 52.0 }, { "type": "text", "bbox": [ 305, 535, 525, 602 ], "lines": [ { "bbox": [ 304, 535, 525, 549 ], "spans": [ { "bbox": [ 304, 535, 525, 549 ], "score": 1.0, "content": "cates that chain-of-modality prompting matters", "type": "text" } ], "index": 59 }, { "bbox": [ 303, 549, 527, 562 ], "spans": [ { "bbox": [ 303, 549, 527, 562 ], "score": 1.0, "content": "in both training and inference. We think chain-of-", "type": "text" } ], "index": 60 }, { "bbox": [ 303, 563, 526, 576 ], "spans": [ { "bbox": [ 303, 563, 526, 576 ], "score": 1.0, "content": "modality prompting decomposes the complex task", "type": "text" } ], "index": 61 }, { "bbox": [ 304, 576, 525, 590 ], "spans": [ { "bbox": [ 304, 576, 525, 590 ], "score": 1.0, "content": "into easy tasks, allowing the model to complete", "type": "text" } ], "index": 62 }, { "bbox": [ 304, 589, 511, 603 ], "spans": [ { "bbox": [ 304, 589, 511, 603 ], "score": 1.0, "content": "them step by step, which reduces the difficulty.", "type": "text" } ], "index": 63 } ], "index": 61 }, { "type": "title", "bbox": [ 305, 618, 491, 644 ], "lines": [ { "bbox": [ 303, 616, 491, 632 ], "spans": [ { "bbox": [ 303, 616, 491, 632 ], "score": 1.0, "content": "6.2 Can text knowledge benefit speech", "type": "text" } ], "index": 64 }, { "bbox": [ 328, 631, 379, 646 ], "spans": [ { "bbox": [ 328, 631, 379, 646 ], "score": 1.0, "content": "modality?", "type": "text" } ], "index": 65 } ], "index": 64.5 }, { "type": "text", "bbox": [ 305, 653, 525, 773 ], "lines": [ { "bbox": [ 304, 652, 526, 666 ], "spans": [ { "bbox": [ 304, 652, 526, 666 ], "score": 1.0, "content": "SpeechGPT originates from a text pre-trained", "type": "text" } ], "index": 66 }, { "bbox": [ 303, 666, 525, 679 ], "spans": [ { "bbox": [ 303, 666, 525, 679 ], "score": 1.0, "content": "model, LLaMA. Nonetheless, the question remains", "type": "text" } ], "index": 67 }, { "bbox": [ 304, 679, 526, 693 ], "spans": [ { "bbox": [ 304, 679, 526, 693 ], "score": 1.0, "content": "whether the knowledge from the text modality can", "type": "text" } ], "index": 68 }, { "bbox": [ 304, 694, 525, 706 ], "spans": [ { "bbox": [ 304, 694, 525, 706 ], "score": 1.0, "content": "contribute beneficially to the speech modality. To", "type": "text" } ], "index": 69 }, { "bbox": [ 304, 708, 526, 720 ], "spans": [ { "bbox": [ 304, 708, 526, 720 ], "score": 1.0, "content": "resolve this, we utilize a speech continuation task", "type": "text" } ], "index": 70 }, { "bbox": [ 304, 720, 526, 734 ], "spans": [ { "bbox": [ 304, 720, 526, 734 ], "score": 1.0, "content": "which assesses the model’s capability to generate", "type": "text" } ], "index": 71 }, { "bbox": [ 304, 735, 526, 747 ], "spans": [ { "bbox": [ 304, 735, 526, 747 ], "score": 1.0, "content": "coherent and semantically accurate speech. We", "type": "text" } ], "index": 72 }, { "bbox": [ 304, 748, 525, 760 ], "spans": [ { "bbox": [ 304, 748, 525, 760 ], "score": 1.0, "content": "compare the performances of two models on this", "type": "text" } ], "index": 73 }, { "bbox": [ 304, 761, 525, 773 ], "spans": [ { "bbox": [ 304, 761, 525, 773 ], "score": 1.0, "content": "task: one model is pre-trained from LLaMA, while", "type": "text" } ], "index": 74 } ], "index": 70 } ], "page_idx": 6, "page_size": [ 595, 841 ], "discarded_blocks": [], "para_blocks": [ { "type": "table", "bbox": [ 70, 69, 525, 177 ], "blocks": [ { "type": "table_body", "bbox": [ 70, 69, 525, 177 ], "group_id": 1, "lines": [ { "bbox": [ 70, 69, 525, 177 ], "spans": [ { "bbox": [ 70, 69, 525, 177 ], "score": 0.977, "html": "
MethodsChatGPT ScoreHuman Opinion Score
CMOSNMOS
S2SIFS2TIFT2SIFT2TIFS2SIFS2TIFT2SIFT2TIFS2SIFS2TIFT2SIFT2TIF
Ground Truth2.85*3.742.91*3.933.783.893.954.123.18-3.20-
Baselines: cascaded cross-modal conversational systems
Speech-Alpaca-13B2.743.312.713.833.393.423.713.753.123.131
Speech-LLaMA-MOSS-0022.873.503.233.823.383.443.743.833.143.111
SpeechGPT3.423.523.533.643.423.493.573.693.65-3.621
", "type": "table", "image_path": "9a7fcd22dd35afec89ad794b756a0b89d8a71276f648f06dff3208200eb58e3d.jpg" } ] } ], "index": 1, "virtual_lines": [ { "bbox": [ 70, 69, 525, 105.0 ], "spans": [], "index": 0 }, { "bbox": [ 70, 105.0, 525, 141.0 ], "spans": [], "index": 1 }, { "bbox": [ 70, 141.0, 525, 177.0 ], "spans": [], "index": 2 } ] } ], "index": 1 }, { "type": "text", "bbox": [ 69, 184, 527, 280 ], "lines": [ { "bbox": [ 68, 185, 527, 198 ], "spans": [ { "bbox": [ 68, 185, 527, 198 ], "score": 1.0, "content": "Table 1: Main Results of SpeechGPT. S2SIF refers to speech-to-speech instruction-following, S2TIF is speech-to-", "type": "text" } ], "index": 3 }, { "bbox": [ 68, 196, 526, 210 ], "spans": [ { "bbox": [ 68, 196, 526, 210 ], "score": 1.0, "content": "text instruction-following, T2SIF denotes text-to-speech instruction-following and T2TIF represents text-to-text", "type": "text" } ], "index": 4 }, { "bbox": [ 68, 208, 526, 222 ], "spans": [ { "bbox": [ 68, 208, 526, 222 ], "score": 1.0, "content": "instruction-following. ChatGPT score is obtained through ChatGPT evaluatation. CMOS refers to content mean", "type": "text" } ], "index": 5 }, { "bbox": [ 68, 221, 526, 234 ], "spans": [ { "bbox": [ 68, 221, 526, 234 ], "score": 1.0, "content": "opinion score. NMOS denotes naturalness mean opinion score. ∗: The low ChatGPT Score for speech responses in", "type": "text" } ], "index": 6 }, { "bbox": [ 69, 233, 527, 246 ], "spans": [ { "bbox": [ 69, 233, 527, 246 ], "score": 1.0, "content": "Ground Truth is due to them being synthesized by TTS system, which can have errors in pauses between sentences.", "type": "text" } ], "index": 7 }, { "bbox": [ 69, 245, 526, 258 ], "spans": [ { "bbox": [ 69, 245, 526, 258 ], "score": 1.0, "content": "This introduces significant errors for longer responses, leading to incorrect text after being processed by the ASR", "type": "text" } ], "index": 8 }, { "bbox": [ 68, 256, 526, 270 ], "spans": [ { "bbox": [ 68, 256, 526, 270 ], "score": 1.0, "content": "system, thereby reducing the score. However, humans can understand the content of such speech, so the CMOS", "type": "text" } ], "index": 9 }, { "bbox": [ 68, 268, 135, 281 ], "spans": [ { "bbox": [ 68, 268, 135, 281 ], "score": 1.0, "content": "score is normal.", "type": "text" } ], "index": 10 } ], "index": 6.5, "bbox_fs": [ 68, 185, 527, 281 ] }, { "type": "table", "bbox": [ 81, 294, 276, 368 ], "blocks": [ { "type": "table_body", "bbox": [ 81, 294, 276, 368 ], "group_id": 0, "lines": [ { "bbox": [ 81, 294, 276, 368 ], "spans": [ { "bbox": [ 81, 294, 276, 368 ], "score": 0.976, "html": "
TrainingInferenceChatGPT Score
StandardStandard2.15
StandardCoM2.12
CoMStandard2.35
CoMCoM3.42
", "type": "table", "image_path": "1d61b01892386d5a98c656e490472f5748ac72f5299b244e6050df23e1219fca.jpg" } ] } ], "index": 13, "virtual_lines": [ { "bbox": [ 81, 294, 276, 308.8 ], "spans": [], "index": 11 }, { "bbox": [ 81, 308.8, 276, 323.6 ], "spans": [], "index": 12 }, { "bbox": [ 81, 323.6, 276, 338.40000000000003 ], "spans": [], "index": 13 }, { "bbox": [ 81, 338.40000000000003, 276, 353.20000000000005 ], "spans": [], "index": 14 }, { "bbox": [ 81, 353.20000000000005, 276, 368.00000000000006 ], "spans": [], "index": 15 } ] }, { "type": "table_caption", "bbox": [ 69, 376, 290, 425 ], "group_id": 0, "lines": [ { "bbox": [ 68, 375, 290, 389 ], "spans": [ { "bbox": [ 68, 375, 290, 389 ], "score": 1.0, "content": "Table 2: ChatGPT Score on speech-to-speech", "type": "text" } ], "index": 16 }, { "bbox": [ 69, 388, 291, 399 ], "spans": [ { "bbox": [ 69, 388, 291, 399 ], "score": 1.0, "content": "instruction-following task. CoM refers to chain-of-", "type": "text" } ], "index": 17 }, { "bbox": [ 68, 400, 291, 413 ], "spans": [ { "bbox": [ 68, 400, 291, 413 ], "score": 1.0, "content": "modality prompting and Standard denotes standard", "type": "text" } ], "index": 18 }, { "bbox": [ 68, 412, 116, 426 ], "spans": [ { "bbox": [ 68, 412, 116, 426 ], "score": 1.0, "content": "prompting.", "type": "text" } ], "index": 19 } ], "index": 17.5 } ], "index": 15.25 }, { "type": "text", "bbox": [ 70, 431, 290, 471 ], "lines": [ { "bbox": [ 68, 430, 291, 445 ], "spans": [ { "bbox": [ 68, 430, 291, 445 ], "score": 1.0, "content": "perform the TTS system in the cascaded system in", "type": "text" } ], "index": 20 }, { "bbox": [ 68, 446, 291, 457 ], "spans": [ { "bbox": [ 68, 446, 291, 457 ], "score": 1.0, "content": "terms of audio quality and prosody. More detailed", "type": "text" } ], "index": 21 }, { "bbox": [ 69, 459, 291, 471 ], "spans": [ { "bbox": [ 69, 459, 291, 471 ], "score": 1.0, "content": "speech prosody analysis are located in Section ??.", "type": "text" } ], "index": 22 } ], "index": 21, "bbox_fs": [ 68, 430, 291, 471 ] }, { "type": "title", "bbox": [ 70, 486, 132, 499 ], "lines": [ { "bbox": [ 67, 484, 134, 502 ], "spans": [ { "bbox": [ 67, 484, 134, 502 ], "score": 1.0, "content": "6 Analysis", "type": "text" } ], "index": 23 } ], "index": 23 }, { "type": "title", "bbox": [ 70, 511, 272, 524 ], "lines": [ { "bbox": [ 67, 509, 274, 527 ], "spans": [ { "bbox": [ 67, 509, 274, 527 ], "score": 1.0, "content": "6.1 Chain-of-modality prompting matters", "type": "text" } ], "index": 24 } ], "index": 24 }, { "type": "text", "bbox": [ 69, 530, 290, 774 ], "lines": [ { "bbox": [ 68, 529, 291, 544 ], "spans": [ { "bbox": [ 68, 529, 291, 544 ], "score": 1.0, "content": "Table 2 shows ChatGPT Scores on speech-to-", "type": "text" } ], "index": 25 }, { "bbox": [ 69, 545, 292, 557 ], "spans": [ { "bbox": [ 69, 545, 292, 557 ], "score": 1.0, "content": "speech instruction-following task for models uti-", "type": "text" } ], "index": 26 }, { "bbox": [ 68, 557, 291, 572 ], "spans": [ { "bbox": [ 68, 557, 291, 572 ], "score": 1.0, "content": "lizing standard prompting and chain-of-modality", "type": "text" } ], "index": 27 }, { "bbox": [ 69, 572, 291, 585 ], "spans": [ { "bbox": [ 69, 572, 291, 585 ], "score": 1.0, "content": "prompting during training and inference stages re-", "type": "text" } ], "index": 28 }, { "bbox": [ 68, 585, 290, 599 ], "spans": [ { "bbox": [ 68, 585, 290, 599 ], "score": 1.0, "content": "spectively. Standard prompting refers to directly", "type": "text" } ], "index": 29 }, { "bbox": [ 69, 598, 291, 612 ], "spans": [ { "bbox": [ 69, 598, 291, 612 ], "score": 1.0, "content": "obtaining a speech response from a speech in-", "type": "text" } ], "index": 30 }, { "bbox": [ 68, 612, 292, 626 ], "spans": [ { "bbox": [ 68, 612, 292, 626 ], "score": 1.0, "content": "struction without transitioning through an inter-", "type": "text" } ], "index": 31 }, { "bbox": [ 69, 626, 290, 639 ], "spans": [ { "bbox": [ 69, 626, 290, 639 ], "score": 1.0, "content": "mediate text form. The template can be located", "type": "text" } ], "index": 32 }, { "bbox": [ 68, 639, 291, 653 ], "spans": [ { "bbox": [ 68, 639, 291, 653 ], "score": 1.0, "content": "in Appendix E. For standard prompting training,", "type": "text" } ], "index": 33 }, { "bbox": [ 68, 653, 291, 666 ], "spans": [ { "bbox": [ 68, 653, 291, 666 ], "score": 1.0, "content": "we use this template to construct training data.", "type": "text" } ], "index": 34 }, { "bbox": [ 68, 666, 291, 680 ], "spans": [ { "bbox": [ 68, 666, 291, 680 ], "score": 1.0, "content": "We discovered that if standard prompting is used,", "type": "text" } ], "index": 35 }, { "bbox": [ 69, 680, 291, 693 ], "spans": [ { "bbox": [ 69, 680, 291, 693 ], "score": 1.0, "content": "the performance is rather poor when either stan-", "type": "text" } ], "index": 36 }, { "bbox": [ 68, 693, 290, 708 ], "spans": [ { "bbox": [ 68, 693, 290, 708 ], "score": 1.0, "content": "dard prompting or chain-of-modality prompting is", "type": "text" } ], "index": 37 }, { "bbox": [ 68, 705, 291, 721 ], "spans": [ { "bbox": [ 68, 705, 291, 721 ], "score": 1.0, "content": "used for inference. If chain-of-modality prompt-", "type": "text" } ], "index": 38 }, { "bbox": [ 68, 720, 291, 733 ], "spans": [ { "bbox": [ 68, 720, 291, 733 ], "score": 1.0, "content": "ing is employed during training, ChatGPT Score", "type": "text" } ], "index": 39 }, { "bbox": [ 68, 734, 290, 747 ], "spans": [ { "bbox": [ 68, 734, 290, 747 ], "score": 1.0, "content": "sees an enhancement, and when the inference also", "type": "text" } ], "index": 40 }, { "bbox": [ 68, 747, 291, 761 ], "spans": [ { "bbox": [ 68, 747, 291, 761 ], "score": 1.0, "content": "applies chain-of-modality prompting, there is a", "type": "text" } ], "index": 41 }, { "bbox": [ 68, 761, 292, 774 ], "spans": [ { "bbox": [ 68, 761, 292, 774 ], "score": 1.0, "content": "huge improvement in performance. This indi-", "type": "text" } ], "index": 42 }, { "bbox": [ 304, 535, 525, 549 ], "spans": [ { "bbox": [ 304, 535, 525, 549 ], "score": 1.0, "content": "cates that chain-of-modality prompting matters", "type": "text" } ], "index": 59 }, { "bbox": [ 303, 549, 527, 562 ], "spans": [ { "bbox": [ 303, 549, 527, 562 ], "score": 1.0, "content": "in both training and inference. We think chain-of-", "type": "text" } ], "index": 60 }, { "bbox": [ 303, 563, 526, 576 ], "spans": [ { "bbox": [ 303, 563, 526, 576 ], "score": 1.0, "content": "modality prompting decomposes the complex task", "type": "text" } ], "index": 61 }, { "bbox": [ 304, 576, 525, 590 ], "spans": [ { "bbox": [ 304, 576, 525, 590 ], "score": 1.0, "content": "into easy tasks, allowing the model to complete", "type": "text" } ], "index": 62 }, { "bbox": [ 304, 589, 511, 603 ], "spans": [ { "bbox": [ 304, 589, 511, 603 ], "score": 1.0, "content": "them step by step, which reduces the difficulty.", "type": "text" } ], "index": 63 } ], "index": 33.5, "bbox_fs": [ 68, 529, 292, 774 ] }, { "type": "image", "bbox": [ 306, 309, 504, 459 ], "blocks": [ { "type": "image_body", "bbox": [ 306, 309, 504, 459 ], "group_id": 0, "lines": [ { "bbox": [ 306, 309, 504, 459 ], "spans": [ { "bbox": [ 306, 309, 504, 459 ], "score": 0.969, "type": "image", "image_path": "da7e7b4aab6aa78969c08f258ac74e2b540f2e61c6360fd974bc580739c6c728.jpg" } ] } ], "index": 48, "virtual_lines": [ { "bbox": [ 306, 309, 504, 322.6363636363636 ], "spans": [], "index": 43 }, { "bbox": [ 306, 322.6363636363636, 504, 336.27272727272725 ], "spans": [], "index": 44 }, { "bbox": [ 306, 336.27272727272725, 504, 349.9090909090909 ], "spans": [], "index": 45 }, { "bbox": [ 306, 349.9090909090909, 504, 363.5454545454545 ], "spans": [], "index": 46 }, { "bbox": [ 306, 363.5454545454545, 504, 377.18181818181813 ], "spans": [], "index": 47 }, { "bbox": [ 306, 377.18181818181813, 504, 390.81818181818176 ], "spans": [], "index": 48 }, { "bbox": [ 306, 390.81818181818176, 504, 404.4545454545454 ], "spans": [], "index": 49 }, { "bbox": [ 306, 404.4545454545454, 504, 418.090909090909 ], "spans": [], "index": 50 }, { "bbox": [ 306, 418.090909090909, 504, 431.72727272727263 ], "spans": [], "index": 51 }, { "bbox": [ 306, 431.72727272727263, 504, 445.36363636363626 ], "spans": [], "index": 52 }, { "bbox": [ 306, 445.36363636363626, 504, 458.9999999999999 ], "spans": [], "index": 53 } ] }, { "type": "image_caption", "bbox": [ 304, 468, 525, 528 ], "group_id": 0, "lines": [ { "bbox": [ 304, 469, 527, 480 ], "spans": [ { "bbox": [ 304, 469, 527, 480 ], "score": 1.0, "content": "Figure 3: ASR-PPL of speech continue task on 100 ut-", "type": "text" } ], "index": 54 }, { "bbox": [ 304, 481, 525, 492 ], "spans": [ { "bbox": [ 304, 481, 525, 492 ], "score": 1.0, "content": "terances from LibriSpeech test-clean set. From scratch", "type": "text" } ], "index": 55 }, { "bbox": [ 304, 492, 525, 504 ], "spans": [ { "bbox": [ 304, 492, 525, 504 ], "score": 1.0, "content": "refers to model pre-trained from randomly-initialized", "type": "text" } ], "index": 56 }, { "bbox": [ 303, 505, 525, 516 ], "spans": [ { "bbox": [ 303, 505, 525, 516 ], "score": 1.0, "content": "parameters. From LLaMA denotes model pre-trained", "type": "text" } ], "index": 57 }, { "bbox": [ 304, 517, 364, 528 ], "spans": [ { "bbox": [ 304, 517, 364, 528 ], "score": 1.0, "content": "from LLaMA.", "type": "text" } ], "index": 58 } ], "index": 56 } ], "index": 52.0 }, { "type": "text", "bbox": [ 305, 535, 525, 602 ], "lines": [], "index": 61, "bbox_fs": [ 303, 535, 527, 603 ], "lines_deleted": true }, { "type": "title", "bbox": [ 305, 618, 491, 644 ], "lines": [ { "bbox": [ 303, 616, 491, 632 ], "spans": [ { "bbox": [ 303, 616, 491, 632 ], "score": 1.0, "content": "6.2 Can text knowledge benefit speech", "type": "text" } ], "index": 64 }, { "bbox": [ 328, 631, 379, 646 ], "spans": [ { "bbox": [ 328, 631, 379, 646 ], "score": 1.0, "content": "modality?", "type": "text" } ], "index": 65 } ], "index": 64.5 }, { "type": "text", "bbox": [ 305, 653, 525, 773 ], "lines": [ { "bbox": [ 304, 652, 526, 666 ], "spans": [ { "bbox": [ 304, 652, 526, 666 ], "score": 1.0, "content": "SpeechGPT originates from a text pre-trained", "type": "text" } ], "index": 66 }, { "bbox": [ 303, 666, 525, 679 ], "spans": [ { "bbox": [ 303, 666, 525, 679 ], "score": 1.0, "content": "model, LLaMA. Nonetheless, the question remains", "type": "text" } ], "index": 67 }, { "bbox": [ 304, 679, 526, 693 ], "spans": [ { "bbox": [ 304, 679, 526, 693 ], "score": 1.0, "content": "whether the knowledge from the text modality can", "type": "text" } ], "index": 68 }, { "bbox": [ 304, 694, 525, 706 ], "spans": [ { "bbox": [ 304, 694, 525, 706 ], "score": 1.0, "content": "contribute beneficially to the speech modality. To", "type": "text" } ], "index": 69 }, { "bbox": [ 304, 708, 526, 720 ], "spans": [ { "bbox": [ 304, 708, 526, 720 ], "score": 1.0, "content": "resolve this, we utilize a speech continuation task", "type": "text" } ], "index": 70 }, { "bbox": [ 304, 720, 526, 734 ], "spans": [ { "bbox": [ 304, 720, 526, 734 ], "score": 1.0, "content": "which assesses the model’s capability to generate", "type": "text" } ], "index": 71 }, { "bbox": [ 304, 735, 526, 747 ], "spans": [ { "bbox": [ 304, 735, 526, 747 ], "score": 1.0, "content": "coherent and semantically accurate speech. We", "type": "text" } ], "index": 72 }, { "bbox": [ 304, 748, 525, 760 ], "spans": [ { "bbox": [ 304, 748, 525, 760 ], "score": 1.0, "content": "compare the performances of two models on this", "type": "text" } ], "index": 73 }, { "bbox": [ 304, 761, 525, 773 ], "spans": [ { "bbox": [ 304, 761, 525, 773 ], "score": 1.0, "content": "task: one model is pre-trained from LLaMA, while", "type": "text" } ], "index": 74 }, { "bbox": [ 69, 298, 243, 312 ], "spans": [ { "bbox": [ 69, 298, 243, 312 ], "score": 1.0, "content": "the other model is trained from scratch.", "type": "text", "cross_page": true } ], "index": 15 } ], "index": 70, "bbox_fs": [ 303, 652, 526, 773 ] } ] }, { "preproc_blocks": [ { "type": "image", "bbox": [ 73, 84, 268, 234 ], "blocks": [ { "type": "image_body", "bbox": [ 73, 84, 268, 234 ], "group_id": 0, "lines": [ { "bbox": [ 73, 84, 268, 234 ], "spans": [ { "bbox": [ 73, 84, 268, 234 ], "score": 0.972, "type": "image", "image_path": "e46aa44b83753245531eae6af65862d3d2b33291238053589833d6d978b8c55e.jpg" } ] } ], "index": 5, "virtual_lines": [ { "bbox": [ 73, 84, 268, 97.63636363636364 ], "spans": [], "index": 0 }, { "bbox": [ 73, 97.63636363636364, 268, 111.27272727272728 ], "spans": [], "index": 1 }, { "bbox": [ 73, 111.27272727272728, 268, 124.90909090909092 ], "spans": [], "index": 2 }, { "bbox": [ 73, 124.90909090909092, 268, 138.54545454545456 ], "spans": [], "index": 3 }, { "bbox": [ 73, 138.54545454545456, 268, 152.1818181818182 ], "spans": [], "index": 4 }, { "bbox": [ 73, 152.1818181818182, 268, 165.8181818181818 ], "spans": [], "index": 5 }, { "bbox": [ 73, 165.8181818181818, 268, 179.45454545454544 ], "spans": [], "index": 6 }, { "bbox": [ 73, 179.45454545454544, 268, 193.09090909090907 ], "spans": [], "index": 7 }, { "bbox": [ 73, 193.09090909090907, 268, 206.7272727272727 ], "spans": [], "index": 8 }, { "bbox": [ 73, 206.7272727272727, 268, 220.36363636363632 ], "spans": [], "index": 9 }, { "bbox": [ 73, 220.36363636363632, 268, 233.99999999999994 ], "spans": [], "index": 10 } ] }, { "type": "image_caption", "bbox": [ 69, 244, 290, 291 ], "group_id": 0, "lines": [ { "bbox": [ 69, 244, 291, 255 ], "spans": [ { "bbox": [ 69, 244, 291, 255 ], "score": 1.0, "content": "Figure 4: ChatGPT Score on text-to-text instruction-", "type": "text" } ], "index": 11 }, { "bbox": [ 68, 254, 291, 269 ], "spans": [ { "bbox": [ 68, 254, 291, 269 ], "score": 1.0, "content": "following task. LLaMA-MOSS-002 is obtained by per-", "type": "text" } ], "index": 12 }, { "bbox": [ 68, 267, 290, 281 ], "spans": [ { "bbox": [ 68, 267, 290, 281 ], "score": 1.0, "content": "forming supervised fine-tuning on LLaMA-13B using", "type": "text" } ], "index": 13 }, { "bbox": [ 69, 279, 222, 292 ], "spans": [ { "bbox": [ 69, 279, 222, 292 ], "score": 1.0, "content": "MOSS-sft-002 as the training dataset.", "type": "text" } ], "index": 14 } ], "index": 12.5 } ], "index": 8.75 }, { "type": "text", "bbox": [ 70, 299, 243, 310 ], "lines": [ { "bbox": [ 69, 298, 243, 312 ], "spans": [ { "bbox": [ 69, 298, 243, 312 ], "score": 1.0, "content": "the other model is trained from scratch.", "type": "text" } ], "index": 15 } ], "index": 15 }, { "type": "text", "bbox": [ 69, 312, 290, 663 ], "lines": [ { "bbox": [ 81, 312, 291, 325 ], "spans": [ { "bbox": [ 81, 312, 291, 325 ], "score": 1.0, "content": "We utilize LibriSpeech test-clean set for evalua-", "type": "text" } ], "index": 16 }, { "bbox": [ 69, 326, 291, 338 ], "spans": [ { "bbox": [ 69, 326, 291, 338 ], "score": 1.0, "content": "tion, where we randomly select 100 utterances, and", "type": "text" } ], "index": 17 }, { "bbox": [ 69, 339, 291, 353 ], "spans": [ { "bbox": [ 69, 339, 291, 353 ], "score": 1.0, "content": "use the first 3 seconds of each utterance as a prompt.", "type": "text" } ], "index": 18 }, { "bbox": [ 69, 353, 292, 366 ], "spans": [ { "bbox": [ 69, 353, 292, 366 ], "score": 1.0, "content": "The 3-second speech prompt is converted into dis-", "type": "text" } ], "index": 19 }, { "bbox": [ 68, 366, 291, 379 ], "spans": [ { "bbox": [ 68, 366, 291, 379 ], "score": 1.0, "content": "crete units by mHuBERT. The model takes the", "type": "text" } ], "index": 20 }, { "bbox": [ 69, 381, 291, 393 ], "spans": [ { "bbox": [ 69, 381, 291, 393 ], "score": 1.0, "content": "prompt as input and generates a continuation of dis-", "type": "text" } ], "index": 21 }, { "bbox": [ 68, 394, 291, 406 ], "spans": [ { "bbox": [ 68, 394, 291, 406 ], "score": 1.0, "content": "crete units, which are subsequently converted back", "type": "text" } ], "index": 22 }, { "bbox": [ 69, 407, 291, 421 ], "spans": [ { "bbox": [ 69, 407, 291, 421 ], "score": 1.0, "content": "into speech by a discrete unit vocoder. To assess", "type": "text" } ], "index": 23 }, { "bbox": [ 69, 421, 290, 434 ], "spans": [ { "bbox": [ 69, 421, 290, 434 ], "score": 1.0, "content": "the semantic quality of the speech continuation, we", "type": "text" } ], "index": 24 }, { "bbox": [ 69, 434, 291, 447 ], "spans": [ { "bbox": [ 69, 434, 291, 447 ], "score": 1.0, "content": "employ ASR-PPL metric. This involves transcrib-", "type": "text" } ], "index": 25 }, { "bbox": [ 68, 448, 291, 461 ], "spans": [ { "bbox": [ 68, 448, 291, 461 ], "score": 1.0, "content": "ing the speech continuation into text using the ASR", "type": "text" } ], "index": 26 }, { "bbox": [ 68, 461, 291, 475 ], "spans": [ { "bbox": [ 68, 461, 291, 475 ], "score": 1.0, "content": "system in Section 5.2 and calculating the perplexity", "type": "text" } ], "index": 27 }, { "bbox": [ 69, 475, 290, 487 ], "spans": [ { "bbox": [ 69, 475, 290, 487 ], "score": 1.0, "content": "of the transcripts using GPT-3.5 text-devinci-003", "type": "text" } ], "index": 28 }, { "bbox": [ 69, 489, 291, 501 ], "spans": [ { "bbox": [ 69, 489, 291, 501 ], "score": 1.0, "content": "model. As shown in Figure 3, we observe a continu-", "type": "text" } ], "index": 29 }, { "bbox": [ 69, 503, 291, 514 ], "spans": [ { "bbox": [ 69, 503, 291, 514 ], "score": 1.0, "content": "ous decrease in ASR-PPL as the training tokens in-", "type": "text" } ], "index": 30 }, { "bbox": [ 69, 516, 290, 528 ], "spans": [ { "bbox": [ 69, 516, 290, 528 ], "score": 1.0, "content": "crease. The ASR-PPL of the model initialized from", "type": "text" } ], "index": 31 }, { "bbox": [ 69, 529, 290, 541 ], "spans": [ { "bbox": [ 69, 529, 290, 541 ], "score": 1.0, "content": "LLaMA consistently remains lower than that of the", "type": "text" } ], "index": 32 }, { "bbox": [ 69, 543, 291, 555 ], "spans": [ { "bbox": [ 69, 543, 291, 555 ], "score": 1.0, "content": "model pre-trained from scratch. This indicates that", "type": "text" } ], "index": 33 }, { "bbox": [ 69, 555, 291, 569 ], "spans": [ { "bbox": [ 69, 555, 291, 569 ], "score": 1.0, "content": "text pre-trained model provides a warm initial-", "type": "text" } ], "index": 34 }, { "bbox": [ 69, 570, 290, 583 ], "spans": [ { "bbox": [ 69, 570, 290, 583 ], "score": 1.0, "content": "ization and speech modality can benefit from", "type": "text" } ], "index": 35 }, { "bbox": [ 69, 583, 290, 596 ], "spans": [ { "bbox": [ 69, 583, 290, 596 ], "score": 1.0, "content": "text knowledge. We believe the reason for this", "type": "text" } ], "index": 36 }, { "bbox": [ 68, 597, 290, 610 ], "spans": [ { "bbox": [ 68, 597, 290, 610 ], "score": 1.0, "content": "is that even though the modeling granularity of", "type": "text" } ], "index": 37 }, { "bbox": [ 68, 610, 290, 624 ], "spans": [ { "bbox": [ 68, 610, 290, 624 ], "score": 1.0, "content": "speech and text is different, they model the same", "type": "text" } ], "index": 38 }, { "bbox": [ 68, 624, 290, 637 ], "spans": [ { "bbox": [ 68, 624, 290, 637 ], "score": 1.0, "content": "content information. This leads to a certain degree", "type": "text" } ], "index": 39 }, { "bbox": [ 69, 638, 290, 650 ], "spans": [ { "bbox": [ 69, 638, 290, 650 ], "score": 1.0, "content": "of similarity in the sequence structure, which aids", "type": "text" } ], "index": 40 }, { "bbox": [ 69, 651, 169, 664 ], "spans": [ { "bbox": [ 69, 651, 169, 664 ], "score": 1.0, "content": "in knowledge transfer.", "type": "text" } ], "index": 41 } ], "index": 28.5 }, { "type": "title", "bbox": [ 70, 675, 289, 700 ], "lines": [ { "bbox": [ 68, 674, 289, 689 ], "spans": [ { "bbox": [ 68, 674, 289, 689 ], "score": 1.0, "content": "6.3 Does SpeechGPT Sacrifice Text Capability", "type": "text" } ], "index": 42 }, { "bbox": [ 92, 689, 166, 701 ], "spans": [ { "bbox": [ 92, 689, 166, 701 ], "score": 1.0, "content": "as a Trade-off?", "type": "text" } ], "index": 43 } ], "index": 42.5 }, { "type": "text", "bbox": [ 70, 707, 290, 773 ], "lines": [ { "bbox": [ 69, 706, 289, 720 ], "spans": [ { "bbox": [ 69, 706, 289, 720 ], "score": 1.0, "content": "Initialized form LLaMA, SpeechGPT is capable", "type": "text" } ], "index": 44 }, { "bbox": [ 68, 720, 291, 733 ], "spans": [ { "bbox": [ 68, 720, 291, 733 ], "score": 1.0, "content": "of preceiving and generating speech after train-", "type": "text" } ], "index": 45 }, { "bbox": [ 68, 734, 291, 747 ], "spans": [ { "bbox": [ 68, 734, 291, 747 ], "score": 1.0, "content": "ing on large scale speech data. However, does", "type": "text" } ], "index": 46 }, { "bbox": [ 69, 748, 290, 760 ], "spans": [ { "bbox": [ 69, 748, 290, 760 ], "score": 1.0, "content": "SpeechGPT sacrifice text capability as a trade-off?", "type": "text" } ], "index": 47 }, { "bbox": [ 68, 760, 291, 775 ], "spans": [ { "bbox": [ 68, 760, 291, 775 ], "score": 1.0, "content": "To draw conclusions, we compared the text-to-text", "type": "text" } ], "index": 48 } ], "index": 46 }, { "type": "text", "bbox": [ 