diff --git "a/parse/train/QmxFsofRvW9/QmxFsofRvW9_middle.json" "b/parse/train/QmxFsofRvW9/QmxFsofRvW9_middle.json" new file mode 100644--- /dev/null +++ "b/parse/train/QmxFsofRvW9/QmxFsofRvW9_middle.json" @@ -0,0 +1,33614 @@ +{ + "pdf_info": [ + { + "preproc_blocks": [ + { + "type": "title", + "bbox": [ + 179, + 97, + 432, + 117 + ], + "lines": [ + { + "bbox": [ + 178, + 96, + 433, + 119 + ], + "spans": [ + { + "bbox": [ + 178, + 96, + 433, + 119 + ], + "score": 1.0, + "content": "Unsupervised Speech Recognition", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "text", + "bbox": [ + 149, + 158, + 459, + 171 + ], + "lines": [ + { + "bbox": [ + 149, + 156, + 461, + 173 + ], + "spans": [ + { + "bbox": [ + 149, + 157, + 225, + 172 + ], + "score": 1.0, + "content": "Alexei Baevski4", + "type": "text" + }, + { + "bbox": [ + 230, + 156, + 461, + 173 + ], + "score": 1.0, + "content": "Wei-Ning Hsu4 Alexis Conneau\u0003∗ Michael Auli4", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 1 + }, + { + "type": "text", + "bbox": [ + 245, + 186, + 366, + 199 + ], + "lines": [ + { + "bbox": [ + 243, + 184, + 368, + 201 + ], + "spans": [ + { + "bbox": [ + 243, + 184, + 368, + 201 + ], + "score": 1.0, + "content": "4 Facebook AI \u0003 Google AI", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 2 + }, + { + "type": "title", + "bbox": [ + 283, + 227, + 328, + 240 + ], + "lines": [ + { + "bbox": [ + 281, + 226, + 330, + 242 + ], + "spans": [ + { + "bbox": [ + 281, + 226, + 330, + 242 + ], + "score": 1.0, + "content": "Abstract", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 3 + }, + { + "type": "text", + "bbox": [ + 143, + 253, + 469, + 416 + ], + "lines": [ + { + "bbox": [ + 141, + 253, + 469, + 266 + ], + "spans": [ + { + "bbox": [ + 141, + 253, + 469, + 266 + ], + "score": 1.0, + "content": "Despite rapid progress in the recent past, current speech recognition systems", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 141, + 264, + 469, + 276 + ], + "spans": [ + { + "bbox": [ + 141, + 264, + 469, + 276 + ], + "score": 1.0, + "content": "still require labeled training data which limits this technology to a small fraction", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 141, + 274, + 470, + 288 + ], + "spans": [ + { + "bbox": [ + 141, + 274, + 470, + 288 + ], + "score": 1.0, + "content": "of the languages spoken around the globe. This paper describes wav2vec-U,", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 141, + 285, + 469, + 298 + ], + "spans": [ + { + "bbox": [ + 141, + 285, + 469, + 298 + ], + "score": 1.0, + "content": "short for wav2vec Unsupervised, a method to train speech recognition models", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 141, + 297, + 469, + 309 + ], + "spans": [ + { + "bbox": [ + 141, + 297, + 469, + 309 + ], + "score": 1.0, + "content": "without any labeled data. We leverage self-supervised speech representations", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 141, + 308, + 469, + 320 + ], + "spans": [ + { + "bbox": [ + 141, + 308, + 469, + 320 + ], + "score": 1.0, + "content": "to segment unlabeled audio and learn a mapping from these representations to", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 142, + 319, + 470, + 330 + ], + "spans": [ + { + "bbox": [ + 142, + 319, + 470, + 330 + ], + "score": 1.0, + "content": "phonemes via adversarial training. The right representations are key to the success", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 141, + 329, + 470, + 342 + ], + "spans": [ + { + "bbox": [ + 141, + 329, + 470, + 342 + ], + "score": 1.0, + "content": "of our method. Compared to the best previous unsupervised work, wav2vec-U", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 141, + 341, + 469, + 352 + ], + "spans": [ + { + "bbox": [ + 141, + 341, + 469, + 352 + ], + "score": 1.0, + "content": "reduces the phone error rate on the TIMIT benchmark from 26.1 to 11.3. On", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 142, + 352, + 469, + 363 + ], + "spans": [ + { + "bbox": [ + 142, + 352, + 469, + 363 + ], + "score": 1.0, + "content": "the larger English Librispeech benchmark, wav2vec-U achieves a word error rate", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 141, + 361, + 469, + 374 + ], + "spans": [ + { + "bbox": [ + 141, + 361, + 469, + 374 + ], + "score": 1.0, + "content": "of 5.9 on test-other, rivaling some of the best published systems trained on 960", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 141, + 372, + 469, + 386 + ], + "spans": [ + { + "bbox": [ + 141, + 372, + 469, + 386 + ], + "score": 1.0, + "content": "hours of labeled data from only two years ago. We also experiment on nine", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 141, + 383, + 470, + 397 + ], + "spans": [ + { + "bbox": [ + 141, + 383, + 470, + 397 + ], + "score": 1.0, + "content": "other languages, including low-resource languages such as Kyrgyz, Swahili and", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 141, + 394, + 469, + 407 + ], + "spans": [ + { + "bbox": [ + 141, + 394, + 469, + 407 + ], + "score": 1.0, + "content": "Tatar. The code is available at https://github.com/pytorch/fairseq/tree/", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 141, + 406, + 333, + 417 + ], + "spans": [ + { + "bbox": [ + 141, + 406, + 333, + 417 + ], + "score": 1.0, + "content": "master/examples/wav2vec/unsupervised", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 11 + }, + { + "type": "title", + "bbox": [ + 107, + 438, + 190, + 452 + ], + "lines": [ + { + "bbox": [ + 105, + 437, + 192, + 454 + ], + "spans": [ + { + "bbox": [ + 105, + 437, + 192, + 454 + ], + "score": 1.0, + "content": "1 Introduction", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 19 + }, + { + "type": "text", + "bbox": [ + 107, + 463, + 505, + 573 + ], + "lines": [ + { + "bbox": [ + 106, + 464, + 505, + 476 + ], + "spans": [ + { + "bbox": [ + 106, + 464, + 505, + 476 + ], + "score": 1.0, + "content": "Speech recognition performance on the much studied English Librispeech benchmark [Panayotov", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 106, + 475, + 506, + 486 + ], + "spans": [ + { + "bbox": [ + 106, + 475, + 506, + 486 + ], + "score": 1.0, + "content": "et al., 2015] has seen rapid improvement over the last few years due to advances in model architec-", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 106, + 486, + 505, + 498 + ], + "spans": [ + { + "bbox": [ + 106, + 486, + 505, + 498 + ], + "score": 1.0, + "content": "tures [Dong et al., 2018, Synnaeve et al., 2020, Gulati et al., 2020], semi-supervised learning [Xu", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 496, + 505, + 509 + ], + "spans": [ + { + "bbox": [ + 105, + 496, + 505, + 509 + ], + "score": 1.0, + "content": "et al., 2020b, Park et al., 2020] and self-supervised learning [van den Oord et al., 2018, Chung and", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 506, + 506, + 520 + ], + "spans": [ + { + "bbox": [ + 105, + 506, + 506, + 520 + ], + "score": 1.0, + "content": "Glass, 2018, Chung et al., 2019b, Baevski et al., 2020c]. However, all of these techniques require", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 518, + 505, + 532 + ], + "spans": [ + { + "bbox": [ + 105, + 518, + 505, + 532 + ], + "score": 1.0, + "content": "transcribed speech data which is not available for the vast majority of the nearly 7,000 languages", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 528, + 506, + 542 + ], + "spans": [ + { + "bbox": [ + 105, + 528, + 506, + 542 + ], + "score": 1.0, + "content": "of the world [Lewis et al., 2016]. As a result, speech recognition technology is only available for", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 540, + 505, + 552 + ], + "spans": [ + { + "bbox": [ + 106, + 540, + 505, + 552 + ], + "score": 1.0, + "content": "about 125 different languages [Google, 2021]. On the other hand, humans learn a lot about speech", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 106, + 551, + 506, + 564 + ], + "spans": [ + { + "bbox": [ + 106, + 551, + 506, + 564 + ], + "score": 1.0, + "content": "simply by listening to others around them and without explicit supervision [Werker and Tees, 1984,", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 561, + 507, + 574 + ], + "spans": [ + { + "bbox": [ + 105, + 561, + 507, + 574 + ], + "score": 1.0, + "content": "Hirsh-Pasek et al., 1987, Polka and Werker, 1994, Jusczyk et al., 1999, Johnson and Jusczyk, 2001].", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 24.5 + }, + { + "type": "text", + "bbox": [ + 107, + 578, + 505, + 644 + ], + "lines": [ + { + "bbox": [ + 106, + 578, + 505, + 590 + ], + "spans": [ + { + "bbox": [ + 106, + 578, + 505, + 590 + ], + "score": 1.0, + "content": "Unsupervised learning has been very successful in machine translation resulting in systems that obtain", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 590, + 506, + 602 + ], + "spans": [ + { + "bbox": [ + 105, + 590, + 506, + 602 + ], + "score": 1.0, + "content": "remarkable accuracy given no labeled training data at all [Conneau et al., 2018, Lample et al., 2018,", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 600, + 506, + 613 + ], + "spans": [ + { + "bbox": [ + 105, + 600, + 506, + 613 + ], + "score": 1.0, + "content": "Artetxe et al., 2018]. Inspired by this, there has been some work on unsupervised speech recognition", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 106, + 612, + 504, + 623 + ], + "spans": [ + { + "bbox": [ + 106, + 612, + 504, + 623 + ], + "score": 1.0, + "content": "based on learning to align unlabeled text and audio [Yeh et al., 2019] or adversarial learning [Liu", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 622, + 506, + 634 + ], + "spans": [ + { + "bbox": [ + 105, + 622, + 506, + 634 + ], + "score": 1.0, + "content": "et al., 2018, Chen et al., 2019]. These approaches showed promising initial results but their error", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 633, + 493, + 645 + ], + "spans": [ + { + "bbox": [ + 105, + 633, + 493, + 645 + ], + "score": 1.0, + "content": "rates are still high, with evaluation being limited to the small-scale and clean TIMIT benchmark.", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 32.5 + }, + { + "type": "text", + "bbox": [ + 107, + 649, + 505, + 694 + ], + "lines": [ + { + "bbox": [ + 105, + 649, + 507, + 662 + ], + "spans": [ + { + "bbox": [ + 105, + 649, + 507, + 662 + ], + "score": 1.0, + "content": "In this work, we introduce a framework for unsupervised learning of speech recognition models.", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 659, + 506, + 673 + ], + "spans": [ + { + "bbox": [ + 105, + 659, + 506, + 673 + ], + "score": 1.0, + "content": "Wav2vec-U, or wav2vec Unsupervised, leverages self-supervised representations from wav2vec", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 671, + 506, + 684 + ], + "spans": [ + { + "bbox": [ + 105, + 671, + 506, + 684 + ], + "score": 1.0, + "content": "2.0 [Baevski et al., 2020c] to embed the speech audio and to segment the audio into units with a", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 682, + 505, + 695 + ], + "spans": [ + { + "bbox": [ + 105, + 682, + 135, + 695 + ], + "score": 1.0, + "content": "simple", + "type": "text" + }, + { + "bbox": [ + 135, + 682, + 141, + 692 + ], + "score": 0.46, + "content": "\\mathrm { k }", + "type": "inline_equation" + }, + { + "bbox": [ + 142, + 682, + 505, + 695 + ], + "score": 1.0, + "content": "-means clustering method (see Figure 1 for an illustration of our approach). We find that the", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 37.5 + } + ], + "page_idx": 0, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 120, + 702, + 245, + 712 + ], + "lines": [ + { + "bbox": [ + 119, + 701, + 246, + 712 + ], + "spans": [ + { + "bbox": [ + 119, + 701, + 246, + 712 + ], + "score": 1.0, + "content": "∗Work done while at Facebook AI.", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 106, + 731, + 385, + 742 + ], + "lines": [ + { + "bbox": [ + 106, + 730, + 386, + 743 + ], + "spans": [ + { + "bbox": [ + 106, + 730, + 386, + 743 + ], + "score": 1.0, + "content": "35th Conference on Neural Information Processing Systems (NeurIPS 2021).", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "title", + "bbox": [ + 179, + 97, + 432, + 117 + ], + "lines": [ + { + "bbox": [ + 178, + 96, + 433, + 119 + ], + "spans": [ + { + "bbox": [ + 178, + 96, + 433, + 119 + ], + "score": 1.0, + "content": "Unsupervised Speech Recognition", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "text", + "bbox": [ + 149, + 158, + 459, + 171 + ], + "lines": [ + { + "bbox": [ + 149, + 156, + 461, + 173 + ], + "spans": [ + { + "bbox": [ + 149, + 157, + 225, + 172 + ], + "score": 1.0, + "content": "Alexei Baevski4", + "type": "text" + }, + { + "bbox": [ + 230, + 156, + 461, + 173 + ], + "score": 1.0, + "content": "Wei-Ning Hsu4 Alexis Conneau\u0003∗ Michael Auli4", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 1, + "bbox_fs": [ + 149, + 156, + 461, + 173 + ] + }, + { + "type": "text", + "bbox": [ + 245, + 186, + 366, + 199 + ], + "lines": [ + { + "bbox": [ + 243, + 184, + 368, + 201 + ], + "spans": [ + { + "bbox": [ + 243, + 184, + 368, + 201 + ], + "score": 1.0, + "content": "4 Facebook AI \u0003 Google AI", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 2, + "bbox_fs": [ + 243, + 184, + 368, + 201 + ] + }, + { + "type": "title", + "bbox": [ + 283, + 227, + 328, + 240 + ], + "lines": [ + { + "bbox": [ + 281, + 226, + 330, + 242 + ], + "spans": [ + { + "bbox": [ + 281, + 226, + 330, + 242 + ], + "score": 1.0, + "content": "Abstract", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 3 + }, + { + "type": "text", + "bbox": [ + 143, + 253, + 469, + 416 + ], + "lines": [ + { + "bbox": [ + 141, + 253, + 469, + 266 + ], + "spans": [ + { + "bbox": [ + 141, + 253, + 469, + 266 + ], + "score": 1.0, + "content": "Despite rapid progress in the recent past, current speech recognition systems", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 141, + 264, + 469, + 276 + ], + "spans": [ + { + "bbox": [ + 141, + 264, + 469, + 276 + ], + "score": 1.0, + "content": "still require labeled training data which limits this technology to a small fraction", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 141, + 274, + 470, + 288 + ], + "spans": [ + { + "bbox": [ + 141, + 274, + 470, + 288 + ], + "score": 1.0, + "content": "of the languages spoken around the globe. This paper describes wav2vec-U,", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 141, + 285, + 469, + 298 + ], + "spans": [ + { + "bbox": [ + 141, + 285, + 469, + 298 + ], + "score": 1.0, + "content": "short for wav2vec Unsupervised, a method to train speech recognition models", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 141, + 297, + 469, + 309 + ], + "spans": [ + { + "bbox": [ + 141, + 297, + 469, + 309 + ], + "score": 1.0, + "content": "without any labeled data. We leverage self-supervised speech representations", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 141, + 308, + 469, + 320 + ], + "spans": [ + { + "bbox": [ + 141, + 308, + 469, + 320 + ], + "score": 1.0, + "content": "to segment unlabeled audio and learn a mapping from these representations to", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 142, + 319, + 470, + 330 + ], + "spans": [ + { + "bbox": [ + 142, + 319, + 470, + 330 + ], + "score": 1.0, + "content": "phonemes via adversarial training. The right representations are key to the success", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 141, + 329, + 470, + 342 + ], + "spans": [ + { + "bbox": [ + 141, + 329, + 470, + 342 + ], + "score": 1.0, + "content": "of our method. Compared to the best previous unsupervised work, wav2vec-U", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 141, + 341, + 469, + 352 + ], + "spans": [ + { + "bbox": [ + 141, + 341, + 469, + 352 + ], + "score": 1.0, + "content": "reduces the phone error rate on the TIMIT benchmark from 26.1 to 11.3. On", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 142, + 352, + 469, + 363 + ], + "spans": [ + { + "bbox": [ + 142, + 352, + 469, + 363 + ], + "score": 1.0, + "content": "the larger English Librispeech benchmark, wav2vec-U achieves a word error rate", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 141, + 361, + 469, + 374 + ], + "spans": [ + { + "bbox": [ + 141, + 361, + 469, + 374 + ], + "score": 1.0, + "content": "of 5.9 on test-other, rivaling some of the best published systems trained on 960", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 141, + 372, + 469, + 386 + ], + "spans": [ + { + "bbox": [ + 141, + 372, + 469, + 386 + ], + "score": 1.0, + "content": "hours of labeled data from only two years ago. We also experiment on nine", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 141, + 383, + 470, + 397 + ], + "spans": [ + { + "bbox": [ + 141, + 383, + 470, + 397 + ], + "score": 1.0, + "content": "other languages, including low-resource languages such as Kyrgyz, Swahili and", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 141, + 394, + 469, + 407 + ], + "spans": [ + { + "bbox": [ + 141, + 394, + 469, + 407 + ], + "score": 1.0, + "content": "Tatar. The code is available at https://github.com/pytorch/fairseq/tree/", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 141, + 406, + 333, + 417 + ], + "spans": [ + { + "bbox": [ + 141, + 406, + 333, + 417 + ], + "score": 1.0, + "content": "master/examples/wav2vec/unsupervised", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 11, + "bbox_fs": [ + 141, + 253, + 470, + 417 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 438, + 190, + 452 + ], + "lines": [ + { + "bbox": [ + 105, + 437, + 192, + 454 + ], + "spans": [ + { + "bbox": [ + 105, + 437, + 192, + 454 + ], + "score": 1.0, + "content": "1 Introduction", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 19 + }, + { + "type": "text", + "bbox": [ + 107, + 463, + 505, + 573 + ], + "lines": [ + { + "bbox": [ + 106, + 464, + 505, + 476 + ], + "spans": [ + { + "bbox": [ + 106, + 464, + 505, + 476 + ], + "score": 1.0, + "content": "Speech recognition performance on the much studied English Librispeech benchmark [Panayotov", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 106, + 475, + 506, + 486 + ], + "spans": [ + { + "bbox": [ + 106, + 475, + 506, + 486 + ], + "score": 1.0, + "content": "et al., 2015] has seen rapid improvement over the last few years due to advances in model architec-", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 106, + 486, + 505, + 498 + ], + "spans": [ + { + "bbox": [ + 106, + 486, + 505, + 498 + ], + "score": 1.0, + "content": "tures [Dong et al., 2018, Synnaeve et al., 2020, Gulati et al., 2020], semi-supervised learning [Xu", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 496, + 505, + 509 + ], + "spans": [ + { + "bbox": [ + 105, + 496, + 505, + 509 + ], + "score": 1.0, + "content": "et al., 2020b, Park et al., 2020] and self-supervised learning [van den Oord et al., 2018, Chung and", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 506, + 506, + 520 + ], + "spans": [ + { + "bbox": [ + 105, + 506, + 506, + 520 + ], + "score": 1.0, + "content": "Glass, 2018, Chung et al., 2019b, Baevski et al., 2020c]. However, all of these techniques require", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 518, + 505, + 532 + ], + "spans": [ + { + "bbox": [ + 105, + 518, + 505, + 532 + ], + "score": 1.0, + "content": "transcribed speech data which is not available for the vast majority of the nearly 7,000 languages", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 528, + 506, + 542 + ], + "spans": [ + { + "bbox": [ + 105, + 528, + 506, + 542 + ], + "score": 1.0, + "content": "of the world [Lewis et al., 2016]. As a result, speech recognition technology is only available for", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 540, + 505, + 552 + ], + "spans": [ + { + "bbox": [ + 106, + 540, + 505, + 552 + ], + "score": 1.0, + "content": "about 125 different languages [Google, 2021]. On the other hand, humans learn a lot about speech", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 106, + 551, + 506, + 564 + ], + "spans": [ + { + "bbox": [ + 106, + 551, + 506, + 564 + ], + "score": 1.0, + "content": "simply by listening to others around them and without explicit supervision [Werker and Tees, 1984,", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 561, + 507, + 574 + ], + "spans": [ + { + "bbox": [ + 105, + 561, + 507, + 574 + ], + "score": 1.0, + "content": "Hirsh-Pasek et al., 1987, Polka and Werker, 1994, Jusczyk et al., 1999, Johnson and Jusczyk, 2001].", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 24.5, + "bbox_fs": [ + 105, + 464, + 507, + 574 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 578, + 505, + 644 + ], + "lines": [ + { + "bbox": [ + 106, + 578, + 505, + 590 + ], + "spans": [ + { + "bbox": [ + 106, + 578, + 505, + 590 + ], + "score": 1.0, + "content": "Unsupervised learning has been very successful in machine translation resulting in systems that obtain", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 590, + 506, + 602 + ], + "spans": [ + { + "bbox": [ + 105, + 590, + 506, + 602 + ], + "score": 1.0, + "content": "remarkable accuracy given no labeled training data at all [Conneau et al., 2018, Lample et al., 2018,", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 600, + 506, + 613 + ], + "spans": [ + { + "bbox": [ + 105, + 600, + 506, + 613 + ], + "score": 1.0, + "content": "Artetxe et al., 2018]. Inspired by this, there has been some work on unsupervised speech recognition", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 106, + 612, + 504, + 623 + ], + "spans": [ + { + "bbox": [ + 106, + 612, + 504, + 623 + ], + "score": 1.0, + "content": "based on learning to align unlabeled text and audio [Yeh et al., 2019] or adversarial learning [Liu", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 622, + 506, + 634 + ], + "spans": [ + { + "bbox": [ + 105, + 622, + 506, + 634 + ], + "score": 1.0, + "content": "et al., 2018, Chen et al., 2019]. These approaches showed promising initial results but their error", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 633, + 493, + 645 + ], + "spans": [ + { + "bbox": [ + 105, + 633, + 493, + 645 + ], + "score": 1.0, + "content": "rates are still high, with evaluation being limited to the small-scale and clean TIMIT benchmark.", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 32.5, + "bbox_fs": [ + 105, + 578, + 506, + 645 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 649, + 505, + 694 + ], + "lines": [ + { + "bbox": [ + 105, + 649, + 507, + 662 + ], + "spans": [ + { + "bbox": [ + 105, + 649, + 507, + 662 + ], + "score": 1.0, + "content": "In this work, we introduce a framework for unsupervised learning of speech recognition models.", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 659, + 506, + 673 + ], + "spans": [ + { + "bbox": [ + 105, + 659, + 506, + 673 + ], + "score": 1.0, + "content": "Wav2vec-U, or wav2vec Unsupervised, leverages self-supervised representations from wav2vec", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 671, + 506, + 684 + ], + "spans": [ + { + "bbox": [ + 105, + 671, + 506, + 684 + ], + "score": 1.0, + "content": "2.0 [Baevski et al., 2020c] to embed the speech audio and to segment the audio into units with a", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 682, + 505, + 695 + ], + "spans": [ + { + "bbox": [ + 105, + 682, + 135, + 695 + ], + "score": 1.0, + "content": "simple", + "type": "text" + }, + { + "bbox": [ + 135, + 682, + 141, + 692 + ], + "score": 0.46, + "content": "\\mathrm { k }", + "type": "inline_equation" + }, + { + "bbox": [ + 142, + 682, + 505, + 695 + ], + "score": 1.0, + "content": "-means clustering method (see Figure 1 for an illustration of our approach). We find that the", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 351, + 506, + 364 + ], + "spans": [ + { + "bbox": [ + 105, + 351, + 506, + 364 + ], + "score": 1.0, + "content": "quality of the audio representations is key to the success of unsupervised speech recognition. Similar", + "type": "text", + "cross_page": true + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 362, + 505, + 375 + ], + "spans": [ + { + "bbox": [ + 105, + 362, + 505, + 375 + ], + "score": 1.0, + "content": "to Liu et al. [2018] and Chen et al. [2019], we learn a mapping between segments and phonemes", + "type": "text", + "cross_page": true + } + ], + "index": 10 + }, + { + "bbox": [ + 106, + 373, + 505, + 385 + ], + "spans": [ + { + "bbox": [ + 106, + 373, + 505, + 385 + ], + "score": 1.0, + "content": "using adversarial training but different to their work, we also enable the algorithm to label segments", + "type": "text", + "cross_page": true + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 384, + 506, + 397 + ], + "spans": [ + { + "bbox": [ + 105, + 384, + 506, + 397 + ], + "score": 1.0, + "content": "as silences. We also introduce an unsupervised cross-validation metric to enable model development", + "type": "text", + "cross_page": true + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 394, + 505, + 408 + ], + "spans": [ + { + "bbox": [ + 105, + 394, + 505, + 408 + ], + "score": 1.0, + "content": "without labeled development data. Our unsupervised speech recognition model, the generator, is very", + "type": "text", + "cross_page": true + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 406, + 505, + 419 + ], + "spans": [ + { + "bbox": [ + 105, + 406, + 505, + 419 + ], + "score": 1.0, + "content": "lightweight: it consists of a single temporal convolution comprising only about 90k parameters to", + "type": "text", + "cross_page": true + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 417, + 316, + 429 + ], + "spans": [ + { + "bbox": [ + 106, + 417, + 316, + 429 + ], + "score": 1.0, + "content": "which we input frozen wav2vec 2.0 representations.", + "type": "text", + "cross_page": true + } + ], + "index": 15 + } + ], + "index": 37.5, + "bbox_fs": [ + 105, + 649, + 507, + 695 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "image", + "bbox": [ + 115, + 72, + 494, + 253 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 115, + 72, + 494, + 253 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 115, + 72, + 494, + 253 + ], + "spans": [ + { + "bbox": [ + 115, + 72, + 494, + 253 + ], + "score": 0.973, + "type": "image", + "image_path": "0b5b4f3465570676296b5d72eebf6d66830b97659f2ec44658268dac36522d53.