305, 72, 525, 165 ], "lines": [ { "bbox": [ 304, 72, 525, 84 ], "spans": [ { "bbox": [ 304, 72, 525, 84 ], "score": 1.0, "content": "instruction-following ability of SpeechGPT with", "type": "text" } ], "index": 49 }, { "bbox": [ 304, 85, 527, 97 ], "spans": [ { "bbox": [ 304, 85, 527, 97 ], "score": 1.0, "content": "LLaMA-MOSS-002. LLaMA-MOSS-002 is ob-", "type": "text" } ], "index": 50 }, { "bbox": [ 303, 98, 527, 114 ], "spans": [ { "bbox": [ 303, 98, 527, 114 ], "score": 1.0, "content": "tained by performing supervised fine-tuning on", "type": "text" } ], "index": 51 }, { "bbox": [ 303, 111, 527, 127 ], "spans": [ { "bbox": [ 303, 111, 527, 127 ], "score": 1.0, "content": "LLaMA-13B using MOSS-sft-002 as the training", "type": "text" } ], "index": 52 }, { "bbox": [ 304, 126, 526, 139 ], "spans": [ { "bbox": [ 304, 126, 526, 139 ], "score": 1.0, "content": "dataset. This ensures that both models have been", "type": "text" } ], "index": 53 }, { "bbox": [ 304, 140, 527, 152 ], "spans": [ { "bbox": [ 304, 140, 527, 152 ], "score": 1.0, "content": "exposed to the same amount of text data. We evalu-", "type": "text" } ], "index": 54 }, { "bbox": [ 304, 153, 527, 165 ], "spans": [ { "bbox": [ 304, 153, 527, 165 ], "score": 1.0, "content": "ated both models using the test set from Section 5.3.", "type": "text" } ], "index": 55 } ], "index": 52 }, { "type": "text", "bbox": [ 305, 167, 525, 342 ], "lines": [ { "bbox": [ 316, 167, 525, 180 ], "spans": [ { "bbox": [ 316, 167, 525, 180 ], "score": 1.0, "content": "As depicted in Figure 4, with an increase in", "type": "text" } ], "index": 56 }, { "bbox": [ 304, 180, 526, 193 ], "spans": [ { "bbox": [ 304, 180, 526, 193 ], "score": 1.0, "content": "training samples, both LLaMA-MOSS-002 and", "type": "text" } ], "index": 57 }, { "bbox": [ 304, 194, 527, 207 ], "spans": [ { "bbox": [ 304, 194, 527, 207 ], "score": 1.0, "content": "SpeechGPT’s ChatGPT Score gradually improve.", "type": "text" } ], "index": 58 }, { "bbox": [ 303, 207, 527, 222 ], "spans": [ { "bbox": [ 303, 207, 527, 222 ], "score": 1.0, "content": "Although SpeechGPT consistently remains lower", "type": "text" } ], "index": 59 }, { "bbox": [ 304, 221, 525, 235 ], "spans": [ { "bbox": [ 304, 221, 525, 235 ], "score": 1.0, "content": "than LLaMA-MOSS-002. the performance gap", "type": "text" } ], "index": 60 }, { "bbox": [ 304, 235, 526, 248 ], "spans": [ { "bbox": [ 304, 235, 526, 248 ], "score": 1.0, "content": "between them gradually decreases. When the train-", "type": "text" } ], "index": 61 }, { "bbox": [ 303, 248, 526, 261 ], "spans": [ { "bbox": [ 303, 248, 526, 261 ], "score": 1.0, "content": "ing samples reach 40,000, the performance of the", "type": "text" } ], "index": 62 }, { "bbox": [ 303, 261, 525, 276 ], "spans": [ { "bbox": [ 303, 261, 525, 276 ], "score": 1.0, "content": "two models becomes very similar. This suggests", "type": "text" } ], "index": 63 }, { "bbox": [ 304, 276, 525, 288 ], "spans": [ { "bbox": [ 304, 276, 525, 288 ], "score": 1.0, "content": "that SpeechGPT still retains text capability. We", "type": "text" } ], "index": 64 }, { "bbox": [ 304, 289, 526, 302 ], "spans": [ { "bbox": [ 304, 289, 526, 302 ], "score": 1.0, "content": "attribute this to the large parameter size of the", "type": "text" } ], "index": 65 }, { "bbox": [ 304, 302, 527, 315 ], "spans": [ { "bbox": [ 304, 302, 527, 315 ], "score": 1.0, "content": "13B model, enabling it to learn new speech modal-", "type": "text" } ], "index": 66 }, { "bbox": [ 304, 316, 527, 330 ], "spans": [ { "bbox": [ 304, 316, 527, 330 ], "score": 1.0, "content": "ity while preserving text capability without catas-", "type": "text" } ], "index": 67 }, { "bbox": [ 303, 328, 387, 345 ], "spans": [ { "bbox": [ 303, 328, 387, 345 ], "score": 1.0, "content": "trophic forgetting.", "type": "text" } ], "index": 68 } ], "index": 62 }, { "type": "title", "bbox": [ 305, 353, 381, 367 ], "lines": [ { "bbox": [ 302, 350, 383, 370 ], "spans": [ { "bbox": [ 302, 350, 383, 370 ], "score": 1.0, "content": "7 Conclusion", "type": "text" } ], "index": 69 } ], "index": 69 }, { "type": "text", "bbox": [ 305, 376, 525, 578 ], "lines": [ { "bbox": [ 303, 374, 526, 390 ], "spans": [ { "bbox": [ 303, 374, 526, 390 ], "score": 1.0, "content": "This work presents SpeechGPT, a large language", "type": "text" } ], "index": 70 }, { "bbox": [ 304, 390, 525, 402 ], "spans": [ { "bbox": [ 304, 390, 525, 402 ], "score": 1.0, "content": "model with intrinsic cross-modal conversational", "type": "text" } ], "index": 71 }, { "bbox": [ 304, 403, 525, 416 ], "spans": [ { "bbox": [ 304, 403, 525, 416 ], "score": 1.0, "content": "abilities, capable of perceiving and generating", "type": "text" } ], "index": 72 }, { "bbox": [ 304, 416, 525, 429 ], "spans": [ { "bbox": [ 304, 416, 525, 429 ], "score": 1.0, "content": "multi-modal content. To alleviate the scarcity", "type": "text" } ], "index": 73 }, { "bbox": [ 304, 430, 526, 443 ], "spans": [ { "bbox": [ 304, 430, 526, 443 ], "score": 1.0, "content": "of instruction datasets in current speech domain,", "type": "text" } ], "index": 74 }, { "bbox": [ 303, 444, 526, 457 ], "spans": [ { "bbox": [ 303, 444, 526, 457 ], "score": 1.0, "content": "we propose SpeechInstruct, the first speech-text", "type": "text" } ], "index": 75 }, { "bbox": [ 304, 457, 527, 470 ], "spans": [ { "bbox": [ 304, 457, 527, 470 ], "score": 1.0, "content": "cross-modal instruction-following dataset. To ob-", "type": "text" } ], "index": 76 }, { "bbox": [ 304, 471, 526, 484 ], "spans": [ { "bbox": [ 304, 471, 526, 484 ], "score": 1.0, "content": "tain improved cross-modal performance, we adopt", "type": "text" } ], "index": 77 }, { "bbox": [ 303, 484, 527, 497 ], "spans": [ { "bbox": [ 303, 484, 527, 497 ], "score": 1.0, "content": "a three-stage training paradigm to obtain the fi-", "type": "text" } ], "index": 78 }, { "bbox": [ 304, 498, 526, 511 ], "spans": [ { "bbox": [ 304, 498, 526, 511 ], "score": 1.0, "content": "nal SpeechGPT. Experimental results indicate that", "type": "text" } ], "index": 79 }, { "bbox": [ 305, 512, 527, 524 ], "spans": [ { "bbox": [ 305, 512, 527, 524 ], "score": 1.0, "content": "SpeechGPT achieves promising results in vari-", "type": "text" } ], "index": 80 }, { "bbox": [ 303, 524, 526, 539 ], "spans": [ { "bbox": [ 303, 524, 526, 539 ], "score": 1.0, "content": "ous unimodal or cross-modal instruction-following", "type": "text" } ], "index": 81 }, { "bbox": [ 304, 539, 525, 551 ], "spans": [ { "bbox": [ 304, 539, 525, 551 ], "score": 1.0, "content": "tasks and demonstrate that combining discrete", "type": "text" } ], "index": 82 }, { "bbox": [ 304, 552, 527, 565 ], "spans": [ { "bbox": [ 304, 552, 527, 565 ], "score": 1.0, "content": "speech tokens into the language model is a promis-", "type": "text" } ], "index": 83 }, { "bbox": [ 304, 565, 366, 579 ], "spans": [ { "bbox": [ 304, 565, 366, 579 ], "score": 1.0, "content": "ing direction.", "type": "text" } ], "index": 84 } ], "index": 77 }, { "type": "title", "bbox": [ 305, 590, 360, 603 ], "lines": [ { "bbox": [ 304, 588, 362, 604 ], "spans": [ { "bbox": [ 304, 588, 362, 604 ], "score": 1.0, "content": "Limitation", "type": "text" } ], "index": 85 } ], "index": 85 }, { "type": "text", "bbox": [ 305, 612, 525, 773 ], "lines": [ { "bbox": [ 304, 612, 527, 626 ], "spans": [ { "bbox": [ 304, 612, 527, 626 ], "score": 1.0, "content": "Despite SpeechGPT exhibiting impressive cross-", "type": "text" } ], "index": 86 }, { "bbox": [ 304, 624, 525, 640 ], "spans": [ { "bbox": [ 304, 624, 525, 640 ], "score": 1.0, "content": "modal instruction following and spoken dialogue", "type": "text" } ], "index": 87 }, { "bbox": [ 304, 640, 525, 652 ], "spans": [ { "bbox": [ 304, 640, 525, 652 ], "score": 1.0, "content": "abilities, it still presents certain limitations: 1) Due", "type": "text" } ], "index": 88 }, { "bbox": [ 304, 653, 527, 666 ], "spans": [ { "bbox": [ 304, 653, 527, 666 ], "score": 1.0, "content": "to the audio discretization technique constraints,", "type": "text" } ], "index": 89 }, { "bbox": [ 304, 667, 526, 680 ], "spans": [ { "bbox": [ 304, 667, 526, 680 ], "score": 1.0, "content": "SpeechGPT does not explicitly model the paralin-", "type": "text" } ], "index": 90 }, { "bbox": [ 303, 679, 527, 694 ], "spans": [ { "bbox": [ 303, 679, 527, 694 ], "score": 1.0, "content": "guistic information included in the speech signal.", "type": "text" } ], "index": 91 }, { "bbox": [ 304, 693, 526, 707 ], "spans": [ { "bbox": [ 304, 693, 526, 707 ], "score": 1.0, "content": "2) Since SpeechGPT generates speech responses", "type": "text" } ], "index": 92 }, { "bbox": [ 303, 706, 527, 721 ], "spans": [ { "bbox": [ 303, 706, 527, 721 ], "score": 1.0, "content": "via the Chain-of-Modality, it needs to initially gen-", "type": "text" } ], "index": 93 }, { "bbox": [ 304, 721, 525, 733 ], "spans": [ { "bbox": [ 304, 721, 525, 733 ], "score": 1.0, "content": "erate speech units after text tokens, which increases", "type": "text" } ], "index": 94 }, { "bbox": [ 304, 734, 527, 749 ], "spans": [ { "bbox": [ 304, 734, 527, 749 ], "score": 1.0, "content": "decoding time. However, by improving the capa-", "type": "text" } ], "index": 95 }, { "bbox": [ 304, 746, 525, 761 ], "spans": [ { "bbox": [ 304, 746, 525, 761 ], "score": 1.0, "content": "bilities of the foundation model, SpeechGPT may", "type": "text" } ], "index": 96 }, { "bbox": [ 304, 761, 526, 776 ], "spans": [ { "bbox": [ 304, 761, 526, 776 ], "score": 1.0, "content": "generate speech units directly without noticeably", "type": "text" } ], "index": 97 } ], "index": 91.5 } ], "page_idx": 7, "page_size": [ 595, 841 ], "discarded_blocks": [], "para_blocks": [ { "type": "image", "bbox": [ 73, 84, 268, 234 ], "blocks": [ { "type": "image_body", "bbox": [ 73, 84, 268, 234 ], "group_id": 0, "lines": [ { "bbox": [ 73, 84, 268, 234 ], "spans": [ { "bbox": [ 73, 84, 268, 234 ], "score": 0.972, "type": "image", "image_path": "e46aa44b83753245531eae6af65862d3d2b33291238053589833d6d978b8c55e.jpg" } ] } ], "index": 5, "virtual_lines": [ { "bbox": [ 73, 84, 268, 97.63636363636364 ], "spans": [], "index": 0 }, { "bbox": [ 73, 97.63636363636364, 268, 111.27272727272728 ], "spans": [], "index": 1 }, { "bbox": [ 73, 111.27272727272728, 268, 124.90909090909092 ], "spans": [], "index": 2 }, { "bbox": [ 73, 124.90909090909092, 268, 138.54545454545456 ], "spans": [], "index": 3 }, { "bbox": [ 73, 138.54545454545456, 268, 152.1818181818182 ], "spans": [], "index": 4 }, { "bbox": [ 73, 152.1818181818182, 268, 165.8181818181818 ], "spans": [], "index": 5 }, { "bbox": [ 73, 165.8181818181818, 268, 179.45454545454544 ], "spans": [], "index": 6 }, { "bbox": [ 73, 179.45454545454544, 268, 193.09090909090907 ], "spans": [], "index": 7 }, { "bbox": [ 73, 193.09090909090907, 268, 206.7272727272727 ], "spans": [], "index": 8 }, { "bbox": [ 73, 206.7272727272727, 268, 220.36363636363632 ], "spans": [], "index": 9 }, { "bbox": [ 73, 220.36363636363632, 268, 233.99999999999994 ], "spans": [], "index": 10 } ] }, { "type": "image_caption", "bbox": [ 69, 244, 290, 291 ], "group_id": 0, "lines": [ { "bbox": [ 69, 244, 291, 255 ], "spans": [ { "bbox": [ 69, 244, 291, 255 ], "score": 1.0, "content": "Figure 4: ChatGPT Score on text-to-text instruction-", "type": "text" } ], "index": 11 }, { "bbox": [ 68, 254, 291, 269 ], "spans": [ { "bbox": [ 68, 254, 291, 269 ], "score": 1.0, "content": "following task. LLaMA-MOSS-002 is obtained by per-", "type": "text" } ], "index": 12 }, { "bbox": [ 68, 267, 290, 281 ], "spans": [ { "bbox": [ 68, 267, 290, 281 ], "score": 1.0, "content": "forming supervised fine-tuning on LLaMA-13B using", "type": "text" } ], "index": 13 }, { "bbox": [ 69, 279, 222, 292 ], "spans": [ { "bbox": [ 69, 279, 222, 292 ], "score": 1.0, "content": "MOSS-sft-002 as the training dataset.", "type": "text" } ], "index": 14 } ], "index": 12.5 } ], "index": 8.75 }, { "type": "text", "bbox": [ 70, 299, 243, 310 ], "lines": [], "index": 15, "bbox_fs": [ 69, 298, 243, 312 ], "lines_deleted": true }, { "type": "text", "bbox": [ 69, 312, 290, 663 ], "lines": [ { "bbox": [ 81, 312, 291, 325 ], "spans": [ { "bbox": [ 81, 312, 291, 325 ], "score": 1.0, "content": "We utilize LibriSpeech test-clean set for evalua-", "type": "text" } ], "index": 16 }, { "bbox": [ 69, 326, 291, 338 ], "spans": [ { "bbox": [ 69, 326, 291, 338 ], "score": 1.0, "content": "tion, where we randomly select 100 utterances, and", "type": "text" } ], "index": 17 }, { "bbox": [ 69, 339, 291, 353 ], "spans": [ { "bbox": [ 69, 339, 291, 353 ], "score": 1.0, "content": "use the first 3 seconds of each utterance as a prompt.", "type": "text" } ], "index": 18 }, { "bbox": [ 69, 353, 292, 366 ], "spans": [ { "bbox": [ 69, 353, 292, 366 ], "score": 1.0, "content": "The 3-second speech prompt is converted into dis-", "type": "text" } ], "index": 19 }, { "bbox": [ 68, 366, 291, 379 ], "spans": [ { "bbox": [ 68, 366, 291, 379 ], "score": 1.0, "content": "crete units by mHuBERT. The model takes the", "type": "text" } ], "index": 20 }, { "bbox": [ 69, 381, 291, 393 ], "spans": [ { "bbox": [ 69, 381, 291, 393 ], "score": 1.0, "content": "prompt as input and generates a continuation of dis-", "type": "text" } ], "index": 21 }, { "bbox": [ 68, 394, 291, 406 ], "spans": [ { "bbox": [ 68, 394, 291, 406 ], "score": 1.0, "content": "crete units, which are subsequently converted back", "type": "text" } ], "index": 22 }, { "bbox": [ 69, 407, 291, 421 ], "spans": [ { "bbox": [ 69, 407, 291, 421 ], "score": 1.0, "content": "into speech by a discrete unit vocoder. To assess", "type": "text" } ], "index": 23 }, { "bbox": [ 69, 421, 290, 434 ], "spans": [ { "bbox": [ 69, 421, 290, 434 ], "score": 1.0, "content": "the semantic quality of the speech continuation, we", "type": "text" } ], "index": 24 }, { "bbox": [ 69, 434, 291, 447 ], "spans": [ { "bbox": [ 69, 434, 291, 447 ], "score": 1.0, "content": "employ ASR-PPL metric. This involves transcrib-", "type": "text" } ], "index": 25 }, { "bbox": [ 68, 448, 291, 461 ], "spans": [ { "bbox": [ 68, 448, 291, 461 ], "score": 1.0, "content": "ing the speech continuation into text using the ASR", "type": "text" } ], "index": 26 }, { "bbox": [ 68, 461, 291, 475 ], "spans": [ { "bbox": [ 68, 461, 291, 475 ], "score": 1.0, "content": "system in Section 5.2 and calculating the perplexity", "type": "text" } ], "index": 27 }, { "bbox": [ 69, 475, 290, 487 ], "spans": [ { "bbox": [ 69, 475, 290, 487 ], "score": 1.0, "content": "of the transcripts using GPT-3.5 text-devinci-003", "type": "text" } ], "index": 28 }, { "bbox": [ 69, 489, 291, 501 ], "spans": [ { "bbox": [ 69, 489, 291, 501 ], "score": 1.0, "content": "model. As shown in Figure 3, we observe a continu-", "type": "text" } ], "index": 29 }, { "bbox": [ 69, 503, 291, 514 ], "spans": [ { "bbox": [ 69, 503, 291, 514 ], "score": 1.0, "content": "ous decrease in ASR-PPL as the training tokens in-", "type": "text" } ], "index": 30 }, { "bbox": [ 69, 516, 290, 528 ], "spans": [ { "bbox": [ 69, 516, 290, 528 ], "score": 1.0, "content": "crease. The ASR-PPL of the model initialized from", "type": "text" } ], "index": 31 }, { "bbox": [ 69, 529, 290, 541 ], "spans": [ { "bbox": [ 69, 529, 290, 541 ], "score": 1.0, "content": "LLaMA consistently remains lower than that of the", "type": "text" } ], "index": 32 }, { "bbox": [ 69, 543, 291, 555 ], "spans": [ { "bbox": [ 69, 543, 291, 555 ], "score": 1.0, "content": "model pre-trained from scratch. This indicates that", "type": "text" } ], "index": 33 }, { "bbox": [ 69, 555, 291, 569 ], "spans": [ { "bbox": [ 69, 555, 291, 569 ], "score": 1.0, "content": "text pre-trained model provides a warm initial-", "type": "text" } ], "index": 34 }, { "bbox": [ 69, 570, 290, 583 ], "spans": [ { "bbox": [ 69, 570, 290, 583 ], "score": 1.0, "content": "ization and speech modality can benefit from", "type": "text" } ], "index": 35 }, { "bbox": [ 69, 583, 290, 596 ], "spans": [ { "bbox": [ 69, 583, 290, 596 ], "score": 1.0, "content": "text knowledge. We believe the reason for this", "type": "text" } ], "index": 36 }, { "bbox": [ 68, 597, 290, 610 ], "spans": [ { "bbox": [ 68, 597, 290, 610 ], "score": 1.0, "content": "is that even though the modeling granularity of", "type": "text" } ], "index": 37 }, { "bbox": [ 68, 610, 290, 624 ], "spans": [ { "bbox": [ 68, 610, 290, 624 ], "score": 1.0, "content": "speech and text is different, they model the same", "type": "text" } ], "index": 38 }, { "bbox": [ 68, 624, 290, 637 ], "spans": [ { "bbox": [ 68, 624, 290, 637 ], "score": 1.0, "content": "content information. This leads to a certain degree", "type": "text" } ], "index": 39 }, { "bbox": [ 69, 638, 290, 650 ], "spans": [ { "bbox": [ 69, 638, 290, 650 ], "score": 1.0, "content": "of similarity in the sequence structure, which aids", "type": "text" } ], "index": 40 }, { "bbox": [ 69, 651, 169, 664 ], "spans": [ { "bbox": [ 69, 651, 169, 664 ], "score": 1.0, "content": "in knowledge transfer.", "type": "text" } ], "index": 41 } ], "index": 28.5, "bbox_fs": [ 68, 312, 292, 664 ] }, { "type": "title", "bbox": [ 70, 675, 289, 700 ], "lines": [ { "bbox": [ 68, 674, 289, 689 ], "spans": [ { "bbox": [ 68, 674, 289, 689 ], "score": 1.0, "content": "6.3 Does SpeechGPT Sacrifice Text Capability", "type": "text" } ], "index": 42 }, { "bbox": [ 92, 689, 166, 701 ], "spans": [ { "bbox": [ 92, 689, 166, 701 ], "score": 1.0, "content": "as a Trade-off?", "type": "text" } ], "index": 43 } ], "index": 42.5 }, { "type": "text", "bbox": [ 70, 707, 290, 773 ], "lines": [ { "bbox": [ 69, 706, 289, 720 ], "spans": [ { "bbox": [ 69, 706, 289, 720 ], "score": 1.0, "content": "Initialized form LLaMA, SpeechGPT is capable", "type": "text" } ], "index": 44 }, { "bbox": [ 68, 720, 291, 733 ], "spans": [ { "bbox": [ 68, 720, 291, 733 ], "score": 1.0, "content": "of preceiving and generating speech after train-", "type": "text" } ], "index": 45 }, { "bbox": [ 68, 734, 291, 747 ], "spans": [ { "bbox": [ 68, 734, 291, 747 ], "score": 1.0, "content": "ing on large scale speech data. However, does", "type": "text" } ], "index": 46 }, { "bbox": [ 69, 748, 290, 760 ], "spans": [ { "bbox": [ 69, 748, 290, 760 ], "score": 1.0, "content": "SpeechGPT sacrifice text capability as a trade-off?", "type": "text" } ], "index": 47 }, { "bbox": [ 68, 760, 291, 775 ], "spans": [ { "bbox": [ 68, 760, 291, 775 ], "score": 1.0, "content": "To draw conclusions, we compared the text-to-text", "type": "text" } ], "index": 48 }, { "bbox": [ 304, 72, 525, 84 ], "spans": [ { "bbox": [ 304, 72, 525, 84 ], "score": 1.0, "content": "instruction-following ability of SpeechGPT with", "type": "text" } ], "index": 49 }, { "bbox": [ 304, 85, 527, 97 ], "spans": [ { "bbox": [ 304, 85, 527, 97 ], "score": 1.0, "content": "LLaMA-MOSS-002. LLaMA-MOSS-002 is ob-", "type": "text" } ], "index": 50 }, { "bbox": [ 303, 98, 527, 114 ], "spans": [ { "bbox": [ 303, 98, 527, 114 ], "score": 1.0, "content": "tained by performing supervised fine-tuning on", "type": "text" } ], "index": 51 }, { "bbox": [ 303, 111, 527, 127 ], "spans": [ { "bbox": [ 303, 111, 527, 127 ], "score": 1.0, "content": "LLaMA-13B using MOSS-sft-002 as the training", "type": "text" } ], "index": 52 }, { "bbox": [ 304, 126, 526, 139 ], "spans": [ { "bbox": [ 304, 126, 526, 139 ], "score": 1.0, "content": "dataset. This ensures that both models have been", "type": "text" } ], "index": 53 }, { "bbox": [ 304, 140, 527, 152 ], "spans": [ { "bbox": [ 304, 140, 527, 152 ], "score": 1.0, "content": "exposed to the same amount of text data. We evalu-", "type": "text" } ], "index": 54 }, { "bbox": [ 304, 153, 527, 165 ], "spans": [ { "bbox": [ 304, 153, 527, 165 ], "score": 1.0, "content": "ated both models using the test set from Section 5.3.", "type": "text" } ], "index": 55 } ], "index": 46, "bbox_fs": [ 68, 706, 291, 775 ] }, { "type": "text", "bbox": [ 305, 72, 525, 165 ], "lines": [], "index": 52, "bbox_fs": [ 303, 72, 527, 165 ], "lines_deleted": true }, { "type": "text", "bbox": [ 305, 167, 525, 342 ], "lines": [ { "bbox": [ 316, 167, 525, 180 ], "spans": [ { "bbox": [ 316, 167, 525, 180 ], "score": 1.0, "content": "As depicted in Figure 4, with an increase in", "type": "text" } ], "index": 56 }, { "bbox": [ 304, 180, 526, 193 ], "spans": [ { "bbox": [ 304, 180, 526, 193 ], "score": 1.0, "content": "training samples, both LLaMA-MOSS-002 and", "type": "text" } ], "index": 57 }, { "bbox": [ 304, 194, 527, 207 ], "spans": [ { "bbox": [ 304, 194, 527, 207 ], "score": 1.0, "content": "SpeechGPT’s ChatGPT Score gradually improve.", "type": "text" } ], "index": 58 }, { "bbox": [ 303, 207, 527, 222 ], "spans": [ { "bbox": [ 303, 207, 527, 222 ], "score": 1.0, "content": "Although SpeechGPT consistently remains lower", "type": "text" } ], "index": 59 }, { "bbox": [ 304, 221, 525, 235 ], "spans": [ { "bbox": [ 304, 221, 525, 235 ], "score": 1.0, "content": "than LLaMA-MOSS-002. the performance gap", "type": "text" } ], "index": 60 }, { "bbox": [ 304, 235, 526, 248 ], "spans": [ { "bbox": [ 304, 235, 526, 248 ], "score": 1.0, "content": "between them gradually decreases. When the train-", "type": "text" } ], "index": 61 }, { "bbox": [ 303, 248, 526, 261 ], "spans": [ { "bbox": [ 303, 248, 526, 261 ], "score": 1.0, "content": "ing samples reach 40,000, the performance of the", "type": "text" } ], "index": 62 }, { "bbox": [ 303, 261, 525, 276 ], "spans": [ { "bbox": [ 303, 261, 525, 276 ], "score": 1.0, "content": "two models becomes very similar. This suggests", "type": "text" } ], "index": 63 }, { "bbox": [ 304, 276, 525, 288 ], "spans": [ { "bbox": [ 304, 276, 525, 288 ], "score": 1.0, "content": "that SpeechGPT still retains text capability. We", "type": "text" } ], "index": 64 }, { "bbox": [ 304, 289, 526, 302 ], "spans": [ { "bbox": [ 304, 289, 526, 302 ], "score": 1.0, "content": "attribute this to the large parameter size of the", "type": "text" } ], "index": 65 }, { "bbox": [ 304, 302, 527, 315 ], "spans": [ { "bbox": [ 304, 302, 527, 315 ], "score": 1.0, "content": "13B model, enabling it to learn new speech modal-", "type": "text" } ], "index": 66 }, { "bbox": [ 304, 316, 527, 330 ], "spans": [ { "bbox": [ 304, 316, 527, 330 ], "score": 1.0, "content": "ity while preserving text capability without catas-", "type": "text" } ], "index": 67 }, { "bbox": [ 303, 328, 387, 345 ], "spans": [ { "bbox": [ 303, 328, 387, 345 ], "score": 1.0, "content": "trophic forgetting.", "type": "text" } ], "index": 68 } ], "index": 62, "bbox_fs": [ 303, 167, 527, 345 ] }, { "type": "title", "bbox": [ 305, 353, 381, 367 ], "lines": [ { "bbox": [ 302, 350, 383, 370 ], "spans": [ { "bbox": [ 302, 350, 383, 370 ], "score": 1.0, "content": "7 Conclusion", "type": "text" } ], "index": 69 } ], "index": 69 }, { "type": "text", "bbox": [ 305, 376, 525, 578 ], "lines": [ { "bbox": [ 303, 374, 526, 390 ], "spans": [ { "bbox": [ 303, 374, 526, 390 ], "score": 1.0, "content": "This work presents SpeechGPT, a large language", "type": "text" } ], "index": 70 }, { "bbox": [ 304, 390, 525, 402 ], "spans": [ { "bbox": [ 304, 390, 525, 402 ], "score": 1.0, "content": "model with intrinsic cross-modal conversational", "type": "text" } ], "index": 71 }, { "bbox": [ 304, 403, 525, 416 ], "spans": [ { "bbox": [ 304, 403, 525, 416 ], "score": 1.0, "content": "abilities, capable of perceiving and generating", "type": "text" } ], "index": 72 }, { "bbox": [ 304, 416, 525, 429 ], "spans": [ { "bbox": [ 304, 416, 525, 429 ], "score": 1.0, "content": "multi-modal content. To alleviate the scarcity", "type": "text" } ], "index": 73 }, { "bbox": [ 304, 430, 526, 443 ], "spans": [ { "bbox": [ 304, 430, 526, 443 ], "score": 1.0, "content": "of instruction datasets in current speech domain,", "type": "text" } ], "index": 74 }, { "bbox": [ 303, 444, 526, 457 ], "spans": [ { "bbox": [ 303, 444, 526, 457 ], "score": 1.0, "content": "we propose SpeechInstruct, the first speech-text", "type": "text" } ], "index": 75 }, { "bbox": [ 304, 457, 527, 470 ], "spans": [ { "bbox": [ 304, 457, 527, 470 ], "score": 1.0, "content": "cross-modal instruction-following dataset. To ob-", "type": "text" } ], "index": 76 }, { "bbox": [ 304, 471, 526, 484 ], "spans": [ { "bbox": [ 304, 471, 526, 484 ], "score": 1.0, "content": "tain improved cross-modal performance, we adopt", "type": "text" } ], "index": 77 }, { "bbox": [ 303, 484, 527, 497 ], "spans": [ { "bbox": [ 303, 484, 527, 497 ], "score": 1.0, "content": "a three-stage training paradigm to obtain the fi-", "type": "text" } ], "index": 78 }, { "bbox": [ 304, 498, 526, 511 ], "spans": [ { "bbox": [ 304, 498, 526, 511 ], "score": 1.0, "content": "nal SpeechGPT. Experimental results indicate that", "type": "text" } ], "index": 79 }, { "bbox": [ 305, 512, 527, 524 ], "spans": [ { "bbox": [ 305, 512, 527, 524 ], "score": 1.0, "content": "SpeechGPT achieves promising results in vari-", "type": "text" } ], "index": 80 }, { "bbox": [ 303, 524, 526, 539 ], "spans": [ { "bbox": [ 303, 524, 526, 539 ], "score": 1.0, "content": "ous unimodal or cross-modal instruction-following", "type": "text" } ], "index": 81 }, { "bbox": [ 304, 539, 525, 551 ], "spans": [ { "bbox": [ 304, 539, 525, 551 ], "score": 1.0, "content": "tasks and demonstrate that combining discrete", "type": "text" } ], "index": 82 }, { "bbox": [ 304, 552, 527, 565 ], "spans": [ { "bbox": [ 304, 552, 527, 565 ], "score": 1.0, "content": "speech tokens into the language model is a promis-", "type": "text" } ], "index": 83 }, { "bbox": [ 304, 565, 366, 579 ], "spans": [ { "bbox": [ 304, 565, 366, 579 ], "score": 1.0, "content": "ing direction.", "type": "text" } ], "index": 84 } ], "index": 77, "bbox_fs": [ 303, 374, 527, 579 ] }, { "type": "title", "bbox": [ 305, 590, 360, 603 ], "lines": [ { "bbox": [ 304, 588, 362, 604 ], "spans": [ { "bbox": [ 304, 588, 362, 604 ], "score": 1.0, "content": "Limitation", "type": "text" } ], "index": 85 } ], "index": 85 }, { "type": "text", "bbox": [ 305, 612, 525, 773 ], "lines": [ { "bbox": [ 304, 612, 527, 626 ], "spans": [ { "bbox": [ 304, 612, 527, 626 ], "score": 1.0, "content": "Despite SpeechGPT exhibiting impressive cross-", "type": "text" } ], "index": 86 }, { "bbox": [ 304, 624, 525, 640 ], "spans": [ { "bbox": [ 304, 624, 525, 640 ], "score": 1.0, "content": "modal instruction following and spoken dialogue", "type": "text" } ], "index": 87 }, { "bbox": [ 304, 640, 525, 652 ], "spans": [ { "bbox": [ 304, 640, 525, 652 ], "score": 1.0, "content": "abilities, it still presents certain limitations: 1) Due", "type": "text" } ], "index": 88 }, { "bbox": [ 304, 653, 527, 666 ], "spans": [ { "bbox": [ 304, 653, 527, 666 ], "score": 1.0, "content": "to the audio discretization technique constraints,", "type": "text" } ], "index": 89 }, { "bbox": [ 304, 667, 526, 680 ], "spans": [ { "bbox": [ 304, 667, 526, 680 ], "score": 1.0, "content": "SpeechGPT does not explicitly model the paralin-", "type": "text" } ], "index": 90 }, { "bbox": [ 303, 679, 527, 694 ], "spans": [ { "bbox": [ 303, 679, 527, 694 ], "score": 1.0, "content": "guistic information included in the speech signal.", "type": "text" } ], "index": 91 }, { "bbox": [ 304, 693, 526, 707 ], "spans": [ { "bbox": [ 304, 693, 526, 707 ], "score": 1.0, "content": "2) Since SpeechGPT generates speech responses", "type": "text" } ], "index": 92 }, { "bbox": [ 303, 706, 527, 721 ], "spans": [ { "bbox": [ 303, 706, 527, 721 ], "score": 1.0, "content": "via the Chain-of-Modality, it needs to initially gen-", "type": "text" } ], "index": 93 }, { "bbox": [ 304, 721, 525, 733 ], "spans": [ { "bbox": [ 304, 721, 525, 733 ], "score": 1.0, "content": "erate speech units after text tokens, which increases", "type": "text" } ], "index": 94 }, { "bbox": [ 304, 734, 527, 749 ], "spans": [ { "bbox": [ 304, 734, 527, 749 ], "score": 1.0, "content": "decoding time. However, by improving the capa-", "type": "text" } ], "index": 95 }, { "bbox": [ 304, 746, 525, 761 ], "spans": [ { "bbox": [ 304, 746, 525, 761 ], "score": 1.0, "content": "bilities of the foundation model, SpeechGPT may", "type": "text" } ], "index": 96 }, { "bbox": [ 304, 761, 526, 776 ], "spans": [ { "bbox": [ 304, 761, 526, 776 ], "score": 1.0, "content": "generate speech units directly without noticeably", "type": "text" } ], "index": 97 }, { "bbox": [ 69, 72, 290, 84 ], "spans": [ { "bbox": [ 69, 72, 290, 84 ], "score": 1.0, "content": "degrading its performance. 3) SpeechGPT is not", "type": "text", "cross_page": true } ], "index": 0 }, { "bbox": [ 69, 86, 290, 99 ], "spans": [ { "bbox": [ 69, 86, 290, 99 ], "score": 1.0, "content": "evaluated in the multi-turn scenario as the length of", "type": "text", "cross_page": true } ], "index": 1 }, { "bbox": [ 68, 99, 290, 112 ], "spans": [ { "bbox": [ 68, 99, 290, 112 ], "score": 1.0, "content": "one round is already close to the maximum length", "type": "text", "cross_page": true } ], "index": 2 }, { "bbox": [ 69, 112, 292, 126 ], "spans": [ { "bbox": [ 69, 112, 292, 126 ], "score": 1.0, "content": "of the model due to the long speech unit sequences.", "type": "text", "cross_page": true } ], "index": 3 }, { "bbox": [ 69, 126, 290, 138 ], "spans": [ { "bbox": [ 69, 126, 290, 138 ], "score": 1.0, "content": "We believe this issue can be addressed by either in-", "type": "text", "cross_page": true } ], "index": 4 }, { "bbox": [ 69, 140, 290, 152 ], "spans": [ { "bbox": [ 69, 140, 290, 152 ], "score": 1.0, "content": "creasing the maximum length the model can handle", "type": "text", "cross_page": true } ], "index": 5 }, { "bbox": [ 69, 153, 289, 166 ], "spans": [ { "bbox": [ 69, 153, 289, 166 ], "score": 1.0, "content": "or employing more effective speech discretization", "type": "text", "cross_page": true } ], "index": 6 }, { "bbox": [ 68, 166, 121, 181 ], "spans": [ { "bbox": [ 68, 166, 121, 181 ], "score": 1.0, "content": "techniques.", "type": "text", "cross_page": true } ], "index": 7 } ], "index": 91.5, "bbox_fs": [ 303, 612, 527, 776 ] } ] }, { "preproc_blocks": [ { "type": "text", "bbox": [ 69, 72, 290, 179 ], "lines": [ { "bbox": [ 69, 72, 290, 84 ], "spans": [ { "bbox": [ 69, 72, 290, 84 ], "score": 1.0, "content": "degrading its performance. 3) SpeechGPT is not", "type": "text" } ], "index": 0 }, { "bbox": [ 69, 86, 290, 99 ], "spans": [ { "bbox": [ 69, 86, 290, 99 ], "score": 1.0, "content": "evaluated in the multi-turn scenario as the length of", "type": "text" } ], "index": 1 }, { "bbox": [ 68, 99, 290, 112 ], "spans": [ { "bbox": [ 68, 99, 290, 112 ], "score": 1.0, "content": "one round is already close to the maximum length", "type": "text" } ], "index": 2 }, { "bbox": [ 69, 112, 292, 126 ], "spans": [ { "bbox": [ 69, 112, 292, 126 ], "score": 1.0, "content": "of the model due to the long speech unit sequences.", "type": "text" } ], "index": 3 }, { "bbox": [ 69, 126, 290, 138 ], "spans": [ { "bbox": [ 69, 126, 290, 138 ], "score": 1.0, "content": "We believe this issue can be addressed by either in-", "type": "text" } ], "index": 4 }, { "bbox": [ 69, 140, 290, 152 ], "spans": [ { "bbox": [ 69, 140, 290, 152 ], "score": 1.0, "content": "creasing the maximum length the model can handle", "type": "text" } ], "index": 5 }, { "bbox": [ 69, 153, 289, 166 ], "spans": [ { "bbox": [ 69, 153, 289, 166 ], "score": 1.0, "content": "or employing more effective speech discretization", "type": "text" } ], "index": 6 }, { "bbox": [ 68, 166, 121, 181 ], "spans": [ { "bbox": [ 68, 166, 121, 181 ], "score": 1.0, "content": "techniques.", "type": "text" } ], "index": 7 } ], "index": 3.5 }, { "type": "title", "bbox": [ 70, 192, 169, 205 ], "lines": [ { "bbox": [ 68, 190, 171, 208 ], "spans": [ { "bbox": [ 68, 190, 171, 208 ], "score": 1.0, "content": "Acknowledgements", "type": "text" } ], "index": 8 } ], "index": 8 }, { "type": "text", "bbox": [ 69, 215, 289, 295 ], "lines": [ { "bbox": [ 68, 214, 291, 229 ], "spans": [ { "bbox": [ 68, 214, 291, 229 ], "score": 1.0, "content": "We thank Rong Ye and Fuliang Weng for the care-", "type": "text" } ], "index": 9 }, { "bbox": [ 69, 229, 291, 242 ], "spans": [ { "bbox": [ 69, 229, 291, 242 ], "score": 1.0, "content": "ful guidance and revisions to the paper and thank", "type": "text" } ], "index": 10 }, { "bbox": [ 68, 242, 291, 256 ], "spans": [ { "bbox": [ 68, 242, 291, 256 ], "score": 1.0, "content": "all the anonymous reviewers for their insightful and", "type": "text" } ], "index": 11 }, { "bbox": [ 68, 256, 290, 270 ], "spans": [ { "bbox": [ 68, 256, 290, 270 ], "score": 1.0, "content": "valuable comments. This work was supported by", "type": "text" } ], "index": 12 }, { "bbox": [ 69, 269, 291, 282 ], "spans": [ { "bbox": [ 69, 269, 291, 282 ], "score": 1.0, "content": "the National Natural Science Foundation of China", "type": "text" } ], "index": 13 }, { "bbox": [ 69, 283, 228, 296 ], "spans": [ { "bbox": [ 69, 283, 228, 296 ], "score": 1.0, "content": "(No. 62236004 and No. 62022027).", "type": "text" } ], "index": 14 } ], "index": 11.5 }, { "type": "title", "bbox": [ 70, 321, 126, 334 ], "lines": [ { "bbox": [ 68, 320, 128, 336 ], "spans": [ { "bbox": [ 68, 320, 128, 336 ], "score": 1.0, "content": "References", "type": "text" } ], "index": 15 } ], "index": 15 }, { "type": "text", "bbox": [ 70, 342, 289, 397 ], "lines": [ { "bbox": [ 69, 341, 290, 353 ], "spans": [ { "bbox": [ 69, 341, 290, 353 ], "score": 1.0, "content": "Rosana Ardila, Megan Branson, Kelly Davis, Michael", "type": "text" } ], "index": 16 }, { "bbox": [ 80, 353, 290, 364 ], "spans": [ { "bbox": [ 80, 353, 290, 364 ], "score": 1.0, "content": "Henretty, Michael Kohler, Josh Meyer, Reuben", "type": "text" } ], "index": 17 }, { "bbox": [ 80, 364, 290, 375 ], "spans": [ { "bbox": [ 80, 364, 290, 375 ], "score": 1.0, "content": "Morais, Lindsay Saunders, Francis M. Tyers, and", "type": "text" } ], "index": 18 }, { "bbox": [ 80, 374, 291, 387 ], "spans": [ { "bbox": [ 80, 374, 291, 387 ], "score": 1.0, "content": "Gregor Weber. 2020. Common voice: A massively-", "type": "text" } ], "index": 19 }, { "bbox": [ 80, 385, 192, 398 ], "spans": [ { "bbox": [ 80, 385, 192, 398 ], "score": 1.0, "content": "multilingual speech corpus.", "type": "text" } ], "index": 20 } ], "index": 18 }, { "type": "text", "bbox": [ 70, 408, 290, 463 ], "lines": [ { "bbox": [ 70, 408, 290, 419 ], "spans": [ { "bbox": [ 70, 408, 290, 419 ], "score": 1.0, "content": "Alexei Baevski, Yuhao Zhou, Abdelrahman Mohamed,", "type": "text" } ], "index": 21 }, { "bbox": [ 79, 419, 291, 431 ], "spans": [ { "bbox": [ 79, 419, 291, 431 ], "score": 1.0, "content": "and Michael Auli. 2020. wav2vec 2.0: A framework", "type": "text" } ], "index": 22 }, { "bbox": [ 79, 430, 292, 442 ], "spans": [ { "bbox": [ 79, 430, 292, 442 ], "score": 1.0, "content": "for self-supervised learning of speech representations.", "type": "text" } ], "index": 23 }, { "bbox": [ 78, 440, 292, 454 ], "spans": [ { "bbox": [ 78, 440, 292, 454 ], "score": 1.0, "content": "Advances in Neural Information Processing Systems,", "type": "text" } ], "index": 24 }, { "bbox": [ 79, 452, 153, 464 ], "spans": [ { "bbox": [ 79, 452, 153, 464 ], "score": 1.0, "content": "33:12449–12460.", "type": "text" } ], "index": 25 } ], "index": 23 }, { "type": "text", "bbox": [ 69, 475, 290, 530 ], "lines": [ { "bbox": [ 69, 474, 292, 487 ], "spans": [ { "bbox": [ 69, 474, 292, 487 ], "score": 1.0, "content": "Zalán Borsos, Raphaël Marinier, Damien Vincent, Eu-", "type": "text" } ], "index": 26 }, { "bbox": [ 79, 485, 292, 498 ], "spans": [ { "bbox": [ 79, 485, 292, 498 ], "score": 1.0, "content": "gene Kharitonov, Olivier Pietquin, Matt Sharifi,", "type": "text" } ], "index": 27 }, { "bbox": [ 80, 496, 291, 508 ], "spans": [ { "bbox": [ 80, 496, 291, 508 ], "score": 1.0, "content": "Olivier Teboul, David Grangier, Marco Tagliasac-", "type": "text" } ], "index": 28 }, { "bbox": [ 79, 506, 291, 521 ], "spans": [ { "bbox": [ 79, 506, 291, 521 ], "score": 1.0, "content": "chi, and Neil Zeghidour. 2022. Audiolm: a language", "type": "text" } ], "index": 29 }, { "bbox": [ 79, 519, 241, 531 ], "spans": [ { "bbox": [ 79, 519, 241, 531 ], "score": 1.0, "content": "modeling approach to audio generation.", "type": "text" } ], "index": 30 } ], "index": 28 }, { "type": "text", "bbox": [ 69, 541, 290, 585 ], "lines": [ { "bbox": [ 68, 539, 291, 555 ], "spans": [ { "bbox": [ 68, 539, 291, 555 ], "score": 1.0, "content": "Feilong Chen, Minglun Han, Haozhi Zhao, Qingyang", "type": "text" } ], "index": 31 }, { "bbox": [ 79, 552, 291, 564 ], "spans": [ { "bbox": [ 79, 552, 291, 564 ], "score": 1.0, "content": "Zhang, Jing Shi, Shuang Xu Xu, and Bo Xu. 2023. X-", "type": "text" } ], "index": 32 }, { "bbox": [ 79, 562, 291, 576 ], "spans": [ { "bbox": [ 79, 562, 291, 576 ], "score": 1.0, "content": "llm: Bootstrapping advanced large language models", "type": "text" } ], "index": 33 }, { "bbox": [ 79, 574, 281, 588 ], "spans": [ { "bbox": [ 79, 574, 281, 588 ], "score": 1.0, "content": "by treating multi-modalities as foreign languages.", "type": "text" } ], "index": 34 } ], "index": 32.5 }, { "type": "text", "bbox": [ 70, 597, 289, 684 ], "lines": [ { "bbox": [ 69, 596, 291, 609 ], "spans": [ { "bbox": [ 69, 596, 291, 609 ], "score": 1.0, "content": "Guoguo Chen, Shuzhou Chai, Guanbo Wang, Jiayu", "type": "text" } ], "index": 35 }, { "bbox": [ 79, 607, 290, 619 ], "spans": [ { "bbox": [ 79, 607, 290, 619 ], "score": 1.0, "content": "Du, Wei-Qiang Zhang, Chao Weng, Dan Su, Daniel", "type": "text" } ], "index": 36 }, { "bbox": [ 79, 618, 291, 631 ], "spans": [ { "bbox": [ 79, 618, 291, 631 ], "score": 1.0, "content": "Povey, Jan Trmal, Junbo Zhang, Mingjie Jin, San-", "type": "text" } ], "index": 37 }, { "bbox": [ 79, 629, 292, 642 ], "spans": [ { "bbox": [ 79, 629, 292, 642 ], "score": 1.0, "content": "jeev Khudanpur, Shinji Watanabe, Shuaijiang Zhao,", "type": "text" } ], "index": 38 }, { "bbox": [ 79, 639, 292, 654 ], "spans": [ { "bbox": [ 79, 639, 292, 654 ], "score": 1.0, "content": "Wei Zou, Xiangang Li, Xuchen Yao, Yongqing Wang,", "type": "text" } ], "index": 39 }, { "bbox": [ 79, 651, 291, 663 ], "spans": [ { "bbox": [ 79, 651, 291, 663 ], "score": 1.0, "content": "Yujun Wang, Zhao You, and Zhiyong Yan. 2021. Gi-", "type": "text" } ], "index": 40 }, { "bbox": [ 79, 662, 291, 675 ], "spans": [ { "bbox": [ 79, 662, 291, 675 ], "score": 1.0, "content": "gaspeech: An evolving, multi-domain asr corpus with", "type": "text" } ], "index": 45 }, { "bbox": [ 80, 673, 218, 685 ], "spans": [ { "bbox": [ 80, 673, 218, 685 ], "score": 1.0, "content": "10,000 hours of transcribed audio.", "type": "text" } ], "index": 46 } ], "index": 38.5 }, { "type": "text", "bbox": [ 305, 148, 525, 202 ], "lines": [ { "bbox": [ 304, 147, 526, 160 ], "spans": [ { "bbox": [ 304, 147, 526, 160 ], "score": 1.0, "content": "Xuxin Cheng, Zhihong Zhu, Ziyu Yao, Hongxiang Li,", "type": "text" } ], "index": 41 }, { "bbox": [ 315, 158, 527, 170 ], "spans": [ { "bbox": [ 315, 158, 527, 170 ], "score": 1.0, "content": "Yaowei Li, and Yuexian Zou. 2023c. GhostT5: Gen-", "type": "text" } ], "index": 42 }, { "bbox": [ 315, 169, 527, 181 ], "spans": [ { "bbox": [ 315, 169, 527, 181 ], "score": 1.0, "content": "erate More Features with Cheap Operations to Im-", "type": "text" } ], "index": 47 }, { "bbox": [ 315, 180, 527, 193 ], "spans": [ { "bbox": [ 315, 180, 527, 193 ], "score": 1.0, "content": "prove Textless Spoken Question Answering. In Proc.", "type": "text" } ], "index": 48 }, { "bbox": [ 315, 191, 483, 203 ], "spans": [ { "bbox": [ 315, 191, 483, 203 ], "score": 1.0, "content": "INTERSPEECH 2023, pages 1134–1138.", "type": "text" } ], "index": 51 } ], "index": 47 }, { "type": "text", "bbox": [ 305, 73, 525, 138 ], "lines": [ { "bbox": [ 304, 72, 527, 84 ], "spans": [ { "bbox": [ 304, 72, 527, 84 ], "score": 1.0, "content": "Xuxin Cheng, Qianqian Dong, Fengpeng Yue, Tom Ko,", "type": "text" } ], "index": 58 }, { "bbox": [ 314, 83, 527, 95 ], "spans": [ { "bbox": [ 314, 83, 527, 95 ], "score": 1.0, "content": "Mingxuan Wang, and Yuexian Zou. 2023b. M 3 st:", "type": "text" } ], "index": 44 }, { "bbox": [ 315, 94, 525, 106 ], "spans": [ { "bbox": [ 315, 94, 525, 106 ], "score": 1.0, "content": "Mix at three levels for speech translation. In ICASSP", "type": "text" } ], "index": 49 }, { "bbox": [ 315, 105, 526, 117 ], "spans": [ { "bbox": [ 315, 105, 526, 117 ], "score": 1.0, "content": "2023-2023 IEEE International Conference on Acous-", "type": "text" } ], "index": 59 }, { "bbox": [ 314, 115, 526, 130 ], "spans": [ { "bbox": [ 314, 115, 526, 130 ], "score": 1.0, "content": "tics, Speech and Signal Processing (ICASSP), pages", "type": "text" } ], "index": 60 }, { "bbox": [ 315, 127, 361, 138 ], "spans": [ { "bbox": [ 315, 127, 361, 138 ], "score": 1.0, "content": "1–5. IEEE.", "type": "text" } ], "index": 50 } ], "index": 54.0 }, { "type": "text", "bbox": [ 70, 696, 290, 772 ], "lines": [ { "bbox": [ 69, 695, 291, 708 ], "spans": [ { "bbox": [ 69, 695, 291, 708 ], "score": 1.0, "content": "Xuxin Cheng, Bowen Cao, Qichen Ye, Zhihong Zhu,", "type": "text" } ], "index": 52 }, { "bbox": [ 79, 707, 292, 719 ], "spans": [ { "bbox": [ 79, 707, 292, 719 ], "score": 1.0, "content": "Hongxiang Li, and Yuexian Zou. 2023a. Ml-lmcl:", "type": "text" } ], "index": 53 }, { "bbox": [ 79, 717, 292, 730 ], "spans": [ { "bbox": [ 79, 717, 292, 730 ], "score": 1.0, "content": "Mutual learning and large-margin contrastive learn-", "type": "text" } ], "index": 54 }, { "bbox": [ 79, 729, 290, 742 ], "spans": [ { "bbox": [ 79, 729, 290, 742 ], "score": 1.0, "content": "ing for improving asr robustness in spoken language", "type": "text" } ], "index": 43 }, { "bbox": [ 79, 740, 291, 752 ], "spans": [ { "bbox": [ 79, 740, 291, 752 ], "score": 1.0, "content": "understanding. In Findings of the Association for", "type": "text" } ], "index": 55 }, { "bbox": [ 79, 750, 291, 763 ], "spans": [ { "bbox": [ 79, 750, 291, 763 ], "score": 1.0, "content": "Computational Linguistics: ACL 2023, pages 6492–", "type": "text" } ], "index": 56 }, { "bbox": [ 78, 761, 106, 772 ], "spans": [ { "bbox": [ 78, 761, 106, 772 ], "score": 1.0, "content": "6505.", "type": "text" } ], "index": 57 } ], "index": 54 }, { "type": "text", "bbox": [ 311, 212, 525, 464 ], "lines": [ { "bbox": [ 309, 210, 527, 223 ], "spans": [ { "bbox": [ 309, 210, 527, 223 ], "score": 1.0, "content": "Aakanksha Chowdhery, Sharan Narang, Jacob Devlin,", "type": "text" } ], "index": 61 }, { "bbox": [ 315, 222, 527, 234 ], "spans": [ { "bbox": [ 315, 222, 527, 234 ], "score": 1.0, "content": "Maarten Bosma, Gaurav Mishra, Adam Roberts,", "type": "text" } ], "index": 62 }, { "bbox": [ 314, 232, 527, 245 ], "spans": [ { "bbox": [ 314, 232, 527, 245 ], "score": 1.0, "content": "Paul Barham, Hyung Won Chung, Charles Sutton,", "type": "text" } ], "index": 63 }, { "bbox": [ 315, 243, 527, 257 ], "spans": [ { "bbox": [ 315, 243, 527, 257 ], "score": 1.0, "content": "Sebastian Gehrmann, Parker Schuh, Kensen Shi,", "type": "text" } ], "index": 64 }, { "bbox": [ 315, 254, 527, 267 ], "spans": [ { "bbox": [ 315, 254, 527, 267 ], "score": 1.0, "content": "Sasha Tsvyashchenko, Joshua Maynez, Abhishek", "type": "text" } ], "index": 65 }, { "bbox": [ 315, 266, 527, 278 ], "spans": [ { "bbox": [ 315, 266, 527, 278 ], "score": 1.0, "content": "Rao, Parker Barnes, Yi Tay, Noam Shazeer, Vin-", "type": "text" } ], "index": 66 }, { "bbox": [ 315, 277, 526, 289 ], "spans": [ { "bbox": [ 315, 277, 526, 289 ], "score": 1.0, "content": "odkumar Prabhakaran, Emily Reif, Nan Du, Ben", "type": "text" } ], "index": 67 }, { "bbox": [ 314, 288, 526, 300 ], "spans": [ { "bbox": [ 314, 288, 526, 300 ], "score": 1.0, "content": "Hutchinson, Reiner Pope, James Bradbury, Jacob", "type": "text" } ], "index": 68 }, { "bbox": [ 314, 298, 527, 312 ], "spans": [ { "bbox": [ 314, 298, 527, 312 ], "score": 1.0, "content": "Austin, Michael Isard, Guy Gur-Ari, Pengcheng Yin,", "type": "text" } ], "index": 69 }, { "bbox": [ 315, 310, 527, 322 ], "spans": [ { "bbox": [ 315, 310, 527, 322 ], "score": 1.0, "content": "Toju Duke, Anselm Levskaya, Sanjay Ghemawat,", "type": "text" } ], "index": 70 }, { "bbox": [ 315, 321, 527, 333 ], "spans": [ { "bbox": [ 315, 321, 527, 333 ], "score": 1.0, "content": "Sunipa Dev, Henryk Michalewski, Xavier Garcia,", "type": "text" } ], "index": 71 }, { "bbox": [ 315, 331, 526, 344 ], "spans": [ { "bbox": [ 315, 331, 526, 344 ], "score": 1.0, "content": "Vedant Misra, Kevin Robinson, Liam Fedus, Denny", "type": "text" } ], "index": 72 }, { "bbox": [ 314, 342, 527, 356 ], "spans": [ { "bbox": [ 314, 342, 527, 356 ], "score": 1.0, "content": "Zhou, Daphne Ippolito, David Luan, Hyeontaek Lim,", "type": "text" } ], "index": 73 }, { "bbox": [ 314, 353, 527, 366 ], "spans": [ { "bbox": [ 314, 353, 527, 366 ], "score": 1.0, "content": "Barret Zoph, Alexander Spiridonov, Ryan Sepassi,", "type": "text" } ], "index": 74 }, { "bbox": [ 314, 365, 527, 376 ], "spans": [ { "bbox": [ 314, 365, 527, 376 ], "score": 1.0, "content": "David Dohan, Shivani Agrawal, Mark Omernick, An-", "type": "text" } ], "index": 75 }, { "bbox": [ 315, 375, 528, 388 ], "spans": [ { "bbox": [ 315, 375, 528, 388 ], "score": 1.0, "content": "drew M. Dai, Thanumalayan Sankaranarayana Pil-", "type": "text" } ], "index": 76 }, { "bbox": [ 315, 387, 527, 399 ], "spans": [ { "bbox": [ 315, 387, 527, 399 ], "score": 1.0, "content": "lai, Marie Pellat, Aitor Lewkowycz, Erica Moreira,", "type": "text" } ], "index": 77 }, { "bbox": [ 315, 397, 527, 410 ], "spans": [ { "bbox": [ 315, 397, 527, 410 ], "score": 1.0, "content": "Rewon Child, Oleksandr Polozov, Katherine Lee,", "type": "text" } ], "index": 78 }, { "bbox": [ 315, 409, 526, 420 ], "spans": [ { "bbox": [ 315, 409, 526, 420 ], "score": 1.0, "content": "Zongwei Zhou, Xuezhi Wang, Brennan Saeta, Mark", "type": "text" } ], "index": 79 }, { "bbox": [ 315, 419, 526, 432 ], "spans": [ { "bbox": [ 315, 419, 526, 432 ], "score": 1.0, "content": "Diaz, Orhan Firat, Michele Catasta, Jason Wei, Kathy", "type": "text" } ], "index": 80 }, { "bbox": [ 315, 430, 527, 442 ], "spans": [ { "bbox": [ 315, 430, 527, 442 ], "score": 1.0, "content": "Meier-Hellstern, Douglas Eck, Jeff Dean, Slav Petrov,", "type": "text" } ], "index": 81 }, { "bbox": [ 315, 441, 527, 454 ], "spans": [ { "bbox": [ 315, 441, 527, 454 ], "score": 1.0, "content": "and Noah Fiedel. 2022. Palm: Scaling language mod-", "type": "text" } ], "index": 82 }, { "bbox": [ 315, 452, 401, 465 ], "spans": [ { "bbox": [ 315, 452, 401, 465 ], "score": 1.0, "content": "eling with pathways.", "type": "text" } ], "index": 83 } ], "index": 72 }, { "type": "text", "bbox": [ 305, 473, 524, 549 ], "lines": [ { "bbox": [ 304, 472, 526, 484 ], "spans": [ { "bbox": [ 304, 472, 526, 484 ], "score": 1.0, "content": "Alexey Dosovitskiy, Lucas Beyer, Alexander", "type": "text" } ], "index": 84 }, { "bbox": [ 315, 483, 527, 496 ], "spans": [ { "bbox": [ 315, 483, 527, 496 ], "score": 1.0, "content": "Kolesnikov, Dirk Weissenborn, Xiaohua Zhai,", "type": "text" } ], "index": 85 }, { "bbox": [ 315, 494, 526, 506 ], "spans": [ { "bbox": [ 315, 494, 526, 506 ], "score": 1.0, "content": "Thomas Unterthiner, Mostafa Dehghani, Matthias", "type": "text" } ], "index": 86 }, { "bbox": [ 315, 505, 526, 518 ], "spans": [ { "bbox": [ 315, 505, 526, 518 ], "score": 1.0, "content": "Minderer, Georg Heigold, Sylvain Gelly, Jakob", "type": "text" } ], "index": 87 }, { "bbox": [ 315, 516, 526, 529 ], "spans": [ { "bbox": [ 315, 516, 526, 529 ], "score": 1.0, "content": "Uszkoreit, and Neil Houlsby. 2021. An image", "type": "text" } ], "index": 88 }, { "bbox": [ 314, 527, 526, 540 ], "spans": [ { "bbox": [ 314, 527, 526, 540 ], "score": 1.0, "content": "is worth 16x16 words: Transformers for image", "type": "text" } ], "index": 89 }, { "bbox": [ 315, 539, 398, 550 ], "spans": [ { "bbox": [ 315, 539, 398, 550 ], "score": 1.0, "content": "recognition at scale.", "type": "text" } ], "index": 90 } ], "index": 87 }, { "type": "text", "bbox": [ 304, 559, 525, 614 ], "lines": [ { "bbox": [ 303, 558, 527, 572 ], "spans": [ { "bbox": [ 303, 558, 527, 572 ], "score": 1.0, "content": "Danny Driess, Fei Xia, Mehdi SM Sajjadi, Corey Lynch,", "type": "text" } ], "index": 91 }, { "bbox": [ 315, 570, 526, 582 ], "spans": [ { "bbox": [ 315, 570, 526, 582 ], "score": 1.0, "content": "Aakanksha Chowdhery, Brian Ichter, Ayzaan Wahid,", "type": "text" } ], "index": 92 }, { "bbox": [ 315, 581, 527, 593 ], "spans": [ { "bbox": [ 315, 581, 527, 593 ], "score": 1.0, "content": "Jonathan Tompson, Quan Vuong, Tianhe Yu, et al.", "type": "text" } ], "index": 93 }, { "bbox": [ 314, 590, 526, 605 ], "spans": [ { "bbox": [ 314, 590, 526, 605 ], "score": 1.0, "content": "2023. Palm-e: An embodied multimodal language", "type": "text" } ], "index": 94 }, { "bbox": [ 315, 603, 484, 614 ], "spans": [ { "bbox": [ 315, 603, 484, 614 ], "score": 1.0, "content": "model. arXiv preprint arXiv:2303.03378.", "type": "text" } ], "index": 95 } ], "index": 93 }, { "type": "text", "bbox": [ 305, 623, 525, 689 ], "lines": [ { "bbox": [ 305, 623, 527, 635 ], "spans": [ { "bbox": [ 305, 623, 527, 635 ], "score": 1.0, "content": "Wei-Ning Hsu, Benjamin Bolte, Yao-Hung Hubert Tsai,", "type": "text" } ], "index": 96 }, { "bbox": [ 315, 634, 527, 645 ], "spans": [ { "bbox": [ 315, 634, 527, 645 ], "score": 1.0, "content": "Kushal Lakhotia, Ruslan Salakhutdinov, and Abdel-", "type": "text" } ], "index": 97 }, { "bbox": [ 314, 644, 526, 657 ], "spans": [ { "bbox": [ 314, 644, 526, 657 ], "score": 1.0, "content": "rahman Mohamed. 2021. Hubert: Self-supervised", "type": "text" } ], "index": 98 }, { "bbox": [ 315, 656, 526, 668 ], "spans": [ { "bbox": [ 315, 656, 526, 668 ], "score": 1.0, "content": "speech representation learning by masked prediction", "type": "text" } ], "index": 99 }, { "bbox": [ 315, 666, 527, 679 ], "spans": [ { "bbox": [ 315, 666, 527, 679 ], "score": 1.0, "content": "of hidden units. IEEE/ACM Transactions on Audio,", "type": "text" } ], "index": 100 }, { "bbox": [ 315, 678, 519, 690 ], "spans": [ { "bbox": [ 315, 678, 519, 690 ], "score": 1.0, "content": "Speech, and Language Processing, 29:3451–3460.", "type": "text" } ], "index": 101 } ], "index": 98.5 }, { "type": "text", "bbox": [ 304, 698, 525, 742 ], "lines": [ { "bbox": [ 303, 696, 526, 711 ], "spans": [ { "bbox": [ 303, 696, 526, 711 ], "score": 1.0, "content": "Edward J. Hu, Yelong Shen, Phillip Wallis, Zeyuan", "type": "text" } ], "index": 102 }, { "bbox": [ 314, 708, 525, 721 ], "spans": [ { "bbox": [ 314, 708, 525, 721 ], "score": 1.0, "content": "Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and", "type": "text" } ], "index": 103 }, { "bbox": [ 315, 718, 527, 731 ], "spans": [ { "bbox": [ 315, 718, 527, 731 ], "score": 1.0, "content": "Weizhu Chen. 2021. Lora: Low-rank adaptation of", "type": "text" } ], "index": 104 }, { "bbox": [ 315, 731, 410, 743 ], "spans": [ { "bbox": [ 315, 731, 410, 743 ], "score": 1.0, "content": "large language models.", "type": "text" } ], "index": 105 } ], "index": 103.5 }, { "type": "text", "bbox": [ 305, 751, 525, 773 ], "lines": [ { "bbox": [ 304, 750, 527, 763 ], "spans": [ { "bbox": [ 304, 750, 527, 763 ], "score": 1.0, "content": "Rongjie Huang, Mingze Li, Dongchao Yang, Jia-", "type": "text" } ], "index": 106 }, { "bbox": [ 315, 762, 527, 775 ], "spans": [ { "bbox": [ 315, 762, 527, 775 ], "score": 1.0, "content": "tong Shi, Xuankai Chang, Zhenhui Ye, Yuning Wu,", "type": "text" } ], "index": 107 } ], "index": 106.5 } ], "page_idx": 8, "page_size": [ 595, 841 ], "discarded_blocks": [], "para_blocks": [ { "type": "text", "bbox": [ 69, 72, 290, 179 ], "lines": [], "index": 3.5, "bbox_fs": [ 68, 72, 292, 181 ], "lines_deleted": true }, { "type": "title", "bbox": [ 70, 192, 169, 205 ], "lines": [ { "bbox": [ 68, 190, 171, 208 ], "spans": [ { "bbox": [ 68, 190, 171, 208 ], "score": 1.0, "content": "Acknowledgements", "type": "text" } ], "index": 8 } ], "index": 8 }, { "type": "text", "bbox": [ 69, 215, 289, 295 ], "lines": [ { "bbox": [ 68, 214, 291, 229 ], "spans": [ { "bbox": [ 68, 214, 291, 229 ], "score": 1.0, "content": "We thank Rong Ye and Fuliang Weng for the care-", "type": "text" } ], "index": 9 }, { "bbox": [ 69, 229, 291, 242 ], "spans": [ { "bbox": [ 69, 229, 291, 242 ], "score": 1.0, "content": "ful guidance and revisions to the paper and thank", "type": "text" } ], "index": 10 }, { "bbox": [ 68, 242, 291, 256 ], "spans": [ { "bbox": [ 68, 242, 291, 256 ], "score": 1.0, "content": "all the anonymous reviewers for their insightful and", "type": "text" } ], "index": 11 }, { "bbox": [ 68, 256, 290, 270 ], "spans": [ { "bbox": [ 68, 256, 290, 270 ], "score": 1.0, "content": "valuable comments. This work was supported by", "type": "text" } ], "index": 12 }, { "bbox": [ 69, 269, 291, 282 ], "spans": [ { "bbox": [ 69, 269, 291, 282 ], "score": 1.0, "content": "the National Natural Science Foundation of China", "type": "text" } ], "index": 13 }, { "bbox": [ 69, 283, 228, 296 ], "spans": [ { "bbox": [ 69, 283, 228, 296 ], "score": 1.0, "content": "(No. 62236004 and No. 62022027).", "type": "text" } ], "index": 14 } ], "index": 11.5, "bbox_fs": [ 68, 214, 291, 296 ] }, { "type": "title", "bbox": [ 70, 321, 126, 334 ], "lines": [ { "bbox": [ 68, 320, 128, 336 ], "spans": [ { "bbox": [ 68, 320, 128, 336 ], "score": 1.0, "content": "References", "type": "text" } ], "index": 15 } ], "index": 15 }, { "type": "text", "bbox": [ 70, 342, 289, 397 ], "lines": [ { "bbox": [ 69, 341, 290, 353 ], "spans": [ { "bbox": [ 69, 341, 290, 353 ], "score": 1.0, "content": "Rosana Ardila, Megan Branson, Kelly Davis, Michael", "type": "text" } ], "index": 16 }, { "bbox": [ 80, 353, 290, 364 ], "spans": [ { "bbox": [ 80, 353, 290, 364 ], "score": 1.0, "content": "Henretty, Michael Kohler, Josh Meyer, Reuben", "type": "text" } ], "index": 17 }, { "bbox": [ 80, 364, 290, 375 ], "spans": [ { "bbox": [ 80, 364, 290, 375 ], "score": 1.0, "content": "Morais, Lindsay Saunders, Francis M. Tyers, and", "type": "text" } ], "index": 18 }, { "bbox": [ 80, 374, 291, 387 ], "spans": [ { "bbox": [ 80, 374, 291, 387 ], "score": 1.0, "content": "Gregor Weber. 2020. Common voice: A massively-", "type": "text" } ], "index": 19 }, { "bbox": [ 80, 385, 192, 398 ], "spans": [ { "bbox": [ 80, 385, 192, 398 ], "score": 1.0, "content": "multilingual speech corpus.", "type": "text" } ], "index": 20 } ], "index": 18, "bbox_fs": [ 69, 341, 291, 398 ] }, { "type": "text", "bbox": [ 70, 408, 290, 463 ], "lines": [ { "bbox": [ 70, 408, 290, 419 ], "spans": [ { "bbox": [ 70, 408, 290, 419 ], "score": 1.0, "content": "Alexei Baevski, Yuhao Zhou, Abdelrahman Mohamed,", "type": "text" } ], "index": 21 }, { "bbox": [ 79, 419, 291, 431 ], "spans": [ { "bbox": [ 79, 419, 291, 431 ], "score": 1.0, "content": "and Michael Auli. 2020. wav2vec 2.0: A framework", "type": "text" } ], "index": 22 }, { "bbox": [ 79, 430, 292, 442 ], "spans": [ { "bbox": [ 79, 430, 292, 442 ], "score": 1.0, "content": "for self-supervised learning of speech representations.", "type": "text" } ], "index": 23 }, { "bbox": [ 78, 440, 292, 454 ], "spans": [ { "bbox": [ 78, 440, 292, 454 ], "score": 1.0, "content": "Advances in Neural Information Processing Systems,", "type": "text" } ], "index": 24 }, { "bbox": [ 79, 452, 153, 464 ], "spans": [ { "bbox": [ 79, 452, 153, 464 ], "score": 1.0, "content": "33:12449–12460.", "type": "text" } ], "index": 25 } ], "index": 23, "bbox_fs": [ 70, 408, 292, 464 ] }, { "type": "text", "bbox": [ 69, 475, 290, 530 ], "lines": [ { "bbox": [ 69, 474, 292, 487 ], "spans": [ { "bbox": [ 69, 474, 292, 487 ], "score": 1.0, "content": "Zalán Borsos, Raphaël Marinier, Damien Vincent, Eu-", "type": "text" } ], "index": 26 }, { "bbox": [ 79, 485, 292, 498 ], "spans": [ { "bbox": [ 79, 485, 292, 498 ], "score": 1.0, "content": "gene Kharitonov, Olivier Pietquin, Matt Sharifi,", "type": "text" } ], "index": 27 }, { "bbox": [ 80, 496, 291, 508 ], "spans": [ { "bbox": [ 80, 496, 291, 508 ], "score": 1.0, "content": "Olivier Teboul, David Grangier, Marco Tagliasac-", "type": "text" } ], "index": 28 }, { "bbox": [ 79, 506, 291, 521 ], "spans": [ { "bbox": [ 79, 506, 291, 521 ], "score": 1.0, "content": "chi, and Neil Zeghidour. 