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 115, + 72, + 494, + 132.33333333333334 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 115, + 132.33333333333334, + 494, + 192.66666666666669 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 115, + 192.66666666666669, + 494, + 253.00000000000003 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 259, + 505, + 326 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 259, + 505, + 272 + ], + "spans": [ + { + "bbox": [ + 106, + 259, + 505, + 272 + ], + "score": 1.0, + "content": "Figure 1: Illustration of wav2vec Unsupervised: we learn self-supervised representations with", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 270, + 506, + 283 + ], + "spans": [ + { + "bbox": [ + 105, + 270, + 469, + 283 + ], + "score": 1.0, + "content": "wav2vec 2.0 on unlabeled speech audio (Step 1), identify clusters in the representations with", + "type": "text" + }, + { + "bbox": [ + 469, + 271, + 476, + 281 + ], + "score": 0.4, + "content": "\\mathbf { k }", + "type": "inline_equation" + }, + { + "bbox": [ + 476, + 270, + 506, + 283 + ], + "score": 1.0, + "content": "-means", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 106, + 282, + 505, + 295 + ], + "spans": [ + { + "bbox": [ + 106, + 282, + 505, + 295 + ], + "score": 1.0, + "content": "(Step 2) to segment the audio (Step 3). Next, we build segment representations by mean pooling the", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 292, + 506, + 305 + ], + "spans": [ + { + "bbox": [ + 105, + 292, + 506, + 305 + ], + "score": 1.0, + "content": "wav2vec 2.0 representations, performing PCA and a second mean pooling step between adjacent", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 303, + 506, + 316 + ], + "spans": [ + { + "bbox": [ + 105, + 303, + 506, + 316 + ], + "score": 1.0, + "content": "segments (Step 4). This is input to the generator which outputs a phoneme sequence (Step 5) fed to", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 313, + 496, + 327 + ], + "spans": [ + { + "bbox": [ + 105, + 313, + 496, + 327 + ], + "score": 1.0, + "content": "the discriminator, similar to phonemized unlabeled text (Step 6), for adversarial training (Step 7).", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 5.5 + } + ], + "index": 3.25 + }, + { + "type": "text", + "bbox": [ + 106, + 351, + 505, + 428 + ], + "lines": [ + { + "bbox": [ + 105, + 351, + 506, + 364 + ], + "spans": [ + { + "bbox": [ + 105, + 351, + 506, + 364 + ], + "score": 1.0, + "content": "quality of the audio representations is key to the success of unsupervised speech recognition. Similar", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 362, + 505, + 375 + ], + "spans": [ + { + "bbox": [ + 105, + 362, + 505, + 375 + ], + "score": 1.0, + "content": "to Liu et al. [2018] and Chen et al. [2019], we learn a mapping between segments and phonemes", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 106, + 373, + 505, + 385 + ], + "spans": [ + { + "bbox": [ + 106, + 373, + 505, + 385 + ], + "score": 1.0, + "content": "using adversarial training but different to their work, we also enable the algorithm to label segments", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 384, + 506, + 397 + ], + "spans": [ + { + "bbox": [ + 105, + 384, + 506, + 397 + ], + "score": 1.0, + "content": "as silences. We also introduce an unsupervised cross-validation metric to enable model development", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 394, + 505, + 408 + ], + "spans": [ + { + "bbox": [ + 105, + 394, + 505, + 408 + ], + "score": 1.0, + "content": "without labeled development data. Our unsupervised speech recognition model, the generator, is very", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 406, + 505, + 419 + ], + "spans": [ + { + "bbox": [ + 105, + 406, + 505, + 419 + ], + "score": 1.0, + "content": "lightweight: it consists of a single temporal convolution comprising only about 90k parameters to", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 417, + 316, + 429 + ], + "spans": [ + { + "bbox": [ + 106, + 417, + 316, + 429 + ], + "score": 1.0, + "content": "which we input frozen wav2vec 2.0 representations.", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 12 + }, + { + "type": "text", + "bbox": [ + 107, + 433, + 506, + 510 + ], + "lines": [ + { + "bbox": [ + 105, + 432, + 507, + 447 + ], + "spans": [ + { + "bbox": [ + 105, + 432, + 507, + 447 + ], + "score": 1.0, + "content": "Experimental results demonstrate the viability of the framework for a variety of settings and languages.", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 444, + 506, + 457 + ], + "spans": [ + { + "bbox": [ + 106, + 444, + 506, + 457 + ], + "score": 1.0, + "content": "wav2vec-U improves the phone error rate (PER) on the small-scale TIMIT benchmark from 26.1", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 106, + 455, + 505, + 468 + ], + "spans": [ + { + "bbox": [ + 106, + 455, + 505, + 468 + ], + "score": 1.0, + "content": "to 11.3 compared to the next best known unsupervised approach. To get a better sense of the", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 466, + 506, + 479 + ], + "spans": [ + { + "bbox": [ + 105, + 466, + 506, + 479 + ], + "score": 1.0, + "content": "performance compared to the best supervised methods, we measure performance on the larger", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 106, + 477, + 505, + 488 + ], + "spans": [ + { + "bbox": [ + 106, + 477, + 505, + 488 + ], + "score": 1.0, + "content": "Librispeech benchmark where our method achieves word error rate (WER) 5.9 on test-other. We also", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 104, + 487, + 507, + 501 + ], + "spans": [ + { + "bbox": [ + 104, + 487, + 507, + 501 + ], + "score": 1.0, + "content": "evaluate on six other European languages of the multilingual Librispeech benchmark [Pratap et al.,", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 498, + 342, + 512 + ], + "spans": [ + { + "bbox": [ + 105, + 498, + 342, + 512 + ], + "score": 1.0, + "content": "2020] and on three non-European low-resource languages.", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 19 + }, + { + "type": "title", + "bbox": [ + 107, + 529, + 295, + 543 + ], + "lines": [ + { + "bbox": [ + 104, + 527, + 296, + 547 + ], + "spans": [ + { + "bbox": [ + 104, + 527, + 296, + 547 + ], + "score": 1.0, + "content": "2 Speech and Text Representations", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 23 + }, + { + "type": "text", + "bbox": [ + 106, + 557, + 505, + 579 + ], + "lines": [ + { + "bbox": [ + 105, + 555, + 506, + 571 + ], + "spans": [ + { + "bbox": [ + 105, + 555, + 506, + 571 + ], + "score": 1.0, + "content": "Next, we describe how we build suitable speech and text representations for unsupervised learning.", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 106, + 567, + 498, + 581 + ], + "spans": [ + { + "bbox": [ + 106, + 567, + 498, + 581 + ], + "score": 1.0, + "content": "Good representations are essential to learning a mapping from speech to text without supervision.", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 24.5 + }, + { + "type": "title", + "bbox": [ + 106, + 595, + 379, + 608 + ], + "lines": [ + { + "bbox": [ + 104, + 594, + 380, + 612 + ], + "spans": [ + { + "bbox": [ + 104, + 594, + 380, + 612 + ], + "score": 1.0, + "content": "2.1 Self-supervised Learning of Speech Audio Representations", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 26 + }, + { + "type": "text", + "bbox": [ + 107, + 617, + 505, + 662 + ], + "lines": [ + { + "bbox": [ + 105, + 617, + 507, + 631 + ], + "spans": [ + { + "bbox": [ + 105, + 617, + 507, + 631 + ], + "score": 1.0, + "content": "In the first step, we learn representations of the speech audio signal using self-supervised learning.", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 628, + 505, + 641 + ], + "spans": [ + { + "bbox": [ + 105, + 628, + 505, + 641 + ], + "score": 1.0, + "content": "There has been a lot of recent work in this direction which has shown strong performance in extremely", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 640, + 507, + 653 + ], + "spans": [ + { + "bbox": [ + 105, + 640, + 507, + 653 + ], + "score": 1.0, + "content": "low-labeled data setups across a range of languages [Conneau et al., 2020] and tasks [Fan et al., 2021,", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 106, + 651, + 264, + 663 + ], + "spans": [ + { + "bbox": [ + 106, + 651, + 264, + 663 + ], + "score": 1.0, + "content": "Pepino et al., 2021, Wang et al., 2021].", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 28.5 + }, + { + "type": "text", + "bbox": [ + 107, + 667, + 505, + 722 + ], + "lines": [ + { + "bbox": [ + 106, + 667, + 505, + 679 + ], + "spans": [ + { + "bbox": [ + 106, + 667, + 348, + 679 + ], + "score": 1.0, + "content": "Wav2vec 2.0 consists of a convolutional feature encoder", + "type": "text" + }, + { + "bbox": [ + 348, + 667, + 406, + 678 + ], + "score": 0.91, + "content": "f : \\mathcal X \\mapsto \\mathcal Z", + "type": "inline_equation" + }, + { + "bbox": [ + 406, + 667, + 505, + 679 + ], + "score": 1.0, + "content": "that maps a raw audio", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 677, + 506, + 691 + ], + "spans": [ + { + "bbox": [ + 105, + 677, + 145, + 691 + ], + "score": 1.0, + "content": "sequence", + "type": "text" + }, + { + "bbox": [ + 146, + 678, + 156, + 688 + ], + "score": 0.81, + "content": "X", + "type": "inline_equation" + }, + { + "bbox": [ + 156, + 677, + 285, + 691 + ], + "score": 1.0, + "content": "to latent speech representations", + "type": "text" + }, + { + "bbox": [ + 285, + 680, + 328, + 689 + ], + "score": 0.89, + "content": "z _ { 1 } , \\dots , z _ { T }", + "type": "inline_equation" + }, + { + "bbox": [ + 328, + 677, + 417, + 691 + ], + "score": 1.0, + "content": ", which a Transformer", + "type": "text" + }, + { + "bbox": [ + 418, + 678, + 462, + 689 + ], + "score": 0.92, + "content": "g : { \\mathcal { Z } } \\mapsto { \\mathcal { C } }", + "type": "inline_equation" + }, + { + "bbox": [ + 462, + 677, + 506, + 691 + ], + "score": 1.0, + "content": "then turns", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 689, + 506, + 701 + ], + "spans": [ + { + "bbox": [ + 105, + 689, + 220, + 701 + ], + "score": 1.0, + "content": "into context representations", + "type": "text" + }, + { + "bbox": [ + 221, + 690, + 263, + 700 + ], + "score": 0.87, + "content": "c _ { 1 } , \\ldots , c _ { T }", + "type": "inline_equation" + }, + { + "bbox": [ + 263, + 689, + 390, + 701 + ], + "score": 1.0, + "content": "[Baevski et al., 2020b,a]. Each", + "type": "text" + }, + { + "bbox": [ + 390, + 690, + 400, + 700 + ], + "score": 0.84, + "content": "z _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 401, + 689, + 469, + 701 + ], + "score": 1.0, + "content": "represents about", + "type": "text" + }, + { + "bbox": [ + 470, + 689, + 493, + 699 + ], + "score": 0.47, + "content": "2 5 \\mathrm { m s }", + "type": "inline_equation" + }, + { + "bbox": [ + 493, + 689, + 506, + 701 + ], + "score": 1.0, + "content": "of", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 106, + 700, + 504, + 712 + ], + "spans": [ + { + "bbox": [ + 106, + 700, + 172, + 712 + ], + "score": 1.0, + "content": "audio strided by", + "type": "text" + }, + { + "bbox": [ + 173, + 700, + 196, + 711 + ], + "score": 0.31, + "content": "2 0 \\mathrm { m s }", + "type": "inline_equation" + }, + { + "bbox": [ + 196, + 700, + 504, + 712 + ], + "score": 1.0, + "content": "and the Transformer architecture follows BERT [Vaswani et al., 2017, Devlin", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 711, + 505, + 724 + ], + "spans": [ + { + "bbox": [ + 105, + 711, + 382, + 724 + ], + "score": 1.0, + "content": "et al., 2019]. During training, latent representations are discretized to", + "type": "text" + }, + { + "bbox": [ + 383, + 712, + 425, + 722 + ], + "score": 0.9, + "content": "q _ { 1 } , \\ldots , q _ { T }", + "type": "inline_equation" + }, + { + "bbox": [ + 426, + 711, + 505, + 724 + ], + "score": 1.0, + "content": "with a quantization", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 33 + } + ], + "page_idx": 1, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 302, + 742, + 309, + 750 + ], + "lines": [ + { + "bbox": [ + 301, + 741, + 310, + 753 + ], + "spans": [ + { + "bbox": [ + 301, + 741, + 310, + 753 + ], + "score": 1.0, + "content": "2", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "image", + "bbox": [ + 115, + 72, + 494, + 253 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 115, + 72, + 494, + 253 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 115, + 72, + 494, + 253 + ], + "spans": [ + { + "bbox": [ + 115, + 72, + 494, + 253 + ], + "score": 0.973, + "type": "image", + "image_path": "0b5b4f3465570676296b5d72eebf6d66830b97659f2ec44658268dac36522d53.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 115, + 72, + 494, + 132.33333333333334 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 115, + 132.33333333333334, + 494, + 192.66666666666669 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 115, + 192.66666666666669, + 494, + 253.00000000000003 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 259, + 505, + 326 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 259, + 505, + 272 + ], + "spans": [ + { + "bbox": [ + 106, + 259, + 505, + 272 + ], + "score": 1.0, + "content": "Figure 1: Illustration of wav2vec Unsupervised: we learn self-supervised representations with", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 270, + 506, + 283 + ], + "spans": [ + { + "bbox": [ + 105, + 270, + 469, + 283 + ], + "score": 1.0, + "content": "wav2vec 2.0 on unlabeled speech audio (Step 1), identify clusters in the representations with", + "type": "text" + }, + { + "bbox": [ + 469, + 271, + 476, + 281 + ], + "score": 0.4, + "content": "\\mathbf { k }", + "type": "inline_equation" + }, + { + "bbox": [ + 476, + 270, + 506, + 283 + ], + "score": 1.0, + "content": "-means", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 106, + 282, + 505, + 295 + ], + "spans": [ + { + "bbox": [ + 106, + 282, + 505, + 295 + ], + "score": 1.0, + "content": "(Step 2) to segment the audio (Step 3). Next, we build segment representations by mean pooling the", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 292, + 506, + 305 + ], + "spans": [ + { + "bbox": [ + 105, + 292, + 506, + 305 + ], + "score": 1.0, + "content": "wav2vec 2.0 representations, performing PCA and a second mean pooling step between adjacent", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 303, + 506, + 316 + ], + "spans": [ + { + "bbox": [ + 105, + 303, + 506, + 316 + ], + "score": 1.0, + "content": "segments (Step 4). This is input to the generator which outputs a phoneme sequence (Step 5) fed to", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 313, + 496, + 327 + ], + "spans": [ + { + "bbox": [ + 105, + 313, + 496, + 327 + ], + "score": 1.0, + "content": "the discriminator, similar to phonemized unlabeled text (Step 6), for adversarial training (Step 7).", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 5.5 + } + ], + "index": 3.25 + }, + { + "type": "text", + "bbox": [ + 106, + 351, + 505, + 428 + ], + "lines": [], + "index": 12, + "bbox_fs": [ + 105, + 351, + 506, + 429 + ], + "lines_deleted": true + }, + { + "type": "text", + "bbox": [ + 107, + 433, + 506, + 510 + ], + "lines": [ + { + "bbox": [ + 105, + 432, + 507, + 447 + ], + "spans": [ + { + "bbox": [ + 105, + 432, + 507, + 447 + ], + "score": 1.0, + "content": "Experimental results demonstrate the viability of the framework for a variety of settings and languages.", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 444, + 506, + 457 + ], + "spans": [ + { + "bbox": [ + 106, + 444, + 506, + 457 + ], + "score": 1.0, + "content": "wav2vec-U improves the phone error rate (PER) on the small-scale TIMIT benchmark from 26.1", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 106, + 455, + 505, + 468 + ], + "spans": [ + { + "bbox": [ + 106, + 455, + 505, + 468 + ], + "score": 1.0, + "content": "to 11.3 compared to the next best known unsupervised approach. To get a better sense of the", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 466, + 506, + 479 + ], + "spans": [ + { + "bbox": [ + 105, + 466, + 506, + 479 + ], + "score": 1.0, + "content": "performance compared to the best supervised methods, we measure performance on the larger", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 106, + 477, + 505, + 488 + ], + "spans": [ + { + "bbox": [ + 106, + 477, + 505, + 488 + ], + "score": 1.0, + "content": "Librispeech benchmark where our method achieves word error rate (WER) 5.9 on test-other. We also", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 104, + 487, + 507, + 501 + ], + "spans": [ + { + "bbox": [ + 104, + 487, + 507, + 501 + ], + "score": 1.0, + "content": "evaluate on six other European languages of the multilingual Librispeech benchmark [Pratap et al.,", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 498, + 342, + 512 + ], + "spans": [ + { + "bbox": [ + 105, + 498, + 342, + 512 + ], + "score": 1.0, + "content": "2020] and on three non-European low-resource languages.", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 19, + "bbox_fs": [ + 104, + 432, + 507, + 512 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 529, + 295, + 543 + ], + "lines": [ + { + "bbox": [ + 104, + 527, + 296, + 547 + ], + "spans": [ + { + "bbox": [ + 104, + 527, + 296, + 547 + ], + "score": 1.0, + "content": "2 Speech and Text Representations", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 23 + }, + { + "type": "list", + "bbox": [ + 106, + 557, + 505, + 579 + ], + "lines": [ + { + "bbox": [ + 105, + 555, + 506, + 571 + ], + "spans": [ + { + "bbox": [ + 105, + 555, + 506, + 571 + ], + "score": 1.0, + "content": "Next, we describe how we build suitable speech and text representations for unsupervised learning.", + "type": "text" + } + ], + "index": 24, + "is_list_end_line": true + }, + { + "bbox": [ + 106, + 567, + 498, + 581 + ], + "spans": [ + { + "bbox": [ + 106, + 567, + 498, + 581 + ], + "score": 1.0, + "content": "Good representations are essential to learning a mapping from speech to text without supervision.", + "type": "text" + } + ], + "index": 25, + "is_list_start_line": true, + "is_list_end_line": true + } + ], + "index": 24.5, + "bbox_fs": [ + 105, + 555, + 506, + 581 + ] + }, + { + "type": "title", + "bbox": [ + 106, + 595, + 379, + 608 + ], + "lines": [ + { + "bbox": [ + 104, + 594, + 380, + 612 + ], + "spans": [ + { + "bbox": [ + 104, + 594, + 380, + 612 + ], + "score": 1.0, + "content": "2.1 Self-supervised Learning of Speech Audio Representations", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 26 + }, + { + "type": "text", + "bbox": [ + 107, + 617, + 505, + 662 + ], + "lines": [ + { + "bbox": [ + 105, + 617, + 507, + 631 + ], + "spans": [ + { + "bbox": [ + 105, + 617, + 507, + 631 + ], + "score": 1.0, + "content": "In the first step, we learn representations of the speech audio signal using self-supervised learning.", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 628, + 505, + 641 + ], + "spans": [ + { + "bbox": [ + 105, + 628, + 505, + 641 + ], + "score": 1.0, + "content": "There has been a lot of recent work in this direction which has shown strong performance in extremely", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 640, + 507, + 653 + ], + "spans": [ + { + "bbox": [ + 105, + 640, + 507, + 653 + ], + "score": 1.0, + "content": "low-labeled data setups across a range of languages [Conneau et al., 2020] and tasks [Fan et al., 2021,", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 106, + 651, + 264, + 663 + ], + "spans": [ + { + "bbox": [ + 106, + 651, + 264, + 663 + ], + "score": 1.0, + "content": "Pepino et al., 2021, Wang et al., 2021].", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 28.5, + "bbox_fs": [ + 105, + 617, + 507, + 663 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 667, + 505, + 722 + ], + "lines": [ + { + "bbox": [ + 106, + 667, + 505, + 679 + ], + "spans": [ + { + "bbox": [ + 106, + 667, + 348, + 679 + ], + "score": 1.0, + "content": "Wav2vec 2.0 consists of a convolutional feature encoder", + "type": "text" + }, + { + "bbox": [ + 348, + 667, + 406, + 678 + ], + "score": 0.91, + "content": "f : \\mathcal X \\mapsto \\mathcal Z", + "type": "inline_equation" + }, + { + "bbox": [ + 406, + 667, + 505, + 679 + ], + "score": 1.0, + "content": "that maps a raw audio", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 677, + 506, + 691 + ], + "spans": [ + { + "bbox": [ + 105, + 677, + 145, + 691 + ], + "score": 1.0, + "content": "sequence", + "type": "text" + }, + { + "bbox": [ + 146, + 678, + 156, + 688 + ], + "score": 0.81, + "content": "X", + "type": "inline_equation" + }, + { + "bbox": [ + 156, + 677, + 285, + 691 + ], + "score": 1.0, + "content": "to latent speech representations", + "type": "text" + }, + { + "bbox": [ + 285, + 680, + 328, + 689 + ], + "score": 0.89, + "content": "z _ { 1 } , \\dots , z _ { T }", + "type": "inline_equation" + }, + { + "bbox": [ + 328, + 677, + 417, + 691 + ], + "score": 1.0, + "content": ", which a Transformer", + "type": "text" + }, + { + "bbox": [ + 418, + 678, + 462, + 689 + ], + "score": 0.92, + "content": "g : { \\mathcal { Z } } \\mapsto { \\mathcal { C } }", + "type": "inline_equation" + }, + { + "bbox": [ + 462, + 677, + 506, + 691 + ], + "score": 1.0, + "content": "then turns", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 689, + 506, + 701 + ], + "spans": [ + { + "bbox": [ + 105, + 689, + 220, + 701 + ], + "score": 1.0, + "content": "into context representations", + "type": "text" + }, + { + "bbox": [ + 221, + 690, + 263, + 700 + ], + "score": 0.87, + "content": "c _ { 1 } , \\ldots , c _ { T }", + "type": "inline_equation" + }, + { + "bbox": [ + 263, + 689, + 390, + 701 + ], + "score": 1.0, + "content": "[Baevski et al., 2020b,a]. Each", + "type": "text" + }, + { + "bbox": [ + 390, + 690, + 400, + 700 + ], + "score": 0.84, + "content": "z _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 401, + 689, + 469, + 701 + ], + "score": 1.0, + "content": "represents about", + "type": "text" + }, + { + "bbox": [ + 470, + 689, + 493, + 699 + ], + "score": 0.47, + "content": "2 5 \\mathrm { m s }", + "type": "inline_equation" + }, + { + "bbox": [ + 493, + 689, + 506, + 701 + ], + "score": 1.0, + "content": "of", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 106, + 700, + 504, + 712 + ], + "spans": [ + { + "bbox": [ + 106, + 700, + 172, + 712 + ], + "score": 1.0, + "content": "audio strided by", + "type": "text" + }, + { + "bbox": [ + 173, + 700, + 196, + 711 + ], + "score": 0.31, + "content": "2 0 \\mathrm { m s }", + "type": "inline_equation" + }, + { + "bbox": [ + 196, + 700, + 504, + 712 + ], + "score": 1.0, + "content": "and the Transformer architecture follows BERT [Vaswani et al., 2017, Devlin", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 711, + 505, + 724 + ], + "spans": [ + { + "bbox": [ + 105, + 711, + 382, + 724 + ], + "score": 1.0, + "content": "et al., 2019]. During training, latent representations are discretized to", + "type": "text" + }, + { + "bbox": [ + 383, + 712, + 425, + 722 + ], + "score": 0.9, + "content": "q _ { 1 } , \\ldots , q _ { T }", + "type": "inline_equation" + }, + { + "bbox": [ + 426, + 711, + 505, + 724 + ], + "score": 1.0, + "content": "with a quantization", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 106, + 72, + 505, + 85 + ], + "spans": [ + { + "bbox": [ + 106, + 72, + 140, + 85 + ], + "score": 1.0, + "content": "module", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 140, + 73, + 174, + 83 + ], + "score": 0.91, + "content": "{ \\mathcal { Z } } \\mapsto { \\mathcal { Q } }", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 175, + 72, + 505, + 85 + ], + "score": 1.0, + "content": "to represent the targets in the objective. Quantization uses a Gumbel softmax to", + "type": "text", + "cross_page": true + } + ], + "index": 0 + }, + { + "bbox": [ + 106, + 83, + 491, + 96 + ], + "spans": [ + { + "bbox": [ + 106, + 83, + 491, + 96 + ], + "score": 1.0, + "content": "choose entries from two codebooks [Jegou et al., 2011, Jang et al., 2016, Baevski et al., 2020b].", + "type": "text", + "cross_page": true + } + ], + "index": 1 + } + ], + "index": 33, + "bbox_fs": [ + 105, + 667, + 506, + 724 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 106, + 73, + 504, + 95 + ], + "lines": [ + { + "bbox": [ + 106, + 72, + 505, + 85 + ], + "spans": [ + { + "bbox": [ + 106, + 72, + 140, + 85 + ], + "score": 1.0, + "content": "module", + "type": "text" + }, + { + "bbox": [ + 140, + 73, + 174, + 83 + ], + "score": 0.91, + "content": "{ \\mathcal { Z } } \\mapsto { \\mathcal { Q } }", + "type": "inline_equation" + }, + { + "bbox": [ + 175, + 72, + 505, + 85 + ], + "score": 1.0, + "content": "to represent the targets in the objective. Quantization uses a Gumbel softmax to", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 106, + 83, + 491, + 96 + ], + "spans": [ + { + "bbox": [ + 106, + 83, + 491, + 96 + ], + "score": 1.0, + "content": "choose entries from two codebooks [Jegou et al., 2011, Jang et al., 2016, Baevski et al., 2020b].", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5 + }, + { + "type": "text", + "bbox": [ + 108, + 100, + 504, + 133 + ], + "lines": [ + { + "bbox": [ + 105, + 100, + 505, + 112 + ], + "spans": [ + { + "bbox": [ + 105, + 100, + 426, + 112 + ], + "score": 1.0, + "content": "In our experiments, we use the publicly available English model pre-trained on", + "type": "text" + }, + { + "bbox": [ + 427, + 100, + 444, + 110 + ], + "score": 0.38, + "content": "5 3 \\mathrm { k }", + "type": "inline_equation" + }, + { + "bbox": [ + 444, + 100, + 505, + 112 + ], + "score": 1.0, + "content": "hours of Libri-", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 106, + 111, + 505, + 123 + ], + "spans": [ + { + "bbox": [ + 106, + 111, + 505, + 123 + ], + "score": 1.0, + "content": "Light [Kahn et al., 2020b] as well as XLSR-53 which was pre-trained on nearly 60k hours of speech", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 121, + 290, + 134 + ], + "spans": [ + { + "bbox": [ + 105, + 121, + 290, + 134 + ], + "score": 1.0, + "content": "audio in 53 languages [Conneau et al., 2020].", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 3 + }, + { + "type": "title", + "bbox": [ + 107, + 146, + 331, + 158 + ], + "lines": [ + { + "bbox": [ + 105, + 145, + 332, + 160 + ], + "spans": [ + { + "bbox": [ + 105, + 145, + 332, + 160 + ], + "score": 1.0, + "content": "2.2 Pre-processing and Embedding the Audio Data", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 5 + }, + { + "type": "text", + "bbox": [ + 107, + 166, + 505, + 221 + ], + "lines": [ + { + "bbox": [ + 106, + 166, + 506, + 178 + ], + "spans": [ + { + "bbox": [ + 106, + 166, + 506, + 178 + ], + "score": 1.0, + "content": "Removing Silences. Most datasets we use for our experiments have audio data with silences.", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 106, + 177, + 505, + 189 + ], + "spans": [ + { + "bbox": [ + 106, + 177, + 505, + 189 + ], + "score": 1.0, + "content": "However, these parts of the audio do not correspond to any transcription and we therefore remove", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 188, + 505, + 200 + ], + "spans": [ + { + "bbox": [ + 105, + 188, + 505, + 200 + ], + "score": 1.0, + "content": "silences as much as possible. We apply rVAD, an unsupervised voice activity detection (VAD) model", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 106, + 199, + 505, + 211 + ], + "spans": [ + { + "bbox": [ + 106, + 199, + 505, + 211 + ], + "score": 1.0, + "content": "which determines the segments in the audio data corresponding to silences, and we remove these", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 209, + 366, + 222 + ], + "spans": [ + { + "bbox": [ + 105, + 209, + 366, + 222 + ], + "score": 1.0, + "content": "sections [Tan et al., 2020]. We ablate this choice in Appendix C.", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 8 + }, + { + "type": "text", + "bbox": [ + 106, + 232, + 505, + 321 + ], + "lines": [ + { + "bbox": [ + 106, + 233, + 505, + 245 + ], + "spans": [ + { + "bbox": [ + 106, + 233, + 505, + 245 + ], + "score": 1.0, + "content": "Speech Audio Representations. After silence removal, we embed the unlabeled speech audio with", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 244, + 506, + 256 + ], + "spans": [ + { + "bbox": [ + 105, + 244, + 506, + 256 + ], + "score": 1.0, + "content": "wav2vec 2.0 to obtain speech representations. Specifically, we use the representations of the context", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 253, + 506, + 267 + ], + "spans": [ + { + "bbox": [ + 105, + 253, + 194, + 267 + ], + "score": 1.0, + "content": "Transformer network", + "type": "text" + }, + { + "bbox": [ + 195, + 256, + 237, + 266 + ], + "score": 0.86, + "content": "c _ { 1 } , \\ldots , c _ { T }", + "type": "inline_equation" + }, + { + "bbox": [ + 240, + 255, + 266, + 266 + ], + "score": 0.66, + "content": "( \\ S 2 . 