2022. Audiolm: a language", "type": "text" } ], "index": 29 }, { "bbox": [ 79, 519, 241, 531 ], "spans": [ { "bbox": [ 79, 519, 241, 531 ], "score": 1.0, "content": "modeling approach to audio generation.", "type": "text" } ], "index": 30 } ], "index": 28, "bbox_fs": [ 69, 474, 292, 531 ] }, { "type": "text", "bbox": [ 69, 541, 290, 585 ], "lines": [ { "bbox": [ 68, 539, 291, 555 ], "spans": [ { "bbox": [ 68, 539, 291, 555 ], "score": 1.0, "content": "Feilong Chen, Minglun Han, Haozhi Zhao, Qingyang", "type": "text" } ], "index": 31 }, { "bbox": [ 79, 552, 291, 564 ], "spans": [ { "bbox": [ 79, 552, 291, 564 ], "score": 1.0, "content": "Zhang, Jing Shi, Shuang Xu Xu, and Bo Xu. 2023. X-", "type": "text" } ], "index": 32 }, { "bbox": [ 79, 562, 291, 576 ], "spans": [ { "bbox": [ 79, 562, 291, 576 ], "score": 1.0, "content": "llm: Bootstrapping advanced large language models", "type": "text" } ], "index": 33 }, { "bbox": [ 79, 574, 281, 588 ], "spans": [ { "bbox": [ 79, 574, 281, 588 ], "score": 1.0, "content": "by treating multi-modalities as foreign languages.", "type": "text" } ], "index": 34 } ], "index": 32.5, "bbox_fs": [ 68, 539, 291, 588 ] }, { "type": "text", "bbox": [ 70, 597, 289, 684 ], "lines": [ { "bbox": [ 69, 596, 291, 609 ], "spans": [ { "bbox": [ 69, 596, 291, 609 ], "score": 1.0, "content": "Guoguo Chen, Shuzhou Chai, Guanbo Wang, Jiayu", "type": "text" } ], "index": 35 }, { "bbox": [ 79, 607, 290, 619 ], "spans": [ { "bbox": [ 79, 607, 290, 619 ], "score": 1.0, "content": "Du, Wei-Qiang Zhang, Chao Weng, Dan Su, Daniel", "type": "text" } ], "index": 36 }, { "bbox": [ 79, 618, 291, 631 ], "spans": [ { "bbox": [ 79, 618, 291, 631 ], "score": 1.0, "content": "Povey, Jan Trmal, Junbo Zhang, Mingjie Jin, San-", "type": "text" } ], "index": 37 }, { "bbox": [ 79, 629, 292, 642 ], "spans": [ { "bbox": [ 79, 629, 292, 642 ], "score": 1.0, "content": "jeev Khudanpur, Shinji Watanabe, Shuaijiang Zhao,", "type": "text" } ], "index": 38 }, { "bbox": [ 79, 639, 292, 654 ], "spans": [ { "bbox": [ 79, 639, 292, 654 ], "score": 1.0, "content": "Wei Zou, Xiangang Li, Xuchen Yao, Yongqing Wang,", "type": "text" } ], "index": 39 }, { "bbox": [ 79, 651, 291, 663 ], "spans": [ { "bbox": [ 79, 651, 291, 663 ], "score": 1.0, "content": "Yujun Wang, Zhao You, and Zhiyong Yan. 2021. Gi-", "type": "text" } ], "index": 40 }, { "bbox": [ 79, 662, 291, 675 ], "spans": [ { "bbox": [ 79, 662, 291, 675 ], "score": 1.0, "content": "gaspeech: An evolving, multi-domain asr corpus with", "type": "text" } ], "index": 45 }, { "bbox": [ 80, 673, 218, 685 ], "spans": [ { "bbox": [ 80, 673, 218, 685 ], "score": 1.0, "content": "10,000 hours of transcribed audio.", "type": "text" } ], "index": 46 } ], "index": 38.5, "bbox_fs": [ 69, 596, 292, 685 ] }, { "type": "text", "bbox": [ 305, 148, 525, 202 ], "lines": [ { "bbox": [ 304, 147, 526, 160 ], "spans": [ { "bbox": [ 304, 147, 526, 160 ], "score": 1.0, "content": "Xuxin Cheng, Zhihong Zhu, Ziyu Yao, Hongxiang Li,", "type": "text" } ], "index": 41 }, { "bbox": [ 315, 158, 527, 170 ], "spans": [ { "bbox": [ 315, 158, 527, 170 ], "score": 1.0, "content": "Yaowei Li, and Yuexian Zou. 2023c. GhostT5: Gen-", "type": "text" } ], "index": 42 }, { "bbox": [ 315, 169, 527, 181 ], "spans": [ { "bbox": [ 315, 169, 527, 181 ], "score": 1.0, "content": "erate More Features with Cheap Operations to Im-", "type": "text" } ], "index": 47 }, { "bbox": [ 315, 180, 527, 193 ], "spans": [ { "bbox": [ 315, 180, 527, 193 ], "score": 1.0, "content": "prove Textless Spoken Question Answering. In Proc.", "type": "text" } ], "index": 48 }, { "bbox": [ 315, 191, 483, 203 ], "spans": [ { "bbox": [ 315, 191, 483, 203 ], "score": 1.0, "content": "INTERSPEECH 2023, pages 1134–1138.", "type": "text" } ], "index": 51 } ], "index": 47, "bbox_fs": [ 304, 147, 527, 203 ] }, { "type": "text", "bbox": [ 305, 73, 525, 138 ], "lines": [ { "bbox": [ 304, 72, 527, 84 ], "spans": [ { "bbox": [ 304, 72, 527, 84 ], "score": 1.0, "content": "Xuxin Cheng, Qianqian Dong, Fengpeng Yue, Tom Ko,", "type": "text" } ], "index": 58 }, { "bbox": [ 314, 83, 527, 95 ], "spans": [ { "bbox": [ 314, 83, 527, 95 ], "score": 1.0, "content": "Mingxuan Wang, and Yuexian Zou. 2023b. M 3 st:", "type": "text" } ], "index": 44 }, { "bbox": [ 315, 94, 525, 106 ], "spans": [ { "bbox": [ 315, 94, 525, 106 ], "score": 1.0, "content": "Mix at three levels for speech translation. In ICASSP", "type": "text" } ], "index": 49 }, { "bbox": [ 315, 105, 526, 117 ], "spans": [ { "bbox": [ 315, 105, 526, 117 ], "score": 1.0, "content": "2023-2023 IEEE International Conference on Acous-", "type": "text" } ], "index": 59 }, { "bbox": [ 314, 115, 526, 130 ], "spans": [ { "bbox": [ 314, 115, 526, 130 ], "score": 1.0, "content": "tics, Speech and Signal Processing (ICASSP), pages", "type": "text" } ], "index": 60 }, { "bbox": [ 315, 127, 361, 138 ], "spans": [ { "bbox": [ 315, 127, 361, 138 ], "score": 1.0, "content": "1–5. IEEE.", "type": "text" } ], "index": 50 } ], "index": 54.0, "bbox_fs": [ 304, 72, 527, 138 ] }, { "type": "text", "bbox": [ 70, 696, 290, 772 ], "lines": [ { "bbox": [ 69, 695, 291, 708 ], "spans": [ { "bbox": [ 69, 695, 291, 708 ], "score": 1.0, "content": "Xuxin Cheng, Bowen Cao, Qichen Ye, Zhihong Zhu,", "type": "text" } ], "index": 52 }, { "bbox": [ 79, 707, 292, 719 ], "spans": [ { "bbox": [ 79, 707, 292, 719 ], "score": 1.0, "content": "Hongxiang Li, and Yuexian Zou. 2023a. Ml-lmcl:", "type": "text" } ], "index": 53 }, { "bbox": [ 79, 717, 292, 730 ], "spans": [ { "bbox": [ 79, 717, 292, 730 ], "score": 1.0, "content": "Mutual learning and large-margin contrastive learn-", "type": "text" } ], "index": 54 }, { "bbox": [ 79, 729, 290, 742 ], "spans": [ { "bbox": [ 79, 729, 290, 742 ], "score": 1.0, "content": "ing for improving asr robustness in spoken language", "type": "text" } ], "index": 43 }, { "bbox": [ 79, 740, 291, 752 ], "spans": [ { "bbox": [ 79, 740, 291, 752 ], "score": 1.0, "content": "understanding. In Findings of the Association for", "type": "text" } ], "index": 55 }, { "bbox": [ 79, 750, 291, 763 ], "spans": [ { "bbox": [ 79, 750, 291, 763 ], "score": 1.0, "content": "Computational Linguistics: ACL 2023, pages 6492–", "type": "text" } ], "index": 56 }, { "bbox": [ 78, 761, 106, 772 ], "spans": [ { "bbox": [ 78, 761, 106, 772 ], "score": 1.0, "content": "6505.", "type": "text" } ], "index": 57 } ], "index": 54, "bbox_fs": [ 69, 695, 292, 772 ] }, { "type": "text", "bbox": [ 311, 212, 525, 464 ], "lines": [ { "bbox": [ 309, 210, 527, 223 ], "spans": [ { "bbox": [ 309, 210, 527, 223 ], "score": 1.0, "content": "Aakanksha Chowdhery, Sharan Narang, Jacob Devlin,", "type": "text" } ], "index": 61 }, { "bbox": [ 315, 222, 527, 234 ], "spans": [ { "bbox": [ 315, 222, 527, 234 ], "score": 1.0, "content": "Maarten Bosma, Gaurav Mishra, Adam Roberts,", "type": "text" } ], "index": 62 }, { "bbox": [ 314, 232, 527, 245 ], "spans": [ { "bbox": [ 314, 232, 527, 245 ], "score": 1.0, "content": "Paul Barham, Hyung Won Chung, Charles Sutton,", "type": "text" } ], "index": 63 }, { "bbox": [ 315, 243, 527, 257 ], "spans": [ { "bbox": [ 315, 243, 527, 257 ], "score": 1.0, "content": "Sebastian Gehrmann, Parker Schuh, Kensen Shi,", "type": "text" } ], "index": 64 }, { "bbox": [ 315, 254, 527, 267 ], "spans": [ { "bbox": [ 315, 254, 527, 267 ], "score": 1.0, "content": "Sasha Tsvyashchenko, Joshua Maynez, Abhishek", "type": "text" } ], "index": 65 }, { "bbox": [ 315, 266, 527, 278 ], "spans": [ { "bbox": [ 315, 266, 527, 278 ], "score": 1.0, "content": "Rao, Parker Barnes, Yi Tay, Noam Shazeer, Vin-", "type": "text" } ], "index": 66 }, { "bbox": [ 315, 277, 526, 289 ], "spans": [ { "bbox": [ 315, 277, 526, 289 ], "score": 1.0, "content": "odkumar Prabhakaran, Emily Reif, Nan Du, Ben", "type": "text" } ], "index": 67 }, { "bbox": [ 314, 288, 526, 300 ], "spans": [ { "bbox": [ 314, 288, 526, 300 ], "score": 1.0, "content": "Hutchinson, Reiner Pope, James Bradbury, Jacob", "type": "text" } ], "index": 68 }, { "bbox": [ 314, 298, 527, 312 ], "spans": [ { "bbox": [ 314, 298, 527, 312 ], "score": 1.0, "content": "Austin, Michael Isard, Guy Gur-Ari, Pengcheng Yin,", "type": "text" } ], "index": 69 }, { "bbox": [ 315, 310, 527, 322 ], "spans": [ { "bbox": [ 315, 310, 527, 322 ], "score": 1.0, "content": "Toju Duke, Anselm Levskaya, Sanjay Ghemawat,", "type": "text" } ], "index": 70 }, { "bbox": [ 315, 321, 527, 333 ], "spans": [ { "bbox": [ 315, 321, 527, 333 ], "score": 1.0, "content": "Sunipa Dev, Henryk Michalewski, Xavier Garcia,", "type": "text" } ], "index": 71 }, { "bbox": [ 315, 331, 526, 344 ], "spans": [ { "bbox": [ 315, 331, 526, 344 ], "score": 1.0, "content": "Vedant Misra, Kevin Robinson, Liam Fedus, Denny", "type": "text" } ], "index": 72 }, { "bbox": [ 314, 342, 527, 356 ], "spans": [ { "bbox": [ 314, 342, 527, 356 ], "score": 1.0, "content": "Zhou, Daphne Ippolito, David Luan, Hyeontaek Lim,", "type": "text" } ], "index": 73 }, { "bbox": [ 314, 353, 527, 366 ], "spans": [ { "bbox": [ 314, 353, 527, 366 ], "score": 1.0, "content": "Barret Zoph, Alexander Spiridonov, Ryan Sepassi,", "type": "text" } ], "index": 74 }, { "bbox": [ 314, 365, 527, 376 ], "spans": [ { "bbox": [ 314, 365, 527, 376 ], "score": 1.0, "content": "David Dohan, Shivani Agrawal, Mark Omernick, An-", "type": "text" } ], "index": 75 }, { "bbox": [ 315, 375, 528, 388 ], "spans": [ { "bbox": [ 315, 375, 528, 388 ], "score": 1.0, "content": "drew M. Dai, Thanumalayan Sankaranarayana Pil-", "type": "text" } ], "index": 76 }, { "bbox": [ 315, 387, 527, 399 ], "spans": [ { "bbox": [ 315, 387, 527, 399 ], "score": 1.0, "content": "lai, Marie Pellat, Aitor Lewkowycz, Erica Moreira,", "type": "text" } ], "index": 77 }, { "bbox": [ 315, 397, 527, 410 ], "spans": [ { "bbox": [ 315, 397, 527, 410 ], "score": 1.0, "content": "Rewon Child, Oleksandr Polozov, Katherine Lee,", "type": "text" } ], "index": 78 }, { "bbox": [ 315, 409, 526, 420 ], "spans": [ { "bbox": [ 315, 409, 526, 420 ], "score": 1.0, "content": "Zongwei Zhou, Xuezhi Wang, Brennan Saeta, Mark", "type": "text" } ], "index": 79 }, { "bbox": [ 315, 419, 526, 432 ], "spans": [ { "bbox": [ 315, 419, 526, 432 ], "score": 1.0, "content": "Diaz, Orhan Firat, Michele Catasta, Jason Wei, Kathy", "type": "text" } ], "index": 80 }, { "bbox": [ 315, 430, 527, 442 ], "spans": [ { "bbox": [ 315, 430, 527, 442 ], "score": 1.0, "content": "Meier-Hellstern, Douglas Eck, Jeff Dean, Slav Petrov,", "type": "text" } ], "index": 81 }, { "bbox": [ 315, 441, 527, 454 ], "spans": [ { "bbox": [ 315, 441, 527, 454 ], "score": 1.0, "content": "and Noah Fiedel. 2022. Palm: Scaling language mod-", "type": "text" } ], "index": 82 }, { "bbox": [ 315, 452, 401, 465 ], "spans": [ { "bbox": [ 315, 452, 401, 465 ], "score": 1.0, "content": "eling with pathways.", "type": "text" } ], "index": 83 } ], "index": 72, "bbox_fs": [ 309, 210, 528, 465 ] }, { "type": "text", "bbox": [ 305, 473, 524, 549 ], "lines": [ { "bbox": [ 304, 472, 526, 484 ], "spans": [ { "bbox": [ 304, 472, 526, 484 ], "score": 1.0, "content": "Alexey Dosovitskiy, Lucas Beyer, Alexander", "type": "text" } ], "index": 84 }, { "bbox": [ 315, 483, 527, 496 ], "spans": [ { "bbox": [ 315, 483, 527, 496 ], "score": 1.0, "content": "Kolesnikov, Dirk Weissenborn, Xiaohua Zhai,", "type": "text" } ], "index": 85 }, { "bbox": [ 315, 494, 526, 506 ], "spans": [ { "bbox": [ 315, 494, 526, 506 ], "score": 1.0, "content": "Thomas Unterthiner, Mostafa Dehghani, Matthias", "type": "text" } ], "index": 86 }, { "bbox": [ 315, 505, 526, 518 ], "spans": [ { "bbox": [ 315, 505, 526, 518 ], "score": 1.0, "content": "Minderer, Georg Heigold, Sylvain Gelly, Jakob", "type": "text" } ], "index": 87 }, { "bbox": [ 315, 516, 526, 529 ], "spans": [ { "bbox": [ 315, 516, 526, 529 ], "score": 1.0, "content": "Uszkoreit, and Neil Houlsby. 2021. An image", "type": "text" } ], "index": 88 }, { "bbox": [ 314, 527, 526, 540 ], "spans": [ { "bbox": [ 314, 527, 526, 540 ], "score": 1.0, "content": "is worth 16x16 words: Transformers for image", "type": "text" } ], "index": 89 }, { "bbox": [ 315, 539, 398, 550 ], "spans": [ { "bbox": [ 315, 539, 398, 550 ], "score": 1.0, "content": "recognition at scale.", "type": "text" } ], "index": 90 } ], "index": 87, "bbox_fs": [ 304, 472, 527, 550 ] }, { "type": "text", "bbox": [ 304, 559, 525, 614 ], "lines": [ { "bbox": [ 303, 558, 527, 572 ], "spans": [ { "bbox": [ 303, 558, 527, 572 ], "score": 1.0, "content": "Danny Driess, Fei Xia, Mehdi SM Sajjadi, Corey Lynch,", "type": "text" } ], "index": 91 }, { "bbox": [ 315, 570, 526, 582 ], "spans": [ { "bbox": [ 315, 570, 526, 582 ], "score": 1.0, "content": "Aakanksha Chowdhery, Brian Ichter, Ayzaan Wahid,", "type": "text" } ], "index": 92 }, { "bbox": [ 315, 581, 527, 593 ], "spans": [ { "bbox": [ 315, 581, 527, 593 ], "score": 1.0, "content": "Jonathan Tompson, Quan Vuong, Tianhe Yu, et al.", "type": "text" } ], "index": 93 }, { "bbox": [ 314, 590, 526, 605 ], "spans": [ { "bbox": [ 314, 590, 526, 605 ], "score": 1.0, "content": "2023. Palm-e: An embodied multimodal language", "type": "text" } ], "index": 94 }, { "bbox": [ 315, 603, 484, 614 ], "spans": [ { "bbox": [ 315, 603, 484, 614 ], "score": 1.0, "content": "model. arXiv preprint arXiv:2303.03378.", "type": "text" } ], "index": 95 } ], "index": 93, "bbox_fs": [ 303, 558, 527, 614 ] }, { "type": "text", "bbox": [ 305, 623, 525, 689 ], "lines": [ { "bbox": [ 305, 623, 527, 635 ], "spans": [ { "bbox": [ 305, 623, 527, 635 ], "score": 1.0, "content": "Wei-Ning Hsu, Benjamin Bolte, Yao-Hung Hubert Tsai,", "type": "text" } ], "index": 96 }, { "bbox": [ 315, 634, 527, 645 ], "spans": [ { "bbox": [ 315, 634, 527, 645 ], "score": 1.0, "content": "Kushal Lakhotia, Ruslan Salakhutdinov, and Abdel-", "type": "text" } ], "index": 97 }, { "bbox": [ 314, 644, 526, 657 ], "spans": [ { "bbox": [ 314, 644, 526, 657 ], "score": 1.0, "content": "rahman Mohamed. 2021. Hubert: Self-supervised", "type": "text" } ], "index": 98 }, { "bbox": [ 315, 656, 526, 668 ], "spans": [ { "bbox": [ 315, 656, 526, 668 ], "score": 1.0, "content": "speech representation learning by masked prediction", "type": "text" } ], "index": 99 }, { "bbox": [ 315, 666, 527, 679 ], "spans": [ { "bbox": [ 315, 666, 527, 679 ], "score": 1.0, "content": "of hidden units. IEEE/ACM Transactions on Audio,", "type": "text" } ], "index": 100 }, { "bbox": [ 315, 678, 519, 690 ], "spans": [ { "bbox": [ 315, 678, 519, 690 ], "score": 1.0, "content": "Speech, and Language Processing, 29:3451–3460.", "type": "text" } ], "index": 101 } ], "index": 98.5, "bbox_fs": [ 305, 623, 527, 690 ] }, { "type": "text", "bbox": [ 304, 698, 525, 742 ], "lines": [ { "bbox": [ 303, 696, 526, 711 ], "spans": [ { "bbox": [ 303, 696, 526, 711 ], "score": 1.0, "content": "Edward J. Hu, Yelong Shen, Phillip Wallis, Zeyuan", "type": "text" } ], "index": 102 }, { "bbox": [ 314, 708, 525, 721 ], "spans": [ { "bbox": [ 314, 708, 525, 721 ], "score": 1.0, "content": "Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and", "type": "text" } ], "index": 103 }, { "bbox": [ 315, 718, 527, 731 ], "spans": [ { "bbox": [ 315, 718, 527, 731 ], "score": 1.0, "content": "Weizhu Chen. 2021. Lora: Low-rank adaptation of", "type": "text" } ], "index": 104 }, { "bbox": [ 315, 731, 410, 743 ], "spans": [ { "bbox": [ 315, 731, 410, 743 ], "score": 1.0, "content": "large language models.", "type": "text" } ], "index": 105 } ], "index": 103.5, "bbox_fs": [ 303, 696, 527, 743 ] }, { "type": "text", "bbox": [ 305, 751, 525, 773 ], "lines": [ { "bbox": [ 304, 750, 527, 763 ], "spans": [ { "bbox": [ 304, 750, 527, 763 ], "score": 1.0, "content": "Rongjie Huang, Mingze Li, Dongchao Yang, Jia-", "type": "text" } ], "index": 106 }, { "bbox": [ 315, 762, 527, 775 ], "spans": [ { "bbox": [ 315, 762, 527, 775 ], "score": 1.0, "content": "tong Shi, Xuankai Chang, Zhenhui Ye, Yuning Wu,", "type": "text" } ], "index": 107 } ], "index": 106.5, "bbox_fs": [ 304, 750, 527, 775 ] } ] }, { "preproc_blocks": [ { "type": "text", "bbox": [ 80, 73, 290, 117 ], "lines": [ { "bbox": [ 79, 72, 292, 85 ], "spans": [ { "bbox": [ 79, 72, 292, 85 ], "score": 1.0, "content": "Zhiqing Hong, Jiawei Huang, Jinglin Liu, Yi Ren,", "type": "text" } ], "index": 0 }, { "bbox": [ 79, 83, 291, 96 ], "spans": [ { "bbox": [ 79, 83, 291, 96 ], "score": 1.0, "content": "Zhou Zhao, and Shinji Watanabe. 2023a. Audiogpt:", "type": "text" } ], "index": 1 }, { "bbox": [ 79, 93, 291, 106 ], "spans": [ { "bbox": [ 79, 93, 291, 106 ], "score": 1.0, "content": "Understanding and generating speech, music, sound,", "type": "text" } ], "index": 2 }, { "bbox": [ 79, 105, 150, 117 ], "spans": [ { "bbox": [ 79, 105, 150, 117 ], "score": 1.0, "content": "and talking head.", "type": "text" } ], "index": 3 } ], "index": 1.5 }, { "type": "text", "bbox": [ 70, 126, 289, 203 ], "lines": [ { "bbox": [ 69, 124, 292, 138 ], "spans": [ { "bbox": [ 69, 124, 292, 138 ], "score": 1.0, "content": "Shaohan Huang, Li Dong, Wenhui Wang, Yaru Hao,", "type": "text" } ], "index": 4 }, { "bbox": [ 79, 137, 290, 149 ], "spans": [ { "bbox": [ 79, 137, 290, 149 ], "score": 1.0, "content": "Saksham Singhal, Shuming Ma, Tengchao Lv, Lei", "type": "text" } ], "index": 5 }, { "bbox": [ 79, 146, 291, 161 ], "spans": [ { "bbox": [ 79, 146, 291, 161 ], "score": 1.0, "content": "Cui, Owais Khan Mohammed, Barun Patra, Qiang", "type": "text" } ], "index": 6 }, { "bbox": [ 79, 158, 291, 171 ], "spans": [ { "bbox": [ 79, 158, 291, 171 ], "score": 1.0, "content": "Liu, Kriti Aggarwal, Zewen Chi, Johan Bjorck,", "type": "text" } ], "index": 7 }, { "bbox": [ 79, 169, 291, 182 ], "spans": [ { "bbox": [ 79, 169, 291, 182 ], "score": 1.0, "content": "Vishrav Chaudhary, Subhojit Som, Xia Song, and", "type": "text" } ], "index": 8 }, { "bbox": [ 78, 179, 292, 194 ], "spans": [ { "bbox": [ 78, 179, 292, 194 ], "score": 1.0, "content": "Furu Wei. 2023b. Language is not all you need:", "type": "text" } ], "index": 9 }, { "bbox": [ 80, 192, 254, 204 ], "spans": [ { "bbox": [ 80, 192, 254, 204 ], "score": 1.0, "content": "Aligning perception with language models.", "type": "text" } ], "index": 10 } ], "index": 7 }, { "type": "text", "bbox": [ 69, 213, 290, 300 ], "lines": [ { "bbox": [ 68, 212, 291, 225 ], "spans": [ { "bbox": [ 68, 212, 291, 225 ], "score": 1.0, "content": "J. Kahn, M. Riviere, W. Zheng, E. Kharitonov, Q. Xu,", "type": "text" } ], "index": 11 }, { "bbox": [ 79, 223, 291, 235 ], "spans": [ { "bbox": [ 79, 223, 291, 235 ], "score": 1.0, "content": "P.E. Mazare, J. Karadayi, V. Liptchinsky, R. Col-", "type": "text" } ], "index": 12 }, { "bbox": [ 79, 234, 292, 247 ], "spans": [ { "bbox": [ 79, 234, 292, 247 ], "score": 1.0, "content": "lobert, C. Fuegen, T. Likhomanenko, G. Synnaeve,", "type": "text" } ], "index": 13 }, { "bbox": [ 80, 246, 291, 257 ], "spans": [ { "bbox": [ 80, 246, 291, 257 ], "score": 1.0, "content": "A. Joulin, A. Mohamed, and E. Dupoux. 2020. Libri-", "type": "text" } ], "index": 14 }, { "bbox": [ 79, 256, 292, 268 ], "spans": [ { "bbox": [ 79, 256, 292, 268 ], "score": 1.0, "content": "light: A benchmark for ASR with limited or no super-", "type": "text" } ], "index": 15 }, { "bbox": [ 79, 267, 290, 279 ], "spans": [ { "bbox": [ 79, 267, 290, 279 ], "score": 1.0, "content": "vision. In ICASSP 2020 - 2020 IEEE International", "type": "text" } ], "index": 16 }, { "bbox": [ 80, 278, 292, 290 ], "spans": [ { "bbox": [ 80, 278, 292, 290 ], "score": 1.0, "content": "Conference on Acoustics, Speech and Signal Process-", "type": "text" } ], "index": 17 }, { "bbox": [ 79, 290, 166, 300 ], "spans": [ { "bbox": [ 79, 290, 166, 300 ], "score": 1.0, "content": "ing (ICASSP). IEEE.", "type": "text" } ], "index": 18 } ], "index": 14.5 }, { "type": "text", "bbox": [ 70, 310, 290, 387 ], "lines": [ { "bbox": [ 68, 308, 292, 323 ], "spans": [ { "bbox": [ 68, 308, 292, 323 ], "score": 1.0, "content": "Kushal Lakhotia, Eugene Kharitonov, Wei-Ning Hsu,", "type": "text" } ], "index": 19 }, { "bbox": [ 79, 320, 290, 333 ], "spans": [ { "bbox": [ 79, 320, 290, 333 ], "score": 1.0, "content": "Yossi Adi, Adam Polyak, Benjamin Bolte, Tu-Anh", "type": "text" } ], "index": 20 }, { "bbox": [ 79, 332, 291, 344 ], "spans": [ { "bbox": [ 79, 332, 291, 344 ], "score": 1.0, "content": "Nguyen, Jade Copet, Alexei Baevski, Abdelrahman", "type": "text" } ], "index": 21 }, { "bbox": [ 79, 343, 291, 354 ], "spans": [ { "bbox": [ 79, 343, 291, 354 ], "score": 1.0, "content": "Mohamed, et al. 2021. On generative spoken lan-", "type": "text" } ], "index": 22 }, { "bbox": [ 79, 354, 291, 366 ], "spans": [ { "bbox": [ 79, 354, 291, 366 ], "score": 1.0, "content": "guage modeling from raw audio. Transactions of the", "type": "text" } ], "index": 23 }, { "bbox": [ 78, 365, 291, 376 ], "spans": [ { "bbox": [ 78, 365, 291, 376 ], "score": 1.0, "content": "Association for Computational Linguistics, 9:1336–", "type": "text" } ], "index": 24 }, { "bbox": [ 79, 375, 105, 387 ], "spans": [ { "bbox": [ 79, 375, 105, 387 ], "score": 1.0, "content": "1354.", "type": "text" } ], "index": 25 } ], "index": 22 }, { "type": "text", "bbox": [ 69, 397, 289, 429 ], "lines": [ { "bbox": [ 69, 396, 290, 409 ], "spans": [ { "bbox": [ 69, 396, 290, 409 ], "score": 1.0, "content": "Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae", "type": "text" } ], "index": 26 }, { "bbox": [ 79, 407, 290, 420 ], "spans": [ { "bbox": [ 79, 407, 290, 420 ], "score": 1.0, "content": "Lee. 2023. Visual instruction tuning. arXiv preprint", "type": "text" } ], "index": 27 }, { "bbox": [ 79, 419, 158, 429 ], "spans": [ { "bbox": [ 79, 419, 158, 429 ], "score": 1.0, "content": "arXiv:2304.08485.", "type": "text" } ], "index": 28 } ], "index": 27 }, { "type": "text", "bbox": [ 70, 439, 290, 494 ], "lines": [ { "bbox": [ 69, 439, 290, 451 ], "spans": [ { "bbox": [ 69, 439, 290, 451 ], "score": 1.0, "content": "Tu Anh Nguyen, Eugene Kharitonov, Jade Copet, Yossi", "type": "text" } ], "index": 29 }, { "bbox": [ 79, 449, 292, 462 ], "spans": [ { "bbox": [ 79, 449, 292, 462 ], "score": 1.0, "content": "Adi, Wei-Ning Hsu, Ali Elkahky, Paden Tomasello,", "type": "text" } ], "index": 30 }, { "bbox": [ 79, 461, 292, 473 ], "spans": [ { "bbox": [ 79, 461, 292, 473 ], "score": 1.0, "content": "Robin Algayres, Benoit Sagot, Abdelrahman Mo-", "type": "text" } ], "index": 31 }, { "bbox": [ 79, 472, 290, 483 ], "spans": [ { "bbox": [ 79, 472, 290, 483 ], "score": 1.0, "content": "hamed, and Emmanuel Dupoux. 2022. Generative", "type": "text" } ], "index": 32 }, { "bbox": [ 79, 482, 229, 496 ], "spans": [ { "bbox": [ 79, 482, 229, 496 ], "score": 1.0, "content": "spoken dialogue language modeling.", "type": "text" } ], "index": 33 } ], "index": 31 }, { "type": "text", "bbox": [ 70, 504, 223, 515 ], "lines": [ { "bbox": [ 69, 502, 226, 516 ], "spans": [ { "bbox": [ 69, 502, 226, 516 ], "score": 1.0, "content": "OpenAI. 2023. Gpt-4 technical report.", "type": "text" } ], "index": 34 } ], "index": 34 }, { "type": "text", "bbox": [ 69, 524, 290, 579 ], "lines": [ { "bbox": [ 69, 523, 291, 536 ], "spans": [ { "bbox": [ 69, 523, 291, 536 ], "score": 1.0, "content": "Vassil Panayotov, Guoguo Chen, Daniel Povey, and San-", "type": "text" } ], "index": 35 }, { "bbox": [ 78, 534, 291, 547 ], "spans": [ { "bbox": [ 78, 534, 291, 547 ], "score": 1.0, "content": "jeev Khudanpur. 2015. Librispeech: An asr corpus", "type": "text" } ], "index": 36 }, { "bbox": [ 79, 546, 290, 557 ], "spans": [ { "bbox": [ 79, 546, 290, 557 ], "score": 1.0, "content": "based on public domain audio books. In 2015 IEEE", "type": "text" } ], "index": 37 }, { "bbox": [ 80, 556, 290, 569 ], "spans": [ { "bbox": [ 80, 556, 290, 569 ], "score": 1.0, "content": "International Conference on Acoustics, Speech and", "type": "text" } ], "index": 38 }, { "bbox": [ 79, 568, 273, 580 ], "spans": [ { "bbox": [ 79, 568, 273, 580 ], "score": 1.0, "content": "Signal Processing (ICASSP), pages 5206–5210.", "type": "text" } ], "index": 39 } ], "index": 37 }, { "type": "text", "bbox": [ 70, 589, 290, 644 ], "lines": [ { "bbox": [ 69, 586, 291, 602 ], "spans": [ { "bbox": [ 69, 586, 291, 602 ], "score": 1.0, "content": "Adam Polyak, Yossi Adi, Jade Copet, Eugene", "type": "text" } ], "index": 40 }, { "bbox": [ 80, 600, 291, 611 ], "spans": [ { "bbox": [ 80, 600, 291, 611 ], "score": 1.0, "content": "Kharitonov, Kushal Lakhotia, Wei-Ning Hsu, Ab-", "type": "text" } ], "index": 41 }, { "bbox": [ 80, 611, 291, 622 ], "spans": [ { "bbox": [ 80, 611, 291, 622 ], "score": 1.0, "content": "delrahman Mohamed, and Emmanuel Dupoux. 2021.", "type": "text" } ], "index": 42 }, { "bbox": [ 80, 622, 292, 633 ], "spans": [ { "bbox": [ 80, 622, 292, 633 ], "score": 1.0, "content": "Speech resynthesis from discrete disentangled self-", "type": "text" } ], "index": 43 }, { "bbox": [ 79, 633, 190, 645 ], "spans": [ { "bbox": [ 79, 633, 190, 645 ], "score": 1.0, "content": "supervised representations.", "type": "text" } ], "index": 44 } ], "index": 42 }, { "type": "text", "bbox": [ 70, 653, 289, 719 ], "lines": [ { "bbox": [ 69, 652, 291, 666 ], "spans": [ { "bbox": [ 69, 652, 291, 666 ], "score": 1.0, "content": "Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya", "type": "text" } ], "index": 45 }, { "bbox": [ 80, 664, 291, 675 ], "spans": [ { "bbox": [ 80, 664, 291, 675 ], "score": 1.0, "content": "Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sas-", "type": "text" } ], "index": 46 }, { "bbox": [ 79, 675, 291, 687 ], "spans": [ { "bbox": [ 79, 675, 291, 687 ], "score": 1.0, "content": "try, Amanda Askell, Pamela Mishkin, Jack Clark,", "type": "text" } ], "index": 47 }, { "bbox": [ 81, 686, 291, 697 ], "spans": [ { "bbox": [ 81, 686, 291, 697 ], "score": 1.0, "content": "Gretchen Krueger, and Ilya Sutskever. 2021. Learn-", "type": "text" } ], "index": 48 }, { "bbox": [ 79, 697, 290, 709 ], "spans": [ { "bbox": [ 79, 697, 290, 709 ], "score": 1.0, "content": "ing transferable visual models from natural language", "type": "text" } ], "index": 49 }, { "bbox": [ 79, 709, 130, 720 ], "spans": [ { "bbox": [ 79, 709, 130, 720 ], "score": 1.0, "content": "supervision.", "type": "text" } ], "index": 50 } ], "index": 47.5 }, { "type": "text", "bbox": [ 69, 729, 290, 774 ], "lines": [ { "bbox": [ 69, 728, 291, 742 ], "spans": [ { "bbox": [ 69, 728, 291, 742 ], "score": 1.0, "content": "Yongliang Shen, Kaitao Song, Xu Tan, Dongsheng Li,", "type": "text" } ], "index": 51 }, { "bbox": [ 79, 738, 291, 753 ], "spans": [ { "bbox": [ 79, 738, 291, 753 ], "score": 1.0, "content": "Weiming Lu, and Yueting Zhuang. 2023. Hugging-", "type": "text" } ], "index": 52 }, { "bbox": [ 79, 751, 290, 762 ], "spans": [ { "bbox": [ 79, 751, 290, 762 ], "score": 1.0, "content": "gpt: Solving ai tasks with chatgpt and its friends in", "type": "text" } ], "index": 53 }, { "bbox": [ 79, 761, 135, 774 ], "spans": [ { "bbox": [ 79, 761, 135, 774 ], "score": 1.0, "content": "huggingface.", "type": "text" } ], "index": 54 } ], "index": 52.5 }, { "type": "text", "bbox": [ 304, 73, 525, 139 ], "lines": [ { "bbox": [ 304, 71, 526, 84 ], "spans": [ { "bbox": [ 304, 71, 526, 84 ], "score": 1.0, "content": "Rohan Taori, Ishaan Gulrajani, Tianyi Zhang, Yann", "type": "text" } ], "index": 55 }, { "bbox": [ 314, 82, 527, 96 ], "spans": [ { "bbox": [ 314, 82, 527, 96 ], "score": 1.0, "content": "Dubois, Xuechen Li, Carlos Guestrin, Percy", "type": "text" } ], "index": 56 }, { "bbox": [ 314, 93, 527, 106 ], "spans": [ { "bbox": [ 314, 93, 527, 106 ], "score": 1.0, "content": "Liang, and Tatsunori B. Hashimoto. 2023. Stan-", "type": "text" } ], "index": 57 }, { "bbox": [ 314, 105, 526, 118 ], "spans": [ { "bbox": [ 314, 105, 526, 118 ], "score": 1.0, "content": "ford alpaca: An instruction-following llama", "type": "text" } ], "index": 58 }, { "bbox": [ 315, 116, 525, 128 ], "spans": [ { "bbox": [ 315, 116, 525, 128 ], "score": 1.0, "content": "model. https://github.com/tatsu-lab/", "type": "text" } ], "index": 59 }, { "bbox": [ 316, 128, 410, 139 ], "spans": [ { "bbox": [ 316, 128, 410, 139 ], "score": 1.0, "content": "stanford_alpaca.", "type": "text" } ], "index": 60 } ], "index": 57.5 }, { "type": "text", "bbox": [ 305, 147, 525, 212 ], "lines": [ { "bbox": [ 303, 146, 527, 159 ], "spans": [ { "bbox": [ 303, 146, 527, 159 ], "score": 1.0, "content": "Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier", "type": "text" } ], "index": 61 }, { "bbox": [ 314, 157, 527, 170 ], "spans": [ { "bbox": [ 314, 157, 527, 170 ], "score": 1.0, "content": "Martinet, Marie-Anne Lachaux, Timothée Lacroix,", "type": "text" } ], "index": 62 }, { "bbox": [ 315, 168, 527, 181 ], "spans": [ { "bbox": [ 315, 168, 527, 181 ], "score": 1.0, "content": "Baptiste Rozière, Naman Goyal, Eric Hambro,", "type": "text" } ], "index": 63 }, { "bbox": [ 315, 180, 527, 192 ], "spans": [ { "bbox": [ 315, 180, 527, 192 ], "score": 1.0, "content": "Faisal Azhar, et al. 2023. Llama: Open and effi-", "type": "text" } ], "index": 64 }, { "bbox": [ 315, 191, 526, 204 ], "spans": [ { "bbox": [ 315, 191, 526, 204 ], "score": 1.0, "content": "cient foundation language models. arXiv preprint", "type": "text" } ], "index": 65 }, { "bbox": [ 315, 202, 394, 213 ], "spans": [ { "bbox": [ 315, 202, 394, 213 ], "score": 1.0, "content": "arXiv:2302.13971.", "type": "text" } ], "index": 66 } ], "index": 63.5 }, { "type": "text", "bbox": [ 304, 222, 525, 277 ], "lines": [ { "bbox": [ 304, 220, 528, 236 ], "spans": [ { "bbox": [ 304, 220, 528, 236 ], "score": 1.0, "content": "Chengyi Wang, Sanyuan Chen, Yu Wu, Ziqiang Zhang,", "type": "text" } ], "index": 67 }, { "bbox": [ 314, 232, 527, 245 ], "spans": [ { "bbox": [ 314, 232, 527, 245 ], "score": 1.0, "content": "Long Zhou, Shujie Liu, Zhuo Chen, Yanqing Liu,", "type": "text" } ], "index": 68 }, { "bbox": [ 315, 244, 525, 255 ], "spans": [ { "bbox": [ 315, 244, 525, 255 ], "score": 1.0, "content": "Huaming Wang, Jinyu Li, Lei He, Sheng Zhao, and", "type": "text" } ], "index": 69 }, { "bbox": [ 315, 254, 525, 267 ], "spans": [ { "bbox": [ 315, 254, 525, 267 ], "score": 1.0, "content": "Furu Wei. 2023. Neural codec language models are", "type": "text" } ], "index": 70 }, { "bbox": [ 315, 266, 465, 277 ], "spans": [ { "bbox": [ 315, 266, 465, 277 ], "score": 1.0, "content": "zero-shot text to speech synthesizers.", "type": "text" } ], "index": 71 } ], "index": 69 }, { "type": "text", "bbox": [ 304, 286, 524, 330 ], "lines": [ { "bbox": [ 305, 285, 525, 298 ], "spans": [ { "bbox": [ 305, 285, 525, 298 ], "score": 1.0, "content": "Yizhong Wang, Yeganeh Kordi, Swaroop Mishra, Alisa", "type": "text" } ], "index": 72 }, { "bbox": [ 315, 296, 526, 308 ], "spans": [ { "bbox": [ 315, 296, 526, 308 ], "score": 1.0, "content": "Liu, Noah A. Smith, Daniel Khashabi, and Hannaneh", "type": "text" } ], "index": 73 }, { "bbox": [ 314, 306, 526, 321 ], "spans": [ { "bbox": [ 314, 306, 526, 321 ], "score": 1.0, "content": "Hajishirzi. 2022. Self-instruct: Aligning language", "type": "text" } ], "index": 74 }, { "bbox": [ 315, 318, 473, 330 ], "spans": [ { "bbox": [ 315, 318, 473, 330 ], "score": 1.0, "content": "model with self generated instructions.", "type": "text" } ], "index": 75 } ], "index": 73.5 }, { "type": "text", "bbox": [ 305, 339, 525, 405 ], "lines": [ { "bbox": [ 303, 336, 528, 353 ], "spans": [ { "bbox": [ 303, 336, 528, 353 ], "score": 1.0, "content": "Dong Zhang, Rong Ye, Tom Ko, Mingxuan Wang,", "type": "text" } ], "index": 76 }, { "bbox": [ 315, 349, 527, 361 ], "spans": [ { "bbox": [ 315, 349, 527, 361 ], "score": 1.0, "content": "and Yaqian Zhou. 2023a. DUB: Discrete unit back-", "type": "text" } ], "index": 77 }, { "bbox": [ 315, 360, 527, 373 ], "spans": [ { "bbox": [ 315, 360, 527, 373 ], "score": 1.0, "content": "translation for speech translation. In Findings of", "type": "text" } ], "index": 78 }, { "bbox": [ 315, 371, 526, 383 ], "spans": [ { "bbox": [ 315, 371, 526, 383 ], "score": 1.0, "content": "the Association for Computational Linguistics: ACL", "type": "text" } ], "index": 79 }, { "bbox": [ 315, 381, 527, 394 ], "spans": [ { "bbox": [ 315, 381, 527, 394 ], "score": 1.0, "content": "2023, pages 7147–7164, Toronto, Canada. Associa-", "type": "text" } ], "index": 80 }, { "bbox": [ 315, 393, 459, 406 ], "spans": [ { "bbox": [ 315, 393, 459, 406 ], "score": 1.0, "content": "tion for Computational Linguistics.", "type": "text" } ], "index": 81 } ], "index": 78.5 }, { "type": "text", "bbox": [ 304, 414, 525, 457 ], "lines": [ { "bbox": [ 304, 412, 527, 425 ], "spans": [ { "bbox": [ 304, 412, 527, 425 ], "score": 1.0, "content": "Renrui Zhang, Jiaming Han, Aojun Zhou, Xiangfei Hu,", "type": "text" } ], "index": 82 }, { "bbox": [ 315, 424, 526, 436 ], "spans": [ { "bbox": [ 315, 424, 526, 436 ], "score": 1.0, "content": "Shilin Yan, Pan Lu, Hongsheng Li, Peng Gao, and", "type": "text" } ], "index": 83 }, { "bbox": [ 315, 433, 527, 448 ], "spans": [ { "bbox": [ 315, 433, 527, 448 ], "score": 1.0, "content": "Yu Qiao. 2023b. Llama-adapter: Efficient fine-tuning", "type": "text" } ], "index": 84 }, { "bbox": [ 315, 446, 492, 458 ], "spans": [ { "bbox": [ 315, 446, 492, 458 ], "score": 1.0, "content": "of language models with zero-init attention.", "type": "text" } ], "index": 85 } ], "index": 83.5 }, { "type": "text", "bbox": [ 304, 466, 524, 500 ], "lines": [ { "bbox": [ 303, 465, 526, 478 ], "spans": [ { "bbox": [ 303, 465, 526, 478 ], "score": 1.0, "content": "Xin Zhang, Dong Zhang, Shimin Li, Yaqian Zhou, and", "type": "text" } ], "index": 86 }, { "bbox": [ 315, 477, 526, 489 ], "spans": [ { "bbox": [ 315, 477, 526, 489 ], "score": 1.0, "content": "Xipeng Qiu. 2023c. Speechtokenizer: Unified speech", "type": "text" } ], "index": 87 }, { "bbox": [ 315, 488, 492, 500 ], "spans": [ { "bbox": [ 315, 488, 492, 500 ], "score": 1.0, "content": "tokenizer for speech large language models.", "type": "text" } ], "index": 88 } ], "index": 87 }, { "type": "text", "bbox": [ 305, 508, 525, 574 ], "lines": [ { "bbox": [ 304, 507, 526, 520 ], "spans": [ { "bbox": [ 304, 507, 526, 520 ], "score": 1.0, "content": "Ziqiang Zhang, Long Zhou, Chengyi Wang, Sanyuan", "type": "text" } ], "index": 89 }, { "bbox": [ 315, 519, 527, 531 ], "spans": [ { "bbox": [ 315, 519, 527, 531 ], "score": 1.0, "content": "Chen, Yu Wu, Shujie Liu, Zhuo Chen, Yanqing Liu,", "type": "text" } ], "index": 90 }, { "bbox": [ 315, 529, 526, 542 ], "spans": [ { "bbox": [ 315, 529, 526, 542 ], "score": 1.0, "content": "Huaming Wang, Jinyu Li, Lei He, Sheng Zhao, and", "type": "text" } ], "index": 91 }, { "bbox": [ 315, 540, 527, 553 ], "spans": [ { "bbox": [ 315, 540, 527, 553 ], "score": 1.0, "content": "Furu Wei. 2023d. Speak foreign languages with your", "type": "text" } ], "index": 92 }, { "bbox": [ 315, 552, 527, 564 ], "spans": [ { "bbox": [ 315, 552, 527, 564 ], "score": 1.0, "content": "own voice: Cross-lingual neural codec language mod-", "type": "text" } ], "index": 93 }, { "bbox": [ 315, 561, 342, 577 ], "spans": [ { "bbox": [ 315, 561, 342, 577 ], "score": 1.0, "content": "eling.", "type": "text" } ], "index": 94 } ], "index": 91.5 }, { "type": "text", "bbox": [ 304, 583, 525, 637 ], "lines": [ { "bbox": [ 305, 582, 525, 595 ], "spans": [ { "bbox": [ 305, 582, 525, 595 ], "score": 1.0, "content": "Chunting Zhou, Pengfei Liu, Puxin Xu, Srini Iyer, Jiao", "type": "text" } ], "index": 95 }, { "bbox": [ 315, 593, 526, 606 ], "spans": [ { "bbox": [ 315, 593, 526, 606 ], "score": 1.0, "content": "Sun, Yuning Mao, Xuezhe Ma, Avia Efrat, Ping Yu,", "type": "text" } ], "index": 96 }, { "bbox": [ 314, 603, 527, 617 ], "spans": [ { "bbox": [ 314, 603, 527, 617 ], "score": 1.0, "content": "Lili Yu, Susan Zhang, Gargi Ghosh, Mike Lewis,", "type": "text" } ], "index": 97 }, { "bbox": [ 315, 615, 526, 627 ], "spans": [ { "bbox": [ 315, 615, 526, 627 ], "score": 1.0, "content": "Luke Zettlemoyer, and Omer Levy. 2023. Lima: Less", "type": "text" } ], "index": 98 }, { "bbox": [ 315, 627, 406, 638 ], "spans": [ { "bbox": [ 315, 627, 406, 638 ], "score": 1.0, "content": "is more for alignment.", "type": "text" } ], "index": 99 } ], "index": 97 } ], "page_idx": 9, "page_size": [ 595, 841 ], "discarded_blocks": [], "para_blocks": [ { "type": "text", "bbox": [ 80, 73, 290, 117 ], "lines": [ { "bbox": [ 79, 72, 292, 85 ], "spans": [ { "bbox": [ 79, 72, 292, 85 ], "score": 1.0, "content": "Zhiqing Hong, Jiawei Huang, Jinglin Liu, Yi Ren,", "type": "text" } ], "index": 0 }, { "bbox": [ 79, 83, 291, 96 ], "spans": [ { "bbox": [ 79, 83, 291, 96 ], "score": 1.0, "content": "Zhou Zhao, and Shinji Watanabe. 2023a. Audiogpt:", "type": "text" } ], "index": 1 }, { "bbox": [ 79, 93, 291, 106 ], "spans": [ { "bbox": [ 79, 93, 291, 106 ], "score": 1.0, "content": "Understanding and generating speech, music, sound,", "type": "text" } ], "index": 2 }, { "bbox": [ 79, 105, 150, 117 ], "spans": [ { "bbox": [ 79, 105, 150, 117 ], "score": 1.0, "content": "and talking head.", "type": "text" } ], "index": 3 } ], "index": 1.5, "bbox_fs": [ 79, 72, 292, 117 ] }, { "type": "text", "bbox": [ 70, 126, 289, 203 ], "lines": [ { "bbox": [ 69, 124, 292, 138 ], "spans": [ { "bbox": [ 69, 124, 292, 138 ], "score": 1.0, "content": "Shaohan Huang, Li Dong, Wenhui Wang, Yaru Hao,", "type": "text" } ], "index": 4 }, { "bbox": [ 79, 137, 290, 149 ], "spans": [ { "bbox": [ 79, 137, 290, 149 ], "score": 1.0, "content": "Saksham Singhal, Shuming Ma, Tengchao Lv, Lei", "type": "text" } ], "index": 5 }, { "bbox": [ 79, 146, 291, 161 ], "spans": [ { "bbox": [ 79, 146, 291, 161 ], "score": 1.0, "content": "Cui, Owais Khan Mohammed, Barun Patra, Qiang", "type": "text" } ], "index": 6 }, { "bbox": [ 79, 158, 291, 171 ], "spans": [ { "bbox": [ 79, 158, 291, 171 ], "score": 1.0, "content": "Liu, Kriti Aggarwal, Zewen Chi, Johan Bjorck,", "type": "text" } ], "index": 7 }, { "bbox": [ 79, 169, 291, 182 ], "spans": [ { "bbox": [ 79, 169, 291, 182 ], "score": 1.0, "content": "Vishrav Chaudhary, Subhojit Som, Xia Song, and", "type": "text" } ], "index": 8 }, { "bbox": [ 78, 179, 292, 194 ], "spans": [ { "bbox": [ 78, 179, 292, 194 ], "score": 1.0, "content": "Furu Wei. 2023b. Language is not all you need:", "type": "text" } ], "index": 9 }, { "bbox": [ 80, 192, 254, 204 ], "spans": [ { "bbox": [ 80, 192, 254, 204 ], "score": 1.0, "content": "Aligning perception with language models.", "type": "text" } ], "index": 10 } ], "index": 7, "bbox_fs": [ 69, 124, 292, 204 ] }, { "type": "text", "bbox": [ 69, 213, 290, 300 ], "lines": [ { "bbox": [ 68, 212, 291, 225 ], "spans": [ { "bbox": [ 68, 212, 291, 225 ], "score": 1.0, "content": "J. Kahn, M. Riviere, W. Zheng, E. Kharitonov, Q. Xu,", "type": "text" } ], "index": 11 }, { "bbox": [ 79, 223, 291, 235 ], "spans": [ { "bbox": [ 79, 223, 291, 235 ], "score": 1.0, "content": "P.E. Mazare, J. Karadayi, V. Liptchinsky, R. Col-", "type": "text" } ], "index": 12 }, { "bbox": [ 79, 234, 292, 247 ], "spans": [ { "bbox": [ 79, 234, 292, 247 ], "score": 1.0, "content": "lobert, C. Fuegen, T. Likhomanenko, G. Synnaeve,", "type": "text" } ], "index": 13 }, { "bbox": [ 80, 246, 291, 257 ], "spans": [ { "bbox": [ 80, 246, 291, 257 ], "score": 1.0, "content": "A. Joulin, A. Mohamed, and E. Dupoux. 2020. Libri-", "type": "text" } ], "index": 14 }, { "bbox": [ 79, 256, 292, 268 ], "spans": [ { "bbox": [ 79, 256, 292, 268 ], "score": 1.0, "content": "light: A benchmark for ASR with limited or no super-", "type": "text" } ], "index": 15 }, { "bbox": [ 79, 267, 290, 279 ], "spans": [ { "bbox": [ 79, 267, 290, 279 ], "score": 1.0, "content": "vision. In ICASSP 2020 - 2020 IEEE International", "type": "text" } ], "index": 16 }, { "bbox": [ 80, 278, 292, 290 ], "spans": [ { "bbox": [ 80, 278, 292, 290 ], "score": 1.0, "content": "Conference on Acoustics, Speech and Signal Process-", "type": "text" } ], "index": 17 }, { "bbox": [ 79, 290, 166, 300 ], "spans": [ { "bbox": [ 79, 290, 166, 300 ], "score": 1.0, "content": "ing (ICASSP). IEEE.", "type": "text" } ], "index": 18 } ], "index": 14.5, "bbox_fs": [ 68, 212, 292, 300 ] }, { "type": "text", "bbox": [ 70, 310, 290, 387 ], "lines": [ { "bbox": [ 68, 308, 292, 323 ], "spans": [ { "bbox": [ 68, 308, 292, 323 ], "score": 1.0, "content": "Kushal Lakhotia, Eugene Kharitonov, Wei-Ning Hsu,", "type": "text" } ], "index": 19 }, { "bbox": [ 79, 320, 290, 333 ], "spans": [ { "bbox": [ 79, 320, 290, 333 ], "score": 1.0, "content": "Yossi Adi, Adam Polyak, Benjamin Bolte, Tu-Anh", "type": "text" } ], "index": 20 }, { "bbox": [ 79, 332, 291, 344 ], "spans": [ { "bbox": [ 79, 332, 291, 344 ], "score": 1.0, "content": "Nguyen, Jade Copet, Alexei Baevski, Abdelrahman", "type": "text" } ], "index": 21 }, { "bbox": [ 79, 343, 291, 354 ], "spans": [ { "bbox": [ 79, 343, 291, 354 ], "score": 1.0, "content": "Mohamed, et al. 2021. On generative spoken lan-", "type": "text" } ], "index": 22 }, { "bbox": [ 79, 354, 291, 366 ], "spans": [ { "bbox": [ 79, 354, 291, 366 ], "score": 1.0, "content": "guage modeling from raw audio. Transactions of the", "type": "text" } ], "index": 23 }, { "bbox": [ 78, 365, 291, 376 ], "spans": [ { "bbox": [ 78, 365, 291, 376 ], "score": 1.0, "content": "Association for Computational Linguistics, 9:1336–", "type": "text" } ], "index": 24 }, { "bbox": [ 79, 375, 105, 387 ], "spans": [ { "bbox": [ 79, 375, 105, 387 ], "score": 1.0, "content": "1354.", "type": "text" } ], "index": 25 } ], "index": 22, "bbox_fs": [ 68, 308, 292, 387 ] }, { "type": "text", "bbox": [ 69, 397, 289, 429 ], "lines": [ { "bbox": [ 69, 396, 290, 409 ], "spans": [ { "bbox": [ 69, 396, 290, 409 ], "score": 1.0, "content": "Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae", "type": "text" } ], "index": 26 }, { "bbox": [ 79, 407, 290, 420 ], "spans": [ { "bbox": [ 79, 407, 290, 420 ], "score": 1.0, "content": "Lee. 2023. Visual instruction tuning. arXiv preprint", "type": "text" } ], "index": 27 }, { "bbox": [ 79, 419, 158, 429 ], "spans": [ { "bbox": [ 79, 419, 158, 429 ], "score": 1.0, "content": "arXiv:2304.08485.", "type": "text" } ], "index": 28 } ], "index": 27, "bbox_fs": [ 69, 396, 290, 429 ] }, { "type": "text", "bbox": [ 70, 439, 290, 494 ], "lines": [ { "bbox": [ 69, 439, 290, 451 ], "spans": [ { "bbox": [ 69, 439, 290, 451 ], "score": 1.0, "content": "Tu Anh Nguyen, Eugene Kharitonov, Jade Copet, Yossi", "type": "text" } ], "index": 29 }, { "bbox": [ 79, 449, 292, 462 ], "spans": [ { "bbox": [ 79, 449, 292, 462 ], "score": 1.0, "content": "Adi, Wei-Ning Hsu, Ali Elkahky, Paden Tomasello,", "type": "text" } ], "index": 30 }, { "bbox": [ 79, 461, 292, 473 ], "spans": [ { "bbox": [ 79, 461, 292, 473 ], "score": 1.0, "content": "Robin Algayres, Benoit Sagot, Abdelrahman Mo-", "type": "text" } ], "index": 31 }, { "bbox": [ 79, 472, 290, 483 ], "spans": [ { "bbox": [ 79, 472, 290, 483 ], "score": 1.0, "content": "hamed, and Emmanuel Dupoux. 2022. Generative", "type": "text" } ], "index": 32 }, { "bbox": [ 79, 482, 229, 496 ], "spans": [ { "bbox": [ 79, 482, 229, 496 ], "score": 1.0, "content": "spoken dialogue language modeling.", "type": "text" } ], "index": 33 } ], "index": 31, "bbox_fs": [ 69, 439, 292, 496 ] }, { "type": "text", "bbox": [ 70, 504, 223, 515 ], "lines": [ { "bbox": [ 69, 502, 226, 516 ], "spans": [ { "bbox": [ 69, 502, 226, 516 ], "score": 1.0, "content": "OpenAI. 2023. Gpt-4 technical report.", "type": "text" } ], "index": 34 } ], "index": 34, "bbox_fs": [ 69, 502, 226, 516 ] }, { "type": "text", "bbox": [ 69, 524, 290, 579 ], "lines": [ { "bbox": [ 69, 523, 291, 536 ], "spans": [ { "bbox": [ 69, 523, 291, 536 ], "score": 1.0, "content": "Vassil Panayotov, Guoguo Chen, Daniel Povey, and San-", "type": "text" } ], "index": 35 }, { "bbox": [ 78, 534, 291, 547 ], "spans": [ { "bbox": [ 78, 534, 291, 547 ], "score": 1.0, "content": "jeev Khudanpur. 2015. Librispeech: An asr corpus", "type": "text" } ], "index": 36 }, { "bbox": [ 79, 546, 290, 557 ], "spans": [ { "bbox": [ 79, 546, 290, 557 ], "score": 1.0, "content": "based on public domain audio books. In 2015 IEEE", "type": "text" } ], "index": 37 }, { "bbox": [ 80, 556, 290, 569 ], "spans": [ { "bbox": [ 80, 556, 290, 569 ], "score": 1.0, "content": "International Conference on Acoustics, Speech and", "type": "text" } ], "index": 38 }, { "bbox": [ 79, 568, 273, 580 ], "spans": [ { "bbox": [ 79, 568, 273, 580 ], "score": 1.0, "content": "Signal Processing (ICASSP), pages 5206–5210.", "type": "text" } ], "index": 39 } ], "index": 37, "bbox_fs": [ 69, 523, 291, 580 ] }, { "type": "text", "bbox": [ 70, 589, 290, 644 ], "lines": [ { "bbox": [ 69, 586, 291, 602 ], "spans": [ { "bbox": [ 69, 586, 291, 602 ], "score": 1.0, "content": "Adam Polyak, Yossi Adi, Jade Copet, Eugene", "type": "text" } ], "index": 40 }, { "bbox": [ 80, 600, 291, 611 ], "spans": [ { "bbox": [ 80, 600, 291, 611 ], "score": 1.0, "content": "Kharitonov, Kushal Lakhotia, Wei-Ning Hsu, Ab-", "type": "text" } ], "index": 41 }, { "bbox": [ 80, 611, 291, 622 ], "spans": [ { "bbox": [ 80, 611, 291, 622 ], "score": 1.0, "content": "delrahman Mohamed, and Emmanuel Dupoux. 2021.", "type": "text" } ], "index": 42 }, { "bbox": [ 80, 622, 292, 633 ], "spans": [ { "bbox": [ 80, 622, 292, 633 ], "score": 1.0, "content": "Speech resynthesis from discrete disentangled self-", "type": "text" } ], "index": 43 }, { "bbox": [ 79, 633, 190, 645 ], "spans": [ { "bbox": [ 79, 633, 190, 645 ], "score": 1.0, "content": "supervised representations.", "type": "text" } ], "index": 44 } ], "index": 42, "bbox_fs": [ 69, 586, 292, 645 ] }, { "type": "text", "bbox": [ 70, 653, 289, 719 ], "lines": [ { "bbox": [ 69, 652, 291, 666 ], "spans": [ { "bbox": [ 69, 652, 291, 666 ], "score": 1.0, "content": "Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya", "type": "text" } ], "index": 45 }, { "bbox": [ 80, 664, 291, 675 ], "spans": [ { "bbox": [ 80, 664, 291, 675 ], "score": 1.0, "content": "Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sas-", "type": "text" } ], "index": 46 }, { "bbox": [ 79, 675, 291, 687 ], "spans": [ { "bbox": [ 79, 675, 291, 687 ], "score": 1.0, "content": "try, Amanda Askell, Pamela Mishkin, Jack Clark,", "type": "text" } ], "index": 47 }, { "bbox": [ 81, 686, 291, 697 ], "spans": [ { "bbox": [ 81, 686, 291, 697 ], "score": 1.0, "content": "Gretchen Krueger, and Ilya Sutskever. 2021. Learn-", "type": "text" } ], "index": 48 }, { "bbox": [ 79, 697, 290, 709 ], "spans": [ { "bbox": [ 79, 697, 290, 709 ], "score": 1.0, "content": "ing transferable visual models from natural language", "type": "text" } ], "index": 49 }, { "bbox": [ 79, 709, 130, 720 ], "spans": [ { "bbox": [ 79, 709, 130, 720 ], "score": 1.0, "content": "supervision.", "type": "text" } ], "index": 50 } ], "index": 47.5, "bbox_fs": [ 69, 652, 291, 720 ] }, { "type": "text", "bbox": [ 69, 729, 290, 774 ], "lines": [ { "bbox": [ 69, 728, 291, 742 ], "spans": [ { "bbox": [ 69, 728, 291, 742 ], "score": 1.0, "content": "Yongliang Shen, Kaitao Song, Xu Tan, Dongsheng Li,", "type": "text" } ], "index": 51 }, { "bbox": [ 79, 738, 291, 753 ], "spans": [ { "bbox": [ 79, 738, 291, 753 ], "score": 1.0, "content": "Weiming Lu, and Yueting Zhuang. 2023. Hugging-", "type": "text" } ], "index": 52 }, { "bbox": [ 79, 751, 290, 762 ], "spans": [ { "bbox": [ 79, 751, 290, 762 ], "score": 1.0, "content": "gpt: Solving ai tasks with chatgpt and its friends in", "type": "text" } ], "index": 53 }, { "bbox": [ 79, 761, 135, 774 ], "spans": [ { "bbox": [ 79, 761, 135, 774 ], "score": 1.0, "content": "huggingface.", "type": "text" } ], "index": 54 } ], "index": 52.5, "bbox_fs": [ 69, 728, 291, 774 ] }, { "type": "text", "bbox": [ 304, 73, 525, 139 ], "lines": [ { "bbox": [ 304, 71, 526, 84 ], "spans": [ { "bbox": [ 304, 71, 526, 84 ], "score": 1.0, "content": "Rohan Taori, Ishaan Gulrajani, Tianyi Zhang, Yann", "type": "text" } ], "index": 55 }, { "bbox": [ 314, 82, 527, 96 ], "spans": [ { "bbox": [ 314, 82, 527, 96 ], "score": 1.0, "content": "Dubois, Xuechen Li, Carlos Guestrin, Percy", "type": "text" } ], "index": 56 }, { "bbox": [ 314, 93, 527, 106 ], "spans": [ { "bbox": [ 314, 93, 527, 106 ], "score": 1.0, "content": "Liang, and Tatsunori B. Hashimoto. 2023. Stan-", "type": "text" } ], "index": 57 }, { "bbox": [ 314, 105, 526, 118 ], "spans": [ { "bbox": [ 314, 105, 526, 118 ], "score": 1.0, "content": "ford alpaca: An instruction-following llama", "type": "text" } ], "index": 58 }, { "bbox": [ 315, 116, 525, 128 ], "spans": [ { "bbox": [ 315, 116, 525, 128 ], "score": 1.0, "content": "model. https://github.com/tatsu-lab/", "type": "text" } ], "index": 59 }, { "bbox": [ 316, 128, 410, 139 ], "spans": [ { "bbox": [ 316, 128, 410, 139 ], "score": 1.0, "content": "stanford_alpaca.", "type": "text" } ], "index": 60 } ], "index": 57.5, "bbox_fs": [ 304, 71, 527, 139 ] }, { "type": "text", "bbox": [ 305, 147, 525, 212 ], "lines": [ { "bbox": [ 303, 146, 527, 159 ], "spans": [ { "bbox": [ 303, 146, 527, 159 ], "score": 1.0, "content": "Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier", "type": "text" } ], "index": 61 }, { "bbox": [ 314, 157, 527, 170 ], "spans": [ { "bbox": [ 314, 157, 527, 170 ], "score": 1.0, "content": "Martinet, Marie-Anne Lachaux, Timothée Lacroix,", "type": "text" } ], "index": 62 }, { "bbox": [ 315, 168, 527, 181 ], "spans": [ { "bbox": [ 315, 168, 527, 181 ], "score": 1.0, "content": "Baptiste Rozière, Naman Goyal, Eric Hambro,", "type": "text" } ], "index": 63 }, { "bbox": [ 315, 180, 527, 192 ], "spans": [ { "bbox": [ 315, 180, 527, 192 ], "score": 1.0, "content": "Faisal Azhar, et al. 2023. Llama: Open and effi-", "type": "text" } ], "index": 64 }, { "bbox": [ 315, 191, 526, 204 ], "spans": [ { "bbox": [ 315, 191, 526, 204 ], "score": 1.0, "content": "cient foundation language models. arXiv preprint", "type": "text" } ], "index": 65 }, { "bbox": [ 315, 202, 394, 213 ], "spans": [ { "bbox": [ 315, 202, 394, 213 ], "score": 1.0, "content": "arXiv:2302.13971.", "type": "text" } ], "index": 66 } ], "index": 63.5, "bbox_fs": [ 303, 146, 527, 213 ] }, { "type": "text", "bbox": [ 304, 222, 525, 277 ], "lines": [ { "bbox": [ 304, 220, 528, 236 ], "spans": [ { "bbox": [ 304, 220, 528, 236 ], "score": 1.0, "content": "Chengyi Wang, Sanyuan Chen, Yu Wu, Ziqiang Zhang,", "type": "text" } ], "index": 67 }, { "bbox": [ 314, 232, 527, 245 ], "spans": [ { "bbox": [ 314, 232, 527, 245 ], "score": 1.0, "content": "Long Zhou, Shujie Liu, Zhuo Chen, Yanqing Liu,", "type": "text" } ], "index": 68 }, { "bbox": [ 315, 244, 525, 255 ], "spans": [ { "bbox": [ 315, 244, 525, 255 ], "score": 1.0, "content": "Huaming Wang, Jinyu Li, Lei He, Sheng Zhao, and", "type": "text" } ], "index": 69 }, { "bbox": [ 315, 254, 525, 267 ], "spans": [ { "bbox": [ 315, 254, 525, 267 ], "score": 1.0, "content": "Furu Wei. 2023. Neural codec language models are", "type": "text" } ], "index": 70 }, { "bbox": [ 315, 266, 465, 277 ], "spans": [ { "bbox": [ 315, 266, 465, 277 ], "score": 1.0, "content": "zero-shot text to speech synthesizers.", "type": "text" } ], "index": 71 } ], "index": 69, "bbox_fs": [ 304, 220, 528, 277 ] }, { "type": "text", "bbox": [ 304, 286, 524, 330 ], "lines": [ { "bbox": [ 305, 285, 525, 298 ], "spans": [ { "bbox": [ 305, 285, 525, 298 ], "score": 1.0, "content": "Yizhong Wang, Yeganeh Kordi, Swaroop Mishra, Alisa", "type": "text" } ], "index": 72 }, { "bbox": [ 315, 296, 526, 308 ], "spans": [ { "bbox": [ 315, 296, 526, 308 ], "score": 1.0, "content": "Liu, Noah A. Smith, Daniel Khashabi, and Hannaneh", "type": "text" } ], "index": 73 }, { "bbox": [ 314, 306, 526, 321 ], "spans": [ { "bbox": [ 314, 306, 526, 321 ], "score": 1.0, "content": "Hajishirzi. 