1 )", + "type": "inline_equation" + }, + { + "bbox": [ + 267, + 253, + 506, + 267 + ], + "score": 1.0, + "content": ". The context network contains 24 Transformer blocks and", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 264, + 505, + 279 + ], + "spans": [ + { + "bbox": [ + 105, + 264, + 232, + 279 + ], + "score": 1.0, + "content": "we denote the output of block", + "type": "text" + }, + { + "bbox": [ + 232, + 266, + 236, + 275 + ], + "score": 0.69, + "content": "l", + "type": "inline_equation" + }, + { + "bbox": [ + 237, + 264, + 289, + 279 + ], + "score": 1.0, + "content": "at time-step", + "type": "text" + }, + { + "bbox": [ + 289, + 266, + 294, + 275 + ], + "score": 0.63, + "content": "t", + "type": "inline_equation" + }, + { + "bbox": [ + 295, + 264, + 307, + 279 + ], + "score": 1.0, + "content": "as", + "type": "text" + }, + { + "bbox": [ + 307, + 265, + 316, + 277 + ], + "score": 0.87, + "content": "c _ { t } ^ { l }", + "type": "inline_equation" + }, + { + "bbox": [ + 316, + 264, + 505, + 279 + ], + "score": 1.0, + "content": ". Our goal is to learn a model which can map", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 276, + 505, + 290 + ], + "spans": [ + { + "bbox": [ + 105, + 277, + 214, + 290 + ], + "score": 1.0, + "content": "from audio representations", + "type": "text" + }, + { + "bbox": [ + 214, + 276, + 223, + 289 + ], + "score": 0.89, + "content": "c _ { t } ^ { l }", + "type": "inline_equation" + }, + { + "bbox": [ + 224, + 277, + 505, + 290 + ], + "score": 1.0, + "content": "to phonemes using no supervision. However, the representations of the", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 288, + 505, + 301 + ], + "spans": [ + { + "bbox": [ + 105, + 288, + 505, + 301 + ], + "score": 1.0, + "content": "uppermost block of wav2vec 2.0 may not be well suited for this task. These features are trained to", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 299, + 506, + 312 + ], + "spans": [ + { + "bbox": [ + 106, + 299, + 326, + 312 + ], + "score": 1.0, + "content": "directly predict masked latent representations spanning", + "type": "text" + }, + { + "bbox": [ + 326, + 299, + 349, + 310 + ], + "score": 0.55, + "content": "2 5 \\mathrm { m s }", + "type": "inline_equation" + }, + { + "bbox": [ + 350, + 299, + 506, + 312 + ], + "score": 1.0, + "content": "of speech audio which is much shorter", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 310, + 266, + 323 + ], + "spans": [ + { + "bbox": [ + 105, + 310, + 266, + 323 + ], + "score": 1.0, + "content": "than the typical duration of a phoneme.", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 14.5 + }, + { + "type": "text", + "bbox": [ + 106, + 326, + 505, + 436 + ], + "lines": [ + { + "bbox": [ + 106, + 327, + 505, + 339 + ], + "spans": [ + { + "bbox": [ + 106, + 327, + 505, + 339 + ], + "score": 1.0, + "content": "To get a better sense of this, we train supervised phoneme recognizers with a CTC loss [Graves", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 337, + 506, + 349 + ], + "spans": [ + { + "bbox": [ + 105, + 337, + 506, + 349 + ], + "score": 1.0, + "content": "et al., 2006] on top of the frozen representations of each of the 24 blocks of the English wav2vec 2.0", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 348, + 506, + 361 + ], + "spans": [ + { + "bbox": [ + 105, + 348, + 506, + 361 + ], + "score": 1.0, + "content": "LARGE model pre-trained on Libri-Light. We then evaluate phone error rate (PER) with respect to", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 358, + 506, + 372 + ], + "spans": [ + { + "bbox": [ + 105, + 358, + 437, + 372 + ], + "score": 1.0, + "content": "the phonemized transcriptions of Librispeech dev-other. The classifier takes as input", + "type": "text" + }, + { + "bbox": [ + 437, + 359, + 446, + 371 + ], + "score": 0.88, + "content": "c _ { t } ^ { l }", + "type": "inline_equation" + }, + { + "bbox": [ + 446, + 358, + 506, + 372 + ], + "score": 1.0, + "content": "and contains a", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 106, + 370, + 506, + 383 + ], + "spans": [ + { + "bbox": [ + 106, + 370, + 506, + 383 + ], + "score": 1.0, + "content": "single softmax-normalized linear layer mapping to the phoneme inventory. Figure 2 shows that most", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 380, + 506, + 394 + ], + "spans": [ + { + "bbox": [ + 105, + 380, + 506, + 394 + ], + "score": 1.0, + "content": "of the first ten blocks as well as the final blocks provide very poor performance, while blocks 15-19", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 391, + 506, + 405 + ], + "spans": [ + { + "bbox": [ + 105, + 391, + 212, + 405 + ], + "score": 1.0, + "content": "provide error rates below", + "type": "text" + }, + { + "bbox": [ + 212, + 392, + 227, + 402 + ], + "score": 0.86, + "content": "9 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 227, + 391, + 416, + 405 + ], + "score": 1.0, + "content": "PER. Block 15 achieves the best error rate of", + "type": "text" + }, + { + "bbox": [ + 416, + 392, + 439, + 402 + ], + "score": 0.87, + "content": "7 . 5 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 439, + 391, + 506, + 405 + ], + "score": 1.0, + "content": "PER. A similar", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 403, + 505, + 415 + ], + "spans": [ + { + "bbox": [ + 105, + 403, + 505, + 415 + ], + "score": 1.0, + "content": "insight has been used in the concurrent work of Hsu et al. [2021b]. Appendix A shows that this", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 414, + 505, + 426 + ], + "spans": [ + { + "bbox": [ + 106, + 414, + 411, + 426 + ], + "score": 1.0, + "content": "choice generalizes to other languages. For brevity we drop the superscript", + "type": "text" + }, + { + "bbox": [ + 411, + 414, + 416, + 424 + ], + "score": 0.49, + "content": "l", + "type": "inline_equation" + }, + { + "bbox": [ + 416, + 414, + 505, + 426 + ], + "score": 1.0, + "content": "and refer to block 15", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 424, + 257, + 438 + ], + "spans": [ + { + "bbox": [ + 105, + 424, + 209, + 438 + ], + "score": 1.0, + "content": "representations simply as", + "type": "text" + }, + { + "bbox": [ + 210, + 426, + 252, + 436 + ], + "score": 0.89, + "content": "c _ { 1 } , \\ldots , c _ { T }", + "type": "inline_equation" + }, + { + "bbox": [ + 253, + 424, + 257, + 438 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 23.5 + }, + { + "type": "image", + "bbox": [ + 108, + 448, + 492, + 532 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 108, + 448, + 492, + 532 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 108, + 448, + 492, + 532 + ], + "spans": [ + { + "bbox": [ + 108, + 448, + 492, + 532 + ], + "score": 0.965, + "type": "image", + "image_path": "4893afedc781e5fe6ccd0b0db23368e07e04f56d069fce1ae7128f54dac18113.jpg" + } + ] + } + ], + "index": 30, + "virtual_lines": [ + { + "bbox": [ + 108, + 448, + 492, + 476.0 + ], + "spans": [], + "index": 29 + }, + { + "bbox": [ + 108, + 476.0, + 492, + 504.0 + ], + "spans": [], + "index": 30 + }, + { + "bbox": [ + 108, + 504.0, + 492, + 532.0 + ], + "spans": [], + "index": 31 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 107, + 538, + 504, + 572 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 538, + 505, + 552 + ], + "spans": [ + { + "bbox": [ + 105, + 538, + 505, + 552 + ], + "score": 1.0, + "content": "Figure 2: Supervised phoneme recognition using representations from different wav2vec 2.0 blocks", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 550, + 505, + 562 + ], + "spans": [ + { + "bbox": [ + 105, + 550, + 505, + 562 + ], + "score": 1.0, + "content": "on dev-other of English Librispeech. Low and high blocks do not provide good features, while as", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 560, + 274, + 573 + ], + "spans": [ + { + "bbox": [ + 105, + 560, + 274, + 573 + ], + "score": 1.0, + "content": "blocks 14-19 do. Block 15 performs best.", + "type": "text" + } + ], + "index": 34 + } + ], + "index": 33 + } + ], + "index": 31.5 + }, + { + "type": "title", + "bbox": [ + 108, + 592, + 253, + 604 + ], + "lines": [ + { + "bbox": [ + 105, + 591, + 254, + 606 + ], + "spans": [ + { + "bbox": [ + 105, + 591, + 254, + 606 + ], + "score": 1.0, + "content": "2.3 Segmenting the Audio Signal", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 35 + }, + { + "type": "text", + "bbox": [ + 107, + 612, + 505, + 722 + ], + "lines": [ + { + "bbox": [ + 106, + 612, + 505, + 625 + ], + "spans": [ + { + "bbox": [ + 106, + 612, + 505, + 625 + ], + "score": 1.0, + "content": "Once the speech signal is embedded, we identify segments corresponding to meaningful units that", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 622, + 506, + 637 + ], + "spans": [ + { + "bbox": [ + 105, + 622, + 506, + 637 + ], + "score": 1.0, + "content": "can be mapped to phonemes. Segmentation has been shown to be crucial in prior work [Chung", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 634, + 505, + 647 + ], + "spans": [ + { + "bbox": [ + 105, + 634, + 505, + 647 + ], + "score": 1.0, + "content": "et al., 2018] since the right boundaries in the input representations make it more aligned to phonetic", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 645, + 506, + 658 + ], + "spans": [ + { + "bbox": [ + 105, + 645, + 506, + 658 + ], + "score": 1.0, + "content": "sequences. There has been a lot of prior work in unsupervised speech segmentation [Kamper et al.,", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 655, + 506, + 668 + ], + "spans": [ + { + "bbox": [ + 105, + 655, + 506, + 668 + ], + "score": 1.0, + "content": "2017a,b, Rasanen et al., 2015, Kreuk et al., 2020] but here we simply use a method based on", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 667, + 506, + 681 + ], + "spans": [ + { + "bbox": [ + 105, + 667, + 304, + 681 + ], + "score": 1.0, + "content": "clustering the wav2vec 2.0 speech representations", + "type": "text" + }, + { + "bbox": [ + 304, + 669, + 347, + 678 + ], + "score": 0.87, + "content": "c _ { 1 } , \\ldots , c _ { T }", + "type": "inline_equation" + }, + { + "bbox": [ + 347, + 667, + 506, + 681 + ], + "score": 1.0, + "content": ". In a first step, we collect all the speech", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 104, + 678, + 504, + 690 + ], + "spans": [ + { + "bbox": [ + 104, + 678, + 465, + 690 + ], + "score": 1.0, + "content": "representations for the unlabeled speech data and perform k-means clustering to identify", + "type": "text" + }, + { + "bbox": [ + 466, + 678, + 504, + 688 + ], + "score": 0.89, + "content": "K = 1 2 8", + "type": "inline_equation" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 689, + 505, + 702 + ], + "spans": [ + { + "bbox": [ + 105, + 689, + 505, + 702 + ], + "score": 1.0, + "content": "clusters. We use the FAISS library to do fast clustering on GPUs [Johnson et al., 2019]. Next, each", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 106, + 700, + 506, + 713 + ], + "spans": [ + { + "bbox": [ + 106, + 701, + 115, + 711 + ], + "score": 0.84, + "content": "c _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 116, + 700, + 297, + 713 + ], + "score": 1.0, + "content": "is labeled with the corresponding cluster ID", + "type": "text" + }, + { + "bbox": [ + 298, + 700, + 364, + 712 + ], + "score": 0.84, + "content": "i _ { t } \\in \\{ 1 , \\ldots , K \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 365, + 700, + 506, + 713 + ], + "score": 1.0, + "content": "and we introduce speech segment", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 710, + 287, + 724 + ], + "spans": [ + { + "bbox": [ + 105, + 710, + 287, + 724 + ], + "score": 1.0, + "content": "boundaries whenever the cluster ID changes.", + "type": "text" + } + ], + "index": 45 + } + ], + "index": 40.5 + } + ], + "page_idx": 2, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 302, + 741, + 309, + 750 + ], + "lines": [ + { + "bbox": [ + 301, + 740, + 310, + 752 + ], + "spans": [ + { + "bbox": [ + 301, + 740, + 310, + 752 + ], + "score": 1.0, + "content": "3", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 106, + 73, + 504, + 95 + ], + "lines": [], + "index": 0.5, + "bbox_fs": [ + 106, + 72, + 505, + 96 + ], + "lines_deleted": true + }, + { + "type": "text", + "bbox": [ + 108, + 100, + 504, + 133 + ], + "lines": [ + { + "bbox": [ + 105, + 100, + 505, + 112 + ], + "spans": [ + { + "bbox": [ + 105, + 100, + 426, + 112 + ], + "score": 1.0, + "content": "In our experiments, we use the publicly available English model pre-trained on", + "type": "text" + }, + { + "bbox": [ + 427, + 100, + 444, + 110 + ], + "score": 0.38, + "content": "5 3 \\mathrm { k }", + "type": "inline_equation" + }, + { + "bbox": [ + 444, + 100, + 505, + 112 + ], + "score": 1.0, + "content": "hours of Libri-", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 106, + 111, + 505, + 123 + ], + "spans": [ + { + "bbox": [ + 106, + 111, + 505, + 123 + ], + "score": 1.0, + "content": "Light [Kahn et al., 2020b] as well as XLSR-53 which was pre-trained on nearly 60k hours of speech", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 121, + 290, + 134 + ], + "spans": [ + { + "bbox": [ + 105, + 121, + 290, + 134 + ], + "score": 1.0, + "content": "audio in 53 languages [Conneau et al., 2020].", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 3, + "bbox_fs": [ + 105, + 100, + 505, + 134 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 146, + 331, + 158 + ], + "lines": [ + { + "bbox": [ + 105, + 145, + 332, + 160 + ], + "spans": [ + { + "bbox": [ + 105, + 145, + 332, + 160 + ], + "score": 1.0, + "content": "2.2 Pre-processing and Embedding the Audio Data", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 5 + }, + { + "type": "text", + "bbox": [ + 107, + 166, + 505, + 221 + ], + "lines": [ + { + "bbox": [ + 106, + 166, + 506, + 178 + ], + "spans": [ + { + "bbox": [ + 106, + 166, + 506, + 178 + ], + "score": 1.0, + "content": "Removing Silences. Most datasets we use for our experiments have audio data with silences.", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 106, + 177, + 505, + 189 + ], + "spans": [ + { + "bbox": [ + 106, + 177, + 505, + 189 + ], + "score": 1.0, + "content": "However, these parts of the audio do not correspond to any transcription and we therefore remove", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 188, + 505, + 200 + ], + "spans": [ + { + "bbox": [ + 105, + 188, + 505, + 200 + ], + "score": 1.0, + "content": "silences as much as possible. We apply rVAD, an unsupervised voice activity detection (VAD) model", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 106, + 199, + 505, + 211 + ], + "spans": [ + { + "bbox": [ + 106, + 199, + 505, + 211 + ], + "score": 1.0, + "content": "which determines the segments in the audio data corresponding to silences, and we remove these", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 209, + 366, + 222 + ], + "spans": [ + { + "bbox": [ + 105, + 209, + 366, + 222 + ], + "score": 1.0, + "content": "sections [Tan et al., 2020]. We ablate this choice in Appendix C.", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 8, + "bbox_fs": [ + 105, + 166, + 506, + 222 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 232, + 505, + 321 + ], + "lines": [ + { + "bbox": [ + 106, + 233, + 505, + 245 + ], + "spans": [ + { + "bbox": [ + 106, + 233, + 505, + 245 + ], + "score": 1.0, + "content": "Speech Audio Representations. After silence removal, we embed the unlabeled speech audio with", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 244, + 506, + 256 + ], + "spans": [ + { + "bbox": [ + 105, + 244, + 506, + 256 + ], + "score": 1.0, + "content": "wav2vec 2.0 to obtain speech representations. Specifically, we use the representations of the context", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 253, + 506, + 267 + ], + "spans": [ + { + "bbox": [ + 105, + 253, + 194, + 267 + ], + "score": 1.0, + "content": "Transformer network", + "type": "text" + }, + { + "bbox": [ + 195, + 256, + 237, + 266 + ], + "score": 0.86, + "content": "c _ { 1 } , \\ldots , c _ { T }", + "type": "inline_equation" + }, + { + "bbox": [ + 240, + 255, + 266, + 266 + ], + "score": 0.66, + "content": "( \\ S 2 . 1 )", + "type": "inline_equation" + }, + { + "bbox": [ + 267, + 253, + 506, + 267 + ], + "score": 1.0, + "content": ". The context network contains 24 Transformer blocks and", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 264, + 505, + 279 + ], + "spans": [ + { + "bbox": [ + 105, + 264, + 232, + 279 + ], + "score": 1.0, + "content": "we denote the output of block", + "type": "text" + }, + { + "bbox": [ + 232, + 266, + 236, + 275 + ], + "score": 0.69, + "content": "l", + "type": "inline_equation" + }, + { + "bbox": [ + 237, + 264, + 289, + 279 + ], + "score": 1.0, + "content": "at time-step", + "type": "text" + }, + { + "bbox": [ + 289, + 266, + 294, + 275 + ], + "score": 0.63, + "content": "t", + "type": "inline_equation" + }, + { + "bbox": [ + 295, + 264, + 307, + 279 + ], + "score": 1.0, + "content": "as", + "type": "text" + }, + { + "bbox": [ + 307, + 265, + 316, + 277 + ], + "score": 0.87, + "content": "c _ { t } ^ { l }", + "type": "inline_equation" + }, + { + "bbox": [ + 316, + 264, + 505, + 279 + ], + "score": 1.0, + "content": ". Our goal is to learn a model which can map", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 276, + 505, + 290 + ], + "spans": [ + { + "bbox": [ + 105, + 277, + 214, + 290 + ], + "score": 1.0, + "content": "from audio representations", + "type": "text" + }, + { + "bbox": [ + 214, + 276, + 223, + 289 + ], + "score": 0.89, + "content": "c _ { t } ^ { l }", + "type": "inline_equation" + }, + { + "bbox": [ + 224, + 277, + 505, + 290 + ], + "score": 1.0, + "content": "to phonemes using no supervision. However, the representations of the", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 288, + 505, + 301 + ], + "spans": [ + { + "bbox": [ + 105, + 288, + 505, + 301 + ], + "score": 1.0, + "content": "uppermost block of wav2vec 2.0 may not be well suited for this task. These features are trained to", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 299, + 506, + 312 + ], + "spans": [ + { + "bbox": [ + 106, + 299, + 326, + 312 + ], + "score": 1.0, + "content": "directly predict masked latent representations spanning", + "type": "text" + }, + { + "bbox": [ + 326, + 299, + 349, + 310 + ], + "score": 0.55, + "content": "2 5 \\mathrm { m s }", + "type": "inline_equation" + }, + { + "bbox": [ + 350, + 299, + 506, + 312 + ], + "score": 1.0, + "content": "of speech audio which is much shorter", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 310, + 266, + 323 + ], + "spans": [ + { + "bbox": [ + 105, + 310, + 266, + 323 + ], + "score": 1.0, + "content": "than the typical duration of a phoneme.", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 14.5, + "bbox_fs": [ + 105, + 233, + 506, + 323 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 326, + 505, + 436 + ], + "lines": [ + { + "bbox": [ + 106, + 327, + 505, + 339 + ], + "spans": [ + { + "bbox": [ + 106, + 327, + 505, + 339 + ], + "score": 1.0, + "content": "To get a better sense of this, we train supervised phoneme recognizers with a CTC loss [Graves", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 337, + 506, + 349 + ], + "spans": [ + { + "bbox": [ + 105, + 337, + 506, + 349 + ], + "score": 1.0, + "content": "et al., 2006] on top of the frozen representations of each of the 24 blocks of the English wav2vec 2.0", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 348, + 506, + 361 + ], + "spans": [ + { + "bbox": [ + 105, + 348, + 506, + 361 + ], + "score": 1.0, + "content": "LARGE model pre-trained on Libri-Light. We then evaluate phone error rate (PER) with respect to", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 358, + 506, + 372 + ], + "spans": [ + { + "bbox": [ + 105, + 358, + 437, + 372 + ], + "score": 1.0, + "content": "the phonemized transcriptions of Librispeech dev-other. The classifier takes as input", + "type": "text" + }, + { + "bbox": [ + 437, + 359, + 446, + 371 + ], + "score": 0.88, + "content": "c _ { t } ^ { l }", + "type": "inline_equation" + }, + { + "bbox": [ + 446, + 358, + 506, + 372 + ], + "score": 1.0, + "content": "and contains a", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 106, + 370, + 506, + 383 + ], + "spans": [ + { + "bbox": [ + 106, + 370, + 506, + 383 + ], + "score": 1.0, + "content": "single softmax-normalized linear layer mapping to the phoneme inventory. Figure 2 shows that most", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 380, + 506, + 394 + ], + "spans": [ + { + "bbox": [ + 105, + 380, + 506, + 394 + ], + "score": 1.0, + "content": "of the first ten blocks as well as the final blocks provide very poor performance, while blocks 15-19", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 391, + 506, + 405 + ], + "spans": [ + { + "bbox": [ + 105, + 391, + 212, + 405 + ], + "score": 1.0, + "content": "provide error rates below", + "type": "text" + }, + { + "bbox": [ + 212, + 392, + 227, + 402 + ], + "score": 0.86, + "content": "9 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 227, + 391, + 416, + 405 + ], + "score": 1.0, + "content": "PER. Block 15 achieves the best error rate of", + "type": "text" + }, + { + "bbox": [ + 416, + 392, + 439, + 402 + ], + "score": 0.87, + "content": "7 . 5 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 439, + 391, + 506, + 405 + ], + "score": 1.0, + "content": "PER. A similar", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 403, + 505, + 415 + ], + "spans": [ + { + "bbox": [ + 105, + 403, + 505, + 415 + ], + "score": 1.0, + "content": "insight has been used in the concurrent work of Hsu et al. [2021b]. Appendix A shows that this", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 414, + 505, + 426 + ], + "spans": [ + { + "bbox": [ + 106, + 414, + 411, + 426 + ], + "score": 1.0, + "content": "choice generalizes to other languages. For brevity we drop the superscript", + "type": "text" + }, + { + "bbox": [ + 411, + 414, + 416, + 424 + ], + "score": 0.49, + "content": "l", + "type": "inline_equation" + }, + { + "bbox": [ + 416, + 414, + 505, + 426 + ], + "score": 1.0, + "content": "and refer to block 15", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 424, + 257, + 438 + ], + "spans": [ + { + "bbox": [ + 105, + 424, + 209, + 438 + ], + "score": 1.0, + "content": "representations simply as", + "type": "text" + }, + { + "bbox": [ + 210, + 426, + 252, + 436 + ], + "score": 0.89, + "content": "c _ { 1 } , \\ldots , c _ { T }", + "type": "inline_equation" + }, + { + "bbox": [ + 253, + 424, + 257, + 438 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 23.5, + "bbox_fs": [ + 105, + 327, + 506, + 438 + ] + }, + { + "type": "image", + "bbox": [ + 108, + 448, + 492, + 532 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 108, + 448, + 492, + 532 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 108, + 448, + 492, + 532 + ], + "spans": [ + { + "bbox": [ + 108, + 448, + 492, + 532 + ], + "score": 0.965, + "type": "image", + "image_path": "4893afedc781e5fe6ccd0b0db23368e07e04f56d069fce1ae7128f54dac18113.jpg" + } + ] + } + ], + "index": 30, + "virtual_lines": [ + { + "bbox": [ + 108, + 448, + 492, + 476.0 + ], + "spans": [], + "index": 29 + }, + { + "bbox": [ + 108, + 476.0, + 492, + 504.0 + ], + "spans": [], + "index": 30 + }, + { + "bbox": [ + 108, + 504.0, + 492, + 532.0 + ], + "spans": [], + "index": 31 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 107, + 538, + 504, + 572 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 538, + 505, + 552 + ], + "spans": [ + { + "bbox": [ + 105, + 538, + 505, + 552 + ], + "score": 1.0, + "content": "Figure 2: Supervised phoneme recognition using representations from different wav2vec 2.0 blocks", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 550, + 505, + 562 + ], + "spans": [ + { + "bbox": [ + 105, + 550, + 505, + 562 + ], + "score": 1.0, + "content": "on dev-other of English Librispeech. Low and high blocks do not provide good features, while as", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 560, + 274, + 573 + ], + "spans": [ + { + "bbox": [ + 105, + 560, + 274, + 573 + ], + "score": 1.0, + "content": "blocks 14-19 do. Block 15 performs best.", + "type": "text" + } + ], + "index": 34 + } + ], + "index": 33 + } + ], + "index": 31.5 + }, + { + "type": "title", + "bbox": [ + 108, + 592, + 253, + 604 + ], + "lines": [ + { + "bbox": [ + 105, + 591, + 254, + 606 + ], + "spans": [ + { + "bbox": [ + 105, + 591, + 254, + 606 + ], + "score": 1.0, + "content": "2.3 Segmenting the Audio Signal", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 35 + }, + { + "type": "text", + "bbox": [ + 107, + 612, + 505, + 722 + ], + "lines": [ + { + "bbox": [ + 106, + 612, + 505, + 625 + ], + "spans": [ + { + "bbox": [ + 106, + 612, + 505, + 625 + ], + "score": 1.0, + "content": "Once the speech signal is embedded, we identify segments corresponding to meaningful units that", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 622, + 506, + 637 + ], + "spans": [ + { + "bbox": [ + 105, + 622, + 506, + 637 + ], + "score": 1.0, + "content": "can be mapped to phonemes. Segmentation has been shown to be crucial in prior work [Chung", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 634, + 505, + 647 + ], + "spans": [ + { + "bbox": [ + 105, + 634, + 505, + 647 + ], + "score": 1.0, + "content": "et al., 2018] since the right boundaries in the input representations make it more aligned to phonetic", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 645, + 506, + 658 + ], + "spans": [ + { + "bbox": [ + 105, + 645, + 506, + 658 + ], + "score": 1.0, + "content": "sequences. There has been a lot of prior work in unsupervised speech segmentation [Kamper et al.,", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 655, + 506, + 668 + ], + "spans": [ + { + "bbox": [ + 105, + 655, + 506, + 668 + ], + "score": 1.0, + "content": "2017a,b, Rasanen et al., 2015, Kreuk et al., 2020] but here we simply use a method based on", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 667, + 506, + 681 + ], + "spans": [ + { + "bbox": [ + 105, + 667, + 304, + 681 + ], + "score": 1.0, + "content": "clustering the wav2vec 2.0 speech representations", + "type": "text" + }, + { + "bbox": [ + 304, + 669, + 347, + 678 + ], + "score": 0.87, + "content": "c _ { 1 } , \\ldots , c _ { T }", + "type": "inline_equation" + }, + { + "bbox": [ + 347, + 667, + 506, + 681 + ], + "score": 1.0, + "content": ". In a first step, we collect all the speech", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 104, + 678, + 504, + 690 + ], + "spans": [ + { + "bbox": [ + 104, + 678, + 465, + 690 + ], + "score": 1.0, + "content": "representations for the unlabeled speech data and perform k-means clustering to identify", + "type": "text" + }, + { + "bbox": [ + 466, + 678, + 504, + 688 + ], + "score": 0.89, + "content": "K = 1 2 8", + "type": "inline_equation" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 689, + 505, + 702 + ], + "spans": [ + { + "bbox": [ + 105, + 689, + 505, + 702 + ], + "score": 1.0, + "content": "clusters. We use the FAISS library to do fast clustering on GPUs [Johnson et al., 2019]. Next, each", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 106, + 700, + 506, + 713 + ], + "spans": [ + { + "bbox": [ + 106, + 701, + 115, + 711 + ], + "score": 0.84, + "content": "c _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 116, + 700, + 297, + 713 + ], + "score": 1.0, + "content": "is labeled with the corresponding cluster ID", + "type": "text" + }, + { + "bbox": [ + 298, + 700, + 364, + 712 + ], + "score": 0.84, + "content": "i _ { t } \\in \\{ 1 , \\ldots , K \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 365, + 700, + 506, + 713 + ], + "score": 1.0, + "content": "and we introduce speech segment", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 710, + 287, + 724 + ], + "spans": [ + { + "bbox": [ + 105, + 710, + 287, + 724 + ], + "score": 1.0, + "content": "boundaries whenever the cluster ID changes.", + "type": "text" + } + ], + "index": 45 + } + ], + "index": 40.5, + "bbox_fs": [ + 104, + 612, + 506, + 724 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 106, + 72, + 505, + 171 + ], + "lines": [ + { + "bbox": [ + 106, + 73, + 506, + 86 + ], + "spans": [ + { + "bbox": [ + 106, + 73, + 506, + 86 + ], + "score": 1.0, + "content": "Once the speech audio representations are segmented, we compute a 512-dimensional PCA over", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 83, + 506, + 96 + ], + "spans": [ + { + "bbox": [ + 105, + 83, + 506, + 96 + ], + "score": 1.0, + "content": "all speech representations output by wav2vec 2.0 for the training set. Next, we mean-pool the PCA", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 95, + 506, + 108 + ], + "spans": [ + { + "bbox": [ + 105, + 95, + 506, + 108 + ], + "score": 1.0, + "content": "representations for a particular segment to obtain an average representation of the segment. The", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 104, + 506, + 118 + ], + "spans": [ + { + "bbox": [ + 105, + 104, + 506, + 118 + ], + "score": 1.0, + "content": "PCA retains only the most important features and we found this to be effective. Segment boundaries", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 116, + 506, + 129 + ], + "spans": [ + { + "bbox": [ + 106, + 116, + 506, + 129 + ], + "score": 1.0, + "content": "are noisy due to the lack of supervision and we therefore found it useful to also mean-pool pairs of", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 106, + 127, + 506, + 141 + ], + "spans": [ + { + "bbox": [ + 106, + 127, + 506, + 141 + ], + "score": 1.0, + "content": "adjacent segment representations to increase robustness. This results in sequences of speech segment", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 138, + 506, + 151 + ], + "spans": [ + { + "bbox": [ + 105, + 138, + 167, + 151 + ], + "score": 1.0, + "content": "representation", + "type": "text" + }, + { + "bbox": [ + 167, + 138, + 263, + 149 + ], + "score": 0.91, + "content": "S = s _ { 1 } , \\ldots , s _ { T } , S \\sim { \\mathcal { S } }", + "type": "inline_equation" + }, + { + "bbox": [ + 263, + 138, + 506, + 151 + ], + "score": 1.0, + "content": "for a given utterance. Appendix B shows an illustration of", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 148, + 506, + 163 + ], + "spans": [ + { + "bbox": [ + 105, + 148, + 506, + 163 + ], + "score": 1.0, + "content": "the segmentation strategy on an actual example as well as a quantitative evaluation of the strategy", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 106, + 160, + 255, + 173 + ], + "spans": [ + { + "bbox": [ + 106, + 160, + 255, + 173 + ], + "score": 1.0, + "content": "compared to human segmented data.", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 4 + }, + { + "type": "title", + "bbox": [ + 108, + 188, + 252, + 200 + ], + "lines": [ + { + "bbox": [ + 105, + 187, + 254, + 202 + ], + "spans": [ + { + "bbox": [ + 105, + 187, + 254, + 202 + ], + "score": 1.0, + "content": "2.4 Pre-processing the Text Data", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 9 + }, + { + "type": "text", + "bbox": [ + 108, + 209, + 504, + 243 + ], + "lines": [ + { + "bbox": [ + 106, + 210, + 505, + 222 + ], + "spans": [ + { + "bbox": [ + 106, + 210, + 505, + 222 + ], + "score": 1.0, + "content": "Similar to how we segment the unlabeled speech audio data into suitable units for unsupervised", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 221, + 505, + 233 + ], + "spans": [ + { + "bbox": [ + 105, + 221, + 505, + 233 + ], + "score": 1.0, + "content": "learning, we do the same for the unlabeled text data. We apply two pre-processing steps to the text", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 107, + 233, + 302, + 244 + ], + "spans": [ + { + "bbox": [ + 107, + 233, + 302, + 244 + ], + "score": 1.0, + "content": "data: phonemization and silence token insertion.", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 11 + }, + { + "type": "text", + "bbox": [ + 107, + 248, + 505, + 325 + ], + "lines": [ + { + "bbox": [ + 106, + 249, + 504, + 260 + ], + "spans": [ + { + "bbox": [ + 106, + 249, + 504, + 260 + ], + "score": 1.0, + "content": "Phonemes characterize the different sounds which distinguish words from each other, e.g., for the", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 106, + 260, + 505, + 271 + ], + "spans": [ + { + "bbox": [ + 106, + 260, + 505, + 271 + ], + "score": 1.0, + "content": "word cat there are three phonemes corresponding to the three distinct sounds in the pronunciation", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 270, + 506, + 282 + ], + "spans": [ + { + "bbox": [ + 105, + 270, + 207, + 282 + ], + "score": 1.0, + "content": "of the word: /K/, /AE/,", + "type": "text" + }, + { + "bbox": [ + 207, + 270, + 222, + 281 + ], + "score": 0.31, + "content": "/ \\mathrm { T } / .", + "type": "inline_equation" + }, + { + "bbox": [ + 222, + 270, + 506, + 282 + ], + "score": 1.0, + "content": ". We phonemize the text data because we found it easier to learn a", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 281, + 506, + 293 + ], + "spans": [ + { + "bbox": [ + 105, + 281, + 506, + 293 + ], + "score": 1.0, + "content": "mapping between speech audio and the different sounds of a word rather than between audio and", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 291, + 505, + 303 + ], + "spans": [ + { + "bbox": [ + 105, + 291, + 370, + 303 + ], + "score": 1.0, + "content": "words or letters. Phonemization converts a sequence of words", + "type": "text" + }, + { + "bbox": [ + 370, + 292, + 380, + 302 + ], + "score": 0.73, + "content": "Y", + "type": "inline_equation" + }, + { + "bbox": [ + 380, + 291, + 505, + 303 + ], + "score": 1.0, + "content": "into a sequence of phonemes", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 107, + 302, + 506, + 316 + ], + "spans": [ + { + "bbox": [ + 107, + 303, + 181, + 315 + ], + "score": 0.91, + "content": "P = [ p _ { 1 } , \\cdots , p _ { M } ]", + "type": "inline_equation" + }, + { + "bbox": [ + 182, + 302, + 212, + 316 + ], + "score": 1.0, + "content": ", where", + "type": "text" + }, + { + "bbox": [ + 213, + 303, + 247, + 314 + ], + "score": 0.92, + "content": "p _ { m } \\in O", + "type": "inline_equation" + }, + { + "bbox": [ + 247, + 302, + 265, + 316 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 266, + 303, + 275, + 313 + ], + "score": 0.82, + "content": "O", + "type": "inline_equation" + }, + { + "bbox": [ + 275, + 302, + 506, + 316 + ], + "score": 1.0, + "content": "is the phoneme inventory. We use off-the-shelf tools for", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 314, + 286, + 326 + ], + "spans": [ + { + "bbox": [ + 105, + 314, + 259, + 326 + ], + "score": 1.0, + "content": "this step which we detail in Appendix", + "type": "text" + }, + { + "bbox": [ + 259, + 314, + 282, + 325 + ], + "score": 0.73, + "content": "\\ S \\operatorname { E } . 2", + "type": "inline_equation" + }, + { + "bbox": [ + 282, + 314, + 286, + 326 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 16 + }, + { + "type": "text", + "bbox": [ + 107, + 330, + 506, + 407 + ], + "lines": [ + { + "bbox": [ + 105, + 329, + 507, + 343 + ], + "spans": [ + { + "bbox": [ + 105, + 329, + 507, + 343 + ], + "score": 1.0, + "content": "The unlabeled speech audio data is pre-processed by applying unsupervised silence removal. However,", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 106, + 341, + 506, + 353 + ], + "spans": [ + { + "bbox": [ + 106, + 341, + 506, + 353 + ], + "score": 1.0, + "content": "this process is not always accurate and many silences in the speech audio remain. To deal with this,", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 106, + 352, + 506, + 364 + ], + "spans": [ + { + "bbox": [ + 106, + 352, + 479, + 364 + ], + "score": 1.0, + "content": "we enable the unsupervised model to label some segments with a phonemic silence token (SIL;", + "type": "text" + }, + { + "bbox": [ + 479, + 352, + 501, + 363 + ], + "score": 0.7, + "content": "\\ S \\ 3 . 1 \\AA .", + "type": "inline_equation" + }, + { + "bbox": [ + 501, + 352, + 506, + 364 + ], + "score": 1.0, + "content": ").", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 362, + 506, + 376 + ], + "spans": [ + { + "bbox": [ + 105, + 362, + 506, + 376 + ], + "score": 1.0, + "content": "However, the phonemized unlabeled text data does not contain any silence tokens and this may pose", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 106, + 373, + 506, + 387 + ], + "spans": [ + { + "bbox": [ + 106, + 373, + 243, + 387 + ], + "score": 1.0, + "content": "difficulties for adversarial learning", + "type": "text" + }, + { + "bbox": [ + 244, + 374, + 262, + 385 + ], + "score": 0.42, + "content": "( \\ S 3 )", + "type": "inline_equation" + }, + { + "bbox": [ + 262, + 373, + 506, + 387 + ], + "score": 1.0, + "content": ". We remedy this by inserting silence markers at the beginning", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 106, + 385, + 506, + 397 + ], + "spans": [ + { + "bbox": [ + 106, + 385, + 506, + 397 + ], + "score": 1.0, + "content": "and end of the phonemized unlabeled text data; we also randomly insert SIL between words, or", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 396, + 504, + 408 + ], + "spans": [ + { + "bbox": [ + 105, + 396, + 331, + 408 + ], + "score": 1.0, + "content": "groups of phonemes corresponding to words at a rate of", + "type": "text" + }, + { + "bbox": [ + 331, + 396, + 351, + 406 + ], + "score": 0.87, + "content": "2 5 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 351, + 396, + 504, + 408 + ], + "score": 1.0, + "content": ". Appendix C evaluates these choices.", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 23 + }, + { + "type": "title", + "bbox": [ + 108, + 425, + 246, + 440 + ], + "lines": [ + { + "bbox": [ + 103, + 423, + 248, + 444 + ], + "spans": [ + { + "bbox": [ + 103, + 423, + 248, + 444 + ], + "score": 1.0, + "content": "3 Unsupervised Learning", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 27 + }, + { + "type": "text", + "bbox": [ + 107, + 453, + 506, + 497 + ], + "lines": [ + { + "bbox": [ + 105, + 452, + 507, + 466 + ], + "spans": [ + { + "bbox": [ + 105, + 452, + 507, + 466 + ], + "score": 1.0, + "content": "We use adversarial training to train an unsupervised speech recognition model using the representa-", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 106, + 464, + 506, + 476 + ], + "spans": [ + { + "bbox": [ + 106, + 464, + 506, + 476 + ], + "score": 1.0, + "content": "tions of the unlabeled speech audio data and the unlabeled phonemized text data [Liu et al., 2018,", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 475, + 506, + 488 + ], + "spans": [ + { + "bbox": [ + 105, + 475, + 506, + 488 + ], + "score": 1.0, + "content": "Chen et al., 2019]. In the following, we detail the model architecture, the training objective as well as", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 106, + 487, + 329, + 498 + ], + "spans": [ + { + "bbox": [ + 106, + 487, + 329, + 498 + ], + "score": 1.0, + "content": "the unsupervised cross-validation metric we developed.", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 29.5 + }, + { + "type": "title", + "bbox": [ + 107, + 514, + 214, + 526 + ], + "lines": [ + { + "bbox": [ + 105, + 513, + 215, + 527 + ], + "spans": [ + { + "bbox": [ + 105, + 513, + 215, + 527 + ], + "score": 1.0, + "content": "3.1 Model Architecture", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 32 + }, + { + "type": "text", + "bbox": [ + 107, + 536, + 505, + 591 + ], + "lines": [ + { + "bbox": [ + 106, + 536, + 505, + 549 + ], + "spans": [ + { + "bbox": [ + 106, + 536, + 478, + 549 + ], + "score": 1.0, + "content": "Generative adversarial networks (GAN; Goodfellow et al. 2014) train a generator network", + "type": "text" + }, + { + "bbox": [ + 478, + 537, + 486, + 547 + ], + "score": 0.83, + "content": "\\mathcal { G }", + "type": "inline_equation" + }, + { + "bbox": [ + 487, + 536, + 505, + 549 + ], + "score": 1.0, + "content": "and", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 546, + 505, + 560 + ], + "spans": [ + { + "bbox": [ + 105, + 546, + 230, + 560 + ], + "score": 1.0, + "content": "a discriminator/critic network", + "type": "text" + }, + { + "bbox": [ + 231, + 547, + 238, + 557 + ], + "score": 0.79, + "content": "\\mathcal { C }", + "type": "inline_equation" + }, + { + "bbox": [ + 238, + 546, + 505, + 560 + ], + "score": 1.0, + "content": "where the generator produces samples which are then judged by", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 557, + 506, + 571 + ], + "spans": [ + { + "bbox": [ + 105, + 557, + 506, + 571 + ], + "score": 1.0, + "content": "the discriminator. The discriminator is trained to classify whether samples are from the generator", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 568, + 505, + 582 + ], + "spans": [ + { + "bbox": [ + 105, + 568, + 505, + 582 + ], + "score": 1.0, + "content": "or from the real data distribution. The objective of the generator is to produce samples that are", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 106, + 579, + 262, + 593 + ], + "spans": [ + { + "bbox": [ + 106, + 579, + 262, + 593 + ], + "score": 1.0, + "content": "indistinguishable by the discriminator.", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 35 + }, + { + "type": "text", + "bbox": [ + 107, + 596, + 505, + 651 + ], + "lines": [ + { + "bbox": [ + 105, + 595, + 506, + 609 + ], + "spans": [ + { + "bbox": [ + 105, + 595, + 153, + 609 + ], + "score": 1.0, + "content": "Concretely,", + "type": "text" + }, + { + "bbox": [ + 154, + 597, + 162, + 607 + ], + "score": 0.8, + "content": "\\mathcal { G }", + "type": "inline_equation" + }, + { + "bbox": [ + 162, + 595, + 273, + 609 + ], + "score": 1.0, + "content": "takes as input a sequence of", + "type": "text" + }, + { + "bbox": [ + 274, + 597, + 282, + 606 + ], + "score": 0.82, + "content": "T", + "type": "inline_equation" + }, + { + "bbox": [ + 283, + 595, + 380, + 609 + ], + "score": 1.0, + "content": "segment representations", + "type": "text" + }, + { + "bbox": [ + 380, + 596, + 449, + 608 + ], + "score": 0.83, + "content": "S = [ s _ { 1 } , \\dotsc , s _ { T } ]", + "type": "inline_equation" + }, + { + "bbox": [ + 449, + 595, + 506, + 609 + ], + "score": 1.0, + "content": "(§ 2.3) which", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 606, + 505, + 619 + ], + "spans": [ + { + "bbox": [ + 105, + 606, + 249, + 619 + ], + "score": 1.0, + "content": "are then mapped to a sequence of", + "type": "text" + }, + { + "bbox": [ + 249, + 608, + 261, + 617 + ], + "score": 0.82, + "content": "M", + "type": "inline_equation" + }, + { + "bbox": [ + 261, + 606, + 307, + 619 + ], + "score": 1.0, + "content": "phonemes", + "type": "text" + }, + { + "bbox": [ + 307, + 607, + 396, + 619 + ], + "score": 0.93, + "content": "\\mathcal { G } ( S ) \\bar { = } [ p _ { 1 } , \\dotsc , p _ { M } ]", + "type": "inline_equation" + }, + { + "bbox": [ + 396, + 606, + 505, + 619 + ], + "score": 1.0, + "content": ". The generator predicts a", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 617, + 505, + 631 + ], + "spans": [ + { + "bbox": [ + 105, + 617, + 246, + 631 + ], + "score": 1.0, + "content": "distribution over the phoneme set", + "type": "text" + }, + { + "bbox": [ + 247, + 618, + 256, + 628 + ], + "score": 0.8, + "content": "O", + "type": "inline_equation" + }, + { + "bbox": [ + 256, + 617, + 505, + 631 + ], + "score": 1.0, + "content": "for each segment and outputs the phoneme with the highest", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 628, + 506, + 642 + ], + "spans": [ + { + "bbox": [ + 105, + 628, + 506, + 642 + ], + "score": 1.0, + "content": "probability. If the argmax prediction of consecutive segments result in the same phoneme, then we", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 639, + 305, + 653 + ], + "spans": [ + { + "bbox": [ + 105, + 639, + 268, + 653 + ], + "score": 1.0, + "content": "sample one of these segments, therefore", + "type": "text" + }, + { + "bbox": [ + 268, + 640, + 300, + 651 + ], + "score": 0.91, + "content": "M \\leq T", + "type": "inline_equation" + }, + { + "bbox": [ + 301, + 639, + 305, + 653 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 42 + } + ], + "index": 40 + }, + { + "type": "text", + "bbox": [ + 107, + 655, + 504, + 722 + ], + "lines": [ + { + "bbox": [ + 105, + 655, + 506, + 669 + ], + "spans": [ + { + "bbox": [ + 105, + 655, + 178, + 669 + ], + "score": 1.0, + "content": "The phoneme set", + "type": "text" + }, + { + "bbox": [ + 178, + 657, + 187, + 666 + ], + "score": 0.74, + "content": "O", + "type": "inline_equation" + }, + { + "bbox": [ + 188, + 655, + 506, + 669 + ], + "score": 1.0, + "content": "includes a silence label SIL to enable labeling silences in the speech audio as", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 666, + 506, + 680 + ], + "spans": [ + { + "bbox": [ + 105, + 666, + 506, + 680 + ], + "score": 1.0, + "content": "such. Without a silence label, we noticed that the model was repurposing a particular phoneme to", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 676, + 506, + 692 + ], + "spans": [ + { + "bbox": [ + 105, + 676, + 506, + 692 + ], + "score": 1.0, + "content": "label silences which resulted in much lower performance since it interfered with subsequent language", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 105, + 689, + 506, + 702 + ], + "spans": [ + { + "bbox": [ + 105, + 689, + 506, + 702 + ], + "score": 1.0, + "content": "model (LM) decoding. In the backward pass, we back-propagate through segments sampled at the", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 105, + 699, + 506, + 713 + ], + "spans": [ + { + "bbox": [ + 105, + 699, + 360, + 713 + ], + "score": 1.0, + "content": "generator output. We do not modify the segment representations", + "type": "text" + }, + { + "bbox": [ + 361, + 701, + 368, + 710 + ], + "score": 0.82, + "content": "S", + "type": "inline_equation" + }, + { + "bbox": [ + 369, + 699, + 506, + 713 + ], + "score": 1.0, + "content": "during unsupervised training. The", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 105, + 711, + 433, + 723 + ], + "spans": [ + { + "bbox": [ + 105, + 711, + 433, + 723 + ], + "score": 1.0, + "content": "generator is parameterized as a single layer convolutional neural network (CNN).", + "type": "text" + } + ], + "index": 48 + } + ], + "index": 45.5 + } + ], + "page_idx": 3, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 302, + 742, + 308, + 750 + ], + "lines": [ + { + "bbox": [ + 301, + 741, + 310, + 752 + ], + "spans": [ + { + "bbox": [ + 301, + 741, + 310, + 752 + ], + "score": 1.0, + "content": "", + "type": "text", + "height": 11, + "width": 9 + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 106, + 72, + 505, + 171 + ], + "lines": [ + { + "bbox": [ + 106, + 73, + 506, + 86 + ], + "spans": [ + { + "bbox": [ + 106, + 73, + 506, + 86 + ], + "score": 1.0, + "content": "Once the speech audio representations are segmented, we compute a 512-dimensional PCA over", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 83, + 506, + 96 + ], + "spans": [ + { + "bbox": [ + 105, + 83, + 506, + 96 + ], + "score": 1.0, + "content": "all speech representations output by wav2vec 2.0 for the training set. Next, we mean-pool the PCA", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 95, + 506, + 108 + ], + "spans": [ + { + "bbox": [ + 105, + 95, + 506, + 108 + ], + "score": 1.0, + "content": "representations for a particular segment to obtain an average representation of the segment. The", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 104, + 506, + 118 + ], + "spans": [ + { + "bbox": [ + 105, + 104, + 506, + 118 + ], + "score": 1.0, + "content": "PCA retains only the most important features and we found this to be effective. Segment boundaries", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 116, + 506, + 129 + ], + "spans": [ + { + "bbox": [ + 106, + 116, + 506, + 129 + ], + "score": 1.0, + "content": "are noisy due to the lack of supervision and we therefore found it useful to also mean-pool pairs of", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 106, + 127, + 506, + 141 + ], + "spans": [ + { + "bbox": [ + 106, + 127, + 506, + 141 + ], + "score": 1.0, + "content": "adjacent segment representations to increase robustness. This results in sequences of speech segment", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 138, + 506, + 151 + ], + "spans": [ + { + "bbox": [ + 105, + 138, + 167, + 151 + ], + "score": 1.0, + "content": "representation", + "type": "text" + }, + { + "bbox": [ + 167, + 138, + 263, + 149 + ], + "score": 0.91, + "content": "S = s _ { 1 } , \\ldots , s _ { T } , S \\sim { \\mathcal { S } }", + "type": "inline_equation" + }, + { + "bbox": [ + 263, + 138, + 506, + 151 + ], + "score": 1.0, + "content": "for a given utterance. Appendix B shows an illustration of", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 148, + 506, + 163 + ], + "spans": [ + { + "bbox": [ + 105, + 148, + 506, + 163 + ], + "score": 1.0, + "content": "the segmentation strategy on an actual example as well as a quantitative evaluation of the strategy", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 106, + 160, + 255, + 173 + ], + "spans": [ + { + "bbox": [ + 106, + 160, + 255, + 173 + ], + "score": 1.0, + "content": "compared to human segmented data.", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 4, + "bbox_fs": [ + 105, + 73, + 506, + 173 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 188, + 252, + 200 + ], + "lines": [ + { + "bbox": [ + 105, + 187, + 254, + 202 + ], + "spans": [ + { + "bbox": [ + 105, + 187, + 254, + 202 + ], + "score": 1.0, + "content": "2.4 Pre-processing the Text Data", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 9 + }, + { + "type": "text", + "bbox": [ + 108, + 209, + 504, + 243 + ], + "lines": [ + { + "bbox": [ + 106, + 210, + 505, + 222 + ], + "spans": [ + { + "bbox": [ + 106, + 210, + 505, + 222 + ], + "score": 1.0, + "content": "Similar to how we segment the unlabeled speech audio data into suitable units for unsupervised", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 221, + 505, + 233 + ], + "spans": [ + { + "bbox": [ + 105, + 221, + 505, + 233 + ], + "score": 1.0, + "content": "learning, we do the same for the unlabeled text data. We apply two pre-processing steps to the text", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 107, + 233, + 302, + 244 + ], + "spans": [ + { + "bbox": [ + 107, + 233, + 302, + 244 + ], + "score": 1.0, + "content": "data: phonemization and silence token insertion.", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 11, + "bbox_fs": [ + 105, + 210, + 505, + 244 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 248, + 505, + 325 + ], + "lines": [ + { + "bbox": [ + 106, + 249, + 504, + 260 + ], + "spans": [ + { + "bbox": [ + 106, + 249, + 504, + 260 + ], + "score": 1.0, + "content": "Phonemes characterize the different sounds which distinguish words from each other, e.g., for the", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 106, + 260, + 505, + 271 + ], + "spans": [ + { + "bbox": [ + 106, + 260, + 505, + 271 + ], + "score": 1.0, + "content": "word cat there are three phonemes corresponding to the three distinct sounds in the pronunciation", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 270, + 506, + 282 + ], + "spans": [ + { + "bbox": [ + 105, + 270, + 207, + 282 + ], + "score": 1.0, + "content": "of the word: /K/, /AE/,", + "type": "text" + }, + { + "bbox": [ + 207, + 270, + 222, + 281 + ], + "score": 0.31, + "content": "/ \\mathrm { T } / .", + "type": "inline_equation" + }, + { + "bbox": [ + 222, + 270, + 506, + 282 + ], + "score": 1.0, + "content": ". We phonemize the text data because we found it easier to learn a", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 281, + 506, + 293 + ], + "spans": [ + { + "bbox": [ + 105, + 281, + 506, + 293 + ], + "score": 1.0, + "content": "mapping between speech audio and the different sounds of a word rather than between audio and", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 291, + 505, + 303 + ], + "spans": [ + { + "bbox": [ + 105, + 291, + 370, + 303 + ], + "score": 1.0, + "content": "words or letters. Phonemization converts a sequence of words", + "type": "text" + }, + { + "bbox": [ + 370, + 292, + 380, + 302 + ], + "score": 0.73, + "content": "Y", + "type": "inline_equation" + }, + { + "bbox": [ + 380, + 291, + 505, + 303 + ], + "score": 1.0, + "content": "into a sequence of phonemes", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 107, + 302, + 506, + 316 + ], + "spans": [ + { + "bbox": [ + 107, + 303, + 181, + 315 + ], + "score": 0.91, + "content": "P = [ p _ { 1 } , \\cdots , p _ { M } ]", + "type": "inline_equation" + }, + { + "bbox": [ + 182, + 302, + 212, + 316 + ], + "score": 1.0, + "content": ", where", + "type": "text" + }, + { + "bbox": [ + 213, + 303, + 247, + 314 + ], + "score": 0.92, + "content": "p _ { m } \\in O", + "type": "inline_equation" + }, + { + "bbox": [ + 247, + 302, + 265, + 316 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 266, + 303, + 275, + 313 + ], + "score": 0.82, + "content": "O", + "type": "inline_equation" + }, + { + "bbox": [ + 275, + 302, + 506, + 316 + ], + "score": 1.0, + "content": "is the phoneme inventory. We use off-the-shelf tools for", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 314, + 286, + 326 + ], + "spans": [ + { + "bbox": [ + 105, + 314, + 259, + 326 + ], + "score": 1.0, + "content": "this step which we detail in Appendix", + "type": "text" + }, + { + "bbox": [ + 259, + 314, + 282, + 325 + ], + "score": 0.73, + "content": "\\ S \\operatorname { E } . 