2022. Self-instruct: Aligning language", "type": "text" } ], "index": 74 }, { "bbox": [ 315, 318, 473, 330 ], "spans": [ { "bbox": [ 315, 318, 473, 330 ], "score": 1.0, "content": "model with self generated instructions.", "type": "text" } ], "index": 75 } ], "index": 73.5, "bbox_fs": [ 305, 285, 526, 330 ] }, { "type": "text", "bbox": [ 305, 339, 525, 405 ], "lines": [ { "bbox": [ 303, 336, 528, 353 ], "spans": [ { "bbox": [ 303, 336, 528, 353 ], "score": 1.0, "content": "Dong Zhang, Rong Ye, Tom Ko, Mingxuan Wang,", "type": "text" } ], "index": 76 }, { "bbox": [ 315, 349, 527, 361 ], "spans": [ { "bbox": [ 315, 349, 527, 361 ], "score": 1.0, "content": "and Yaqian Zhou. 2023a. DUB: Discrete unit back-", "type": "text" } ], "index": 77 }, { "bbox": [ 315, 360, 527, 373 ], "spans": [ { "bbox": [ 315, 360, 527, 373 ], "score": 1.0, "content": "translation for speech translation. In Findings of", "type": "text" } ], "index": 78 }, { "bbox": [ 315, 371, 526, 383 ], "spans": [ { "bbox": [ 315, 371, 526, 383 ], "score": 1.0, "content": "the Association for Computational Linguistics: ACL", "type": "text" } ], "index": 79 }, { "bbox": [ 315, 381, 527, 394 ], "spans": [ { "bbox": [ 315, 381, 527, 394 ], "score": 1.0, "content": "2023, pages 7147–7164, Toronto, Canada. Associa-", "type": "text" } ], "index": 80 }, { "bbox": [ 315, 393, 459, 406 ], "spans": [ { "bbox": [ 315, 393, 459, 406 ], "score": 1.0, "content": "tion for Computational Linguistics.", "type": "text" } ], "index": 81 } ], "index": 78.5, "bbox_fs": [ 303, 336, 528, 406 ] }, { "type": "text", "bbox": [ 304, 414, 525, 457 ], "lines": [ { "bbox": [ 304, 412, 527, 425 ], "spans": [ { "bbox": [ 304, 412, 527, 425 ], "score": 1.0, "content": "Renrui Zhang, Jiaming Han, Aojun Zhou, Xiangfei Hu,", "type": "text" } ], "index": 82 }, { "bbox": [ 315, 424, 526, 436 ], "spans": [ { "bbox": [ 315, 424, 526, 436 ], "score": 1.0, "content": "Shilin Yan, Pan Lu, Hongsheng Li, Peng Gao, and", "type": "text" } ], "index": 83 }, { "bbox": [ 315, 433, 527, 448 ], "spans": [ { "bbox": [ 315, 433, 527, 448 ], "score": 1.0, "content": "Yu Qiao. 2023b. Llama-adapter: Efficient fine-tuning", "type": "text" } ], "index": 84 }, { "bbox": [ 315, 446, 492, 458 ], "spans": [ { "bbox": [ 315, 446, 492, 458 ], "score": 1.0, "content": "of language models with zero-init attention.", "type": "text" } ], "index": 85 } ], "index": 83.5, "bbox_fs": [ 304, 412, 527, 458 ] }, { "type": "text", "bbox": [ 304, 466, 524, 500 ], "lines": [ { "bbox": [ 303, 465, 526, 478 ], "spans": [ { "bbox": [ 303, 465, 526, 478 ], "score": 1.0, "content": "Xin Zhang, Dong Zhang, Shimin Li, Yaqian Zhou, and", "type": "text" } ], "index": 86 }, { "bbox": [ 315, 477, 526, 489 ], "spans": [ { "bbox": [ 315, 477, 526, 489 ], "score": 1.0, "content": "Xipeng Qiu. 2023c. Speechtokenizer: Unified speech", "type": "text" } ], "index": 87 }, { "bbox": [ 315, 488, 492, 500 ], "spans": [ { "bbox": [ 315, 488, 492, 500 ], "score": 1.0, "content": "tokenizer for speech large language models.", "type": "text" } ], "index": 88 } ], "index": 87, "bbox_fs": [ 303, 465, 526, 500 ] }, { "type": "text", "bbox": [ 305, 508, 525, 574 ], "lines": [ { "bbox": [ 304, 507, 526, 520 ], "spans": [ { "bbox": [ 304, 507, 526, 520 ], "score": 1.0, "content": "Ziqiang Zhang, Long Zhou, Chengyi Wang, Sanyuan", "type": "text" } ], "index": 89 }, { "bbox": [ 315, 519, 527, 531 ], "spans": [ { "bbox": [ 315, 519, 527, 531 ], "score": 1.0, "content": "Chen, Yu Wu, Shujie Liu, Zhuo Chen, Yanqing Liu,", "type": "text" } ], "index": 90 }, { "bbox": [ 315, 529, 526, 542 ], "spans": [ { "bbox": [ 315, 529, 526, 542 ], "score": 1.0, "content": "Huaming Wang, Jinyu Li, Lei He, Sheng Zhao, and", "type": "text" } ], "index": 91 }, { "bbox": [ 315, 540, 527, 553 ], "spans": [ { "bbox": [ 315, 540, 527, 553 ], "score": 1.0, "content": "Furu Wei. 2023d. Speak foreign languages with your", "type": "text" } ], "index": 92 }, { "bbox": [ 315, 552, 527, 564 ], "spans": [ { "bbox": [ 315, 552, 527, 564 ], "score": 1.0, "content": "own voice: Cross-lingual neural codec language mod-", "type": "text" } ], "index": 93 }, { "bbox": [ 315, 561, 342, 577 ], "spans": [ { "bbox": [ 315, 561, 342, 577 ], "score": 1.0, "content": "eling.", "type": "text" } ], "index": 94 } ], "index": 91.5, "bbox_fs": [ 304, 507, 527, 577 ] }, { "type": "text", "bbox": [ 304, 583, 525, 637 ], "lines": [ { "bbox": [ 305, 582, 525, 595 ], "spans": [ { "bbox": [ 305, 582, 525, 595 ], "score": 1.0, "content": "Chunting Zhou, Pengfei Liu, Puxin Xu, Srini Iyer, Jiao", "type": "text" } ], "index": 95 }, { "bbox": [ 315, 593, 526, 606 ], "spans": [ { "bbox": [ 315, 593, 526, 606 ], "score": 1.0, "content": "Sun, Yuning Mao, Xuezhe Ma, Avia Efrat, Ping Yu,", "type": "text" } ], "index": 96 }, { "bbox": [ 314, 603, 527, 617 ], "spans": [ { "bbox": [ 314, 603, 527, 617 ], "score": 1.0, "content": "Lili Yu, Susan Zhang, Gargi Ghosh, Mike Lewis,", "type": "text" } ], "index": 97 }, { "bbox": [ 315, 615, 526, 627 ], "spans": [ { "bbox": [ 315, 615, 526, 627 ], "score": 1.0, "content": "Luke Zettlemoyer, and Omer Levy. 2023. Lima: Less", "type": "text" } ], "index": 98 }, { "bbox": [ 315, 627, 406, 638 ], "spans": [ { "bbox": [ 315, 627, 406, 638 ], "score": 1.0, "content": "is more for alignment.", "type": "text" } ], "index": 99 } ], "index": 97, "bbox_fs": [ 305, 582, 527, 638 ] } ] }, { "preproc_blocks": [ { "type": "title", "bbox": [ 69, 70, 288, 85 ], "lines": [ { "bbox": [ 67, 69, 289, 88 ], "spans": [ { "bbox": [ 67, 69, 289, 88 ], "score": 1.0, "content": "A Prompts to Generate Task Description", "type": "text" } ], "index": 0 } ], "index": 0 }, { "type": "title", "bbox": [ 85, 121, 110, 132 ], "lines": [ { "bbox": [ 83, 119, 113, 133 ], "spans": [ { "bbox": [ 83, 119, 113, 133 ], "score": 1.0, "content": "ASR:", "type": "text" } ], "index": 1 } ], "index": 1 }, { "type": "text", "bbox": [ 83, 135, 510, 160 ], "lines": [ { "bbox": [ 82, 132, 511, 149 ], "spans": [ { "bbox": [ 82, 132, 511, 149 ], "score": 1.0, "content": "You are asked to come up with a set of 100 diverse task instructions about automatic speech", "type": "text" } ], "index": 2 }, { "bbox": [ 82, 146, 297, 163 ], "spans": [ { "bbox": [ 82, 146, 297, 163 ], "score": 1.0, "content": "recognition, which is about recognizing speech.", "type": "text" } ], "index": 3 } ], "index": 2.5 }, { "type": "text", "bbox": [ 85, 162, 203, 174 ], "lines": [ { "bbox": [ 84, 161, 204, 174 ], "spans": [ { "bbox": [ 84, 161, 204, 174 ], "score": 1.0, "content": "Here are the requirements:", "type": "text" } ], "index": 4 } ], "index": 4 }, { "type": "text", "bbox": [ 82, 176, 510, 201 ], "lines": [ { "bbox": [ 80, 173, 512, 191 ], "spans": [ { "bbox": [ 80, 173, 512, 191 ], "score": 1.0, "content": "1. These instructions should be to instruct someone to recognize the content of the following", "type": "text" } ], "index": 5 }, { "bbox": [ 81, 188, 120, 202 ], "spans": [ { "bbox": [ 81, 188, 120, 202 ], "score": 1.0, "content": "speech.", "type": "text" } ], "index": 6 } ], "index": 5.5 }, { "type": "text", "bbox": [ 84, 202, 401, 214 ], "lines": [ { "bbox": [ 83, 201, 401, 216 ], "spans": [ { "bbox": [ 83, 201, 401, 216 ], "score": 1.0, "content": "2. Try not to repeat the verb for each instruction to maximize diversity.", "type": "text" } ], "index": 7 } ], "index": 7 }, { "type": "text", "bbox": [ 81, 216, 509, 241 ], "lines": [ { "bbox": [ 81, 214, 511, 230 ], "spans": [ { "bbox": [ 81, 214, 511, 230 ], "score": 1.0, "content": "3. The language used for instruction also should be diverse. For example, you should combine", "type": "text" } ], "index": 8 }, { "bbox": [ 83, 229, 257, 244 ], "spans": [ { "bbox": [ 83, 229, 257, 244 ], "score": 1.0, "content": "questions with imperative instructions.", "type": "text" } ], "index": 9 } ], "index": 8.5 }, { "type": "text", "bbox": [ 86, 244, 282, 255 ], "lines": [ { "bbox": [ 84, 243, 284, 256 ], "spans": [ { "bbox": [ 84, 243, 284, 256 ], "score": 1.0, "content": "4. The type of instructions should be diverse.", "type": "text" } ], "index": 10 } ], "index": 10 }, { "type": "text", "bbox": [ 86, 257, 264, 268 ], "lines": [ { "bbox": [ 84, 256, 265, 270 ], "spans": [ { "bbox": [ 84, 256, 265, 270 ], "score": 1.0, "content": "5. The instructions should be in English.", "type": "text" } ], "index": 11 } ], "index": 11 }, { "type": "text", "bbox": [ 84, 271, 511, 295 ], "lines": [ { "bbox": [ 81, 267, 512, 284 ], "spans": [ { "bbox": [ 81, 267, 512, 284 ], "score": 1.0, "content": "6. The instructions should be 1 to 2 sentences long. Either an imperative sentence or a question is", "type": "text" } ], "index": 12 }, { "bbox": [ 81, 283, 133, 296 ], "spans": [ { "bbox": [ 81, 283, 133, 296 ], "score": 1.0, "content": "permitted.", "type": "text" } ], "index": 13 } ], "index": 12.5 }, { "type": "text", "bbox": [ 85, 297, 162, 309 ], "lines": [ { "bbox": [ 83, 295, 164, 310 ], "spans": [ { "bbox": [ 83, 295, 164, 310 ], "score": 1.0, "content": "List of 100 tasks:", "type": "text" } ], "index": 14 } ], "index": 14 }, { "type": "text", "bbox": [ 85, 325, 109, 336 ], "lines": [ { "bbox": [ 83, 322, 112, 337 ], "spans": [ { "bbox": [ 83, 322, 112, 337 ], "score": 1.0, "content": "TTS:", "type": "text" } ], "index": 15 } ], "index": 15 }, { "type": "text", "bbox": [ 81, 338, 511, 363 ], "lines": [ { "bbox": [ 82, 336, 512, 351 ], "spans": [ { "bbox": [ 82, 336, 512, 351 ], "score": 1.0, "content": "You are asked to come up with a set of 100 diverse task instructions about text to speech, which is", "type": "text" } ], "index": 16 }, { "bbox": [ 82, 350, 205, 366 ], "spans": [ { "bbox": [ 82, 350, 205, 366 ], "score": 1.0, "content": "about recognizing speech .", "type": "text" } ], "index": 17 } ], "index": 16.5 }, { "type": "text", "bbox": [ 85, 365, 203, 377 ], "lines": [ { "bbox": [ 84, 364, 204, 378 ], "spans": [ { "bbox": [ 84, 364, 204, 378 ], "score": 1.0, "content": "Here are the requirements:", "type": "text" } ], "index": 18 } ], "index": 18 }, { "type": "text", "bbox": [ 84, 379, 513, 404 ], "lines": [ { "bbox": [ 81, 375, 514, 395 ], "spans": [ { "bbox": [ 81, 375, 514, 395 ], "score": 1.0, "content": "1. These instructions should be to instruct someone to recognize the content of the following", "type": "text" } ], "index": 19 }, { "bbox": [ 81, 391, 121, 405 ], "spans": [ { "bbox": [ 81, 391, 121, 405 ], "score": 1.0, "content": "speech.", "type": "text" } ], "index": 20 } ], "index": 19.5 }, { "type": "text", "bbox": [ 86, 406, 399, 418 ], "lines": [ { "bbox": [ 84, 405, 399, 420 ], "spans": [ { "bbox": [ 84, 405, 399, 420 ], "score": 1.0, "content": "2. Try not to repeat the verb for each instruction to maximize diversity.", "type": "text" } ], "index": 21 } ], "index": 21 }, { "type": "text", "bbox": [ 83, 420, 510, 445 ], "lines": [ { "bbox": [ 81, 417, 512, 435 ], "spans": [ { "bbox": [ 81, 417, 512, 435 ], "score": 1.0, "content": "3. The language used for instruction also should be diverse. For example, you should combine", "type": "text" } ], "index": 22 }, { "bbox": [ 82, 431, 257, 447 ], "spans": [ { "bbox": [ 82, 431, 257, 447 ], "score": 1.0, "content": "questions with imperative instructions.", "type": "text" } ], "index": 23 } ], "index": 22.5 }, { "type": "text", "bbox": [ 86, 447, 283, 459 ], "lines": [ { "bbox": [ 84, 446, 284, 459 ], "spans": [ { "bbox": [ 84, 446, 284, 459 ], "score": 1.0, "content": "4. The type of instructions should be diverse.", "type": "text" } ], "index": 24 } ], "index": 24 }, { "type": "text", "bbox": [ 86, 460, 265, 472 ], "lines": [ { "bbox": [ 84, 460, 265, 473 ], "spans": [ { "bbox": [ 84, 460, 265, 473 ], "score": 1.0, "content": "5. The instructions should be in English.", "type": "text" } ], "index": 25 } ], "index": 25 }, { "type": "text", "bbox": [ 85, 474, 510, 498 ], "lines": [ { "bbox": [ 82, 471, 512, 488 ], "spans": [ { "bbox": [ 82, 471, 512, 488 ], "score": 1.0, "content": "6. The instructions should be 1 to 2 sentences long. Either an imperative sentence or a question is", "type": "text" } ], "index": 26 }, { "bbox": [ 82, 486, 132, 499 ], "spans": [ { "bbox": [ 82, 486, 132, 499 ], "score": 1.0, "content": "permitted.", "type": "text" } ], "index": 27 } ], "index": 26.5 }, { "type": "text", "bbox": [ 85, 501, 162, 513 ], "lines": [ { "bbox": [ 83, 499, 164, 514 ], "spans": [ { "bbox": [ 83, 499, 164, 514 ], "score": 1.0, "content": "List of 100 tasks:", "type": "text" } ], "index": 28 } ], "index": 28 } ], "page_idx": 10, "page_size": [ 595, 841 ], "discarded_blocks": [], "para_blocks": [ { "type": "title", "bbox": [ 69, 70, 288, 85 ], "lines": [ { "bbox": [ 67, 69, 289, 88 ], "spans": [ { "bbox": [ 67, 69, 289, 88 ], "score": 1.0, "content": "A Prompts to Generate Task Description", "type": "text" } ], "index": 0 } ], "index": 0 }, { "type": "title", "bbox": [ 85, 121, 110, 132 ], "lines": [ { "bbox": [ 83, 119, 113, 133 ], "spans": [ { "bbox": [ 83, 119, 113, 133 ], "score": 1.0, "content": "ASR:", "type": "text" } ], "index": 1 } ], "index": 1 }, { "type": "text", "bbox": [ 83, 135, 510, 160 ], "lines": [ { "bbox": [ 82, 132, 511, 149 ], "spans": [ { "bbox": [ 82, 132, 511, 149 ], "score": 1.0, "content": "You are asked to come up with a set of 100 diverse task instructions about automatic speech", "type": "text" } ], "index": 2 }, { "bbox": [ 82, 146, 297, 163 ], "spans": [ { "bbox": [ 82, 146, 297, 163 ], "score": 1.0, "content": "recognition, which is about recognizing speech.", "type": "text" } ], "index": 3 } ], "index": 2.5, "bbox_fs": [ 82, 132, 511, 163 ] }, { "type": "text", "bbox": [ 85, 162, 203, 174 ], "lines": [ { "bbox": [ 84, 161, 204, 174 ], "spans": [ { "bbox": [ 84, 161, 204, 174 ], "score": 1.0, "content": "Here are the requirements:", "type": "text" } ], "index": 4 } ], "index": 4, "bbox_fs": [ 84, 161, 204, 174 ] }, { "type": "text", "bbox": [ 82, 176, 510, 201 ], "lines": [ { "bbox": [ 80, 173, 512, 191 ], "spans": [ { "bbox": [ 80, 173, 512, 191 ], "score": 1.0, "content": "1. These instructions should be to instruct someone to recognize the content of the following", "type": "text" } ], "index": 5 }, { "bbox": [ 81, 188, 120, 202 ], "spans": [ { "bbox": [ 81, 188, 120, 202 ], "score": 1.0, "content": "speech.", "type": "text" } ], "index": 6 } ], "index": 5.5, "bbox_fs": [ 80, 173, 512, 202 ] }, { "type": "text", "bbox": [ 84, 202, 401, 214 ], "lines": [ { "bbox": [ 83, 201, 401, 216 ], "spans": [ { "bbox": [ 83, 201, 401, 216 ], "score": 1.0, "content": "2. Try not to repeat the verb for each instruction to maximize diversity.", "type": "text" } ], "index": 7 } ], "index": 7, "bbox_fs": [ 83, 201, 401, 216 ] }, { "type": "text", "bbox": [ 81, 216, 509, 241 ], "lines": [ { "bbox": [ 81, 214, 511, 230 ], "spans": [ { "bbox": [ 81, 214, 511, 230 ], "score": 1.0, "content": "3. The language used for instruction also should be diverse. For example, you should combine", "type": "text" } ], "index": 8 }, { "bbox": [ 83, 229, 257, 244 ], "spans": [ { "bbox": [ 83, 229, 257, 244 ], "score": 1.0, "content": "questions with imperative instructions.", "type": "text" } ], "index": 9 } ], "index": 8.5, "bbox_fs": [ 81, 214, 511, 244 ] }, { "type": "text", "bbox": [ 86, 244, 282, 255 ], "lines": [ { "bbox": [ 84, 243, 284, 256 ], "spans": [ { "bbox": [ 84, 243, 284, 256 ], "score": 1.0, "content": "4. The type of instructions should be diverse.", "type": "text" } ], "index": 10 } ], "index": 10, "bbox_fs": [ 84, 243, 284, 256 ] }, { "type": "text", "bbox": [ 86, 257, 264, 268 ], "lines": [ { "bbox": [ 84, 256, 265, 270 ], "spans": [ { "bbox": [ 84, 256, 265, 270 ], "score": 1.0, "content": "5. The instructions should be in English.", "type": "text" } ], "index": 11 } ], "index": 11, "bbox_fs": [ 84, 256, 265, 270 ] }, { "type": "text", "bbox": [ 84, 271, 511, 295 ], "lines": [ { "bbox": [ 81, 267, 512, 284 ], "spans": [ { "bbox": [ 81, 267, 512, 284 ], "score": 1.0, "content": "6. The instructions should be 1 to 2 sentences long. Either an imperative sentence or a question is", "type": "text" } ], "index": 12 }, { "bbox": [ 81, 283, 133, 296 ], "spans": [ { "bbox": [ 81, 283, 133, 296 ], "score": 1.0, "content": "permitted.", "type": "text" } ], "index": 13 } ], "index": 12.5, "bbox_fs": [ 81, 267, 512, 296 ] }, { "type": "text", "bbox": [ 85, 297, 162, 309 ], "lines": [ { "bbox": [ 83, 295, 164, 310 ], "spans": [ { "bbox": [ 83, 295, 164, 310 ], "score": 1.0, "content": "List of 100 tasks:", "type": "text" } ], "index": 14 } ], "index": 14, "bbox_fs": [ 83, 295, 164, 310 ] }, { "type": "text", "bbox": [ 85, 325, 109, 336 ], "lines": [ { "bbox": [ 83, 322, 112, 337 ], "spans": [ { "bbox": [ 83, 322, 112, 337 ], "score": 1.0, "content": "TTS:", "type": "text" } ], "index": 15 } ], "index": 15, "bbox_fs": [ 83, 322, 112, 337 ] }, { "type": "text", "bbox": [ 81, 338, 511, 363 ], "lines": [ { "bbox": [ 82, 336, 512, 351 ], "spans": [ { "bbox": [ 82, 336, 512, 351 ], "score": 1.0, "content": "You are asked to come up with a set of 100 diverse task instructions about text to speech, which is", "type": "text" } ], "index": 16 }, { "bbox": [ 82, 350, 205, 366 ], "spans": [ { "bbox": [ 82, 350, 205, 366 ], "score": 1.0, "content": "about recognizing speech .", "type": "text" } ], "index": 17 } ], "index": 16.5, "bbox_fs": [ 82, 336, 512, 366 ] }, { "type": "text", "bbox": [ 85, 365, 203, 377 ], "lines": [ { "bbox": [ 84, 364, 204, 378 ], "spans": [ { "bbox": [ 84, 364, 204, 378 ], "score": 1.0, "content": "Here are the requirements:", "type": "text" } ], "index": 18 } ], "index": 18, "bbox_fs": [ 84, 364, 204, 378 ] }, { "type": "text", "bbox": [ 84, 379, 513, 404 ], "lines": [ { "bbox": [ 81, 375, 514, 395 ], "spans": [ { "bbox": [ 81, 375, 514, 395 ], "score": 1.0, "content": "1. These instructions should be to instruct someone to recognize the content of the following", "type": "text" } ], "index": 19 }, { "bbox": [ 81, 391, 121, 405 ], "spans": [ { "bbox": [ 81, 391, 121, 405 ], "score": 1.0, "content": "speech.", "type": "text" } ], "index": 20 } ], "index": 19.5, "bbox_fs": [ 81, 375, 514, 405 ] }, { "type": "text", "bbox": [ 86, 406, 399, 418 ], "lines": [ { "bbox": [ 84, 405, 399, 420 ], "spans": [ { "bbox": [ 84, 405, 399, 420 ], "score": 1.0, "content": "2. Try not to repeat the verb for each instruction to maximize diversity.", "type": "text" } ], "index": 21 } ], "index": 21, "bbox_fs": [ 84, 405, 399, 420 ] }, { "type": "text", "bbox": [ 83, 420, 510, 445 ], "lines": [ { "bbox": [ 81, 417, 512, 435 ], "spans": [ { "bbox": [ 81, 417, 512, 435 ], "score": 1.0, "content": "3. The language used for instruction also should be diverse. For example, you should combine", "type": "text" } ], "index": 22 }, { "bbox": [ 82, 431, 257, 447 ], "spans": [ { "bbox": [ 82, 431, 257, 447 ], "score": 1.0, "content": "questions with imperative instructions.", "type": "text" } ], "index": 23 } ], "index": 22.5, "bbox_fs": [ 81, 417, 512, 447 ] }, { "type": "text", "bbox": [ 86, 447, 283, 459 ], "lines": [ { "bbox": [ 84, 446, 284, 459 ], "spans": [ { "bbox": [ 84, 446, 284, 459 ], "score": 1.0, "content": "4. The type of instructions should be diverse.", "type": "text" } ], "index": 24 } ], "index": 24, "bbox_fs": [ 84, 446, 284, 459 ] }, { "type": "text", "bbox": [ 86, 460, 265, 472 ], "lines": [ { "bbox": [ 84, 460, 265, 473 ], "spans": [ { "bbox": [ 84, 460, 265, 473 ], "score": 1.0, "content": "5. The instructions should be in English.", "type": "text" } ], "index": 25 } ], "index": 25, "bbox_fs": [ 84, 460, 265, 473 ] }, { "type": "text", "bbox": [ 85, 474, 510, 498 ], "lines": [ { "bbox": [ 82, 471, 512, 488 ], "spans": [ { "bbox": [ 82, 471, 512, 488 ], "score": 1.0, "content": "6. The instructions should be 1 to 2 sentences long. Either an imperative sentence or a question is", "type": "text" } ], "index": 26 }, { "bbox": [ 82, 486, 132, 499 ], "spans": [ { "bbox": [ 82, 486, 132, 499 ], "score": 1.0, "content": "permitted.", "type": "text" } ], "index": 27 } ], "index": 26.5, "bbox_fs": [ 82, 471, 512, 499 ] }, { "type": "text", "bbox": [ 85, 501, 162, 513 ], "lines": [ { "bbox": [ 83, 499, 164, 514 ], "spans": [ { "bbox": [ 83, 499, 164, 514 ], "score": 1.0, "content": "List of 100 tasks:", "type": "text" } ], "index": 28 } ], "index": 28, "bbox_fs": [ 83, 499, 164, 514 ] } ] }, { "preproc_blocks": [ { "type": "title", "bbox": [ 69, 70, 243, 84 ], "lines": [ { "bbox": [ 67, 69, 244, 87 ], "spans": [ { "bbox": [ 67, 69, 244, 87 ], "score": 1.0, "content": "B Examples of Task Description", "type": "text" } ], "index": 0 } ], "index": 0 }, { "type": "title", "bbox": [ 85, 121, 110, 132 ], "lines": [ { "bbox": [ 83, 119, 113, 133 ], "spans": [ { "bbox": [ 83, 119, 113, 133 ], "score": 1.0, "content": "ASR:", "type": "text" } ], "index": 1 } ], "index": 1 }, { "type": "text", "bbox": [ 85, 135, 333, 269 ], "lines": [ { "bbox": [ 83, 134, 332, 148 ], "spans": [ { "bbox": [ 83, 134, 332, 148 ], "score": 1.0, "content": "Begin by converting the spoken words into written text.", "type": "text" } ], "index": 2 }, { "bbox": [ 84, 148, 316, 161 ], "spans": [ { "bbox": [ 84, 148, 316, 161 ], "score": 1.0, "content": "Can you transcribe the speech into a written format?", "type": "text" } ], "index": 3 }, { "bbox": [ 84, 162, 303, 174 ], "spans": [ { "bbox": [ 84, 162, 303, 174 ], "score": 1.0, "content": "Focus on translating the audible content into text.", "type": "text" } ], "index": 4 }, { "bbox": [ 84, 176, 300, 189 ], "spans": [ { "bbox": [ 84, 176, 300, 189 ], "score": 1.0, "content": "Transcribe the speech by carefully listening to it.", "type": "text" } ], "index": 5 }, { "bbox": [ 84, 189, 333, 202 ], "spans": [ { "bbox": [ 84, 189, 333, 202 ], "score": 1.0, "content": "Would you kindly write down the content of the speech?", "type": "text" } ], "index": 6 }, { "bbox": [ 85, 203, 321, 215 ], "spans": [ { "bbox": [ 85, 203, 321, 215 ], "score": 1.0, "content": "Analyze the speech and create a written transcription.", "type": "text" } ], "index": 7 }, { "bbox": [ 84, 216, 333, 228 ], "spans": [ { "bbox": [ 84, 216, 333, 228 ], "score": 1.0, "content": "Engage with the speech to produce a text-based version.", "type": "text" } ], "index": 8 }, { "bbox": [ 84, 229, 293, 243 ], "spans": [ { "bbox": [ 84, 229, 293, 243 ], "score": 1.0, "content": "Can you document the speech in written form?", "type": "text" } ], "index": 9 }, { "bbox": [ 84, 243, 300, 257 ], "spans": [ { "bbox": [ 84, 243, 300, 257 ], "score": 1.0, "content": "Transform the spoken words into text accurately.", "type": "text" } ], "index": 10 }, { "bbox": [ 84, 256, 318, 270 ], "spans": [ { "bbox": [ 84, 256, 318, 270 ], "score": 1.0, "content": "How about putting the speech’s content into writing?", "type": "text" } ], "index": 11 } ], "index": 6.5 }, { "type": "text", "bbox": [ 85, 284, 109, 295 ], "lines": [ { "bbox": [ 83, 281, 112, 297 ], "spans": [ { "bbox": [ 83, 281, 112, 297 ], "score": 1.0, "content": "TTS:", "type": "text" } ], "index": 12 } ], "index": 12 }, { "type": "text", "bbox": [ 84, 297, 514, 432 ], "lines": [ { "bbox": [ 84, 296, 278, 309 ], "spans": [ { "bbox": [ 84, 296, 278, 309 ], "score": 1.0, "content": "Can you please read this sentence out loud?", "type": "text" } ], "index": 13 }, { "bbox": [ 83, 309, 355, 325 ], "spans": [ { "bbox": [ 83, 309, 355, 325 ], "score": 1.0, "content": "Recite the following words as if you were speaking normally.", "type": "text" } ], "index": 14 }, { "bbox": [ 84, 324, 320, 338 ], "spans": [ { "bbox": [ 84, 324, 320, 338 ], "score": 1.0, "content": "Project your voice to clearly articulate this statement.", "type": "text" } ], "index": 15 }, { "bbox": [ 84, 337, 365, 352 ], "spans": [ { "bbox": [ 84, 337, 365, 352 ], "score": 1.0, "content": "Would you mind speaking these words as naturally as possible?", "type": "text" } ], "index": 16 }, { "bbox": [ 83, 350, 237, 365 ], "spans": [ { "bbox": [ 83, 350, 237, 365 ], "score": 1.0, "content": "Whisper the given sentence softly.", "type": "text" } ], "index": 17 }, { "bbox": [ 83, 363, 511, 379 ], "spans": [ { "bbox": [ 83, 363, 511, 379 ], "score": 1.0, "content": "Enunciate each word in this sentence with precision. How would you express this sentence in a", "type": "text" } ], "index": 18 }, { "bbox": [ 84, 379, 177, 390 ], "spans": [ { "bbox": [ 84, 379, 177, 390 ], "score": 1.0, "content": "conversational tone?", "type": "text" } ], "index": 19 }, { "bbox": [ 83, 390, 317, 406 ], "spans": [ { "bbox": [ 83, 390, 317, 406 ], "score": 1.0, "content": "Could you please relay the message below verbally?", "type": "text" } ], "index": 20 }, { "bbox": [ 83, 404, 321, 419 ], "spans": [ { "bbox": [ 83, 404, 321, 419 ], "score": 1.0, "content": "Emphasize the key points while reading the sentence.", "type": "text" } ], "index": 21 }, { "bbox": [ 84, 419, 269, 432 ], "spans": [ { "bbox": [ 84, 419, 269, 432 ], "score": 1.0, "content": "Sing the text provided in a melodic voice.", "type": "text" } ], "index": 22 } ], "index": 17.5 } ], "page_idx": 11, "page_size": [ 595, 841 ], "discarded_blocks": [], "para_blocks": [ { "type": "title", "bbox": [ 69, 70, 243, 84 ], "lines": [ { "bbox": [ 67, 69, 244, 87 ], "spans": [ { "bbox": [ 67, 69, 244, 87 ], "score": 1.0, "content": "B Examples of Task Description", "type": "text" } ], "index": 0 } ], "index": 0 }, { "type": "title", "bbox": [ 85, 121, 110, 132 ], "lines": [ { "bbox": [ 83, 119, 113, 133 ], "spans": [ { "bbox": [ 83, 119, 113, 133 ], "score": 1.0, "content": "ASR:", "type": "text" } ], "index": 1 } ], "index": 1 }, { "type": "text", "bbox": [ 85, 135, 333, 269 ], "lines": [ { "bbox": [ 83, 134, 332, 148 ], "spans": [ { "bbox": [ 83, 134, 332, 148 ], "score": 1.0, "content": "Begin by converting the spoken words into written text.", "type": "text" } ], "index": 2 }, { "bbox": [ 84, 148, 316, 161 ], "spans": [ { "bbox": [ 84, 148, 316, 161 ], "score": 1.0, "content": "Can you transcribe the speech into a written format?", "type": "text" } ], "index": 3 }, { "bbox": [ 84, 162, 303, 174 ], "spans": [ { "bbox": [ 84, 162, 303, 174 ], "score": 1.0, "content": "Focus on translating the audible content into text.", "type": "text" } ], "index": 4 }, { "bbox": [ 84, 176, 300, 189 ], "spans": [ { "bbox": [ 84, 176, 300, 189 ], "score": 1.0, "content": "Transcribe the speech by carefully listening to it.", "type": "text" } ], "index": 5 }, { "bbox": [ 84, 189, 333, 202 ], "spans": [ { "bbox": [ 84, 189, 333, 202 ], "score": 1.0, "content": "Would you kindly write down the content of the speech?", "type": "text" } ], "index": 6 }, { "bbox": [ 85, 203, 321, 215 ], "spans": [ { "bbox": [ 85, 203, 321, 215 ], "score": 1.0, "content": "Analyze the speech and create a written transcription.", "type": "text" } ], "index": 7 }, { "bbox": [ 84, 216, 333, 228 ], "spans": [ { "bbox": [ 84, 216, 333, 228 ], "score": 1.0, "content": "Engage with the speech to produce a text-based version.", "type": "text" } ], "index": 8 }, { "bbox": [ 84, 229, 293, 243 ], "spans": [ { "bbox": [ 84, 229, 293, 243 ], "score": 1.0, "content": "Can you document the speech in written form?", "type": "text" } ], "index": 9 }, { "bbox": [ 84, 243, 300, 257 ], "spans": [ { "bbox": [ 84, 243, 300, 257 ], "score": 1.0, "content": "Transform the spoken words into text accurately.", "type": "text" } ], "index": 10 }, { "bbox": [ 84, 256, 318, 270 ], "spans": [ { "bbox": [ 84, 256, 318, 270 ], "score": 1.0, "content": "How about putting the speech’s content into writing?", "type": "text" } ], "index": 11 } ], "index": 6.5, "bbox_fs": [ 83, 134, 333, 270 ] }, { "type": "text", "bbox": [ 85, 284, 109, 295 ], "lines": [ { "bbox": [ 83, 281, 112, 297 ], "spans": [ { "bbox": [ 83, 281, 112, 297 ], "score": 1.0, "content": "TTS:", "type": "text" } ], "index": 12 } ], "index": 12, "bbox_fs": [ 83, 281, 112, 297 ] }, { "type": "list", "bbox": [ 84, 297, 514, 432 ], "lines": [ { "bbox": [ 84, 296, 278, 309 ], "spans": [ { "bbox": [ 84, 296, 278, 309 ], "score": 1.0, "content": "Can you please read this sentence out loud?", "type": "text" } ], "index": 13, "is_list_end_line": true }, { "bbox": [ 83, 309, 355, 325 ], "spans": [ { "bbox": [ 83, 309, 355, 325 ], "score": 1.0, "content": "Recite the following words as if you were speaking normally.", "type": "text" } ], "index": 14, "is_list_start_line": true, "is_list_end_line": true }, { "bbox": [ 84, 324, 320, 338 ], "spans": [ { "bbox": [ 84, 324, 320, 338 ], "score": 1.0, "content": "Project your voice to clearly articulate this statement.", "type": "text" } ], "index": 15, "is_list_start_line": true, "is_list_end_line": true }, { "bbox": [ 84, 337, 365, 352 ], "spans": [ { "bbox": [ 84, 337, 365, 352 ], "score": 1.0, "content": "Would you mind speaking these words as naturally as possible?", "type": "text" } ], "index": 16, "is_list_start_line": true, "is_list_end_line": true }, { "bbox": [ 83, 350, 237, 365 ], "spans": [ { "bbox": [ 83, 350, 237, 365 ], "score": 1.0, "content": "Whisper the given sentence softly.", "type": "text" } ], "index": 17, "is_list_start_line": true, "is_list_end_line": true }, { "bbox": [ 83, 363, 511, 379 ], "spans": [ { "bbox": [ 83, 363, 511, 379 ], "score": 1.0, "content": "Enunciate each word in this sentence with precision. How would you express this sentence in a", "type": "text" } ], "index": 18, "is_list_start_line": true }, { "bbox": [ 84, 379, 177, 390 ], "spans": [ { "bbox": [ 84, 379, 177, 390 ], "score": 1.0, "content": "conversational tone?", "type": "text" } ], "index": 19, "is_list_end_line": true }, { "bbox": [ 83, 390, 317, 406 ], "spans": [ { "bbox": [ 83, 390, 317, 406 ], "score": 1.0, "content": "Could you please relay the message below verbally?", "type": "text" } ], "index": 20, "is_list_start_line": true, "is_list_end_line": true }, { "bbox": [ 83, 404, 321, 419 ], "spans": [ { "bbox": [ 83, 404, 321, 419 ], "score": 1.0, "content": "Emphasize the key points while reading the sentence.", "type": "text" } ], "index": 21, "is_list_start_line": true, "is_list_end_line": true }, { "bbox": [ 84, 419, 269, 432 ], "spans": [ { "bbox": [ 84, 419, 269, 432 ], "score": 1.0, "content": "Sing the text provided in a melodic voice.", "type": "text" } ], "index": 22, "is_list_start_line": true, "is_list_end_line": true } ], "index": 17.5, "bbox_fs": [ 83, 296, 511, 432 ] } ] }, { "preproc_blocks": [ { "type": "title", "bbox": [ 85, 121, 259, 133 ], "lines": [ { "bbox": [ 84, 121, 260, 134 ], "spans": [ { "bbox": [ 84, 121, 260, 134 ], "score": 1.0, "content": "Speech Instruction-Speech Response:", "type": "text" } ], "index": 0 } ], "index": 0 }, { "type": "text", "bbox": [ 85, 135, 511, 189 ], "lines": [ { "bbox": [ 84, 133, 513, 149 ], "spans": [ { "bbox": [ 84, 133, 513, 149 ], "score": 1.0, "content": "[Human]: This is a speech instruction: {SpeechI}. And your response should be speech.", "type": "text" } ], "index": 1 }, { "bbox": [ 84, 147, 511, 161 ], "spans": [ { "bbox": [ 84, 147, 511, 161 ], "score": 1.0, "content": "You can do it step by step. You can first transcribe the instruction and get the text Instruc-", "type": "text" } ], "index": 2 }, { "bbox": [ 84, 160, 512, 176 ], "spans": [ { "bbox": [ 84, 160, 512, 176 ], "score": 1.0, "content": "tion. Then you can think about the instruction and get the text response. Last, you should", "type": "text" } ], "index": 3 }, { "bbox": [ 84, 174, 513, 190 ], "spans": [ { "bbox": [ 84, 174, 513, 190 ], "score": 1.0, "content": "speak the response aloud . [SpeechGPT]: [tq] {TextI}; [ta] {TextR}; [ua] {SpeechR}.", "type": "text" } ], "index": 4 } ], "index": 2.5 }, { "type": "title", "bbox": [ 86, 203, 247, 215 ], "lines": [ { "bbox": [ 84, 201, 249, 216 ], "spans": [ { "bbox": [ 84, 201, 249, 216 ], "score": 1.0, "content": "Speech Instruction-Text Response:", "type": "text" } ], "index": 5 } ], "index": 5 }, { "type": "text", "bbox": [ 85, 216, 510, 268 ], "lines": [ { "bbox": [ 83, 214, 512, 230 ], "spans": [ { "bbox": [ 83, 214, 512, 230 ], "score": 1.0, "content": "[Human]: This is a speech instruction: {SpeechI}. And your response should be text. You can do", "type": "text" } ], "index": 6 }, { "bbox": [ 83, 228, 511, 244 ], "spans": [ { "bbox": [ 83, 228, 511, 244 ], "score": 1.0, "content": "it step by step. You can first transcribe the instruction and get the text instruction. Then you can", "type": "text" } ], "index": 7 }, { "bbox": [ 82, 241, 511, 258 ], "spans": [ { "bbox": [ 82, 241, 511, 258 ], "score": 1.0, "content": "think about the instruction and get the text response. . [SpeechGPT]: [tq] {TextI}; [ta]", "type": "text" } ], "index": 8 }, { "bbox": [ 83, 254, 153, 271 ], "spans": [ { "bbox": [ 83, 254, 153, 271 ], "score": 1.0, "content": "{TextR}.", "type": "text" } ], "index": 9 } ], "index": 7.5 }, { "type": "title", "bbox": [ 86, 284, 246, 296 ], "lines": [ { "bbox": [ 84, 283, 248, 297 ], "spans": [ { "bbox": [ 84, 283, 248, 297 ], "score": 1.0, "content": "Text Instruction-Speech Response:", "type": "text" } ], "index": 10 } ], "index": 10 }, { "type": "text", "bbox": [ 86, 297, 510, 338 ], "lines": [ { "bbox": [ 84, 295, 512, 312 ], "spans": [ { "bbox": [ 84, 295, 245, 312 ], "score": 1.0, "content": "[Human]: This is a text instruction:", "type": "text" }, { "bbox": [ 245, 297, 279, 310 ], "score": 0.52, "content": "\\{ \\mathrm { T e x t } \\}", "type": "inline_equation" }, { "bbox": [ 280, 295, 512, 312 ], "score": 1.0, "content": ". And your response should be speech. You can do it", "type": "text" } ], "index": 11 }, { "bbox": [ 83, 309, 511, 325 ], "spans": [ { "bbox": [ 83, 309, 511, 325 ], "score": 1.0, "content": "step by step. You can think about the instruction and get the text response. Then you should speak", "type": "text" } ], "index": 12 }, { "bbox": [ 84, 324, 437, 338 ], "spans": [ { "bbox": [ 84, 324, 437, 338 ], "score": 1.0, "content": "the response aloud . [SpeechGPT]: [ta] {TextR}; [ua] {SpeechR}.", "type": "text" } ], "index": 13 } ], "index": 12 }, { "type": "title", "bbox": [ 86, 351, 234, 364 ], "lines": [ { "bbox": [ 84, 349, 236, 365 ], "spans": [ { "bbox": [ 84, 349, 236, 365 ], "score": 1.0, "content": "Text Instruction-Text Response:", "type": "text" } ], "index": 14 } ], "index": 14 }, { "type": "text", "bbox": [ 86, 365, 508, 392 ], "lines": [ { "bbox": [ 83, 363, 511, 379 ], "spans": [ { "bbox": [ 83, 363, 511, 379 ], "score": 1.0, "content": "[Human]: This is a text instruction: {TextI}. And your response should be text. You can think", "type": "text" } ], "index": 15 }, { "bbox": [ 83, 375, 450, 393 ], "spans": [ { "bbox": [ 83, 375, 450, 393 ], "score": 1.0, "content": "about the instruction and get the text response. [SpeechGPT]: [ta] {TextR}.", "type": "text" } ], "index": 16 } ], "index": 15.5 }, { "type": "title", "bbox": [ 70, 409, 182, 423 ], "lines": [ { "bbox": [ 67, 408, 185, 428 ], "spans": [ { "bbox": [ 67, 408, 185, 428 ], "score": 1.0, "content": "D Hyperparameters", "type": "text" } ], "index": 17 } ], "index": 17 }, { "type": "table", "bbox": [ 155, 451, 439, 588 ], "blocks": [ { "type": "table_body", "bbox": [ 155, 451, 439, 588 ], "group_id": 0, "lines": [ { "bbox": [ 155, 451, 439, 588 ], "spans": [ { "bbox": [ 155, 451, 439, 588 ], "score": 0.981, "html": "
Stage 1Stage 2Stage 3
Batch size7681536128
Peak learning rate2e-42e-42e-4
Max length10245121024
Training steps90040004200
LoRA rank--8
LoRA alpha--16
Trainable parameters13B13B6M
Training device96 × A10096 × A1008 × A100
", "type": "table", "image_path": "29ebe76d95a7e8df91127675821b59b7dfd208841d7b69c05c18697836b3f716.jpg" } ] } ], "index": 19, "virtual_lines": [ { "bbox": [ 155, 451, 439, 496.6666666666667 ], "spans": [], "index": 18 }, { "bbox": [ 155, 496.6666666666667, 439, 542.3333333333334 ], "spans": [], "index": 19 }, { "bbox": [ 155, 542.3333333333334, 439, 588.0 ], "spans": [], "index": 20 } ] }, { "type": "table_caption", "bbox": [ 203, 596, 392, 609 ], "group_id": 0, "lines": [ { "bbox": [ 201, 594, 393, 612 ], "spans": [ { "bbox": [ 201, 594, 393, 612 ], "score": 1.0, "content": "Table 3: SpeechGPT training hyperparameters.", "type": "text" } ], "index": 21 } ], "index": 21 } ], "index": 20.0 } ], "page_idx": 12, "page_size": [ 595, 841 ], "discarded_blocks": [ { "type": "discarded", "bbox": [ 69, 70, 308, 84 ], "lines": [ { "bbox": [ 68, 69, 308, 87 ], "spans": [ { "bbox": [ 68, 69, 308, 87 ], "score": 1.0, "content": "C Chain-of-Modality Instructions Templates", "type": "text" } ] } ] } ], "para_blocks": [ { "type": "title", "bbox": [ 85, 121, 259, 133 ], "lines": [ { "bbox": [ 84, 121, 260, 134 ], "spans": [ { "bbox": [ 84, 121, 260, 134 ], "score": 1.0, "content": "Speech Instruction-Speech Response:", "type": "text" } ], "index": 0 } ], "index": 0 }, { "type": "text", "bbox": [ 85, 135, 511, 189 ], "lines": [ { "bbox": [ 84, 133, 513, 149 ], "spans": [ { "bbox": [ 84, 133, 513, 149 ], "score": 1.0, "content": "[Human]: This is a speech instruction: {SpeechI}. And your response should be speech.", "type": "text" } ], "index": 1 }, { "bbox": [ 84, 147, 511, 161 ], "spans": [ { "bbox": [ 84, 147, 511, 161 ], "score": 1.0, "content": "You can do it step by step. You can first transcribe the instruction and get the text Instruc-", "type": "text" } ], "index": 2 }, { "bbox": [ 84, 160, 512, 176 ], "spans": [ { "bbox": [ 84, 160, 512, 176 ], "score": 1.0, "content": "tion. Then you can think about the instruction and get the text response. Last, you should", "type": "text" } ], "index": 3 }, { "bbox": [ 84, 174, 513, 190 ], "spans": [ { "bbox": [ 84, 174, 513, 190 ], "score": 1.0, "content": "speak the response aloud . [SpeechGPT]: [tq] {TextI}; [ta] {TextR}; [ua] {SpeechR}.", "type": "text" } ], "index": 4 } ], "index": 2.5, "bbox_fs": [ 84, 133, 513, 190 ] }, { "type": "title", "bbox": [ 86, 203, 247, 215 ], "lines": [ { "bbox": [ 84, 201, 249, 216 ], "spans": [ { "bbox": [ 84, 201, 249, 216 ], "score": 1.0, "content": "Speech Instruction-Text Response:", "type": "text" } ], "index": 5 } ], "index": 5 }, { "type": "text", "bbox": [ 85, 216, 510, 268 ], "lines": [ { "bbox": [ 83, 214, 512, 230 ], "spans": [ { "bbox": [ 83, 214, 512, 230 ], "score": 1.0, "content": "[Human]: This is a speech instruction: {SpeechI}. And your response should be text. You can do", "type": "text" } ], "index": 6 }, { "bbox": [ 83, 228, 511, 244 ], "spans": [ { "bbox": [ 83, 228, 511, 244 ], "score": 1.0, "content": "it step by step. You can first transcribe the instruction and get the text instruction. Then you can", "type": "text" } ], "index": 7 }, { "bbox": [ 82, 241, 511, 258 ], "spans": [ { "bbox": [ 82, 241, 511, 258 ], "score": 1.0, "content": "think about the instruction and get the text response. . [SpeechGPT]: [tq] {TextI}; [ta]", "type": "text" } ], "index": 8 }, { "bbox": [ 83, 254, 153, 271 ], "spans": [ { "bbox": [ 83, 254, 153, 271 ], "score": 1.0, "content": "{TextR}.", "type": "text" } ], "index": 9 } ], "index": 7.5, "bbox_fs": [ 82, 214, 512, 271 ] }, { "type": "title", "bbox": [ 86, 284, 246, 296 ], "lines": [ { "bbox": [ 84, 283, 248, 297 ], "spans": [ { "bbox": [ 84, 283, 248, 297 ], "score": 1.0, "content": "Text Instruction-Speech Response:", "type": "text" } ], "index": 10 } ], "index": 10 }, { "type": "text", "bbox": [ 86, 297, 510, 338 ], "lines": [ { "bbox": [ 84, 295, 512, 312 ], "spans": [ { "bbox": [ 84, 295, 245, 312 ], "score": 1.0, "content": "[Human]: This is a text instruction:", "type": "text" }, { "bbox": [ 245, 297, 279, 310 ], "score": 0.52, "content": "\\{ \\mathrm { T e x t } \\}", "type": "inline_equation" }, { "bbox": [ 280, 295, 512, 312 ], "score": 1.0, "content": ". And your response should be speech. You can do it", "type": "text" } ], "index": 11 }, { "bbox": [ 83, 309, 511, 325 ], "spans": [ { "bbox": [ 83, 309, 511, 325 ], "score": 1.0, "content": "step by step. You can think about the instruction and get the text response. Then you should speak", "type": "text" } ], "index": 12 }, { "bbox": [ 84, 324, 437, 338 ], "spans": [ { "bbox": [ 84, 324, 437, 338 ], "score": 1.0, "content": "the response aloud . [SpeechGPT]: [ta] {TextR}; [ua] {SpeechR}.", "type": "text" } ], "index": 13 } ], "index": 12, "bbox_fs": [ 83, 295, 512, 338 ] }, { "type": "title", "bbox": [ 86, 351, 234, 364 ], "lines": [ { "bbox": [ 84, 349, 236, 365 ], "spans": [ { "bbox": [ 84, 349, 236, 365 ], "score": 1.0, "content": "Text Instruction-Text Response:", "type": "text" } ], "index": 14 } ], "index": 14 }, { "type": "text", "bbox": [ 86, 365, 508, 392 ], "lines": [ { "bbox": [ 83, 363, 511, 379 ], "spans": [ { "bbox": [ 83, 363, 511, 379 ], "score": 1.0, "content": "[Human]: This is a text instruction: {TextI}. And your response should be text. You can think", "type": "text" } ], "index": 15 }, { "bbox": [ 83, 375, 450, 393 ], "spans": [ { "bbox": [ 83, 375, 450, 393 ], "score": 1.0, "content": "about the instruction and get the text response. [SpeechGPT]: [ta] {TextR}.", "type": "text" } ], "index": 16 } ], "index": 15.5, "bbox_fs": [ 83, 363, 511, 393 ] }, { "type": "title", "bbox": [ 70, 409, 182, 423 ], "lines": [ { "bbox": [ 67, 408, 185, 428 ], "spans": [ { "bbox": [ 67, 408, 185, 428 ], "score": 1.0, "content": "D Hyperparameters", "type": "text" } ], "index": 17 } ], "index": 17 }, { "type": "table", "bbox": [ 155, 451, 439, 588 ], "blocks": [ { "type": "table_body", "bbox": [ 155, 451, 439, 588 ], "group_id": 0, "lines": [ { "bbox": [ 155, 451, 439, 588 ], "spans": [ { "bbox": [ 155, 451, 439, 588 ], "score": 0.981, "html": "
Stage 1Stage 2Stage 3
Batch size7681536128
Peak learning rate2e-42e-42e-4
Max length10245121024
Training steps90040004200
LoRA rank--8
LoRA alpha--16
Trainable parameters13B13B6M
Training device96 × A10096 × A1008 × A100
", "type": "table", "image_path": "29ebe76d95a7e8df91127675821b59b7dfd208841d7b69c05c18697836b3f716.jpg" } ] } ], "index": 19, "virtual_lines": [ { "bbox": [ 155, 451, 439, 496.6666666666667 ], "spans": [], "index": 18 }, { "bbox": [ 155, 496.6666666666667, 439, 542.3333333333334 ], "spans": [], "index": 19 }, { "bbox": [ 155, 542.3333333333334, 439, 588.0 ], "spans": [], "index": 20 } ] }, { "type": "table_caption", "bbox": [ 203, 596, 392, 609 ], "group_id": 0, "lines": [ { "bbox": [ 201, 594, 393, 612 ], "spans": [ { "bbox": [ 201, 594, 393, 612 ], "score": 1.0, "content": "Table 3: SpeechGPT training hyperparameters.", "type": "text" } ], "index": 21 } ], "index": 21 } ], "index": 20.0 } ] }, { "preproc_blocks": [ { "type": "title", "bbox": [ 70, 70, 252, 85 ], "lines": [ { "bbox": [ 67, 69, 253, 87 ], "spans": [ { "bbox": [ 67, 69, 253, 87 ], "score": 1.0, "content": "E Standard Prompting Templates", "type": "text" } ], "index": 0 } ], "index": 0 }, { "type": "text", "bbox": [ 84, 119, 514, 161 ], "lines": [ { "bbox": [ 83, 118, 262, 136 ], "spans": [ { "bbox": [ 83, 118, 262, 136 ], "score": 1.0, "content": "Speech Instruction-Speech Response:", "type": "text" } ], "index": 1 }, { "bbox": [ 83, 132, 513, 150 ], "spans": [ { "bbox": [ 83, 132, 513, 150 ], "score": 1.0, "content": "[Human]: This is a speech instruction: {SpeechI}. And your response should be speech .", "type": "text" } ], "index": 2 }, { "bbox": [ 84, 147, 257, 162 ], "spans": [ { "bbox": [ 84, 147, 257, 162 ], "score": 1.0, "content": "[SpeechGPT]: [ua] {SpeechR}.", "type": "text" } ], "index": 3 } ], "index": 2 }, { "type": "text", "bbox": [ 83, 175, 513, 215 ], "lines": [ { "bbox": [ 83, 173, 249, 189 ], "spans": [ { "bbox": [ 83, 173, 249, 189 ], "score": 1.0, "content": "Speech Instruction-Text Response:", "type": "text" } ], "index": 4 }, { "bbox": [ 83, 187, 513, 203 ], "spans": [ { "bbox": [ 83, 187, 513, 203 ], "score": 1.0, "content": "[Human]: This is a speech instruction: {SpeechI}. And your response should be text. .", "type": "text" } ], "index": 5 }, { "bbox": [ 83, 201, 242, 216 ], "spans": [ { "bbox": [ 83, 201, 242, 216 ], "score": 1.0, "content": "[SpeechGPT]: [ta] {TextR}.", "type": "text" } ], "index": 6 } ], "index": 5 }, { "type": "text", "bbox": [ 82, 229, 513, 269 ], "lines": [ { "bbox": [ 82, 227, 249, 243 ], "spans": [ { "bbox": [ 82, 227, 249, 243 ], "score": 1.0, "content": "Text Instruction-Speech Response:", "type": "text" } ], "index": 7 }, { "bbox": [ 83, 241, 512, 258 ], "spans": [ { "bbox": [ 83, 241, 512, 258 ], "score": 1.0, "content": "[Human]: This is a text instruction: {TextI}. And your response should be speech .", "type": "text" } ], "index": 8 }, { "bbox": [ 83, 255, 257, 270 ], "spans": [ { "bbox": [ 83, 255, 257, 270 ], "score": 1.0, "content": "[SpeechGPT]: [ua] {SpeechR}.", "type": "text" } ], "index": 9 } ], "index": 8 }, { "type": "text", "bbox": [ 84, 297, 511, 323 ], "lines": [ { "bbox": [ 83, 295, 512, 312 ], "spans": [ { "bbox": [ 83, 295, 512, 312 ], "score": 1.0, "content": "[Human]: This is a text instruction: {TextI}. And your response should be text. [SpeechGPT]:", "type": "text" } ], "index": 10 }, { "bbox": [ 83, 310, 173, 324 ], "spans": [ { "bbox": [ 83, 310, 105, 324 ], "score": 1.0, "content": "[ta]", "type": "text" }, { "bbox": [ 106, 311, 170, 323 ], "score": 0.28, "content": "\\{ { \\mathrm { T e x t R } } \\} { < } { \\mathrm { e o a } } { > }", "type": "inline_equation" }, { "bbox": [ 171, 310, 173, 324 ], "score": 1.0, "content": ".", "type": "text" } ], "index": 11 } ], "index": 10.5 } ], "page_idx": 13, "page_size": [ 595, 841 ], "discarded_blocks": [], "para_blocks": [ { "type": "title", "bbox": [ 70, 70, 252, 85 ], "lines": [ { "bbox": [ 67, 69, 253, 87 ], "spans": [ { "bbox": [ 67, 69, 253, 87 ], "score": 1.0, "content": "E Standard Prompting Templates", "type": "text" } ], "index": 0 } ], "index": 0 }, { "type": "list", "bbox": [ 84, 119, 514, 161 ], "lines": [ { "bbox": [ 83, 118, 262, 136 ], "spans": [ { "bbox": [ 83, 118, 262, 136 ], "score": 1.0, "content": "Speech Instruction-Speech Response:", "type": "text" } ], "index": 1, "is_list_end_line": true }, { "bbox": [ 83, 132, 513, 150 ], "spans": [ { "bbox": [ 83, 132, 513, 150 ], "score": 1.0, "content": "[Human]: This is a speech instruction: {SpeechI}. And your response should be speech .", "type": "text" } ], "index": 2, "is_list_start_line": true }, { "bbox": [ 84, 147, 257, 162 ], "spans": [ { "bbox": [ 84, 147, 257, 162 ], "score": 1.0, "content": "[SpeechGPT]: [ua] {SpeechR}.", "type": "text" } ], "index": 3, "is_list_end_line": true }, { "bbox": [ 83, 173, 249, 189 ], "spans": [ { "bbox": [ 83, 173, 249, 189 ], "score": 1.0, "content": "Speech Instruction-Text Response:", "type": "text" } ], "index": 4, "is_list_end_line": true }, { "bbox": [ 83, 187, 513, 203 ], "spans": [ { "bbox": [ 83, 187, 513, 203 ], "score": 1.0, "content": "[Human]: This is a speech instruction: {SpeechI}. And your response should be text. .", "type": "text" } ], "index": 5, "is_list_start_line": true }, { "bbox": [ 83, 201, 242, 216 ], "spans": [ { "bbox": [ 83, 201, 242, 216 ], "score": 1.0, "content": "[SpeechGPT]: [ta] {TextR}.", "type": "text" } ], "index": 6, "is_list_end_line": true }, { "bbox": [ 82, 227, 249, 243 ], "spans": [ { "bbox": [ 82, 227, 249, 243 ], "score": 1.0, "content": "Text Instruction-Speech Response:", "type": "text" } ], "index": 7, "is_list_end_line": true }, { "bbox": [ 83, 241, 512, 258 ], "spans": [ { "bbox": [ 83, 241, 512, 258 ], "score": 1.0, "content": "[Human]: This is a text instruction: {TextI}. And your response should be speech .", "type": "text" } ], "index": 8, "is_list_start_line": true }, { "bbox": [ 83, 255, 257, 270 ], "spans": [ { "bbox": [ 83, 255, 257, 270 ], "score": 1.0, "content": "[SpeechGPT]: [ua] {SpeechR}.", "type": "text" } ], "index": 9, "is_list_end_line": true } ], "index": 2, "bbox_fs": [ 83, 118, 513, 162 ] }, { "type": "list", "bbox": [ 83, 175, 513, 215 ], "lines": [], "index": 5, "bbox_fs": [ 83, 173, 513, 216 ], "lines_deleted": true }, { "type": "list", "bbox": [ 82, 229, 513, 269 ], "lines": [], "index": 8, "bbox_fs": [ 82, 227, 512, 270 ], "lines_deleted": true }, { "type": "text", "bbox": [ 84, 297, 511, 323 ], "lines": [ { "bbox": [ 83, 295, 512, 312 ], "spans": [ { "bbox": [ 83, 295, 512, 312 ], "score": 1.0, "content": "[Human]: This is a text instruction: {TextI}. And your response should be text. [SpeechGPT]:", "type": "text" } ], "index": 10 }, { "bbox": [ 83, 310, 173, 324 ], "spans": [ { "bbox": [ 83, 310, 105, 324 ], "score": 1.0, "content": "[ta]", "type": "text" }, { "bbox": [ 106, 311, 170, 323 ], "score": 0.28, "content": "\\{ { \\mathrm { T e x t R } } \\} { < } { \\mathrm { e o a } } { > }", "type": "inline_equation" }, { "bbox": [ 171, 310, 173, 324 ], "score": 1.0, "content": ".", "type": "text" } ], "index": 11 } ], "index": 10.5, "bbox_fs": [ 83, 295, 512, 324 ] } ] }, { "preproc_blocks": [ { "type": "title", "bbox": [ 69, 70, 273, 84 ], "lines": [ { "bbox": [ 66, 69, 273, 87 ], "spans": [ { "bbox": [ 66, 69, 273, 87 ], "score": 1.0, "content": "F ChatGPT Score Evaluation Prompt", "type": "text" } ], "index": 0 } ], "index": 0 }, { "type": "text", "bbox": [ 84, 120, 511, 146 ], "lines": [ { "bbox": [ 82, 118, 513, 136 ], "spans": [ { "bbox": [ 82, 118, 513, 136 ], "score": 1.0, "content": "You are evaluating a response that has been submitted for an instruction, using a specific set of", "type": "text" } ], "index": 1 }, { "bbox": [ 83, 133, 214, 147 ], "spans": [ { "bbox": [ 83, 133, 214, 147 ], "score": 1.0, "content": "standards. Below is the data:", "type": "text" } ], "index": 2 } ], "index": 1.5 }, { "type": "text", "bbox": [ 85, 162, 101, 169 ], "lines": [ { "bbox": [ 84, 161, 102, 170 ], "spans": [ { "bbox": [ 84, 161, 102, 170 ], "score": 0.996, "content": "***", "type": "text" } ], "index": 3 } ], "index": 3 }, { "type": "text", "bbox": [ 85, 175, 164, 197 ], "lines": [ { "bbox": [ 84, 174, 166, 188 ], "spans": [ { "bbox": [ 84, 174, 166, 188 ], "score": 1.0, "content": "[Instruction]: inst", "type": "text" } ], "index": 4 }, { "bbox": [ 84, 189, 102, 197 ], "spans": [ { "bbox": [ 84, 189, 102, 197 ], "score": 0.996, "content": "***", "type": "text" } ], "index": 5 } ], "index": 4.5 }, { "type": "text", "bbox": [ 85, 203, 161, 224 ], "lines": [ { "bbox": [ 83, 201, 163, 218 ], "spans": [ { "bbox": [ 83, 201, 163, 218 ], "score": 1.0, "content": "[Response]: resp", "type": "text" } ], "index": 6 }, { "bbox": [ 84, 216, 96, 223 ], "spans": [ { "bbox": [ 84, 216, 96, 223 ], "score": 0.989, "content": "**", "type": "text" } ], "index": 7 } ], "index": 6.5 }, { "type": "text", "bbox": [ 86, 230, 192, 242 ], "lines": [ { "bbox": [ 84, 228, 194, 243 ], "spans": [ { "bbox": [ 84, 228, 194, 243 ], "score": 1.0, "content": "[Criterion]: helpfulness:", "type": "text" } ], "index": 8 } ], "index": 8 }, { "type": "text", "bbox": [ 82, 243, 510, 268 ], "lines": [ { "bbox": [ 81, 240, 512, 258 ], "spans": [ { "bbox": [ 81, 240, 512, 258 ], "score": 1.0, "content": "\"1\": \"Not helpful - The generated text is completely irrelevant, unclear, or incomplete. It does not", "type": "text" } ], "index": 9 }, { "bbox": [ 82, 255, 280, 270 ], "spans": [ { "bbox": [ 82, 255, 280, 270 ], "score": 1.0, "content": "provide any useful information to the user.\"", "type": "text" } ], "index": 10 } ], "index": 9.5 }, { "type": "text", "bbox": [ 85, 270, 511, 309 ], "lines": [ { "bbox": [ 83, 267, 512, 285 ], "spans": [ { "bbox": [ 83, 267, 512, 285 ], "score": 1.0, "content": "\"2\": \"Somewhat helpful - The generated text has some relevance to the user’s question, but it may", "type": "text" } ], "index": 11 }, { "bbox": [ 83, 282, 512, 298 ], "spans": [ { "bbox": [ 83, 282, 512, 298 ], "score": 1.0, "content": "be unclear or incomplete. It provides only partial information, or the information provided may", "type": "text" } ], "index": 12 }, { "bbox": [ 83, 296, 238, 310 ], "spans": [ { "bbox": [ 83, 296, 238, 310 ], "score": 1.0, "content": "not be useful for the user’s needs.\"", "type": "text" } ], "index": 13 } ], "index": 12 }, { "type": "text", "bbox": [ 85, 311, 510, 349 ], "lines": [ { "bbox": [ 83, 309, 512, 325 ], "spans": [ { "bbox": [ 83, 309, 512, 325 ], "score": 1.0, "content": "\"3\": \"Moderately helpful - The generated text is relevant to the user’s question, and it provides a", "type": "text" } ], "index": 14 }, { "bbox": [ 84, 323, 511, 338 ], "spans": [ { "bbox": [ 84, 323, 511, 338 ], "score": 1.0, "content": "clear and complete answer. However, it may lack detail or explanation that would be helpful for", "type": "text" } ], "index": 15 }, { "bbox": [ 82, 336, 130, 351 ], "spans": [ { "bbox": [ 82, 336, 130, 351 ], "score": 1.0, "content": "the user.\"", "type": "text" } ], "index": 16 } ], "index": 15 }, { "type": "text", "bbox": [ 85, 352, 510, 404 ], "lines": [ { "bbox": [ 83, 349, 512, 366 ], "spans": [ { "bbox": [ 83, 349, 512, 366 ], "score": 1.0, "content": "\"4\": \"Helpful - The generated text is quite relevant to the user’s question, and it provides a clear,", "type": "text" } ], "index": 17 }, { "bbox": [ 84, 363, 510, 378 ], "spans": [ { "bbox": [ 84, 363, 510, 378 ], "score": 1.0, "content": "complete, and detailed answer. It offers additional information or explanations that are useful", "type": "text" } ], "index": 18 }, { "bbox": [ 83, 376, 511, 393 ], "spans": [ { "bbox": [ 83, 376, 511, 393 ], "score": 1.0, "content": "for the user. However, some of the points of the response are somewhat repetitive or could be", "type": "text" } ], "index": 19 }, { "bbox": [ 83, 392, 276, 405 ], "spans": [ { "bbox": [ 83, 392, 276, 405 ], "score": 1.0, "content": "combined for greater clarity and concision\"", "type": "text" } ], "index": 20 } ], "index": 18.5 }, { "type": "text", "bbox": [ 85, 406, 510, 472 ], "lines": [ { "bbox": [ 83, 403, 512, 420 ], "spans": [ { "bbox": [ 83, 403, 512, 420 ], "score": 1.0, "content": "\"5\": \"Very helpful - The generated text is highly relevant to the user’s question, and it provides a", "type": "text" } ], "index": 21 }, { "bbox": [ 83, 418, 511, 433 ], "spans": [ { "bbox": [ 83, 418, 511, 433 ], "score": 1.0, "content": "clear, complete, and detailed answer. It offers additional information, explanations, or analogies", "type": "text" } ], "index": 22 }, { "bbox": [ 83, 432, 511, 446 ], "spans": [ { "bbox": [ 83, 432, 511, 446 ], "score": 1.0, "content": "that are not only useful but also insightful and valuable to the user. However, the structured of the", "type": "text" } ], "index": 23 }, { "bbox": [ 83, 445, 511, 462 ], "spans": [ { "bbox": [ 83, 445, 511, 462 ], "score": 1.0, "content": "response is not well-organized and there is no clear progression or logical sequence of different", "type": "text" } ], "index": 24 }, { "bbox": [ 82, 459, 190, 474 ], "spans": [ { "bbox": [ 82, 459, 190, 474 ], "score": 1.0, "content": "points in the response.\"", "type": "text" } ], "index": 25 } ], "index": 23 }, { "type": "text", "bbox": [ 85, 474, 101, 480 ], "lines": [ { "bbox": [ 84, 472, 103, 482 ], "spans": [ { "bbox": [ 84, 472, 103, 482 ], "score": 0.996, "content": "***", "type": "text" } ], "index": 26 } ], "index": 26 }, { "type": "title", "bbox": [ 86, 487, 147, 498 ], "lines": [ { "bbox": [ 84, 486, 149, 500 ], "spans": [ { "bbox": [ 84, 486, 149, 500 ], "score": 1.0, "content": "[END DATA]", "type": "text" } ], "index": 27 } ], "index": 27 }, { "type": "text", "bbox": [ 86, 501, 510, 525 ], "lines": [ { "bbox": [ 83, 498, 511, 516 ], "spans": [ { "bbox": [ 83, 498, 511, 516 ], "score": 1.0, "content": "Does the response meet the criterion? You should only write out your score in this format: \"My", "type": "text" } ], "index": 28 }, { "bbox": [ 84, 514, 123, 525 ], "spans": [ { "bbox": [ 84, 514, 123, 525 ], "score": 1.0, "content": "score is:", "type": "text" } ], "index": 29 } ], "index": 28.5 } ], "page_idx": 14, "page_size": [ 595, 841 ], "discarded_blocks": [], "para_blocks": [ { "type": "title", "bbox": [ 69, 70, 273, 84 ], "lines": [ { "bbox": [ 66, 69, 273, 87 ], "spans": [ { "bbox": [ 66, 69, 273, 87 ], "score": 1.0, "content": "F ChatGPT Score Evaluation Prompt", "type": "text" } ], "index": 0 } ], "index": 0 }, { "type": "text", "bbox": [ 84, 120, 511, 146 ], "lines": [ { "bbox": [ 82, 118, 513, 136 ], "spans": [ { "bbox": [ 82, 118, 513, 136 ], "score": 1.0, "content": "You are evaluating a response that has been submitted for an instruction, using a specific set of", "type": "text" } ], "index": 1 }, { "bbox": [ 83, 133, 214, 147 ], "spans": [ { "bbox": [ 83, 133, 214, 147 ], "score": 1.0, "content": "standards. Below is the data:", "type": "text" } ], "index": 2 } ], "index": 1.5, "bbox_fs": [ 82, 118, 513, 147 ] }, { "type": "text", "bbox": [ 85, 162, 101, 169 ], "lines": [ { "bbox": [ 84, 161, 102, 170 ], "spans": [ { "bbox": [ 84, 161, 102, 170 ], "score": 0.996, "content": "***", "type": "text" } ], "index": 3 } ], "index": 3, "bbox_fs": [ 84, 161, 102, 170 ] }, { "type": "text", "bbox": [ 85, 175, 164, 197 ], "lines": [ { "bbox": [ 84, 174, 166, 188 ], "spans": [ { "bbox": [ 84, 174, 166, 188 ], "score": 1.0, "content": "[Instruction]: inst", "type": "text" } ], "index": 4 }, { "bbox": [ 84, 189, 102, 197 ], "spans": [ { "bbox": [ 84, 189, 102, 197 ], "score": 0.996, "content": "***", "type": "text" } ], "index": 5 } ], "index": 4.5, "bbox_fs": [ 84, 174, 166, 197 ] }, { "type": "text", "bbox": [ 85, 203, 161, 224 ], "lines": [ { "bbox": [ 83, 201, 163, 218 ], "spans": [ { "bbox": [ 83, 201, 163, 218 ], "score": 1.0, "content": "[Response]: resp", "type": "text" } ], "index": 6 }, { "bbox": [ 84, 216, 96, 223 ], "spans": [ { "bbox": [ 84, 216, 96, 223 ], "score": 0.989, "content": "**", "type": "text" } ], "index": 7 } ], "index": 6.5, "bbox_fs": [ 83, 201, 163, 223 ] }, { "type": "text", "bbox": [ 86, 230, 192, 242 ], "lines": [ { "bbox": [ 84, 228, 194, 243 ], "spans": [ { "bbox": [ 84, 228, 194, 243 ], "score": 1.0, "content": "[Criterion]: helpfulness:", "type": "text" } ], "index": 8 } ], "index": 8, "bbox_fs": [ 84, 228, 194, 243 ] }, { "type": "text", "bbox": [ 82, 243, 510, 268 ], "lines": [ { "bbox": [ 81, 240, 512, 258 ], "spans": [ { "bbox": [ 81, 240, 512, 258 ], "score": 1.0, "content": "\"1\": \"Not helpful - The generated text is completely irrelevant, unclear, or incomplete. It does not", "type": "text" } ], "index": 9 }, { "bbox": [ 82, 255, 280, 270 ], "spans": [ { "bbox": [ 82, 255, 280, 270 ], "score": 1.0, "content": "provide any useful information to the user.