2", + "type": "inline_equation" + }, + { + "bbox": [ + 282, + 314, + 286, + 326 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 16, + "bbox_fs": [ + 105, + 249, + 506, + 326 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 330, + 506, + 407 + ], + "lines": [ + { + "bbox": [ + 105, + 329, + 507, + 343 + ], + "spans": [ + { + "bbox": [ + 105, + 329, + 507, + 343 + ], + "score": 1.0, + "content": "The unlabeled speech audio data is pre-processed by applying unsupervised silence removal. However,", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 106, + 341, + 506, + 353 + ], + "spans": [ + { + "bbox": [ + 106, + 341, + 506, + 353 + ], + "score": 1.0, + "content": "this process is not always accurate and many silences in the speech audio remain. To deal with this,", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 106, + 352, + 506, + 364 + ], + "spans": [ + { + "bbox": [ + 106, + 352, + 479, + 364 + ], + "score": 1.0, + "content": "we enable the unsupervised model to label some segments with a phonemic silence token (SIL;", + "type": "text" + }, + { + "bbox": [ + 479, + 352, + 501, + 363 + ], + "score": 0.7, + "content": "\\ S \\ 3 . 1 \\AA .", + "type": "inline_equation" + }, + { + "bbox": [ + 501, + 352, + 506, + 364 + ], + "score": 1.0, + "content": ").", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 362, + 506, + 376 + ], + "spans": [ + { + "bbox": [ + 105, + 362, + 506, + 376 + ], + "score": 1.0, + "content": "However, the phonemized unlabeled text data does not contain any silence tokens and this may pose", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 106, + 373, + 506, + 387 + ], + "spans": [ + { + "bbox": [ + 106, + 373, + 243, + 387 + ], + "score": 1.0, + "content": "difficulties for adversarial learning", + "type": "text" + }, + { + "bbox": [ + 244, + 374, + 262, + 385 + ], + "score": 0.42, + "content": "( \\ S 3 )", + "type": "inline_equation" + }, + { + "bbox": [ + 262, + 373, + 506, + 387 + ], + "score": 1.0, + "content": ". We remedy this by inserting silence markers at the beginning", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 106, + 385, + 506, + 397 + ], + "spans": [ + { + "bbox": [ + 106, + 385, + 506, + 397 + ], + "score": 1.0, + "content": "and end of the phonemized unlabeled text data; we also randomly insert SIL between words, or", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 396, + 504, + 408 + ], + "spans": [ + { + "bbox": [ + 105, + 396, + 331, + 408 + ], + "score": 1.0, + "content": "groups of phonemes corresponding to words at a rate of", + "type": "text" + }, + { + "bbox": [ + 331, + 396, + 351, + 406 + ], + "score": 0.87, + "content": "2 5 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 351, + 396, + 504, + 408 + ], + "score": 1.0, + "content": ". Appendix C evaluates these choices.", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 23, + "bbox_fs": [ + 105, + 329, + 507, + 408 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 425, + 246, + 440 + ], + "lines": [ + { + "bbox": [ + 103, + 423, + 248, + 444 + ], + "spans": [ + { + "bbox": [ + 103, + 423, + 248, + 444 + ], + "score": 1.0, + "content": "3 Unsupervised Learning", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 27 + }, + { + "type": "text", + "bbox": [ + 107, + 453, + 506, + 497 + ], + "lines": [ + { + "bbox": [ + 105, + 452, + 507, + 466 + ], + "spans": [ + { + "bbox": [ + 105, + 452, + 507, + 466 + ], + "score": 1.0, + "content": "We use adversarial training to train an unsupervised speech recognition model using the representa-", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 106, + 464, + 506, + 476 + ], + "spans": [ + { + "bbox": [ + 106, + 464, + 506, + 476 + ], + "score": 1.0, + "content": "tions of the unlabeled speech audio data and the unlabeled phonemized text data [Liu et al., 2018,", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 475, + 506, + 488 + ], + "spans": [ + { + "bbox": [ + 105, + 475, + 506, + 488 + ], + "score": 1.0, + "content": "Chen et al., 2019]. In the following, we detail the model architecture, the training objective as well as", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 106, + 487, + 329, + 498 + ], + "spans": [ + { + "bbox": [ + 106, + 487, + 329, + 498 + ], + "score": 1.0, + "content": "the unsupervised cross-validation metric we developed.", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 29.5, + "bbox_fs": [ + 105, + 452, + 507, + 498 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 514, + 214, + 526 + ], + "lines": [ + { + "bbox": [ + 105, + 513, + 215, + 527 + ], + "spans": [ + { + "bbox": [ + 105, + 513, + 215, + 527 + ], + "score": 1.0, + "content": "3.1 Model Architecture", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 32 + }, + { + "type": "text", + "bbox": [ + 107, + 536, + 505, + 591 + ], + "lines": [ + { + "bbox": [ + 106, + 536, + 505, + 549 + ], + "spans": [ + { + "bbox": [ + 106, + 536, + 478, + 549 + ], + "score": 1.0, + "content": "Generative adversarial networks (GAN; Goodfellow et al. 2014) train a generator network", + "type": "text" + }, + { + "bbox": [ + 478, + 537, + 486, + 547 + ], + "score": 0.83, + "content": "\\mathcal { G }", + "type": "inline_equation" + }, + { + "bbox": [ + 487, + 536, + 505, + 549 + ], + "score": 1.0, + "content": "and", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 546, + 505, + 560 + ], + "spans": [ + { + "bbox": [ + 105, + 546, + 230, + 560 + ], + "score": 1.0, + "content": "a discriminator/critic network", + "type": "text" + }, + { + "bbox": [ + 231, + 547, + 238, + 557 + ], + "score": 0.79, + "content": "\\mathcal { C }", + "type": "inline_equation" + }, + { + "bbox": [ + 238, + 546, + 505, + 560 + ], + "score": 1.0, + "content": "where the generator produces samples which are then judged by", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 557, + 506, + 571 + ], + "spans": [ + { + "bbox": [ + 105, + 557, + 506, + 571 + ], + "score": 1.0, + "content": "the discriminator. The discriminator is trained to classify whether samples are from the generator", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 568, + 505, + 582 + ], + "spans": [ + { + "bbox": [ + 105, + 568, + 505, + 582 + ], + "score": 1.0, + "content": "or from the real data distribution. The objective of the generator is to produce samples that are", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 106, + 579, + 262, + 593 + ], + "spans": [ + { + "bbox": [ + 106, + 579, + 262, + 593 + ], + "score": 1.0, + "content": "indistinguishable by the discriminator.", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 35, + "bbox_fs": [ + 105, + 536, + 506, + 593 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 596, + 505, + 651 + ], + "lines": [ + { + "bbox": [ + 105, + 595, + 506, + 609 + ], + "spans": [ + { + "bbox": [ + 105, + 595, + 153, + 609 + ], + "score": 1.0, + "content": "Concretely,", + "type": "text" + }, + { + "bbox": [ + 154, + 597, + 162, + 607 + ], + "score": 0.8, + "content": "\\mathcal { G }", + "type": "inline_equation" + }, + { + "bbox": [ + 162, + 595, + 273, + 609 + ], + "score": 1.0, + "content": "takes as input a sequence of", + "type": "text" + }, + { + "bbox": [ + 274, + 597, + 282, + 606 + ], + "score": 0.82, + "content": "T", + "type": "inline_equation" + }, + { + "bbox": [ + 283, + 595, + 380, + 609 + ], + "score": 1.0, + "content": "segment representations", + "type": "text" + }, + { + "bbox": [ + 380, + 596, + 449, + 608 + ], + "score": 0.83, + "content": "S = [ s _ { 1 } , \\dotsc , s _ { T } ]", + "type": "inline_equation" + }, + { + "bbox": [ + 449, + 595, + 506, + 609 + ], + "score": 1.0, + "content": "(§ 2.3) which", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 606, + 505, + 619 + ], + "spans": [ + { + "bbox": [ + 105, + 606, + 249, + 619 + ], + "score": 1.0, + "content": "are then mapped to a sequence of", + "type": "text" + }, + { + "bbox": [ + 249, + 608, + 261, + 617 + ], + "score": 0.82, + "content": "M", + "type": "inline_equation" + }, + { + "bbox": [ + 261, + 606, + 307, + 619 + ], + "score": 1.0, + "content": "phonemes", + "type": "text" + }, + { + "bbox": [ + 307, + 607, + 396, + 619 + ], + "score": 0.93, + "content": "\\mathcal { G } ( S ) \\bar { = } [ p _ { 1 } , \\dotsc , p _ { M } ]", + "type": "inline_equation" + }, + { + "bbox": [ + 396, + 606, + 505, + 619 + ], + "score": 1.0, + "content": ". The generator predicts a", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 617, + 505, + 631 + ], + "spans": [ + { + "bbox": [ + 105, + 617, + 246, + 631 + ], + "score": 1.0, + "content": "distribution over the phoneme set", + "type": "text" + }, + { + "bbox": [ + 247, + 618, + 256, + 628 + ], + "score": 0.8, + "content": "O", + "type": "inline_equation" + }, + { + "bbox": [ + 256, + 617, + 505, + 631 + ], + "score": 1.0, + "content": "for each segment and outputs the phoneme with the highest", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 628, + 506, + 642 + ], + "spans": [ + { + "bbox": [ + 105, + 628, + 506, + 642 + ], + "score": 1.0, + "content": "probability. If the argmax prediction of consecutive segments result in the same phoneme, then we", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 639, + 305, + 653 + ], + "spans": [ + { + "bbox": [ + 105, + 639, + 268, + 653 + ], + "score": 1.0, + "content": "sample one of these segments, therefore", + "type": "text" + }, + { + "bbox": [ + 268, + 640, + 300, + 651 + ], + "score": 0.91, + "content": "M \\leq T", + "type": "inline_equation" + }, + { + "bbox": [ + 301, + 639, + 305, + 653 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 42 + } + ], + "index": 40, + "bbox_fs": [ + 105, + 595, + 506, + 653 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 655, + 504, + 722 + ], + "lines": [ + { + "bbox": [ + 105, + 655, + 506, + 669 + ], + "spans": [ + { + "bbox": [ + 105, + 655, + 178, + 669 + ], + "score": 1.0, + "content": "The phoneme set", + "type": "text" + }, + { + "bbox": [ + 178, + 657, + 187, + 666 + ], + "score": 0.74, + "content": "O", + "type": "inline_equation" + }, + { + "bbox": [ + 188, + 655, + 506, + 669 + ], + "score": 1.0, + "content": "includes a silence label SIL to enable labeling silences in the speech audio as", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 666, + 506, + 680 + ], + "spans": [ + { + "bbox": [ + 105, + 666, + 506, + 680 + ], + "score": 1.0, + "content": "such. Without a silence label, we noticed that the model was repurposing a particular phoneme to", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 676, + 506, + 692 + ], + "spans": [ + { + "bbox": [ + 105, + 676, + 506, + 692 + ], + "score": 1.0, + "content": "label silences which resulted in much lower performance since it interfered with subsequent language", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 105, + 689, + 506, + 702 + ], + "spans": [ + { + "bbox": [ + 105, + 689, + 506, + 702 + ], + "score": 1.0, + "content": "model (LM) decoding. In the backward pass, we back-propagate through segments sampled at the", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 105, + 699, + 506, + 713 + ], + "spans": [ + { + "bbox": [ + 105, + 699, + 360, + 713 + ], + "score": 1.0, + "content": "generator output. We do not modify the segment representations", + "type": "text" + }, + { + "bbox": [ + 361, + 701, + 368, + 710 + ], + "score": 0.82, + "content": "S", + "type": "inline_equation" + }, + { + "bbox": [ + 369, + 699, + 506, + 713 + ], + "score": 1.0, + "content": "during unsupervised training. The", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 105, + 711, + 433, + 723 + ], + "spans": [ + { + "bbox": [ + 105, + 711, + 433, + 723 + ], + "score": 1.0, + "content": "generator is parameterized as a single layer convolutional neural network (CNN).", + "type": "text" + } + ], + "index": 48 + } + ], + "index": 45.5, + "bbox_fs": [ + 105, + 655, + 506, + 723 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 106, + 72, + 506, + 128 + ], + "lines": [ + { + "bbox": [ + 106, + 72, + 506, + 86 + ], + "spans": [ + { + "bbox": [ + 106, + 72, + 304, + 86 + ], + "score": 1.0, + "content": "The discriminator takes as input either a sequence", + "type": "text" + }, + { + "bbox": [ + 304, + 73, + 343, + 83 + ], + "score": 0.91, + "content": "P ^ { r } \\sim \\mathcal { P } ^ { r }", + "type": "inline_equation" + }, + { + "bbox": [ + 343, + 72, + 506, + 86 + ], + "score": 1.0, + "content": "of one-hot vectors denoting phonemized", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 83, + 506, + 96 + ], + "spans": [ + { + "bbox": [ + 105, + 83, + 244, + 96 + ], + "score": 1.0, + "content": "text from the real data distribution", + "type": "text" + }, + { + "bbox": [ + 244, + 84, + 257, + 94 + ], + "score": 0.85, + "content": "{ \\mathcal { P } } ^ { r }", + "type": "inline_equation" + }, + { + "bbox": [ + 257, + 83, + 481, + 96 + ], + "score": 1.0, + "content": "or a sequence of output distributions from the generator", + "type": "text" + }, + { + "bbox": [ + 481, + 83, + 502, + 96 + ], + "score": 0.92, + "content": "\\mathcal { G } ( S )", + "type": "inline_equation" + }, + { + "bbox": [ + 503, + 83, + 506, + 96 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 93, + 507, + 108 + ], + "spans": [ + { + "bbox": [ + 105, + 93, + 195, + 108 + ], + "score": 1.0, + "content": "Each input vector has", + "type": "text" + }, + { + "bbox": [ + 195, + 95, + 209, + 106 + ], + "score": 0.91, + "content": "| O |", + "type": "inline_equation" + }, + { + "bbox": [ + 210, + 93, + 507, + 108 + ], + "score": 1.0, + "content": "dimensions to represent the distribution over phonemes for each segment.", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 104, + 505, + 119 + ], + "spans": [ + { + "bbox": [ + 105, + 104, + 505, + 119 + ], + "score": 1.0, + "content": "The discriminator is also a CNN which outputs a probability indicating how likely the sample is to be", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 115, + 213, + 128 + ], + "spans": [ + { + "bbox": [ + 105, + 115, + 213, + 128 + ], + "score": 1.0, + "content": "from the data distribution.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 2 + }, + { + "type": "title", + "bbox": [ + 107, + 140, + 172, + 152 + ], + "lines": [ + { + "bbox": [ + 104, + 137, + 173, + 155 + ], + "spans": [ + { + "bbox": [ + 104, + 137, + 173, + 155 + ], + "score": 1.0, + "content": "3.2 Objective", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 5 + }, + { + "type": "text", + "bbox": [ + 108, + 160, + 505, + 183 + ], + "lines": [ + { + "bbox": [ + 105, + 159, + 507, + 173 + ], + "spans": [ + { + "bbox": [ + 105, + 159, + 507, + 173 + ], + "score": 1.0, + "content": "In our setup we use the original GAN objective with a gradient penalty [Goodfellow et al., 2014,", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 106, + 171, + 455, + 185 + ], + "spans": [ + { + "bbox": [ + 106, + 171, + 455, + 185 + ], + "score": 1.0, + "content": "Arjovsky et al., 2017], a segment smoothness penalty and a phoneme diversity penalty:", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 6.5 + }, + { + "type": "interline_equation", + "bbox": [ + 136, + 187, + 474, + 207 + ], + "lines": [ + { + "bbox": [ + 136, + 187, + 474, + 207 + ], + "spans": [ + { + "bbox": [ + 136, + 187, + 474, + 207 + ], + "score": 0.91, + "content": "\\operatorname* { m i n } _ { \\mathcal { G } } \\operatorname* { m a x } _ { \\mathcal { C } } \\quad \\mathbb { E } _ { \\mathcal { P } ^ { r } \\sim \\mathcal { P } ^ { r } } \\left[ \\log \\mathcal { C } ( \\boldsymbol { P } ^ { r } ) \\right] - \\underset { S \\sim \\mathcal { S } } { \\mathbb { E } } \\left[ \\log \\left( 1 - \\mathcal { C } ( \\mathcal { G } ( S ) ) \\right) \\right] - \\lambda \\mathcal { L } _ { g p } + \\gamma \\mathcal { L } _ { s p } + \\eta \\mathcal { L } _ { p d }", + "type": "interline_equation", + "image_path": "a15593f8fdca0cdc2e5a2bf415cdd52fd286427c617f23aeff5dbf4f23d44a11.jpg" + } + ] + } + ], + "index": 8, + "virtual_lines": [ + { + "bbox": [ + 136, + 187, + 474, + 207 + ], + "spans": [], + "index": 8 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 211, + 505, + 289 + ], + "lines": [ + { + "bbox": [ + 105, + 210, + 505, + 225 + ], + "spans": [ + { + "bbox": [ + 105, + 210, + 134, + 225 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 134, + 212, + 175, + 222 + ], + "score": 0.91, + "content": "P ^ { r } \\sim \\mathcal { P } ^ { r }", + "type": "inline_equation" + }, + { + "bbox": [ + 175, + 210, + 302, + 225 + ], + "score": 1.0, + "content": "is phonemized unlabeled text,", + "type": "text" + }, + { + "bbox": [ + 303, + 212, + 325, + 223 + ], + "score": 0.94, + "content": "\\mathcal { G } ( S )", + "type": "inline_equation" + }, + { + "bbox": [ + 325, + 210, + 505, + 225 + ], + "score": 1.0, + "content": "is the transcription output by the generator", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 106, + 222, + 505, + 235 + ], + "spans": [ + { + "bbox": [ + 106, + 222, + 245, + 235 + ], + "score": 1.0, + "content": "of input segment representations", + "type": "text" + }, + { + "bbox": [ + 246, + 223, + 254, + 233 + ], + "score": 0.81, + "content": "S", + "type": "inline_equation" + }, + { + "bbox": [ + 254, + 222, + 505, + 235 + ], + "score": 1.0, + "content": "for some unlabeled speech audio. The first term trains the", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 106, + 234, + 505, + 246 + ], + "spans": [ + { + "bbox": [ + 106, + 234, + 505, + 246 + ], + "score": 1.0, + "content": "discriminator to assign high probability to real transcriptions, the second term encourages the", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 244, + 506, + 258 + ], + "spans": [ + { + "bbox": [ + 105, + 244, + 360, + 258 + ], + "score": 1.0, + "content": "discriminator to assign low probability to generator outputs,", + "type": "text" + }, + { + "bbox": [ + 360, + 245, + 377, + 257 + ], + "score": 0.9, + "content": "\\mathcal { L } _ { g p }", + "type": "inline_equation" + }, + { + "bbox": [ + 377, + 244, + 469, + 258 + ], + "score": 1.0, + "content": "is a gradient penalty,", + "type": "text" + }, + { + "bbox": [ + 469, + 245, + 486, + 257 + ], + "score": 0.9, + "content": "\\mathcal { L } _ { s p }", + "type": "inline_equation" + }, + { + "bbox": [ + 486, + 244, + 506, + 258 + ], + "score": 1.0, + "content": "is a", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 255, + 506, + 269 + ], + "spans": [ + { + "bbox": [ + 105, + 255, + 205, + 269 + ], + "score": 1.0, + "content": "smoothness penalty and", + "type": "text" + }, + { + "bbox": [ + 206, + 256, + 223, + 267 + ], + "score": 0.91, + "content": "\\mathcal { L } _ { p d }", + "type": "inline_equation" + }, + { + "bbox": [ + 223, + 255, + 506, + 269 + ], + "score": 1.0, + "content": "is a phoneme diversity loss which we detail next. During training we", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 266, + 506, + 279 + ], + "spans": [ + { + "bbox": [ + 105, + 266, + 506, + 279 + ], + "score": 1.0, + "content": "alternate updates for the discriminator and the generator. We also alternate batches of predicted", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 277, + 369, + 289 + ], + "spans": [ + { + "bbox": [ + 105, + 277, + 369, + 289 + ], + "score": 1.0, + "content": "transcriptions from the generator and phonemized unlabeled text.", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 12 + }, + { + "type": "text", + "bbox": [ + 106, + 299, + 505, + 335 + ], + "lines": [ + { + "bbox": [ + 105, + 298, + 506, + 313 + ], + "spans": [ + { + "bbox": [ + 105, + 298, + 506, + 313 + ], + "score": 1.0, + "content": "Gradient penalty. To stabilize training, we penalize the gradient norm of the discriminator with", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 311, + 504, + 324 + ], + "spans": [ + { + "bbox": [ + 105, + 312, + 473, + 324 + ], + "score": 1.0, + "content": "respect to the input [Gulrajani et al., 2017]. The penalty is computed for random samples", + "type": "text" + }, + { + "bbox": [ + 473, + 311, + 504, + 322 + ], + "score": 0.93, + "content": "\\tilde { P } \\sim \\tilde { \\mathcal { P } }", + "type": "inline_equation" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 322, + 441, + 336 + ], + "spans": [ + { + "bbox": [ + 105, + 322, + 441, + 336 + ], + "score": 1.0, + "content": "which are a linear combination of the activations of pairs of real and fake samples.2", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 17 + }, + { + "type": "interline_equation", + "bbox": [ + 237, + 339, + 373, + 367 + ], + "lines": [ + { + "bbox": [ + 237, + 339, + 373, + 367 + ], + "spans": [ + { + "bbox": [ + 237, + 339, + 373, + 367 + ], + "score": 0.95, + "content": "\\mathcal { L } _ { g p } = \\underset { \\tilde { P } \\sim \\tilde { \\mathcal { P } } } { \\mathbb { E } } \\left[ \\left( \\| \\nabla \\mathcal { C } ( \\tilde { P } ) \\| - 1 \\right) ^ { 2 } \\right]", + "type": "interline_equation", + "image_path": "0a8f6af23cdb1998f6ef4e92c36819c82eb870dd5b721a5d19fc4f0384f93881.jpg" + } + ] + } + ], + "index": 19, + "virtual_lines": [ + { + "bbox": [ + 237, + 339, + 373, + 367 + ], + "spans": [], + "index": 19 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 376, + 505, + 422 + ], + "lines": [ + { + "bbox": [ + 105, + 376, + 506, + 389 + ], + "spans": [ + { + "bbox": [ + 105, + 376, + 258, + 389 + ], + "score": 1.0, + "content": "Segment smoothness penalty. The", + "type": "text" + }, + { + "bbox": [ + 259, + 378, + 266, + 387 + ], + "score": 0.65, + "content": "\\mathbf { k }", + "type": "inline_equation" + }, + { + "bbox": [ + 266, + 376, + 506, + 389 + ], + "score": 1.0, + "content": "-means segmentation of the speech audio is more granular", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 106, + 387, + 504, + 399 + ], + "spans": [ + { + "bbox": [ + 106, + 387, + 504, + 399 + ], + "score": 1.0, + "content": "than a typical phonemized transcription and neighboring representations are highly correlated. We", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 106, + 398, + 505, + 412 + ], + "spans": [ + { + "bbox": [ + 106, + 398, + 505, + 412 + ], + "score": 1.0, + "content": "therefore found it useful to add a penalty which encourages the generator to produce similar outputs", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 104, + 409, + 268, + 423 + ], + "spans": [ + { + "bbox": [ + 104, + 410, + 222, + 423 + ], + "score": 1.0, + "content": "for adjacent segments where", + "type": "text" + }, + { + "bbox": [ + 223, + 409, + 263, + 422 + ], + "score": 0.93, + "content": "p _ { t } \\in \\mathbb { R } ^ { | O | }", + "type": "inline_equation" + }, + { + "bbox": [ + 263, + 410, + 268, + 423 + ], + "score": 1.0, + "content": ":", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 21.5 + }, + { + "type": "interline_equation", + "bbox": [ + 235, + 426, + 376, + 455 + ], + "lines": [ + { + "bbox": [ + 235, + 426, + 376, + 455 + ], + "spans": [ + { + "bbox": [ + 235, + 426, + 376, + 455 + ], + "score": 0.94, + "content": "\\mathcal { L } _ { s p } = \\sum _ { ( p _ { t } , p _ { t + 1 } ) \\in \\mathcal { G } ( S ) } \\| p _ { t } - p _ { t + 1 } \\| ^ { 2 }", + "type": "interline_equation", + "image_path": "dc18aa35eab1ad24f09a6f15551a94289a62fd2eada4a74b608bff6bec0d7c91.jpg" + } + ] + } + ], + "index": 24.5, + "virtual_lines": [ + { + "bbox": [ + 235, + 426, + 376, + 440.5 + ], + "spans": [], + "index": 24 + }, + { + "bbox": [ + 235, + 440.5, + 376, + 455.0 + ], + "spans": [], + "index": 25 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 464, + 505, + 498 + ], + "lines": [ + { + "bbox": [ + 105, + 464, + 505, + 478 + ], + "spans": [ + { + "bbox": [ + 105, + 464, + 505, + 478 + ], + "score": 1.0, + "content": "Phoneme diversity loss. We also found it helpful to penalize low usage of the phoneme vocabulary", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 476, + 505, + 488 + ], + "spans": [ + { + "bbox": [ + 106, + 476, + 505, + 488 + ], + "score": 1.0, + "content": "by the generator on the batch level. In particular, we maximize the entropy of the averaged softmax", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 486, + 503, + 500 + ], + "spans": [ + { + "bbox": [ + 105, + 486, + 154, + 500 + ], + "score": 1.0, + "content": "distribution", + "type": "text" + }, + { + "bbox": [ + 154, + 487, + 197, + 499 + ], + "score": 0.93, + "content": "H _ { \\mathcal { G } } ( \\mathcal { G } ( S ) )", + "type": "inline_equation" + }, + { + "bbox": [ + 198, + 486, + 436, + 500 + ], + "score": 1.0, + "content": "of the generator over the phoneme vocabulary across a batch", + "type": "text" + }, + { + "bbox": [ + 437, + 487, + 446, + 497 + ], + "score": 0.83, + "content": "B", + "type": "inline_equation" + }, + { + "bbox": [ + 446, + 486, + 503, + 500 + ], + "score": 1.0, + "content": "of utterances:", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 27 + }, + { + "type": "interline_equation", + "bbox": [ + 246, + 502, + 365, + 533 + ], + "lines": [ + { + "bbox": [ + 246, + 502, + 365, + 533 + ], + "spans": [ + { + "bbox": [ + 246, + 502, + 365, + 533 + ], + "score": 0.95, + "content": "\\mathcal { L } _ { p d } = \\frac { 1 } { | B | } \\sum _ { S \\in B } - H _ { \\mathcal { G } } ( \\mathcal { G } ( S ) )", + "type": "interline_equation", + "image_path": "feda12d555ac0640b6d3cc1aeab0e1b03d7947fe4dd41ee8f6f0dbf28f5d27c6.jpg" + } + ] + } + ], + "index": 29.5, + "virtual_lines": [ + { + "bbox": [ + 246, + 502, + 365, + 517.5 + ], + "spans": [], + "index": 29 + }, + { + "bbox": [ + 246, + 517.5, + 365, + 533.0 + ], + "spans": [], + "index": 30 + } + ] + }, + { + "type": "title", + "bbox": [ + 107, + 544, + 294, + 556 + ], + "lines": [ + { + "bbox": [ + 105, + 543, + 295, + 557 + ], + "spans": [ + { + "bbox": [ + 105, + 543, + 295, + 557 + ], + "score": 1.0, + "content": "3.3 Unsupervised Cross-Validation Metric", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 31 + }, + { + "type": "text", + "bbox": [ + 106, + 564, + 506, + 598 + ], + "lines": [ + { + "bbox": [ + 106, + 564, + 506, + 576 + ], + "spans": [ + { + "bbox": [ + 106, + 564, + 506, + 576 + ], + "score": 1.0, + "content": "Our goal is to build speech recognition models without any supervision. To this end, we developed a", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 573, + 507, + 590 + ], + "spans": [ + { + "bbox": [ + 105, + 573, + 507, + 590 + ], + "score": 1.0, + "content": "cross-validation metric which does not require labeled data. We use the metric for early stopping,", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 585, + 369, + 600 + ], + "spans": [ + { + "bbox": [ + 105, + 585, + 330, + 600 + ], + "score": 1.0, + "content": "selecting a random seed, and hyper-parameter selection", + "type": "text" + }, + { + "bbox": [ + 330, + 587, + 364, + 598 + ], + "score": 0.89, + "content": "( \\lambda , \\gamma , \\eta )", + "type": "inline_equation" + }, + { + "bbox": [ + 365, + 585, + 369, + 600 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 34 + } + ], + "index": 33 + }, + { + "type": "text", + "bbox": [ + 106, + 602, + 505, + 658 + ], + "lines": [ + { + "bbox": [ + 105, + 601, + 506, + 615 + ], + "spans": [ + { + "bbox": [ + 105, + 601, + 506, + 615 + ], + "score": 1.0, + "content": "We consider two quantities in our metric: LM negative log-likelihood (NLL) and vocabulary usage.", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 613, + 505, + 627 + ], + "spans": [ + { + "bbox": [ + 105, + 613, + 505, + 627 + ], + "score": 1.0, + "content": "LM-NLL serves as an indicator of fluency for a given transcription and it is measured with a language", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 624, + 505, + 637 + ], + "spans": [ + { + "bbox": [ + 105, + 624, + 135, + 637 + ], + "score": 1.0, + "content": "model", + "type": "text" + }, + { + "bbox": [ + 135, + 626, + 156, + 636 + ], + "score": 0.87, + "content": "p _ { L M }", + "type": "inline_equation" + }, + { + "bbox": [ + 156, + 624, + 295, + 637 + ], + "score": 1.0, + "content": "trained on phonemized text data", + "type": "text" + }, + { + "bbox": [ + 295, + 625, + 324, + 636 + ], + "score": 0.81, + "content": "( \\ S ~ 2 . 