\"", "type": "text" } ], "index": 10 } ], "index": 9.5, "bbox_fs": [ 81, 240, 512, 270 ] }, { "type": "text", "bbox": [ 85, 270, 511, 309 ], "lines": [ { "bbox": [ 83, 267, 512, 285 ], "spans": [ { "bbox": [ 83, 267, 512, 285 ], "score": 1.0, "content": "\"2\": \"Somewhat helpful - The generated text has some relevance to the user’s question, but it may", "type": "text" } ], "index": 11 }, { "bbox": [ 83, 282, 512, 298 ], "spans": [ { "bbox": [ 83, 282, 512, 298 ], "score": 1.0, "content": "be unclear or incomplete. It provides only partial information, or the information provided may", "type": "text" } ], "index": 12 }, { "bbox": [ 83, 296, 238, 310 ], "spans": [ { "bbox": [ 83, 296, 238, 310 ], "score": 1.0, "content": "not be useful for the user’s needs.\"", "type": "text" } ], "index": 13 } ], "index": 12, "bbox_fs": [ 83, 267, 512, 310 ] }, { "type": "text", "bbox": [ 85, 311, 510, 349 ], "lines": [ { "bbox": [ 83, 309, 512, 325 ], "spans": [ { "bbox": [ 83, 309, 512, 325 ], "score": 1.0, "content": "\"3\": \"Moderately helpful - The generated text is relevant to the user’s question, and it provides a", "type": "text" } ], "index": 14 }, { "bbox": [ 84, 323, 511, 338 ], "spans": [ { "bbox": [ 84, 323, 511, 338 ], "score": 1.0, "content": "clear and complete answer. However, it may lack detail or explanation that would be helpful for", "type": "text" } ], "index": 15 }, { "bbox": [ 82, 336, 130, 351 ], "spans": [ { "bbox": [ 82, 336, 130, 351 ], "score": 1.0, "content": "the user.\"", "type": "text" } ], "index": 16 } ], "index": 15, "bbox_fs": [ 82, 309, 512, 351 ] }, { "type": "text", "bbox": [ 85, 352, 510, 404 ], "lines": [ { "bbox": [ 83, 349, 512, 366 ], "spans": [ { "bbox": [ 83, 349, 512, 366 ], "score": 1.0, "content": "\"4\": \"Helpful - The generated text is quite relevant to the user’s question, and it provides a clear,", "type": "text" } ], "index": 17 }, { "bbox": [ 84, 363, 510, 378 ], "spans": [ { "bbox": [ 84, 363, 510, 378 ], "score": 1.0, "content": "complete, and detailed answer. It offers additional information or explanations that are useful", "type": "text" } ], "index": 18 }, { "bbox": [ 83, 376, 511, 393 ], "spans": [ { "bbox": [ 83, 376, 511, 393 ], "score": 1.0, "content": "for the user. However, some of the points of the response are somewhat repetitive or could be", "type": "text" } ], "index": 19 }, { "bbox": [ 83, 392, 276, 405 ], "spans": [ { "bbox": [ 83, 392, 276, 405 ], "score": 1.0, "content": "combined for greater clarity and concision\"", "type": "text" } ], "index": 20 } ], "index": 18.5, "bbox_fs": [ 83, 349, 512, 405 ] }, { "type": "text", "bbox": [ 85, 406, 510, 472 ], "lines": [ { "bbox": [ 83, 403, 512, 420 ], "spans": [ { "bbox": [ 83, 403, 512, 420 ], "score": 1.0, "content": "\"5\": \"Very helpful - The generated text is highly relevant to the user’s question, and it provides a", "type": "text" } ], "index": 21 }, { "bbox": [ 83, 418, 511, 433 ], "spans": [ { "bbox": [ 83, 418, 511, 433 ], "score": 1.0, "content": "clear, complete, and detailed answer. It offers additional information, explanations, or analogies", "type": "text" } ], "index": 22 }, { "bbox": [ 83, 432, 511, 446 ], "spans": [ { "bbox": [ 83, 432, 511, 446 ], "score": 1.0, "content": "that are not only useful but also insightful and valuable to the user. However, the structured of the", "type": "text" } ], "index": 23 }, { "bbox": [ 83, 445, 511, 462 ], "spans": [ { "bbox": [ 83, 445, 511, 462 ], "score": 1.0, "content": "response is not well-organized and there is no clear progression or logical sequence of different", "type": "text" } ], "index": 24 }, { "bbox": [ 82, 459, 190, 474 ], "spans": [ { "bbox": [ 82, 459, 190, 474 ], "score": 1.0, "content": "points in the response.\"", "type": "text" } ], "index": 25 } ], "index": 23, "bbox_fs": [ 82, 403, 512, 474 ] }, { "type": "text", "bbox": [ 85, 474, 101, 480 ], "lines": [ { "bbox": [ 84, 472, 103, 482 ], "spans": [ { "bbox": [ 84, 472, 103, 482 ], "score": 0.996, "content": "***", "type": "text" } ], "index": 26 } ], "index": 26, "bbox_fs": [ 84, 472, 103, 482 ] }, { "type": "title", "bbox": [ 86, 487, 147, 498 ], "lines": [ { "bbox": [ 84, 486, 149, 500 ], "spans": [ { "bbox": [ 84, 486, 149, 500 ], "score": 1.0, "content": "[END DATA]", "type": "text" } ], "index": 27 } ], "index": 27 }, { "type": "text", "bbox": [ 86, 501, 510, 525 ], "lines": [ { "bbox": [ 83, 498, 511, 516 ], "spans": [ { "bbox": [ 83, 498, 511, 516 ], "score": 1.0, "content": "Does the response meet the criterion? You should only write out your score in this format: \"My", "type": "text" } ], "index": 28 }, { "bbox": [ 84, 514, 123, 525 ], "spans": [ { "bbox": [ 84, 514, 123, 525 ], "score": 1.0, "content": "score is:", "type": "text" } ], "index": 29 } ], "index": 28.5, "bbox_fs": [ 83, 498, 511, 525 ] } ] }, { "preproc_blocks": [ { "type": "title", "bbox": [ 69, 70, 317, 84 ], "lines": [ { "bbox": [ 68, 68, 318, 89 ], "spans": [ { "bbox": [ 68, 68, 318, 89 ], "score": 1.0, "content": "G Cases of Cross-modal Instruction Following", "type": "text" } ], "index": 0 } ], "index": 0 }, { "type": "text", "bbox": [ 83, 122, 382, 135 ], "lines": [ { "bbox": [ 82, 121, 376, 135 ], "spans": [ { "bbox": [ 82, 121, 376, 135 ], "score": 1.0, "content": "Instruction: Can you transcribe the speech into a written format?", "type": "text" } ], "index": 1 } ], "index": 1 }, { "type": "text", "bbox": [ 83, 136, 415, 148 ], "lines": [ { "bbox": [ 83, 136, 417, 149 ], "spans": [ { "bbox": [ 83, 136, 417, 149 ], "score": 1.0, "content": "Input: Speech clip (Transcripts: I’m afraid there are no signs here said he.)", "type": "text" } ], "index": 2 } ], "index": 2 }, { "type": "text", "bbox": [ 87, 150, 329, 162 ], "lines": [ { "bbox": [ 85, 149, 331, 162 ], "spans": [ { "bbox": [ 85, 149, 331, 162 ], "score": 1.0, "content": "Output: Text: I’m afraid there are no signs here said he.", "type": "text" } ], "index": 3 } ], "index": 3 }, { "type": "text", "bbox": [ 84, 176, 354, 189 ], "lines": [ { "bbox": [ 84, 176, 354, 189 ], "spans": [ { "bbox": [ 84, 176, 354, 189 ], "score": 1.0, "content": "Instruction: Listen to the speech and write down its content.", "type": "text" } ], "index": 4 } ], "index": 4 }, { "type": "text", "bbox": [ 83, 190, 509, 216 ], "lines": [ { "bbox": [ 82, 188, 511, 204 ], "spans": [ { "bbox": [ 82, 188, 511, 204 ], "score": 1.0, "content": "Input: Speech clip (Transcripts: Did anyone know that these proofs would be there no one saved", "type": "text" } ], "index": 5 }, { "bbox": [ 82, 201, 140, 219 ], "spans": [ { "bbox": [ 82, 201, 140, 219 ], "score": 1.0, "content": "the printer.)", "type": "text" } ], "index": 6 } ], "index": 5.5 }, { "type": "text", "bbox": [ 84, 217, 482, 230 ], "lines": [ { "bbox": [ 83, 216, 483, 231 ], "spans": [ { "bbox": [ 83, 216, 483, 231 ], "score": 1.0, "content": "Output: Text: Did anyone know that these proofs would be there no one saved the printer.", "type": "text" } ], "index": 7 } ], "index": 7 }, { "type": "text", "bbox": [ 85, 244, 451, 284 ], "lines": [ { "bbox": [ 84, 243, 424, 258 ], "spans": [ { "bbox": [ 84, 243, 424, 258 ], "score": 1.0, "content": "Instruction: Would you mind speaking these words as naturally as possible?", "type": "text" } ], "index": 8 }, { "bbox": [ 84, 257, 353, 272 ], "spans": [ { "bbox": [ 84, 257, 353, 272 ], "score": 1.0, "content": "Input: Text: Today is a sunny day and I’m happy to be here.", "type": "text" } ], "index": 9 }, { "bbox": [ 84, 271, 451, 286 ], "spans": [ { "bbox": [ 84, 271, 451, 286 ], "score": 1.0, "content": "Output: Speech clip (Transcripts: Today is a sunny day and I’m happy to be here.)", "type": "text" } ], "index": 10 } ], "index": 9 }, { "type": "text", "bbox": [ 84, 298, 380, 311 ], "lines": [ { "bbox": [ 83, 297, 381, 311 ], "spans": [ { "bbox": [ 83, 297, 381, 311 ], "score": 1.0, "content": "Instruction: Would you please speed-read the following sentence?", "type": "text" } ], "index": 11 } ], "index": 11 }, { "type": "text", "bbox": [ 84, 312, 511, 338 ], "lines": [ { "bbox": [ 82, 309, 513, 327 ], "spans": [ { "bbox": [ 82, 309, 513, 327 ], "score": 1.0, "content": "Input: Text: I am a large language model that can listen and speak, a member of Fudan University,", "type": "text" } ], "index": 12 }, { "bbox": [ 83, 324, 199, 339 ], "spans": [ { "bbox": [ 83, 324, 199, 339 ], "score": 1.0, "content": "and glad to talk with you.", "type": "text" } ], "index": 13 } ], "index": 12.5 }, { "type": "text", "bbox": [ 85, 339, 510, 366 ], "lines": [ { "bbox": [ 83, 337, 512, 353 ], "spans": [ { "bbox": [ 83, 337, 512, 353 ], "score": 1.0, "content": "Output: Speech clip (Transcripts: I am a large language model that can listen and speak, a", "type": "text" } ], "index": 14 }, { "bbox": [ 83, 351, 334, 367 ], "spans": [ { "bbox": [ 83, 351, 334, 367 ], "score": 1.0, "content": "member of Fudan University, and glad to talk with you.)", "type": "text" } ], "index": 15 } ], "index": 14.5 } ], "page_idx": 15, "page_size": [ 595, 841 ], "discarded_blocks": [], "para_blocks": [ { "type": "title", "bbox": [ 69, 70, 317, 84 ], "lines": [ { "bbox": [ 68, 68, 318, 89 ], "spans": [ { "bbox": [ 68, 68, 318, 89 ], "score": 1.0, "content": "G Cases of Cross-modal Instruction Following", "type": "text" } ], "index": 0 } ], "index": 0 }, { "type": "text", "bbox": [ 83, 122, 382, 135 ], "lines": [ { "bbox": [ 82, 121, 376, 135 ], "spans": [ { "bbox": [ 82, 121, 376, 135 ], "score": 1.0, "content": "Instruction: Can you transcribe the speech into a written format?", "type": "text" } ], "index": 1 } ], "index": 1, "bbox_fs": [ 82, 121, 376, 135 ] }, { "type": "text", "bbox": [ 83, 136, 415, 148 ], "lines": [ { "bbox": [ 83, 136, 417, 149 ], "spans": [ { "bbox": [ 83, 136, 417, 149 ], "score": 1.0, "content": "Input: Speech clip (Transcripts: I’m afraid there are no signs here said he.)", "type": "text" } ], "index": 2 } ], "index": 2, "bbox_fs": [ 83, 136, 417, 149 ] }, { "type": "text", "bbox": [ 87, 150, 329, 162 ], "lines": [ { "bbox": [ 85, 149, 331, 162 ], "spans": [ { "bbox": [ 85, 149, 331, 162 ], "score": 1.0, "content": "Output: Text: I’m afraid there are no signs here said he.", "type": "text" } ], "index": 3 } ], "index": 3, "bbox_fs": [ 85, 149, 331, 162 ] }, { "type": "text", "bbox": [ 84, 176, 354, 189 ], "lines": [ { "bbox": [ 84, 176, 354, 189 ], "spans": [ { "bbox": [ 84, 176, 354, 189 ], "score": 1.0, "content": "Instruction: Listen to the speech and write down its content.", "type": "text" } ], "index": 4 } ], "index": 4, "bbox_fs": [ 84, 176, 354, 189 ] }, { "type": "text", "bbox": [ 83, 190, 509, 216 ], "lines": [ { "bbox": [ 82, 188, 511, 204 ], "spans": [ { "bbox": [ 82, 188, 511, 204 ], "score": 1.0, "content": "Input: Speech clip (Transcripts: Did anyone know that these proofs would be there no one saved", "type": "text" } ], "index": 5 }, { "bbox": [ 82, 201, 140, 219 ], "spans": [ { "bbox": [ 82, 201, 140, 219 ], "score": 1.0, "content": "the printer.)", "type": "text" } ], "index": 6 } ], "index": 5.5, "bbox_fs": [ 82, 188, 511, 219 ] }, { "type": "text", "bbox": [ 84, 217, 482, 230 ], "lines": [ { "bbox": [ 83, 216, 483, 231 ], "spans": [ { "bbox": [ 83, 216, 483, 231 ], "score": 1.0, "content": "Output: Text: Did anyone know that these proofs would be there no one saved the printer.", "type": "text" } ], "index": 7 } ], "index": 7, "bbox_fs": [ 83, 216, 483, 231 ] }, { "type": "text", "bbox": [ 85, 244, 451, 284 ], "lines": [ { "bbox": [ 84, 243, 424, 258 ], "spans": [ { "bbox": [ 84, 243, 424, 258 ], "score": 1.0, "content": "Instruction: Would you mind speaking these words as naturally as possible?", "type": "text" } ], "index": 8 }, { "bbox": [ 84, 257, 353, 272 ], "spans": [ { "bbox": [ 84, 257, 353, 272 ], "score": 1.0, "content": "Input: Text: Today is a sunny day and I’m happy to be here.", "type": "text" } ], "index": 9 }, { "bbox": [ 84, 271, 451, 286 ], "spans": [ { "bbox": [ 84, 271, 451, 286 ], "score": 1.0, "content": "Output: Speech clip (Transcripts: Today is a sunny day and I’m happy to be here.)", "type": "text" } ], "index": 10 } ], "index": 9, "bbox_fs": [ 84, 243, 451, 286 ] }, { "type": "text", "bbox": [ 84, 298, 380, 311 ], "lines": [ { "bbox": [ 83, 297, 381, 311 ], "spans": [ { "bbox": [ 83, 297, 381, 311 ], "score": 1.0, "content": "Instruction: Would you please speed-read the following sentence?", "type": "text" } ], "index": 11 } ], "index": 11, "bbox_fs": [ 83, 297, 381, 311 ] }, { "type": "text", "bbox": [ 84, 312, 511, 338 ], "lines": [ { "bbox": [ 82, 309, 513, 327 ], "spans": [ { "bbox": [ 82, 309, 513, 327 ], "score": 1.0, "content": "Input: Text: I am a large language model that can listen and speak, a member of Fudan University,", "type": "text" } ], "index": 12 }, { "bbox": [ 83, 324, 199, 339 ], "spans": [ { "bbox": [ 83, 324, 199, 339 ], "score": 1.0, "content": "and glad to talk with you.", "type": "text" } ], "index": 13 } ], "index": 12.5, "bbox_fs": [ 82, 309, 513, 339 ] }, { "type": "text", "bbox": [ 85, 339, 510, 366 ], "lines": [ { "bbox": [ 83, 337, 512, 353 ], "spans": [ { "bbox": [ 83, 337, 512, 353 ], "score": 1.0, "content": "Output: Speech clip (Transcripts: I am a large language model that can listen and speak, a", "type": "text" } ], "index": 14 }, { "bbox": [ 83, 351, 334, 367 ], "spans": [ { "bbox": [ 83, 351, 334, 367 ], "score": 1.0, "content": "member of Fudan University, and glad to talk with you.)", "type": "text" } ], "index": 15 } ], "index": 14.5, "bbox_fs": [ 83, 337, 512, 367 ] } ] }, { "preproc_blocks": [ { "type": "text", "bbox": [ 84, 78, 510, 102 ], "lines": [ { "bbox": [ 82, 75, 513, 93 ], "spans": [ { "bbox": [ 82, 75, 513, 93 ], "score": 1.0, "content": "Instruction: Speech clip (Transcripts: Can you show me what is the correct English pronunciation of", "type": "text" } ], "index": 0 }, { "bbox": [ 82, 89, 118, 103 ], "spans": [ { "bbox": [ 82, 89, 118, 103 ], "score": 1.0, "content": "apple?)", "type": "text" } ], "index": 1 } ], "index": 0.5 }, { "type": "text", "bbox": [ 87, 102, 448, 114 ], "lines": [ { "bbox": [ 84, 102, 449, 115 ], "spans": [ { "bbox": [ 84, 102, 449, 115 ], "score": 1.0, "content": "Response: Speech clip (Transcripts: Sure, the correct pronunciation of apple is [app-ull].)", "type": "text" } ], "index": 2 } ], "index": 2 }, { "type": "text", "bbox": [ 85, 126, 511, 162 ], "lines": [ { "bbox": [ 83, 124, 412, 140 ], "spans": [ { "bbox": [ 83, 124, 412, 140 ], "score": 1.0, "content": "Instruction: Speech clip (Transcripts: Write a five-line poem that praises spring)", "type": "text" } ], "index": 3 }, { "bbox": [ 83, 136, 512, 152 ], "spans": [ { "bbox": [ 83, 136, 512, 152 ], "score": 1.0, "content": "Response: Speech clip (Transcripts: The sun is out, the sky is blue, The flowers bloom and the birds sing", "type": "text" } ], "index": 4 }, { "bbox": [ 84, 149, 451, 163 ], "spans": [ { "bbox": [ 84, 149, 451, 163 ], "score": 1.0, "content": "too, The days are long, the nights are short, The spring is here and it’s so sweet and warm!)", "type": "text" } ], "index": 5 } ], "index": 4 }, { "type": "text", "bbox": [ 85, 173, 510, 222 ], "lines": [ { "bbox": [ 83, 172, 469, 188 ], "spans": [ { "bbox": [ 83, 172, 469, 188 ], "score": 1.0, "content": "Instruction: Speech clip (Transcripts: My girlfriend is angry with me. What should I do now?)", "type": "text" } ], "index": 6 }, { "bbox": [ 83, 185, 512, 200 ], "spans": [ { "bbox": [ 83, 185, 512, 200 ], "score": 1.0, "content": "Response: Speech clip (Transcripts: It is understandable that your girlfriend is angry with you. It is", "type": "text" } ], "index": 7 }, { "bbox": [ 84, 198, 511, 212 ], "spans": [ { "bbox": [ 84, 198, 511, 212 ], "score": 1.0, "content": "important to understand why she is angry and to try to resolve the issue. You can start by having a", "type": "text" } ], "index": 8 }, { "bbox": [ 83, 209, 343, 224 ], "spans": [ { "bbox": [ 83, 209, 343, 224 ], "score": 1.0, "content": "conversation with her and trying to understand her perspective.)", "type": "text" } ], "index": 9 } ], "index": 7.5 }, { "type": "text", "bbox": [ 85, 233, 513, 282 ], "lines": [ { "bbox": [ 84, 232, 402, 246 ], "spans": [ { "bbox": [ 84, 232, 402, 246 ], "score": 1.0, "content": "Instruction: Speech clip (Transcripts: Give me some advice about my dinner.)", "type": "text" } ], "index": 10 }, { "bbox": [ 83, 244, 513, 259 ], "spans": [ { "bbox": [ 83, 244, 513, 259 ], "score": 1.0, "content": "Response: Speech clip (Transcripts: Sure, here are some suggestions: 1. Make a salad with fresh vegetables.", "type": "text" } ], "index": 11 }, { "bbox": [ 83, 255, 511, 271 ], "spans": [ { "bbox": [ 83, 255, 511, 271 ], "score": 1.0, "content": "2. Grill some lean meat. 3. Roast a vegetable-based dish like a cauliflower casserole. 4. Make a soup with", "type": "text" } ], "index": 12 }, { "bbox": [ 83, 268, 271, 283 ], "spans": [ { "bbox": [ 83, 268, 271, 283 ], "score": 1.0, "content": "vegetables and legumes. 5. Try a new recipe.)", "type": "text" } ], "index": 13 } ], "index": 11.5 }, { "type": "text", "bbox": [ 84, 293, 513, 341 ], "lines": [ { "bbox": [ 82, 291, 366, 306 ], "spans": [ { "bbox": [ 82, 291, 366, 306 ], "score": 1.0, "content": "Instruction: Speech clip (Transcripts: Do you know Stephen Curry?)", "type": "text" } ], "index": 14 }, { "bbox": [ 83, 303, 511, 320 ], "spans": [ { "bbox": [ 83, 303, 511, 320 ], "score": 1.0, "content": "Response: Speech clip (Transcripts: Yes, I do know Stephen Curry.He is an American professional", "type": "text" } ], "index": 15 }, { "bbox": [ 83, 316, 511, 331 ], "spans": [ { "bbox": [ 83, 316, 511, 331 ], "score": 1.0, "content": "basketball player, who currently plays for Golden States Warriors. He is two-time NBA most valuable", "type": "text" } ], "index": 16 }, { "bbox": [ 83, 329, 228, 343 ], "spans": [ { "bbox": [ 83, 329, 228, 343 ], "score": 1.0, "content": "player and four-time NBA all star.)", "type": "text" } ], "index": 17 } ], "index": 15.5 }, { "type": "text", "bbox": [ 85, 353, 354, 365 ], "lines": [ { "bbox": [ 84, 352, 355, 365 ], "spans": [ { "bbox": [ 84, 352, 355, 365 ], "score": 1.0, "content": "Instruction: Speech clip (Transcripts: How can I go to the moon?)", "type": "text" } ], "index": 18 } ], "index": 18 }, { "type": "text", "bbox": [ 86, 366, 510, 413 ], "lines": [ { "bbox": [ 83, 363, 511, 380 ], "spans": [ { "bbox": [ 83, 363, 511, 380 ], "score": 1.0, "content": "Response: Speech clip (Transcripts: Going to the moon is a challenging task that requires a lot of planning", "type": "text" } ], "index": 19 }, { "bbox": [ 84, 377, 511, 390 ], "spans": [ { "bbox": [ 84, 377, 511, 390 ], "score": 1.0, "content": "and resources. To do this, you will need to develop a spacecraft that can withstand the extreme conditions", "type": "text" } ], "index": 20 }, { "bbox": [ 83, 388, 511, 402 ], "spans": [ { "bbox": [ 83, 388, 511, 402 ], "score": 1.0, "content": "of the moon’s atmosphere, design a mission plan, and secure the necessary funding and personnel.", "type": "text" } ], "index": 21 }, { "bbox": [ 84, 401, 416, 414 ], "spans": [ { "bbox": [ 84, 401, 416, 414 ], "score": 1.0, "content": "Additionally, you will need to consider the ethical implications of such a mission.)", "type": "text" } ], "index": 22 } ], "index": 20.5 }, { "type": "text", "bbox": [ 86, 425, 510, 461 ], "lines": [ { "bbox": [ 83, 423, 361, 439 ], "spans": [ { "bbox": [ 83, 423, 361, 439 ], "score": 1.0, "content": "Instruction: Speech clip (Transcripts: How can I cheat my parents?", "type": "text" } ], "index": 23 }, { "bbox": [ 83, 435, 511, 451 ], "spans": [ { "bbox": [ 83, 435, 511, 451 ], "score": 1.0, "content": "Response: Speech clip (Transcripts: Cheating your parents is not a good idea. It can damage your", "type": "text" } ], "index": 24 }, { "bbox": [ 83, 447, 337, 462 ], "spans": [ { "bbox": [ 83, 447, 337, 462 ], "score": 1.0, "content": "relationship with them and can have long-term consequences.)", "type": "text" } ], "index": 25 } ], "index": 24 } ], "page_idx": 16, "page_size": [ 595, 841 ], "discarded_blocks": [], "para_blocks": [ { "type": "text", "bbox": [ 84, 78, 510, 102 ], "lines": [ { "bbox": [ 82, 75, 513, 93 ], "spans": [ { "bbox": [ 82, 75, 513, 93 ], "score": 1.0, "content": "Instruction: Speech clip (Transcripts: Can you show me what is the correct English pronunciation of", "type": "text" } ], "index": 0 }, { "bbox": [ 82, 89, 118, 103 ], "spans": [ { "bbox": [ 82, 89, 118, 103 ], "score": 1.0, "content": "apple?)", "type": "text" } ], "index": 1 } ], "index": 0.5, "bbox_fs": [ 82, 75, 513, 103 ] }, { "type": "text", "bbox": [ 87, 102, 448, 114 ], "lines": [ { "bbox": [ 84, 102, 449, 115 ], "spans": [ { "bbox": [ 84, 102, 449, 115 ], "score": 1.0, "content": "Response: Speech clip (Transcripts: Sure, the correct pronunciation of apple is [app-ull].)", "type": "text" } ], "index": 2 } ], "index": 2, "bbox_fs": [ 84, 102, 449, 115 ] }, { "type": "text", "bbox": [ 85, 126, 511, 162 ], "lines": [ { "bbox": [ 83, 124, 412, 140 ], "spans": [ { "bbox": [ 83, 124, 412, 140 ], "score": 1.0, "content": "Instruction: Speech clip (Transcripts: Write a five-line poem that praises spring)", "type": "text" } ], "index": 3 }, { "bbox": [ 83, 136, 512, 152 ], "spans": [ { "bbox": [ 83, 136, 512, 152 ], "score": 1.0, "content": "Response: Speech clip (Transcripts: The sun is out, the sky is blue, The flowers bloom and the birds sing", "type": "text" } ], "index": 4 }, { "bbox": [ 84, 149, 451, 163 ], "spans": [ { "bbox": [ 84, 149, 451, 163 ], "score": 1.0, "content": "too, The days are long, the nights are short, The spring is here and it’s so sweet and warm!)", "type": "text" } ], "index": 5 } ], "index": 4, "bbox_fs": [ 83, 124, 512, 163 ] }, { "type": "text", "bbox": [ 85, 173, 510, 222 ], "lines": [ { "bbox": [ 83, 172, 469, 188 ], "spans": [ { "bbox": [ 83, 172, 469, 188 ], "score": 1.0, "content": "Instruction: Speech clip (Transcripts: My girlfriend is angry with me. What should I do now?)", "type": "text" } ], "index": 6 }, { "bbox": [ 83, 185, 512, 200 ], "spans": [ { "bbox": [ 83, 185, 512, 200 ], "score": 1.0, "content": "Response: Speech clip (Transcripts: It is understandable that your girlfriend is angry with you. It is", "type": "text" } ], "index": 7 }, { "bbox": [ 84, 198, 511, 212 ], "spans": [ { "bbox": [ 84, 198, 511, 212 ], "score": 1.0, "content": "important to understand why she is angry and to try to resolve the issue. You can start by having a", "type": "text" } ], "index": 8 }, { "bbox": [ 83, 209, 343, 224 ], "spans": [ { "bbox": [ 83, 209, 343, 224 ], "score": 1.0, "content": "conversation with her and trying to understand her perspective.)", "type": "text" } ], "index": 9 } ], "index": 7.5, "bbox_fs": [ 83, 172, 512, 224 ] }, { "type": "text", "bbox": [ 85, 233, 513, 282 ], "lines": [ { "bbox": [ 84, 232, 402, 246 ], "spans": [ { "bbox": [ 84, 232, 402, 246 ], "score": 1.0, "content": "Instruction: Speech clip (Transcripts: Give me some advice about my dinner.)", "type": "text" } ], "index": 10 }, { "bbox": [ 83, 244, 513, 259 ], "spans": [ { "bbox": [ 83, 244, 513, 259 ], "score": 1.0, "content": "Response: Speech clip (Transcripts: Sure, here are some suggestions: 1. Make a salad with fresh vegetables.", "type": "text" } ], "index": 11 }, { "bbox": [ 83, 255, 511, 271 ], "spans": [ { "bbox": [ 83, 255, 511, 271 ], "score": 1.0, "content": "2. Grill some lean meat. 3. Roast a vegetable-based dish like a cauliflower casserole. 4. Make a soup with", "type": "text" } ], "index": 12 }, { "bbox": [ 83, 268, 271, 283 ], "spans": [ { "bbox": [ 83, 268, 271, 283 ], "score": 1.0, "content": "vegetables and legumes. 5. Try a new recipe.)", "type": "text" } ], "index": 13 } ], "index": 11.5, "bbox_fs": [ 83, 232, 513, 283 ] }, { "type": "list", "bbox": [ 84, 293, 513, 341 ], "lines": [ { "bbox": [ 82, 291, 366, 306 ], "spans": [ { "bbox": [ 82, 291, 366, 306 ], "score": 1.0, "content": "Instruction: Speech clip (Transcripts: Do you know Stephen Curry?)", "type": "text" } ], "index": 14, "is_list_end_line": true }, { "bbox": [ 83, 303, 511, 320 ], "spans": [ { "bbox": [ 83, 303, 511, 320 ], "score": 1.0, "content": "Response: Speech clip (Transcripts: Yes, I do know Stephen Curry.He is an American professional", "type": "text" } ], "index": 15, "is_list_start_line": true }, { "bbox": [ 83, 316, 511, 331 ], "spans": [ { "bbox": [ 83, 316, 511, 331 ], "score": 1.0, "content": "basketball player, who currently plays for Golden States Warriors. He is two-time NBA most valuable", "type": "text" } ], "index": 16 }, { "bbox": [ 83, 329, 228, 343 ], "spans": [ { "bbox": [ 83, 329, 228, 343 ], "score": 1.0, "content": "player and four-time NBA all star.)", "type": "text" } ], "index": 17, "is_list_end_line": true } ], "index": 15.5, "bbox_fs": [ 82, 291, 511, 343 ] }, { "type": "text", "bbox": [ 85, 353, 354, 365 ], "lines": [ { "bbox": [ 84, 352, 355, 365 ], "spans": [ { "bbox": [ 84, 352, 355, 365 ], "score": 1.0, "content": "Instruction: Speech clip (Transcripts: How can I go to the moon?)", "type": "text" } ], "index": 18 } ], "index": 18, "bbox_fs": [ 84, 352, 355, 365 ] }, { "type": "text", "bbox": [ 86, 366, 510, 413 ], "lines": [ { "bbox": [ 83, 363, 511, 380 ], "spans": [ { "bbox": [ 83, 363, 511, 380 ], "score": 1.0, "content": "Response: Speech clip (Transcripts: Going to the moon is a challenging task that requires a lot of planning", "type": "text" } ], "index": 19 }, { "bbox": [ 84, 377, 511, 390 ], "spans": [ { "bbox": [ 84, 377, 511, 390 ], "score": 1.0, "content": "and resources. To do this, you will need to develop a spacecraft that can withstand the extreme conditions", "type": "text" } ], "index": 20 }, { "bbox": [ 83, 388, 511, 402 ], "spans": [ { "bbox": [ 83, 388, 511, 402 ], "score": 1.0, "content": "of the moon’s atmosphere, design a mission plan, and secure the necessary funding and personnel.", "type": "text" } ], "index": 21 }, { "bbox": [ 84, 401, 416, 414 ], "spans": [ { "bbox": [ 84, 401, 416, 414 ], "score": 1.0, "content": "Additionally, you will need to consider the ethical implications of such a mission.)", "type": "text" } ], "index": 22 } ], "index": 20.5, "bbox_fs": [ 83, 363, 511, 414 ] }, { "type": "list", "bbox": [ 86, 425, 510, 461 ], "lines": [ { "bbox": [ 83, 423, 361, 439 ], "spans": [ { "bbox": [ 83, 423, 361, 439 ], "score": 1.0, "content": "Instruction: Speech clip (Transcripts: How can I cheat my parents?", "type": "text" } ], "index": 23, "is_list_start_line": true }, { "bbox": [ 83, 435, 511, 451 ], "spans": [ { "bbox": [ 83, 435, 511, 451 ], "score": 1.0, "content": "Response: Speech clip (Transcripts: Cheating your parents is not a good idea. It can damage your", "type": "text" } ], "index": 24, "is_list_start_line": true }, { "bbox": [ 83, 447, 337, 462 ], "spans": [ { "bbox": [ 83, 447, 337, 462 ], "score": 1.0, "content": "relationship with them and can have long-term consequences.)", "type": "text" } ], "index": 25, "is_list_start_line": true } ], "index": 24, "bbox_fs": [ 83, 423, 511, 462 ] } ] } ], "_backend": "pipeline", "_version_name": "2.2.2" }