4 )", + "type": "inline_equation" + }, + { + "bbox": [ + 324, + 624, + 505, + 637 + ], + "score": 1.0, + "content": ". Vocabulary usage is the proportion of the", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 633, + 506, + 650 + ], + "spans": [ + { + "bbox": [ + 105, + 633, + 506, + 650 + ], + "score": 1.0, + "content": "phoneme vocabulary being output by the model via Viterbi decoding. Measuring vocabulary usage", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 646, + 402, + 660 + ], + "spans": [ + { + "bbox": [ + 105, + 646, + 402, + 660 + ], + "score": 1.0, + "content": "identifies degenerate models which output fluent but trivial transcriptions.", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 37 + }, + { + "type": "text", + "bbox": [ + 107, + 662, + 506, + 704 + ], + "lines": [ + { + "bbox": [ + 101, + 662, + 511, + 694 + ], + "spans": [ + { + "bbox": [ + 101, + 662, + 130, + 694 + ], + "score": 1.0, + "content": "We deaudio", + "type": "text" + }, + { + "bbox": [ + 168, + 662, + 179, + 694 + ], + "score": 1.0, + "content": "rbias", + "type": "text" + }, + { + "bbox": [ + 236, + 662, + 410, + 694 + ], + "score": 1.0, + "content": "criptions for a given generator configuration . LM-NLL is measured in the standard way o", + "type": "text" + }, + { + "bbox": [ + 410, + 663, + 418, + 673 + ], + "score": 0.84, + "content": "\\mathcal { G }", + "type": "inline_equation" + }, + { + "bbox": [ + 419, + 662, + 511, + 694 + ], + "score": 1.0, + "content": "and unlabeled speechr the phonemized tran-", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 131, + 673, + 236, + 689 + ], + "spans": [ + { + "bbox": [ + 131, + 673, + 168, + 689 + ], + "score": 0.93, + "content": "\\{ X _ { j } \\} _ { j = 1 } ^ { N _ { s } }", + "type": "inline_equation" + }, + { + "bbox": [ + 179, + 673, + 236, + 689 + ], + "score": 0.93, + "content": "\\mathbf { \\bar { \\mathcal { P } } } = \\{ P _ { j } \\} _ { j = 1 } ^ { N _ { s } }", + "type": "inline_equation" + } + ], + "index": 41 + }, + { + "bbox": [ + 106, + 686, + 505, + 706 + ], + "spans": [ + { + "bbox": [ + 106, + 689, + 151, + 703 + ], + "score": 1.0, + "content": "scriptions:", + "type": "text" + }, + { + "bbox": [ + 151, + 688, + 314, + 704 + ], + "score": 0.89, + "content": "\\begin{array} { r } { N L L _ { L M } ( \\mathcal { P } ) = \\frac { 1 } { N _ { s } } \\sum _ { j = 1 } ^ { N _ { s } } N L L _ { L M } ( P _ { j } ) } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 314, + 686, + 342, + 706 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 342, + 687, + 505, + 703 + ], + "score": 0.92, + "content": "\\begin{array} { r } { N L L _ { L M } ( P ) = - \\frac { 1 } { M } \\sum _ { t = 1 } ^ { M } \\log p _ { L M } ( p _ { t } ) } \\end{array}", + "type": "inline_equation" + } + ], + "index": 42 + } + ], + "index": 41 + } + ], + "page_idx": 4, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 119, + 711, + 324, + 722 + ], + "lines": [ + { + "bbox": [ + 118, + 709, + 325, + 725 + ], + "spans": [ + { + "bbox": [ + 118, + 709, + 325, + 725 + ], + "score": 1.0, + "content": "2We simply shorten longer sequence if the lengths differ.", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 741, + 309, + 750 + ], + "lines": [ + { + "bbox": [ + 301, + 740, + 310, + 753 + ], + "spans": [ + { + "bbox": [ + 301, + 740, + 310, + 753 + ], + "score": 1.0, + "content": "5", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 106, + 72, + 506, + 128 + ], + "lines": [ + { + "bbox": [ + 106, + 72, + 506, + 86 + ], + "spans": [ + { + "bbox": [ + 106, + 72, + 304, + 86 + ], + "score": 1.0, + "content": "The discriminator takes as input either a sequence", + "type": "text" + }, + { + "bbox": [ + 304, + 73, + 343, + 83 + ], + "score": 0.91, + "content": "P ^ { r } \\sim \\mathcal { P } ^ { r }", + "type": "inline_equation" + }, + { + "bbox": [ + 343, + 72, + 506, + 86 + ], + "score": 1.0, + "content": "of one-hot vectors denoting phonemized", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 83, + 506, + 96 + ], + "spans": [ + { + "bbox": [ + 105, + 83, + 244, + 96 + ], + "score": 1.0, + "content": "text from the real data distribution", + "type": "text" + }, + { + "bbox": [ + 244, + 84, + 257, + 94 + ], + "score": 0.85, + "content": "{ \\mathcal { P } } ^ { r }", + "type": "inline_equation" + }, + { + "bbox": [ + 257, + 83, + 481, + 96 + ], + "score": 1.0, + "content": "or a sequence of output distributions from the generator", + "type": "text" + }, + { + "bbox": [ + 481, + 83, + 502, + 96 + ], + "score": 0.92, + "content": "\\mathcal { G } ( S )", + "type": "inline_equation" + }, + { + "bbox": [ + 503, + 83, + 506, + 96 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 93, + 507, + 108 + ], + "spans": [ + { + "bbox": [ + 105, + 93, + 195, + 108 + ], + "score": 1.0, + "content": "Each input vector has", + "type": "text" + }, + { + "bbox": [ + 195, + 95, + 209, + 106 + ], + "score": 0.91, + "content": "| O |", + "type": "inline_equation" + }, + { + "bbox": [ + 210, + 93, + 507, + 108 + ], + "score": 1.0, + "content": "dimensions to represent the distribution over phonemes for each segment.", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 104, + 505, + 119 + ], + "spans": [ + { + "bbox": [ + 105, + 104, + 505, + 119 + ], + "score": 1.0, + "content": "The discriminator is also a CNN which outputs a probability indicating how likely the sample is to be", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 115, + 213, + 128 + ], + "spans": [ + { + "bbox": [ + 105, + 115, + 213, + 128 + ], + "score": 1.0, + "content": "from the data distribution.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 2, + "bbox_fs": [ + 105, + 72, + 507, + 128 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 140, + 172, + 152 + ], + "lines": [ + { + "bbox": [ + 104, + 137, + 173, + 155 + ], + "spans": [ + { + "bbox": [ + 104, + 137, + 173, + 155 + ], + "score": 1.0, + "content": "3.2 Objective", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 5 + }, + { + "type": "text", + "bbox": [ + 108, + 160, + 505, + 183 + ], + "lines": [ + { + "bbox": [ + 105, + 159, + 507, + 173 + ], + "spans": [ + { + "bbox": [ + 105, + 159, + 507, + 173 + ], + "score": 1.0, + "content": "In our setup we use the original GAN objective with a gradient penalty [Goodfellow et al., 2014,", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 106, + 171, + 455, + 185 + ], + "spans": [ + { + "bbox": [ + 106, + 171, + 455, + 185 + ], + "score": 1.0, + "content": "Arjovsky et al., 2017], a segment smoothness penalty and a phoneme diversity penalty:", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 6.5, + "bbox_fs": [ + 105, + 159, + 507, + 185 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 136, + 187, + 474, + 207 + ], + "lines": [ + { + "bbox": [ + 136, + 187, + 474, + 207 + ], + "spans": [ + { + "bbox": [ + 136, + 187, + 474, + 207 + ], + "score": 0.91, + "content": "\\operatorname* { m i n } _ { \\mathcal { G } } \\operatorname* { m a x } _ { \\mathcal { C } } \\quad \\mathbb { E } _ { \\mathcal { P } ^ { r } \\sim \\mathcal { P } ^ { r } } \\left[ \\log \\mathcal { C } ( \\boldsymbol { P } ^ { r } ) \\right] - \\underset { S \\sim \\mathcal { S } } { \\mathbb { E } } \\left[ \\log \\left( 1 - \\mathcal { C } ( \\mathcal { G } ( S ) ) \\right) \\right] - \\lambda \\mathcal { L } _ { g p } + \\gamma \\mathcal { L } _ { s p } + \\eta \\mathcal { L } _ { p d }", + "type": "interline_equation", + "image_path": "a15593f8fdca0cdc2e5a2bf415cdd52fd286427c617f23aeff5dbf4f23d44a11.jpg" + } + ] + } + ], + "index": 8, + "virtual_lines": [ + { + "bbox": [ + 136, + 187, + 474, + 207 + ], + "spans": [], + "index": 8 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 211, + 505, + 289 + ], + "lines": [ + { + "bbox": [ + 105, + 210, + 505, + 225 + ], + "spans": [ + { + "bbox": [ + 105, + 210, + 134, + 225 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 134, + 212, + 175, + 222 + ], + "score": 0.91, + "content": "P ^ { r } \\sim \\mathcal { P } ^ { r }", + "type": "inline_equation" + }, + { + "bbox": [ + 175, + 210, + 302, + 225 + ], + "score": 1.0, + "content": "is phonemized unlabeled text,", + "type": "text" + }, + { + "bbox": [ + 303, + 212, + 325, + 223 + ], + "score": 0.94, + "content": "\\mathcal { G } ( S )", + "type": "inline_equation" + }, + { + "bbox": [ + 325, + 210, + 505, + 225 + ], + "score": 1.0, + "content": "is the transcription output by the generator", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 106, + 222, + 505, + 235 + ], + "spans": [ + { + "bbox": [ + 106, + 222, + 245, + 235 + ], + "score": 1.0, + "content": "of input segment representations", + "type": "text" + }, + { + "bbox": [ + 246, + 223, + 254, + 233 + ], + "score": 0.81, + "content": "S", + "type": "inline_equation" + }, + { + "bbox": [ + 254, + 222, + 505, + 235 + ], + "score": 1.0, + "content": "for some unlabeled speech audio. The first term trains the", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 106, + 234, + 505, + 246 + ], + "spans": [ + { + "bbox": [ + 106, + 234, + 505, + 246 + ], + "score": 1.0, + "content": "discriminator to assign high probability to real transcriptions, the second term encourages the", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 244, + 506, + 258 + ], + "spans": [ + { + "bbox": [ + 105, + 244, + 360, + 258 + ], + "score": 1.0, + "content": "discriminator to assign low probability to generator outputs,", + "type": "text" + }, + { + "bbox": [ + 360, + 245, + 377, + 257 + ], + "score": 0.9, + "content": "\\mathcal { L } _ { g p }", + "type": "inline_equation" + }, + { + "bbox": [ + 377, + 244, + 469, + 258 + ], + "score": 1.0, + "content": "is a gradient penalty,", + "type": "text" + }, + { + "bbox": [ + 469, + 245, + 486, + 257 + ], + "score": 0.9, + "content": "\\mathcal { L } _ { s p }", + "type": "inline_equation" + }, + { + "bbox": [ + 486, + 244, + 506, + 258 + ], + "score": 1.0, + "content": "is a", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 255, + 506, + 269 + ], + "spans": [ + { + "bbox": [ + 105, + 255, + 205, + 269 + ], + "score": 1.0, + "content": "smoothness penalty and", + "type": "text" + }, + { + "bbox": [ + 206, + 256, + 223, + 267 + ], + "score": 0.91, + "content": "\\mathcal { L } _ { p d }", + "type": "inline_equation" + }, + { + "bbox": [ + 223, + 255, + 506, + 269 + ], + "score": 1.0, + "content": "is a phoneme diversity loss which we detail next. During training we", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 266, + 506, + 279 + ], + "spans": [ + { + "bbox": [ + 105, + 266, + 506, + 279 + ], + "score": 1.0, + "content": "alternate updates for the discriminator and the generator. We also alternate batches of predicted", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 277, + 369, + 289 + ], + "spans": [ + { + "bbox": [ + 105, + 277, + 369, + 289 + ], + "score": 1.0, + "content": "transcriptions from the generator and phonemized unlabeled text.", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 12, + "bbox_fs": [ + 105, + 210, + 506, + 289 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 299, + 505, + 335 + ], + "lines": [ + { + "bbox": [ + 105, + 298, + 506, + 313 + ], + "spans": [ + { + "bbox": [ + 105, + 298, + 506, + 313 + ], + "score": 1.0, + "content": "Gradient penalty. To stabilize training, we penalize the gradient norm of the discriminator with", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 311, + 504, + 324 + ], + "spans": [ + { + "bbox": [ + 105, + 312, + 473, + 324 + ], + "score": 1.0, + "content": "respect to the input [Gulrajani et al., 2017]. The penalty is computed for random samples", + "type": "text" + }, + { + "bbox": [ + 473, + 311, + 504, + 322 + ], + "score": 0.93, + "content": "\\tilde { P } \\sim \\tilde { \\mathcal { P } }", + "type": "inline_equation" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 322, + 441, + 336 + ], + "spans": [ + { + "bbox": [ + 105, + 322, + 441, + 336 + ], + "score": 1.0, + "content": "which are a linear combination of the activations of pairs of real and fake samples.2", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 17, + "bbox_fs": [ + 105, + 298, + 506, + 336 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 237, + 339, + 373, + 367 + ], + "lines": [ + { + "bbox": [ + 237, + 339, + 373, + 367 + ], + "spans": [ + { + "bbox": [ + 237, + 339, + 373, + 367 + ], + "score": 0.95, + "content": "\\mathcal { L } _ { g p } = \\underset { \\tilde { P } \\sim \\tilde { \\mathcal { P } } } { \\mathbb { E } } \\left[ \\left( \\| \\nabla \\mathcal { C } ( \\tilde { P } ) \\| - 1 \\right) ^ { 2 } \\right]", + "type": "interline_equation", + "image_path": "0a8f6af23cdb1998f6ef4e92c36819c82eb870dd5b721a5d19fc4f0384f93881.jpg" + } + ] + } + ], + "index": 19, + "virtual_lines": [ + { + "bbox": [ + 237, + 339, + 373, + 367 + ], + "spans": [], + "index": 19 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 376, + 505, + 422 + ], + "lines": [ + { + "bbox": [ + 105, + 376, + 506, + 389 + ], + "spans": [ + { + "bbox": [ + 105, + 376, + 258, + 389 + ], + "score": 1.0, + "content": "Segment smoothness penalty. The", + "type": "text" + }, + { + "bbox": [ + 259, + 378, + 266, + 387 + ], + "score": 0.65, + "content": "\\mathbf { k }", + "type": "inline_equation" + }, + { + "bbox": [ + 266, + 376, + 506, + 389 + ], + "score": 1.0, + "content": "-means segmentation of the speech audio is more granular", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 106, + 387, + 504, + 399 + ], + "spans": [ + { + "bbox": [ + 106, + 387, + 504, + 399 + ], + "score": 1.0, + "content": "than a typical phonemized transcription and neighboring representations are highly correlated. We", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 106, + 398, + 505, + 412 + ], + "spans": [ + { + "bbox": [ + 106, + 398, + 505, + 412 + ], + "score": 1.0, + "content": "therefore found it useful to add a penalty which encourages the generator to produce similar outputs", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 104, + 409, + 268, + 423 + ], + "spans": [ + { + "bbox": [ + 104, + 410, + 222, + 423 + ], + "score": 1.0, + "content": "for adjacent segments where", + "type": "text" + }, + { + "bbox": [ + 223, + 409, + 263, + 422 + ], + "score": 0.93, + "content": "p _ { t } \\in \\mathbb { R } ^ { | O | }", + "type": "inline_equation" + }, + { + "bbox": [ + 263, + 410, + 268, + 423 + ], + "score": 1.0, + "content": ":", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 21.5, + "bbox_fs": [ + 104, + 376, + 506, + 423 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 235, + 426, + 376, + 455 + ], + "lines": [ + { + "bbox": [ + 235, + 426, + 376, + 455 + ], + "spans": [ + { + "bbox": [ + 235, + 426, + 376, + 455 + ], + "score": 0.94, + "content": "\\mathcal { L } _ { s p } = \\sum _ { ( p _ { t } , p _ { t + 1 } ) \\in \\mathcal { G } ( S ) } \\| p _ { t } - p _ { t + 1 } \\| ^ { 2 }", + "type": "interline_equation", + "image_path": "dc18aa35eab1ad24f09a6f15551a94289a62fd2eada4a74b608bff6bec0d7c91.jpg" + } + ] + } + ], + "index": 24.5, + "virtual_lines": [ + { + "bbox": [ + 235, + 426, + 376, + 440.5 + ], + "spans": [], + "index": 24 + }, + { + "bbox": [ + 235, + 440.5, + 376, + 455.0 + ], + "spans": [], + "index": 25 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 464, + 505, + 498 + ], + "lines": [ + { + "bbox": [ + 105, + 464, + 505, + 478 + ], + "spans": [ + { + "bbox": [ + 105, + 464, + 505, + 478 + ], + "score": 1.0, + "content": "Phoneme diversity loss. We also found it helpful to penalize low usage of the phoneme vocabulary", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 476, + 505, + 488 + ], + "spans": [ + { + "bbox": [ + 106, + 476, + 505, + 488 + ], + "score": 1.0, + "content": "by the generator on the batch level. In particular, we maximize the entropy of the averaged softmax", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 486, + 503, + 500 + ], + "spans": [ + { + "bbox": [ + 105, + 486, + 154, + 500 + ], + "score": 1.0, + "content": "distribution", + "type": "text" + }, + { + "bbox": [ + 154, + 487, + 197, + 499 + ], + "score": 0.93, + "content": "H _ { \\mathcal { G } } ( \\mathcal { G } ( S ) )", + "type": "inline_equation" + }, + { + "bbox": [ + 198, + 486, + 436, + 500 + ], + "score": 1.0, + "content": "of the generator over the phoneme vocabulary across a batch", + "type": "text" + }, + { + "bbox": [ + 437, + 487, + 446, + 497 + ], + "score": 0.83, + "content": "B", + "type": "inline_equation" + }, + { + "bbox": [ + 446, + 486, + 503, + 500 + ], + "score": 1.0, + "content": "of utterances:", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 27, + "bbox_fs": [ + 105, + 464, + 505, + 500 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 246, + 502, + 365, + 533 + ], + "lines": [ + { + "bbox": [ + 246, + 502, + 365, + 533 + ], + "spans": [ + { + "bbox": [ + 246, + 502, + 365, + 533 + ], + "score": 0.95, + "content": "\\mathcal { L } _ { p d } = \\frac { 1 } { | B | } \\sum _ { S \\in B } - H _ { \\mathcal { G } } ( \\mathcal { G } ( S ) )", + "type": "interline_equation", + "image_path": "feda12d555ac0640b6d3cc1aeab0e1b03d7947fe4dd41ee8f6f0dbf28f5d27c6.jpg" + } + ] + } + ], + "index": 29.5, + "virtual_lines": [ + { + "bbox": [ + 246, + 502, + 365, + 517.5 + ], + "spans": [], + "index": 29 + }, + { + "bbox": [ + 246, + 517.5, + 365, + 533.0 + ], + "spans": [], + "index": 30 + } + ] + }, + { + "type": "title", + "bbox": [ + 107, + 544, + 294, + 556 + ], + "lines": [ + { + "bbox": [ + 105, + 543, + 295, + 557 + ], + "spans": [ + { + "bbox": [ + 105, + 543, + 295, + 557 + ], + "score": 1.0, + "content": "3.3 Unsupervised Cross-Validation Metric", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 31 + }, + { + "type": "text", + "bbox": [ + 106, + 564, + 506, + 598 + ], + "lines": [ + { + "bbox": [ + 106, + 564, + 506, + 576 + ], + "spans": [ + { + "bbox": [ + 106, + 564, + 506, + 576 + ], + "score": 1.0, + "content": "Our goal is to build speech recognition models without any supervision. To this end, we developed a", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 573, + 507, + 590 + ], + "spans": [ + { + "bbox": [ + 105, + 573, + 507, + 590 + ], + "score": 1.0, + "content": "cross-validation metric which does not require labeled data. We use the metric for early stopping,", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 585, + 369, + 600 + ], + "spans": [ + { + "bbox": [ + 105, + 585, + 330, + 600 + ], + "score": 1.0, + "content": "selecting a random seed, and hyper-parameter selection", + "type": "text" + }, + { + "bbox": [ + 330, + 587, + 364, + 598 + ], + "score": 0.89, + "content": "( \\lambda , \\gamma , \\eta )", + "type": "inline_equation" + }, + { + "bbox": [ + 365, + 585, + 369, + 600 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 34 + } + ], + "index": 33, + "bbox_fs": [ + 105, + 564, + 507, + 600 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 602, + 505, + 658 + ], + "lines": [ + { + "bbox": [ + 105, + 601, + 506, + 615 + ], + "spans": [ + { + "bbox": [ + 105, + 601, + 506, + 615 + ], + "score": 1.0, + "content": "We consider two quantities in our metric: LM negative log-likelihood (NLL) and vocabulary usage.", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 613, + 505, + 627 + ], + "spans": [ + { + "bbox": [ + 105, + 613, + 505, + 627 + ], + "score": 1.0, + "content": "LM-NLL serves as an indicator of fluency for a given transcription and it is measured with a language", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 624, + 505, + 637 + ], + "spans": [ + { + "bbox": [ + 105, + 624, + 135, + 637 + ], + "score": 1.0, + "content": "model", + "type": "text" + }, + { + "bbox": [ + 135, + 626, + 156, + 636 + ], + "score": 0.87, + "content": "p _ { L M }", + "type": "inline_equation" + }, + { + "bbox": [ + 156, + 624, + 295, + 637 + ], + "score": 1.0, + "content": "trained on phonemized text data", + "type": "text" + }, + { + "bbox": [ + 295, + 625, + 324, + 636 + ], + "score": 0.81, + "content": "( \\ S ~ 2 . 4 )", + "type": "inline_equation" + }, + { + "bbox": [ + 324, + 624, + 505, + 637 + ], + "score": 1.0, + "content": ". Vocabulary usage is the proportion of the", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 633, + 506, + 650 + ], + "spans": [ + { + "bbox": [ + 105, + 633, + 506, + 650 + ], + "score": 1.0, + "content": "phoneme vocabulary being output by the model via Viterbi decoding. Measuring vocabulary usage", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 646, + 402, + 660 + ], + "spans": [ + { + "bbox": [ + 105, + 646, + 402, + 660 + ], + "score": 1.0, + "content": "identifies degenerate models which output fluent but trivial transcriptions.", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 37, + "bbox_fs": [ + 105, + 601, + 506, + 660 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 662, + 506, + 704 + ], + "lines": [ + { + "bbox": [ + 101, + 662, + 511, + 694 + ], + "spans": [ + { + "bbox": [ + 101, + 662, + 130, + 694 + ], + "score": 1.0, + "content": "We deaudio", + "type": "text" + }, + { + "bbox": [ + 168, + 662, + 179, + 694 + ], + "score": 1.0, + "content": "rbias", + "type": "text" + }, + { + "bbox": [ + 236, + 662, + 410, + 694 + ], + "score": 1.0, + "content": "criptions for a given generator configuration . LM-NLL is measured in the standard way o", + "type": "text" + }, + { + "bbox": [ + 410, + 663, + 418, + 673 + ], + "score": 0.84, + "content": "\\mathcal { G }", + "type": "inline_equation" + }, + { + "bbox": [ + 419, + 662, + 511, + 694 + ], + "score": 1.0, + "content": "and unlabeled speechr the phonemized tran-", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 131, + 673, + 236, + 689 + ], + "spans": [ + { + "bbox": [ + 131, + 673, + 168, + 689 + ], + "score": 0.93, + "content": "\\{ X _ { j } \\} _ { j = 1 } ^ { N _ { s } }", + "type": "inline_equation" + }, + { + "bbox": [ + 179, + 673, + 236, + 689 + ], + "score": 0.93, + "content": "\\mathbf { \\bar { \\mathcal { P } } } = \\{ P _ { j } \\} _ { j = 1 } ^ { N _ { s } }", + "type": "inline_equation" + } + ], + "index": 41 + }, + { + "bbox": [ + 106, + 686, + 505, + 706 + ], + "spans": [ + { + "bbox": [ + 106, + 689, + 151, + 703 + ], + "score": 1.0, + "content": "scriptions:", + "type": "text" + }, + { + "bbox": [ + 151, + 688, + 314, + 704 + ], + "score": 0.89, + "content": "\\begin{array} { r } { N L L _ { L M } ( \\mathcal { P } ) = \\frac { 1 } { N _ { s } } \\sum _ { j = 1 } ^ { N _ { s } } N L L _ { L M } ( P _ { j } ) } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 314, + 686, + 342, + 706 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 342, + 687, + 505, + 703 + ], + "score": 0.92, + "content": "\\begin{array} { r } { N L L _ { L M } ( P ) = - \\frac { 1 } { M } \\sum _ { t = 1 } ^ { M } \\log p _ { L M } ( p _ { t } ) } \\end{array}", + "type": "inline_equation" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 70, + 505, + 86 + ], + "spans": [ + { + "bbox": [ + 105, + 70, + 132, + 86 + ], + "score": 1.0, + "content": "using", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 132, + 73, + 169, + 84 + ], + "score": 0.9, + "content": "p _ { L M } ( p _ { t } )", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 169, + 70, + 245, + 86 + ], + "score": 1.0, + "content": "as shorthand for", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 245, + 73, + 334, + 85 + ], + "score": 0.9, + "content": "p _ { L M } ( p _ { t } | p _ { t - 1 } , . . . , p _ { 1 } )", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 335, + 70, + 465, + 86 + ], + "score": 1.0, + "content": ".3 On the other hand, we use", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 466, + 73, + 505, + 85 + ], + "score": 0.9, + "content": "U ( { \\mathcal { P } } ) =", + "type": "inline_equation", + "cross_page": true + } + ], + "index": 0 + }, + { + "bbox": [ + 107, + 83, + 362, + 99 + ], + "spans": [ + { + "bbox": [ + 107, + 83, + 210, + 99 + ], + "score": 0.9, + "content": "\\begin{array} { r } { \\frac { 1 } { | O | } \\bar { \\sum _ { o \\in O } } [ o \\in \\mathcal { P } ] \\in [ 0 , 1 ] } \\end{array}", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 210, + 83, + 348, + 96 + ], + "score": 1.0, + "content": "to denote the vocabulary usage of", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 348, + 84, + 356, + 93 + ], + "score": 0.84, + "content": "\\mathcal { P }", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 357, + 83, + 362, + 96 + ], + "score": 1.0, + "content": ".", + "type": "text", + "cross_page": true + } + ], + "index": 1 + } + ], + "index": 41, + "bbox_fs": [ + 101, + 662, + 511, + 706 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 104, + 72, + 504, + 96 + ], + "lines": [ + { + "bbox": [ + 105, + 70, + 505, + 86 + ], + "spans": [ + { + "bbox": [ + 105, + 70, + 132, + 86 + ], + "score": 1.0, + "content": "using", + "type": "text" + }, + { + "bbox": [ + 132, + 73, + 169, + 84 + ], + "score": 0.9, + "content": "p _ { L M } ( p _ { t } )", + "type": "inline_equation" + }, + { + "bbox": [ + 169, + 70, + 245, + 86 + ], + "score": 1.0, + "content": "as shorthand for", + "type": "text" + }, + { + "bbox": [ + 245, + 73, + 334, + 85 + ], + "score": 0.9, + "content": "p _ { L M } ( p _ { t } | p _ { t - 1 } , . . . , p _ { 1 } )", + "type": "inline_equation" + }, + { + "bbox": [ + 335, + 70, + 465, + 86 + ], + "score": 1.0, + "content": ".3 On the other hand, we use", + "type": "text" + }, + { + "bbox": [ + 466, + 73, + 505, + 85 + ], + "score": 0.9, + "content": "U ( { \\mathcal { P } } ) =", + "type": "inline_equation" + } + ], + "index": 0 + }, + { + "bbox": [ + 107, + 83, + 362, + 99 + ], + "spans": [ + { + "bbox": [ + 107, + 83, + 210, + 99 + ], + "score": 0.9, + "content": "\\begin{array} { r } { \\frac { 1 } { | O | } \\bar { \\sum _ { o \\in O } } [ o \\in \\mathcal { P } ] \\in [ 0 , 1 ] } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 210, + 83, + 348, + 96 + ], + "score": 1.0, + "content": "to denote the vocabulary usage of", + "type": "text" + }, + { + "bbox": [ + 348, + 84, + 356, + 93 + ], + "score": 0.84, + "content": "\\mathcal { P }", + "type": "inline_equation" + }, + { + "bbox": [ + 357, + 83, + 362, + 96 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5 + }, + { + "type": "text", + "bbox": [ + 107, + 101, + 505, + 150 + ], + "lines": [ + { + "bbox": [ + 105, + 102, + 506, + 114 + ], + "spans": [ + { + "bbox": [ + 105, + 102, + 506, + 114 + ], + "score": 1.0, + "content": "In a first step, we generate phoneme transcriptions for different training checkpoints or hyper-", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 112, + 505, + 126 + ], + "spans": [ + { + "bbox": [ + 105, + 112, + 505, + 126 + ], + "score": 1.0, + "content": "parameter settings and denote the transcriptions of the configuration with the lowest vocabulary-usage", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 124, + 505, + 138 + ], + "spans": [ + { + "bbox": [ + 105, + 124, + 174, + 138 + ], + "score": 1.0, + "content": "adjusted NLL as", + "type": "text" + }, + { + "bbox": [ + 175, + 124, + 340, + 137 + ], + "score": 0.84, + "content": "\\begin{array} { r } { \\hat { \\mathcal { P } } = \\arg \\operatorname* { m i n } _ { \\mathcal { P } } N L L _ { L M } ( \\mathcal { P } ) - \\log U ( \\mathcal { P } ) } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 340, + 124, + 505, + 138 + ], + "score": 1.0, + "content": ".4 Next, we discard model configurations", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 137, + 336, + 151 + ], + "spans": [ + { + "bbox": [ + 105, + 138, + 269, + 151 + ], + "score": 1.0, + "content": "which do not satisfy the following using", + "type": "text" + }, + { + "bbox": [ + 269, + 137, + 278, + 149 + ], + "score": 0.85, + "content": "\\hat { \\mathcal { P } }", + "type": "inline_equation" + }, + { + "bbox": [ + 278, + 138, + 336, + 151 + ], + "score": 1.0, + "content": "as the anchor:", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 3.5 + }, + { + "type": "interline_equation", + "bbox": [ + 191, + 156, + 420, + 190 + ], + "lines": [ + { + "bbox": [ + 191, + 156, + 420, + 190 + ], + "spans": [ + { + "bbox": [ + 191, + 156, + 420, + 190 + ], + "score": 0.93, + "content": "N L L _ { L M } ( \\mathcal { P } ) < N L L _ { L M } ( \\hat { \\mathcal { P } } ) + \\log \\left( \\frac { U ( \\mathcal { P } ) } { U ( \\hat { \\mathcal { P } } ) } \\right) + \\log 1 . 2", + "type": "interline_equation", + "image_path": "b07976d07eb1264788bb86ec8c503a84fc3c183351681c48efba85a9d06b98dd.jpg" + } + ] + } + ], + "index": 6.5, + "virtual_lines": [ + { + "bbox": [ + 191, + 156, + 420, + 173.0 + ], + "spans": [], + "index": 6 + }, + { + "bbox": [ + 191, + 173.0, + 420, + 190.0 + ], + "spans": [], + "index": 7 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 196, + 505, + 267 + ], + "lines": [ + { + "bbox": [ + 105, + 196, + 506, + 210 + ], + "spans": [ + { + "bbox": [ + 105, + 196, + 506, + 210 + ], + "score": 1.0, + "content": "The second term on the right hand side introduces a margin over the NLL of the anchor transcription", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 107, + 208, + 506, + 223 + ], + "spans": [ + { + "bbox": [ + 107, + 208, + 160, + 222 + ], + "score": 0.93, + "content": "N L L _ { L M } ( \\hat { \\mathcal { P } } )", + "type": "inline_equation" + }, + { + "bbox": [ + 160, + 208, + 294, + 223 + ], + "score": 1.0, + "content": "based on the vocabulary usage of", + "type": "text" + }, + { + "bbox": [ + 294, + 210, + 303, + 220 + ], + "score": 0.82, + "content": "\\mathcal { P }", + "type": "inline_equation" + }, + { + "bbox": [ + 303, + 208, + 320, + 223 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 321, + 208, + 329, + 220 + ], + "score": 0.84, + "content": "\\hat { \\mathcal { P } }", + "type": "inline_equation" + }, + { + "bbox": [ + 330, + 208, + 343, + 223 + ], + "score": 1.0, + "content": ": If", + "type": "text" + }, + { + "bbox": [ + 344, + 208, + 368, + 222 + ], + "score": 0.93, + "content": "U ( \\hat { \\mathcal { P } } )", + "type": "inline_equation" + }, + { + "bbox": [ + 368, + 208, + 478, + 223 + ], + "score": 1.0, + "content": "is much lower compared to", + "type": "text" + }, + { + "bbox": [ + 479, + 212, + 502, + 222 + ], + "score": 0.93, + "content": "U ( \\mathcal { P } )", + "type": "inline_equation" + }, + { + "bbox": [ + 502, + 208, + 506, + 223 + ], + "score": 1.0, + "content": ",", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 106, + 220, + 506, + 234 + ], + "spans": [ + { + "bbox": [ + 106, + 220, + 506, + 234 + ], + "score": 1.0, + "content": "then we allow model configurations which produce transcriptions with higher NLL compared to", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 106, + 231, + 505, + 246 + ], + "spans": [ + { + "bbox": [ + 106, + 231, + 115, + 244 + ], + "score": 0.82, + "content": "\\hat { \\mathcal { P } }", + "type": "inline_equation" + }, + { + "bbox": [ + 116, + 232, + 171, + 246 + ], + "score": 1.0, + "content": ". However, if", + "type": "text" + }, + { + "bbox": [ + 172, + 232, + 196, + 245 + ], + "score": 0.93, + "content": "U ( \\hat { \\mathcal { P } } )", + "type": "inline_equation" + }, + { + "bbox": [ + 196, + 232, + 277, + 246 + ], + "score": 1.0, + "content": "is a lot higher than", + "type": "text" + }, + { + "bbox": [ + 277, + 233, + 301, + 245 + ], + "score": 0.92, + "content": "U ( \\mathcal { P } )", + "type": "inline_equation" + }, + { + "bbox": [ + 302, + 232, + 505, + 246 + ], + "score": 1.0, + "content": ", then the model configuration will not satisfy the", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 244, + 505, + 258 + ], + "spans": [ + { + "bbox": [ + 105, + 244, + 172, + 258 + ], + "score": 1.0, + "content": "constraint. The", + "type": "text" + }, + { + "bbox": [ + 172, + 245, + 201, + 256 + ], + "score": 0.62, + "content": "\\log 1 . 2", + "type": "inline_equation" + }, + { + "bbox": [ + 201, + 244, + 505, + 258 + ], + "score": 1.0, + "content": "factor serves as another margin allowing checkpoints with slightly worse", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 106, + 255, + 297, + 268 + ], + "spans": [ + { + "bbox": [ + 106, + 255, + 297, + 268 + ], + "score": 1.0, + "content": "vocabulary-usage adjusted NLL to be included.", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 10.5 + }, + { + "type": "text", + "bbox": [ + 107, + 271, + 505, + 305 + ], + "lines": [ + { + "bbox": [ + 105, + 270, + 504, + 284 + ], + "spans": [ + { + "bbox": [ + 105, + 270, + 492, + 284 + ], + "score": 1.0, + "content": "In a final step, we take into account the length of the transcriptions: out of the configurations", + "type": "text" + }, + { + "bbox": [ + 492, + 273, + 504, + 282 + ], + "score": 0.83, + "content": "{ \\mathcal { P } } ^ { \\prime }", + "type": "inline_equation" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 282, + 505, + 296 + ], + "spans": [ + { + "bbox": [ + 105, + 282, + 505, + 296 + ], + "score": 1.0, + "content": "which satisfy the above constraint, we select the one which has the highest sum of log probability", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 106, + 294, + 234, + 306 + ], + "spans": [ + { + "bbox": [ + 106, + 294, + 234, + 306 + ], + "score": 1.0, + "content": "without normalizing the length:", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 15 + }, + { + "type": "interline_equation", + "bbox": [ + 169, + 312, + 443, + 348 + ], + "lines": [ + { + "bbox": [ + 169, + 312, + 443, + 348 + ], + "spans": [ + { + "bbox": [ + 169, + 312, + 443, + 348 + ], + "score": 0.94, + "content": "\\mathcal { P } ^ { * } = \\arg \\operatorname* { m a x } _ { \\mathcal { P } ^ { \\prime } } \\sum _ { j = 1 } ^ { N _ { s } } \\sum _ { t = 1 } ^ { M } \\log p _ { L M } ( p _ { t } ^ { j } ) , M = | P ^ { j } | , P ^ { j } = [ p _ { 1 } ^ { j } , \\dots , p _ { M } ^ { j } ]", + "type": "interline_equation", + "image_path": "2930d2c4de9bfc9242467fc6ca39076d0e51025e6eb1f2125a1e21c64638989c.jpg" + } + ] + } + ], + "index": 18, + "virtual_lines": [ + { + "bbox": [ + 169, + 312, + 443, + 324.0 + ], + "spans": [], + "index": 17 + }, + { + "bbox": [ + 169, + 324.0, + 443, + 336.0 + ], + "spans": [], + "index": 18 + }, + { + "bbox": [ + 169, + 336.0, + 443, + 348.0 + ], + "spans": [], + "index": 19 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 354, + 506, + 388 + ], + "lines": [ + { + "bbox": [ + 105, + 354, + 505, + 367 + ], + "spans": [ + { + "bbox": [ + 105, + 354, + 505, + 367 + ], + "score": 1.0, + "content": "This selects model configurations which produce phoneme sequences that score high under the", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 106, + 366, + 505, + 378 + ], + "spans": [ + { + "bbox": [ + 106, + 366, + 505, + 378 + ], + "score": 1.0, + "content": "language model but are not too long. Appendix D compares accuracy when developing with this", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 106, + 377, + 295, + 389 + ], + "spans": [ + { + "bbox": [ + 106, + 377, + 295, + 389 + ], + "score": 1.0, + "content": "metric compared to a labeled development set.", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 21 + }, + { + "type": "title", + "bbox": [ + 107, + 406, + 163, + 419 + ], + "lines": [ + { + "bbox": [ + 105, + 404, + 165, + 421 + ], + "spans": [ + { + "bbox": [ + 105, + 404, + 165, + 421 + ], + "score": 1.0, + "content": "4 Results", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 23 + }, + { + "type": "title", + "bbox": [ + 107, + 431, + 395, + 444 + ], + "lines": [ + { + "bbox": [ + 104, + 430, + 397, + 447 + ], + "spans": [ + { + "bbox": [ + 104, + 430, + 397, + 447 + ], + "score": 1.0, + "content": "4.1 Comparison to Supervised Speech Recognition on Librispeech", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 24 + }, + { + "type": "text", + "bbox": [ + 106, + 452, + 505, + 573 + ], + "lines": [ + { + "bbox": [ + 106, + 453, + 504, + 466 + ], + "spans": [ + { + "bbox": [ + 106, + 453, + 504, + 466 + ], + "score": 1.0, + "content": "We first test our approach on Librispeech to get a sense of how unsupervised speech recognition", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 464, + 505, + 476 + ], + "spans": [ + { + "bbox": [ + 105, + 464, + 505, + 476 + ], + "score": 1.0, + "content": "compares to the best supervised systems trained on a large amount of labeled data. Librispeech", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 475, + 506, + 487 + ], + "spans": [ + { + "bbox": [ + 105, + 475, + 506, + 487 + ], + "score": 1.0, + "content": "is a standard benchmark in the speech recognition community which provides about 960 hours of", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 106, + 486, + 506, + 498 + ], + "spans": [ + { + "bbox": [ + 106, + 486, + 506, + 498 + ], + "score": 1.0, + "content": "transcribed read audiobooks. We use the language modeling data of Librispeech as unlabeled text", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 106, + 496, + 505, + 509 + ], + "spans": [ + { + "bbox": [ + 106, + 496, + 505, + 509 + ], + "score": 1.0, + "content": "data for unsupervised training. In Appendix G we show that far less unlabeled text and speech audio", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 507, + 506, + 520 + ], + "spans": [ + { + "bbox": [ + 105, + 507, + 506, + 520 + ], + "score": 1.0, + "content": "are sufficient to reach a similar level of performance. We experiment with the frozen representations", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 518, + 505, + 531 + ], + "spans": [ + { + "bbox": [ + 105, + 518, + 291, + 531 + ], + "score": 1.0, + "content": "of a wav2vec 2.0 LARGE model trained on the", + "type": "text" + }, + { + "bbox": [ + 291, + 519, + 315, + 529 + ], + "score": 0.81, + "content": "5 3 . 2 \\mathrm { k }", + "type": "inline_equation" + }, + { + "bbox": [ + 315, + 518, + 505, + 531 + ], + "score": 1.0, + "content": "hours of Libri-Light (LL-60k) which we denote", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 529, + 505, + 541 + ], + "spans": [ + { + "bbox": [ + 105, + 529, + 505, + 541 + ], + "score": 1.0, + "content": "as wav2vec-U LARGE. We also consider self-training over three iterations by first training an HMM", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 106, + 541, + 505, + 553 + ], + "spans": [ + { + "bbox": [ + 106, + 541, + 505, + 553 + ], + "score": 1.0, + "content": "on the labels generated by the GANm then fine-tuning the original wav2vec 2.0 model on the labels", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 550, + 505, + 565 + ], + "spans": [ + { + "bbox": [ + 105, + 550, + 505, + 565 + ], + "score": 1.0, + "content": "of the HMM for Librispeech followed by then fine-tuning on Libri-Light; Appendix F investigates", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 561, + 157, + 574 + ], + "spans": [ + { + "bbox": [ + 105, + 561, + 157, + 574 + ], + "score": 1.0, + "content": "alternatives.", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 30 + }, + { + "type": "text", + "bbox": [ + 106, + 578, + 506, + 687 + ], + "lines": [ + { + "bbox": [ + 105, + 578, + 505, + 591 + ], + "spans": [ + { + "bbox": [ + 105, + 578, + 295, + 591 + ], + "score": 1.0, + "content": "wav2vec-U LARGE with self-training (wav2vec-", + "type": "text" + }, + { + "bbox": [ + 296, + 579, + 327, + 590 + ], + "score": 0.7, + "content": "\\mathbf { \\partial } . \\mathbf { U } + \\mathbf { S } \\mathbf { T } ,", + "type": "inline_equation" + }, + { + "bbox": [ + 327, + 578, + 505, + 591 + ], + "score": 1.0, + "content": ") and a Transformer language model achieves", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 589, + 506, + 602 + ], + "spans": [ + { + "bbox": [ + 105, + 589, + 506, + 602 + ], + "score": 1.0, + "content": "WER 5.9 on test-other, the noisy test set. This shows that unsupervised speech recognition can", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 104, + 600, + 506, + 613 + ], + "spans": [ + { + "bbox": [ + 104, + 600, + 506, + 613 + ], + "score": 1.0, + "content": "perform remarkably well compared to the best supervised systems of the recent past on this much", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 106, + 612, + 505, + 623 + ], + "spans": [ + { + "bbox": [ + 106, + 612, + 505, + 623 + ], + "score": 1.0, + "content": "studied benchmark. Also, self-training is effective even when the teacher model is unsupervised as", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 622, + 506, + 635 + ], + "spans": [ + { + "bbox": [ + 105, + 622, + 506, + 635 + ], + "score": 1.0, + "content": "per the improvement over GAN training (wav2vec-U). Interestingly, self-training on just Librispeech,", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 632, + 506, + 646 + ], + "spans": [ + { + "bbox": [ + 105, + 632, + 506, + 646 + ], + "score": 1.0, + "content": "or 960 hours of unlabeled speech audio, achieves already very good performance of WER 6.4 on", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 106, + 645, + 506, + 656 + ], + "spans": [ + { + "bbox": [ + 106, + 645, + 506, + 656 + ], + "score": 1.0, + "content": "dev-other compared to self-training on all of Libri-Light (53.2k hours) which compares at 6.0 WER.", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 654, + 506, + 668 + ], + "spans": [ + { + "bbox": [ + 105, + 654, + 506, + 668 + ], + "score": 1.0, + "content": "We note that the number of parameters trained during adversarial training is very small: the generator", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 666, + 506, + 679 + ], + "spans": [ + { + "bbox": [ + 105, + 666, + 506, + 679 + ], + "score": 1.0, + "content": "contains only about 90k parameters for a single temporal convolution mapping to the phoneme set", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 106, + 676, + 274, + 689 + ], + "spans": [ + { + "bbox": [ + 106, + 676, + 274, + 689 + ], + "score": 1.0, + "content": "from frozen wav2vec 2.0 representations.", + "type": "text" + } + ], + "index": 45 + } + ], + "index": 40.5 + } + ], + "page_idx": 5, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 118, + 700, + 469, + 722 + ], + "lines": [ + { + "bbox": [ + 118, + 697, + 448, + 714 + ], + "spans": [ + { + "bbox": [ + 118, + 697, + 224, + 714 + ], + "score": 1.0, + "content": "3We remove SIL labels from", + "type": "text" + }, + { + "bbox": [ + 225, + 703, + 232, + 710 + ], + "score": 0.8, + "content": "\\mathcal { P }", + "type": "inline_equation" + }, + { + "bbox": [ + 232, + 697, + 424, + 714 + ], + "score": 1.0, + "content": "when computing the NLL because SIL is not used in", + "type": "text" + }, + { + "bbox": [ + 425, + 702, + 443, + 711 + ], + "score": 0.79, + "content": "p _ { L M }", + "type": "inline_equation" + }, + { + "bbox": [ + 444, + 697, + 448, + 714 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ] + }, + { + "bbox": [ + 119, + 709, + 468, + 725 + ], + "spans": [ + { + "bbox": [ + 119, + 709, + 468, + 725 + ], + "score": 1.0, + "content": "4In practice, we used language model perplexity which is equivalent to NLL after taking the log.", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 742, + 309, + 750 + ], + "lines": [ + { + "bbox": [ + 302, + 741, + 310, + 752 + ], + "spans": [ + { + "bbox": [ + 302, + 741, + 310, + 752 + ], + "score": 1.0, + "content": "6", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 104, + 72, + 504, + 96 + ], + "lines": [], + "index": 0.5, + "bbox_fs": [ + 105, + 70, + 505, + 99 + ], + "lines_deleted": true + }, + { + "type": "text", + "bbox": [ + 107, + 101, + 505, + 150 + ], + "lines": [ + { + "bbox": [ + 105, + 102, + 506, + 114 + ], + "spans": [ + { + "bbox": [ + 105, + 102, + 506, + 114 + ], + "score": 1.0, + "content": "In a first step, we generate phoneme transcriptions for different training checkpoints or hyper-", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 112, + 505, + 126 + ], + "spans": [ + { + "bbox": [ + 105, + 112, + 505, + 126 + ], + "score": 1.0, + "content": "parameter settings and denote the transcriptions of the configuration with the lowest vocabulary-usage", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 124, + 505, + 138 + ], + "spans": [ + { + "bbox": [ + 105, + 124, + 174, + 138 + ], + "score": 1.0, + "content": "adjusted NLL as", + "type": "text" + }, + { + "bbox": [ + 175, + 124, + 340, + 137 + ], + "score": 0.84, + "content": "\\begin{array} { r } { \\hat { \\mathcal { P } } = \\arg \\operatorname* { m i n } _ { \\mathcal { P } } N L L _ { L M } ( \\mathcal { P } ) - \\log U ( \\mathcal { P } ) } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 340, + 124, + 505, + 138 + ], + "score": 1.0, + "content": ".4 Next, we discard model configurations", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 137, + 336, + 151 + ], + "spans": [ + { + "bbox": [ + 105, + 138, + 269, + 151 + ], + "score": 1.0, + "content": "which do not satisfy the following using", + "type": "text" + }, + { + "bbox": [ + 269, + 137, + 278, + 149 + ], + "score": 0.85, + "content": "\\hat { \\mathcal { P } }", + "type": "inline_equation" + }, + { + "bbox": [ + 278, + 138, + 336, + 151 + ], + "score": 1.0, + "content": "as the anchor:", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 3.5, + "bbox_fs": [ + 105, + 102, + 506, + 151 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 191, + 156, + 420, + 190 + ], + "lines": [ + { + "bbox": [ + 191, + 156, + 420, + 190 + ], + "spans": [ + { + "bbox": [ + 191, + 156, + 420, + 190 + ], + "score": 0.93, + "content": "N L L _ { L M } ( \\mathcal { P } ) < N L L _ { L M } ( \\hat { \\mathcal { P } } ) + \\log \\left( \\frac { U ( \\mathcal { P } ) } { U ( \\hat { \\mathcal { P } } ) } \\right) + \\log 1 . 2", + "type": "interline_equation", + "image_path": "b07976d07eb1264788bb86ec8c503a84fc3c183351681c48efba85a9d06b98dd.jpg" + } + ] + } + ], + "index": 6.5, + "virtual_lines": [ + { + "bbox": [ + 191, + 156, + 420, + 173.0 + ], + "spans": [], + "index": 6 + }, + { + "bbox": [ + 191, + 173.0, + 420, + 190.0 + ], + "spans": [], + "index": 7 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 196, + 505, + 267 + ], + "lines": [ + { + "bbox": [ + 105, + 196, + 506, + 210 + ], + "spans": [ + { + "bbox": [ + 105, + 196, + 506, + 210 + ], + "score": 1.0, + "content": "The second term on the right hand side introduces a margin over the NLL of the anchor transcription", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 107, + 208, + 506, + 223 + ], + "spans": [ + { + "bbox": [ + 107, + 208, + 160, + 222 + ], + "score": 0.93, + "content": "N L L _ { L M } ( \\hat { \\mathcal { P } } )", + "type": "inline_equation" + }, + { + "bbox": [ + 160, + 208, + 294, + 223 + ], + "score": 1.0, + "content": "based on the vocabulary usage of", + "type": "text" + }, + { + "bbox": [ + 294, + 210, + 303, + 220 + ], + "score": 0.82, + "content": "\\mathcal { P }", + "type": "inline_equation" + }, + { + "bbox": [ + 303, + 208, + 320, + 223 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 321, + 208, + 329, + 220 + ], + "score": 0.84, + "content": "\\hat { \\mathcal { P } }", + "type": "inline_equation" + }, + { + "bbox": [ + 330, + 208, + 343, + 223 + ], + "score": 1.0, + "content": ": If", + "type": "text" + }, + { + "bbox": [ + 344, + 208, + 368, + 222 + ], + "score": 0.93, + "content": "U ( \\hat { \\mathcal { P } } )", + "type": "inline_equation" + }, + { + "bbox": [ + 368, + 208, + 478, + 223 + ], + "score": 1.0, + "content": "is much lower compared to", + "type": "text" + }, + { + "bbox": [ + 479, + 212, + 502, + 222 + ], + "score": 0.93, + "content": "U ( \\mathcal { P } )", + "type": "inline_equation" + }, + { + "bbox": [ + 502, + 208, + 506, + 223 + ], + "score": 1.0, + "content": ",", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 106, + 220, + 506, + 234 + ], + "spans": [ + { + "bbox": [ + 106, + 220, + 506, + 234 + ], + "score": 1.0, + "content": "then we allow model configurations which produce transcriptions with higher NLL compared to", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 106, + 231, + 505, + 246 + ], + "spans": [ + { + "bbox": [ + 106, + 231, + 115, + 244 + ], + "score": 0.82, + "content": "\\hat { \\mathcal { P } }", + "type": "inline_equation" + }, + { + "bbox": [ + 116, + 232, + 171, + 246 + ], + "score": 1.0, + "content": ". However, if", + "type": "text" + }, + { + "bbox": [ + 172, + 232, + 196, + 245 + ], + "score": 0.93, + "content": "U ( \\hat { \\mathcal { P } } )", + "type": "inline_equation" + }, + { + "bbox": [ + 196, + 232, + 277, + 246 + ], + "score": 1.0, + "content": "is a lot higher than", + "type": "text" + }, + { + "bbox": [ + 277, + 233, + 301, + 245 + ], + "score": 0.92, + "content": "U ( \\mathcal { P } )", + "type": "inline_equation" + }, + { + "bbox": [ + 302, + 232, + 505, + 246 + ], + "score": 1.0, + "content": ", then the model configuration will not satisfy the", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 244, + 505, + 258 + ], + "spans": [ + { + "bbox": [ + 105, + 244, + 172, + 258 + ], + "score": 1.0, + "content": "constraint. The", + "type": "text" + }, + { + "bbox": [ + 172, + 245, + 201, + 256 + ], + "score": 0.62, + "content": "\\log 1 . 2", + "type": "inline_equation" + }, + { + "bbox": [ + 201, + 244, + 505, + 258 + ], + "score": 1.0, + "content": "factor serves as another margin allowing checkpoints with slightly worse", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 106, + 255, + 297, + 268 + ], + "spans": [ + { + "bbox": [ + 106, + 255, + 297, + 268 + ], + "score": 1.0, + "content": "vocabulary-usage adjusted NLL to be included.", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 10.5, + "bbox_fs": [ + 105, + 196, + 506, + 268 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 271, + 505, + 305 + ], + "lines": [ + { + "bbox": [ + 105, + 270, + 504, + 284 + ], + "spans": [ + { + "bbox": [ + 105, + 270, + 492, + 284 + ], + "score": 1.0, + "content": "In a final step, we take into account the length of the transcriptions: out of the configurations", + "type": "text" + }, + { + "bbox": [ + 492, + 273, + 504, + 282 + ], + "score": 0.83, + "content": "{ \\mathcal { P } } ^ { \\prime }", + "type": "inline_equation" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 282, + 505, + 296 + ], + "spans": [ + { + "bbox": [ + 105, + 282, + 505, + 296 + ], + "score": 1.0, + "content": "which satisfy the above constraint, we select the one which has the highest sum of log probability", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 106, + 294, + 234, + 306 + ], + "spans": [ + { + "bbox": [ + 106, + 294, + 234, + 306 + ], + "score": 1.0, + "content": "without normalizing the length:", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 15, + "bbox_fs": [ + 105, + 270, + 505, + 306 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 169, + 312, + 443, + 348 + ], + "lines": [ + { + "bbox": [ + 169, + 312, + 443, + 348 + ], + "spans": [ + { + "bbox": [ + 169, + 312, + 443, + 348 + ], + "score": 0.94, + "content": "\\mathcal { P } ^ { * } = \\arg \\operatorname* { m a x } _ { \\mathcal { P } ^ { \\prime } } \\sum _ { j = 1 } ^ { N _ { s } } \\sum _ { t = 1 } ^ { M } \\log p _ { L M } ( p _ { t } ^ { j } ) , M = | P ^ { j } | , P ^ { j } = [ p _ { 1 } ^ { j } , \\dots , p _ { M } ^ { j } ]", + "type": "interline_equation", + "image_path": "2930d2c4de9bfc9242467fc6ca39076d0e51025e6eb1f2125a1e21c64638989c.jpg" + } + ] + } + ], + "index": 18, + "virtual_lines": [ + { + "bbox": [ + 169, + 312, + 443, + 324.0 + ], + "spans": [], + "index": 17 + }, + { + "bbox": [ + 169, + 324.0, + 443, + 336.0 + ], + "spans": [], + "index": 18 + }, + { + "bbox": [ + 169, + 336.0, + 443, + 348.0 + ], + "spans": [], + "index": 19 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 354, + 506, + 388 + ], + "lines": [ + { + "bbox": [ + 105, + 354, + 505, + 367 + ], + "spans": [ + { + "bbox": [ + 105, + 354, + 505, + 367 + ], + "score": 1.0, + "content": "This selects model configurations which produce phoneme sequences that score high under the", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 106, + 366, + 505, + 378 + ], + "spans": [ + { + "bbox": [ + 106, + 366, + 505, + 378 + ], + "score": 1.0, + "content": "language model but are not too long. Appendix D compares accuracy when developing with this", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 106, + 377, + 295, + 389 + ], + "spans": [ + { + "bbox": [ + 106, + 377, + 295, + 389 + ], + "score": 1.0, + "content": "metric compared to a labeled development set.", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 21, + "bbox_fs": [ + 105, + 354, + 505, + 389 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 406, + 163, + 419 + ], + "lines": [ + { + "bbox": [ + 105, + 404, + 165, + 421 + ], + "spans": [ + { + "bbox": [ + 105, + 404, + 165, + 421 + ], + "score": 1.0, + "content": "4 Results", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 23 + }, + { + "type": "title", + "bbox": [ + 107, + 431, + 395, + 444 + ], + "lines": [ + { + "bbox": [ + 104, + 430, + 397, + 447 + ], + "spans": [ + { + "bbox": [ + 104, + 430, + 397, + 447 + ], + "score": 1.0, + "content": "4.1 Comparison to Supervised Speech Recognition on Librispeech", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 24 + }, + { + "type": "text", + "bbox": [ + 106, + 452, + 505, + 573 + ], + "lines": [ + { + "bbox": [ + 106, + 453, + 504, + 466 + ], + "spans": [ + { + "bbox": [ + 106, + 453, + 504, + 466 + ], + "score": 1.0, + "content": "We first test our approach on Librispeech to get a sense of how unsupervised speech recognition", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 464, + 505, + 476 + ], + "spans": [ + { + "bbox": [ + 105, + 464, + 505, + 476 + ], + "score": 1.0, + "content": "compares to the best supervised systems trained on a large amount of labeled data. Librispeech", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 475, + 506, + 487 + ], + "spans": [ + { + "bbox": [ + 105, + 475, + 506, + 487 + ], + "score": 1.0, + "content": "is a standard benchmark in the speech recognition community which provides about 960 hours of", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 106, + 486, + 506, + 498 + ], + "spans": [ + { + "bbox": [ + 106, + 486, + 506, + 498 + ], + "score": 1.0, + "content": "transcribed read audiobooks. We use the language modeling data of Librispeech as unlabeled text", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 106, + 496, + 505, + 509 + ], + "spans": [ + { + "bbox": [ + 106, + 496, + 505, + 509 + ], + "score": 1.0, + "content": "data for unsupervised training. In Appendix G we show that far less unlabeled text and speech audio", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 507, + 506, + 520 + ], + "spans": [ + { + "bbox": [ + 105, + 507, + 506, + 520 + ], + "score": 1.0, + "content": "are sufficient to reach a similar level of performance. We experiment with the frozen representations", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 518, + 505, + 531 + ], + "spans": [ + { + "bbox": [ + 105, + 518, + 291, + 531 + ], + "score": 1.0, + "content": "of a wav2vec 2.0 LARGE model trained on the", + "type": "text" + }, + { + "bbox": [ + 291, + 519, + 315, + 529 + ], + "score": 0.81, + "content": "5 3 . 2 \\mathrm { k }", + "type": "inline_equation" + }, + { + "bbox": [ + 315, + 518, + 505, + 531 + ], + "score": 1.0, + "content": "hours of Libri-Light (LL-60k) which we denote", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 529, + 505, + 541 + ], + "spans": [ + { + "bbox": [ + 105, + 529, + 505, + 541 + ], + "score": 1.0, + "content": "as wav2vec-U LARGE. We also consider self-training over three iterations by first training an HMM", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 106, + 541, + 505, + 553 + ], + "spans": [ + { + "bbox": [ + 106, + 541, + 505, + 553 + ], + "score": 1.0, + "content": "on the labels generated by the GANm then fine-tuning the original wav2vec 2.0 model on the labels", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 550, + 505, + 565 + ], + "spans": [ + { + "bbox": [ + 105, + 550, + 505, + 565 + ], + "score": 1.0, + "content": "of the HMM for Librispeech followed by then fine-tuning on Libri-Light; Appendix F investigates", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 561, + 157, + 574 + ], + "spans": [ + { + "bbox": [ + 105, + 561, + 157, + 574 + ], + "score": 1.0, + "content": "alternatives.", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 30, + "bbox_fs": [ + 105, + 453, + 506, + 574 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 578, + 506, + 687 + ], + "lines": [ + { + "bbox": [ + 105, + 578, + 505, + 591 + ], + "spans": [ + { + "bbox": [ + 105, + 578, + 295, + 591 + ], + "score": 1.0, + "content": "wav2vec-U LARGE with self-training (wav2vec-", + "type": "text" + }, + { + "bbox": [ + 296, + 579, + 327, + 590 + ], + "score": 0.7, + "content": "\\mathbf { \\partial } . \\mathbf { U } + \\mathbf { S } \\mathbf { T } ,", + "type": "inline_equation" + }, + { + "bbox": [ + 327, + 578, + 505, + 591 + ], + "score": 1.0, + "content": ") and a Transformer language model achieves", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 589, + 506, + 602 + ], + "spans": [ + { + "bbox": [ + 105, + 589, + 506, + 602 + ], + "score": 1.0, + "content": "WER 5.9 on test-other, the noisy test set. This shows that unsupervised speech recognition can", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 104, + 600, + 506, + 613 + ], + "spans": [ + { + "bbox": [ + 104, + 600, + 506, + 613 + ], + "score": 1.0, + "content": "perform remarkably well compared to the best supervised systems of the recent past on this much", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 106, + 612, + 505, + 623 + ], + "spans": [ + { + "bbox": [ + 106, + 612, + 505, + 623 + ], + "score": 1.0, + "content": "studied benchmark. Also, self-training is effective even when the teacher model is unsupervised as", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 622, + 506, + 635 + ], + "spans": [ + { + "bbox": [ + 105, + 622, + 506, + 635 + ], + "score": 1.0, + "content": "per the improvement over GAN training (wav2vec-U). Interestingly, self-training on just Librispeech,", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 632, + 506, + 646 + ], + "spans": [ + { + "bbox": [ + 105, + 632, + 506, + 646 + ], + "score": 1.0, + "content": "or 960 hours of unlabeled speech audio, achieves already very good performance of WER 6.4 on", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 106, + 645, + 506, + 656 + ], + "spans": [ + { + "bbox": [ + 106, + 645, + 506, + 656 + ], + "score": 1.0, + "content": "dev-other compared to self-training on all of Libri-Light (53.2k hours) which compares at 6.0 WER.", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 654, + 506, + 668 + ], + "spans": [ + { + "bbox": [ + 105, + 654, + 506, + 668 + ], + "score": 1.0, + "content": "We note that the number of parameters trained during adversarial training is very small: the generator", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 666, + 506, + 679 + ], + "spans": [ + { + "bbox": [ + 105, + 666, + 506, + 679 + ], + "score": 1.0, + "content": "contains only about 90k parameters for a single temporal convolution mapping to the phoneme set", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 106, + 676, + 274, + 689 + ], + "spans": [ + { + "bbox": [ + 106, + 676, + 274, + 689 + ], + "score": 1.0, + "content": "from frozen wav2vec 2.0 representations.", + "type": "text" + } + ], + "index": 45 + } + ], + "index": 40.5, + "bbox_fs": [ + 104, + 578, + 506, + 689 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "table", + "bbox": [ + 107, + 122, + 504, + 331 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 106, + 70, + 505, + 115 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 70, + 505, + 82 + ], + "spans": [ + { + "bbox": [ + 105, + 70, + 505, + 82 + ], + "score": 1.0, + "content": "Table 1: WER on Librispeech dev/test sets when using 960 hours of unlabeled audio from Librispeech", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 82, + 506, + 93 + ], + "spans": [ + { + "bbox": [ + 105, + 82, + 155, + 93 + ], + "score": 1.0, + "content": "(LS-960) or", + "type": "text" + }, + { + "bbox": [ + 155, + 82, + 180, + 92 + ], + "score": 0.31, + "content": "5 3 . 2 \\mathrm { k }", + "type": "inline_equation" + }, + { + "bbox": [ + 180, + 82, + 506, + 93 + ], + "score": 1.0, + "content": "hours from Libri-Light (LL-60k) using representations from wav2vec 2.0 LARGE.", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 104, + 92, + 506, + 105 + ], + "spans": [ + { + "bbox": [ + 104, + 92, + 506, + 105 + ], + "score": 1.0, + "content": "Librispeech provides clean dev/test sets which are less challenging than the other sets. We report", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 102, + 495, + 115 + ], + "spans": [ + { + "bbox": [ + 105, + 102, + 458, + 115 + ], + "score": 1.0, + "content": "results for GAN training only (wav2vec-U) and with subsequent self-training (wav2vec-", + "type": "text" + }, + { + "bbox": [ + 458, + 103, + 490, + 114 + ], + "score": 0.78, + "content": "\\mathbf { \\partial } . \\mathbf { U } + \\mathbf { S } \\mathbf { T } ,", + "type": "inline_equation" + }, + { + "bbox": [ + 491, + 102, + 495, + 115 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 1.5 + }, + { + "type": "table_body", + "bbox": [ + 107, + 122, + 504, + 331 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 107, + 122, + 504, + 331 + ], + "spans": [ + { + "bbox": [ + 107, + 122, + 504, + 331 + ], + "score": 0.983, + "html": "
ModelUnlabeled dataLMdevtest
cleanothercleanother
960h - Supervised learning
DeepSpeech 2 [Amodei et al.,2016]5-gram5.3313.25
Fully Conv [Zeghidour et al., 2018]ConvLM3.089.943.2610.47
TDNN+Kaldi [Xu et al., 2018]4-gram2.717.373.127.63
SpecAugment [Park et al., 2019]RNN1-2.55.8
ContextNet [Han et al.,2020]LSTM1.93.91.94.1
Conformer [Gulati et al.,2020]LSTM2.14.31.93.9
960h - Self and semi-supervised learning
Transf.+ PL [Synnaeve et al.,2020]LL-60kCLM+Transf.2.003.652.094.11
IPL [Xu et al., 2020b]LL-60k4-gram+Transf.1.853.262.104.01
NST [Park et al., 2020]LL-60kLSTM1.63.41.73.4
wav2vec 2.0 [Baevski et al.,2020c]LL-60kTransf.1.63.01.83.3
wav2vec 2.0 + NST [Zhang et al.,2020b]LL-60kLSTM1.32.61.42.6
Unsupervised learning
wav2vec-ULARGELL-60k4-gram13.315.113.818.0
wav2vec-ULARGE+ STLL-60k4-gram3.46.03.86.5
LL-60kTransf.3.25.53.45.9
", + "type": "table", + "image_path": "7d5f6a9816458b5d9d90dd620c97a44d70ecc200d9e2e077d3e3e3e887499080.jpg" + } + ] + } + ], + "index": 5, + "virtual_lines": [ + { + "bbox": [ + 107, + 122, + 504, + 191.66666666666669 + ], + "spans": [], + "index": 4 + }, + { + "bbox": [ + 107, + 191.66666666666669, + 504, + 261.33333333333337 + ], + "spans": [], + "index": 5 + }, + { + "bbox": [ + 107, + 261.33333333333337, + 504, + 331.00000000000006 + ], + "spans": [], + "index": 6 + } + ] + } + ], + "index": 3.25 + }, + { + "type": "title", + "bbox": [ + 106, + 345, + 306, + 358 + ], + "lines": [ + { + "bbox": [ + 105, + 345, + 308, + 360 + ], + "spans": [ + { + "bbox": [ + 105, + 345, + 308, + 360 + ], + "score": 1.0, + "content": "4.2 Comparison to Prior Unsupervised Work", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 7 + }, + { + "type": "text", + "bbox": [ + 107, + 366, + 505, + 421 + ], + "lines": [ + { + "bbox": [ + 106, + 366, + 505, + 378 + ], + "spans": [ + { + "bbox": [ + 106, + 366, + 505, + 378 + ], + "score": 1.0, + "content": "Prior work on unsupervised speech recognition focused on the TIMIT benchmark. In order to perform", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 106, + 377, + 505, + 389 + ], + "spans": [ + { + "bbox": [ + 106, + 377, + 505, + 389 + ], + "score": 1.0, + "content": "a direct comparison to these approaches, we report results on this benchmark as well. We consider", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 387, + 505, + 401 + ], + "spans": [ + { + "bbox": [ + 105, + 387, + 505, + 401 + ], + "score": 1.0, + "content": "two setups to compare to previous work: in the matched setting, the unlabeled text data is simply the", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 106, + 399, + 505, + 411 + ], + "spans": [ + { + "bbox": [ + 106, + 399, + 505, + 411 + ], + "score": 1.0, + "content": "transcriptions of the unlabeled audio data but unpaired. In the unmatched setup, the unlabeled text", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 408, + 469, + 423 + ], + "spans": [ + { + "bbox": [ + 105, + 408, + 469, + 423 + ], + "score": 1.0, + "content": "data does not contain the transcriptions for the audio data which is a more realistic setting.", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 10 + }, + { + "type": "text", + "bbox": [ + 107, + 426, + 505, + 481 + ], + "lines": [ + { + "bbox": [ + 106, + 426, + 506, + 438 + ], + "spans": [ + { + "bbox": [ + 106, + 426, + 506, + 438 + ], + "score": 1.0, + "content": "We measure performance on the standard Kaldi dev and test sets (core-dev/core-test) as well as a", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 106, + 437, + 506, + 449 + ], + "spans": [ + { + "bbox": [ + 106, + 437, + 506, + 449 + ], + "score": 1.0, + "content": "slightly larger version of the test set (all-test) to be able to compare to Liu et al. [2018] and Chen et al.", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 447, + 506, + 461 + ], + "spans": [ + { + "bbox": [ + 106, + 447, + 371, + 461 + ], + "score": 1.0, + "content": "[2019]. Further details of the two setups can be found in Appendix", + "type": "text" + }, + { + "bbox": [ + 372, + 448, + 393, + 459 + ], + "score": 0.66, + "content": "\\ S \\operatorname { E } . 1", + "type": "inline_equation" + }, + { + "bbox": [ + 394, + 447, + 506, + 461 + ], + "score": 1.0, + "content": ". We report performance for", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 459, + 506, + 472 + ], + "spans": [ + { + "bbox": [ + 105, + 459, + 506, + 472 + ], + "score": 1.0, + "content": "wav2vec-U with a 4-gram language model trained on the language modeling data of TIMIT and we", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 469, + 291, + 482 + ], + "spans": [ + { + "bbox": [ + 105, + 469, + 254, + 482 + ], + "score": 1.0, + "content": "also consider self-training (wav2vec-", + "type": "text" + }, + { + "bbox": [ + 254, + 470, + 285, + 481 + ], + "score": 0.47, + "content": "\\mathbf { \\partial } . \\mathbf { U } + \\mathbf { S } \\mathbf { T }", + "type": "inline_equation" + }, + { + "bbox": [ + 286, + 469, + 291, + 482 + ], + "score": 1.0, + "content": ").", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 15 + }, + { + "type": "text", + "bbox": [ + 107, + 486, + 505, + 541 + ], + "lines": [ + { + "bbox": [ + 105, + 485, + 505, + 498 + ], + "spans": [ + { + "bbox": [ + 105, + 485, + 505, + 498 + ], + "score": 1.0, + "content": "Table 2 shows that wav2vec-U outperforms prior unsupervised work in both the matched and", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 497, + 505, + 509 + ], + "spans": [ + { + "bbox": [ + 105, + 498, + 379, + 509 + ], + "score": 1.0, + "content": "unmatched settings, reducing PER on all-test in the matched setup by", + "type": "text" + }, + { + "bbox": [ + 379, + 497, + 399, + 508 + ], + "score": 0.87, + "content": "57 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 399, + 498, + 505, + 509 + ], + "score": 1.0, + "content": "relative compared to Chen", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 106, + 508, + 505, + 520 + ], + "spans": [ + { + "bbox": [ + 106, + 508, + 505, + 520 + ], + "score": 1.0, + "content": "et al. [2019]. Our method has lower performance than the best supervised methods but it performs", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 519, + 505, + 532 + ], + "spans": [ + { + "bbox": [ + 105, + 519, + 505, + 532 + ], + "score": 1.0, + "content": "still very well at PER 12 on core-test in the matched setup compared to PER 8.3 for the state of the", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 529, + 214, + 543 + ], + "spans": [ + { + "bbox": [ + 105, + 529, + 214, + 543 + ], + "score": 1.0, + "content": "art [Baevski et al., 2020c].", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 20 + }, + { + "type": "title", + "bbox": [ + 108, + 555, + 297, + 567 + ], + "lines": [ + { + "bbox": [ + 105, + 553, + 298, + 570 + ], + "spans": [ + { + "bbox": [ + 105, + 553, + 298, + 570 + ], + "score": 1.0, + "content": "4.3 Performance on non-English languages", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 23 + }, + { + "type": "text", + "bbox": [ + 107, + 575, + 505, + 663 + ], + "lines": [ + { + "bbox": [ + 105, + 574, + 506, + 589 + ], + "spans": [ + { + "bbox": [ + 105, + 574, + 506, + 589 + ], + "score": 1.0, + "content": "To get a sense of how well the method works on non-English data, we experiment on six languages", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 586, + 506, + 599 + ], + "spans": [ + { + "bbox": [ + 105, + 586, + 506, + 599 + ], + "score": 1.0, + "content": "of the multilingual Librispeech corpus (MLS; Pratap et al. 2020). As baseline we consider the", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 597, + 506, + 610 + ], + "spans": [ + { + "bbox": [ + 105, + 597, + 364, + 610 + ], + "score": 1.0, + "content": "supervised systems of Pratap et al. [2020] trained on between", + "type": "text" + }, + { + "bbox": [ + 365, + 598, + 377, + 608 + ], + "score": 0.31, + "content": "2 \\mathrm { k }", + "type": "inline_equation" + }, + { + "bbox": [ + 377, + 597, + 506, + 610 + ], + "score": 1.0, + "content": "and 161 hours of labeled data,", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 608, + 506, + 621 + ], + "spans": [ + { + "bbox": [ + 106, + 608, + 506, + 621 + ], + "score": 1.0, + "content": "depending on the language. For adversarial learning we use 100 hours of unlabeled audio data from", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 618, + 506, + 632 + ], + "spans": [ + { + "bbox": [ + 105, + 618, + 506, + 632 + ], + "score": 1.0, + "content": "MLS for every language as well as the MLS language modeling data. As input to wav2vec-U we use", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 630, + 505, + 643 + ], + "spans": [ + { + "bbox": [ + 105, + 630, + 505, + 643 + ], + "score": 1.0, + "content": "the representations from XLSR-53 [Conneau et al., 2020], a wav2vec 2.0 model pre-trained on 53", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 106, + 641, + 506, + 653 + ], + "spans": [ + { + "bbox": [ + 106, + 641, + 506, + 653 + ], + "score": 1.0, + "content": "languages. Table 3 shows that wav2vec-U generalizes across a range of languages. Performance is", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 651, + 412, + 666 + ], + "spans": [ + { + "bbox": [ + 105, + 651, + 412, + 666 + ], + "score": 1.0, + "content": "lower than supervised systems but it shows the viability for other languages.", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 27.5 + }, + { + "type": "text", + "bbox": [ + 108, + 668, + 506, + 702 + ], + "lines": [ + { + "bbox": [ + 105, + 667, + 506, + 681 + ], + "spans": [ + { + "bbox": [ + 105, + 667, + 506, + 681 + ], + "score": 1.0, + "content": "Next, we turn to three low-resource languages, Swahili, Kyrgyz, and Tatar. Swahili is an African lan-", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 679, + 505, + 691 + ], + "spans": [ + { + "bbox": [ + 105, + 679, + 351, + 691 + ], + "score": 1.0, + "content": "guage, Kyrgyz and Tatar are Turkic languages with only about", + "type": "text" + }, + { + "bbox": [ + 351, + 680, + 372, + 690 + ], + "score": 0.58, + "content": "4 . 3 \\mathrm { m }", + "type": "inline_equation" + }, + { + "bbox": [ + 373, + 679, + 389, + 691 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 389, + 680, + 411, + 690 + ], + "score": 0.68, + "content": "5 . 2 \\mathrm { m }", + "type": "inline_equation" + }, + { + "bbox": [ + 411, + 679, + 505, + 691 + ], + "score": 1.0, + "content": "speakers, respectively.5", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 689, + 506, + 703 + ], + "spans": [ + { + "bbox": [ + 105, + 689, + 480, + 703 + ], + "score": 1.0, + "content": "We use between 1.8 hours (Kyrgyz) and 9.2 hours of unlabeled audio (Swahili), see Appendix", + "type": "text" + }, + { + "bbox": [ + 481, + 690, + 503, + 702 + ], + "score": 0.83, + "content": "\\ S \\operatorname { E } . 1", + "type": "inline_equation" + }, + { + "bbox": [ + 503, + 689, + 506, + 703 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 34 + } + ], + "index": 33 + } + ], + "page_idx": 6, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 118, + 712, + 374, + 722 + ], + "lines": [ + { + "bbox": [ + 118, + 709, + 374, + 725 + ], + "spans": [ + { + "bbox": [ + 118, + 709, + 374, + 725 + ], + "score": 1.0, + "content": "5https://en.wikipedia.org/wiki/{Kyrgyz,Tatar}_language", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 741, + 308, + 750 + ], + "lines": [ + { + "bbox": [ + 302, + 741, + 309, + 752 + ], + "spans": [ + { + "bbox": [ + 302, + 741, + 309, + 752 + ], + "score": 1.0, + "content": "7", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "table", + "bbox": [ + 107, + 122, + 504, + 331 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 106, + 70, + 505, + 115 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 70, + 505, + 82 + ], + "spans": [ + { + "bbox": [ + 105, + 70, + 505, + 82 + ], + "score": 1.0, + "content": "Table 1: WER on Librispeech dev/test sets when using 960 hours of unlabeled audio from Librispeech", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 82, + 506, + 93 + ], + "spans": [ + { + "bbox": [ + 105, + 82, + 155, + 93 + ], + "score": 1.0, + "content": "(LS-960) or", + "type": "text" + }, + { + "bbox": [ + 155, + 82, + 180, + 92 + ], + "score": 0.31, + "content": "5 3 . 2 \\mathrm { k }", + "type": "inline_equation" + }, + { + "bbox": [ + 180, + 82, + 506, + 93 + ], + "score": 1.0, + "content": "hours from Libri-Light (LL-60k) using representations from wav2vec 2.0 LARGE.", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 104, + 92, + 506, + 105 + ], + "spans": [ + { + "bbox": [ + 104, + 92, + 506, + 105 + ], + "score": 1.0, + "content": "Librispeech provides clean dev/test sets which are less challenging than the other sets. We report", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 102, + 495, + 115 + ], + "spans": [ + { + "bbox": [ + 105, + 102, + 458, + 115 + ], + "score": 1.0, + "content": "results for GAN training only (wav2vec-U) and with subsequent self-training (wav2vec-", + "type": "text" + }, + { + "bbox": [ + 458, + 103, + 490, + 114 + ], + "score": 0.78, + "content": "\\mathbf { \\partial } . \\mathbf { U } + \\mathbf { S } \\mathbf { T } ,", + "type": "inline_equation" + }, + { + "bbox": [ + 491, + 102, + 495, + 115 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 1.5 + }, + { + "type": "table_body", + "bbox": [ + 107, + 122, + 504, + 331 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 107, + 122, + 504, + 331 + ], + "spans": [ + { + "bbox": [ + 107, + 122, + 504, + 331 + ], + "score": 0.983, + "html": "
ModelUnlabeled dataLMdevtest
cleanothercleanother
960h - Supervised learning
DeepSpeech 2 [Amodei et al.,2016]5-gram5.3313.25
Fully Conv [Zeghidour et al., 2018]ConvLM3.089.943.2610.47
TDNN+Kaldi [Xu et al., 2018]4-gram2.717.373.127.63
SpecAugment [Park et al., 2019]RNN1-2.55.8
ContextNet [Han et al.,2020]LSTM1.93.91.94.1
Conformer [Gulati et al.,2020]LSTM2.14.31.93.9
960h - Self and semi-supervised learning
Transf.+ PL [Synnaeve et al.,2020]LL-60kCLM+Transf.2.003.652.094.11
IPL [Xu et al., 2020b]LL-60k4-gram+Transf.1.853.262.104.01
NST [Park et al., 2020]LL-60kLSTM1.63.41.73.4
wav2vec 2.0 [Baevski et al.,2020c]LL-60kTransf.1.63.01.83.3
wav2vec 2.0 + NST [Zhang et al.,2020b]LL-60kLSTM1.32.61.42.6
Unsupervised learning
wav2vec-ULARGELL-60k4-gram13.315.113.818.0
wav2vec-ULARGE+ STLL-60k4-gram3.46.03.86.5
LL-60kTransf.3.25.53.45.9
", + "type": "table", + "image_path": "7d5f6a9816458b5d9d90dd620c97a44d70ecc200d9e2e077d3e3e3e887499080.jpg" + } + ] + } + ], + "index": 5, + "virtual_lines": [ + { + "bbox": [ + 107, + 122, + 504, + 191.66666666666669 + ], + "spans": [], + "index": 4 + }, + { + "bbox": [ + 107, + 191.66666666666669, + 504, + 261.33333333333337 + ], + "spans": [], + "index": 5 + }, + { + "bbox": [ + 107, + 261.33333333333337, + 504, + 331.00000000000006 + ], + "spans": [], + "index": 6 + } + ] + } + ], + "index": 3.25 + }, + { + "type": "title", + "bbox": [ + 106, + 345, + 306, + 358 + ], + "lines": [ + { + "bbox": [ + 105, + 345, + 308, + 360 + ], + "spans": [ + { + "bbox": [ + 105, + 345, + 308, + 360 + ], + "score": 1.0, + "content": "4.2 Comparison to Prior Unsupervised Work", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 7 + }, + { + "type": "text", + "bbox": [ + 107, + 366, + 505, + 421 + ], + "lines": [ + { + "bbox": [ + 106, + 366, + 505, + 378 + ], + "spans": [ + { + "bbox": [ + 106, + 366, + 505, + 378 + ], + "score": 1.0, + "content": "Prior work on unsupervised speech recognition focused on the TIMIT benchmark. In order to perform", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 106, + 377, + 505, + 389 + ], + "spans": [ + { + "bbox": [ + 106, + 377, + 505, + 389 + ], + "score": 1.0, + "content": "a direct comparison to these approaches, we report results on this benchmark as well. We consider", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 387, + 505, + 401 + ], + "spans": [ + { + "bbox": [ + 105, + 387, + 505, + 401 + ], + "score": 1.0, + "content": "two setups to compare to previous work: in the matched setting, the unlabeled text data is simply the", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 106, + 399, + 505, + 411 + ], + "spans": [ + { + "bbox": [ + 106, + 399, + 505, + 411 + ], + "score": 1.0, + "content": "transcriptions of the unlabeled audio data but unpaired. In the unmatched setup, the unlabeled text", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 408, + 469, + 423 + ], + "spans": [ + { + "bbox": [ + 105, + 408, + 469, + 423 + ], + "score": 1.0, + "content": "data does not contain the transcriptions for the audio data which is a more realistic setting.", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 10, + "bbox_fs": [ + 105, + 366, + 505, + 423 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 426, + 505, + 481 + ], + "lines": [ + { + "bbox": [ + 106, + 426, + 506, + 438 + ], + "spans": [ + { + "bbox": [ + 106, + 426, + 506, + 438 + ], + "score": 1.0, + "content": "We measure performance on the standard Kaldi dev and test sets (core-dev/core-test) as well as a", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 106, + 437, + 506, + 449 + ], + "spans": [ + { + "bbox": [ + 106, + 437, + 506, + 449 + ], + "score": 1.0, + "content": "slightly larger version of the test set (all-test) to be able to compare to Liu et al. [2018] and Chen et al.", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 447, + 506, + 461 + ], + "spans": [ + { + "bbox": [ + 106, + 447, + 371, + 461 + ], + "score": 1.0, + "content": "[2019]. Further details of the two setups can be found in Appendix", + "type": "text" + }, + { + "bbox": [ + 372, + 448, + 393, + 459 + ], + "score": 0.66, + "content": "\\ S \\operatorname { E } . 1", + "type": "inline_equation" + }, + { + "bbox": [ + 394, + 447, + 506, + 461 + ], + "score": 1.0, + "content": ". We report performance for", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 459, + 506, + 472 + ], + "spans": [ + { + "bbox": [ + 105, + 459, + 506, + 472 + ], + "score": 1.0, + "content": "wav2vec-U with a 4-gram language model trained on the language modeling data of TIMIT and we", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 469, + 291, + 482 + ], + "spans": [ + { + "bbox": [ + 105, + 469, + 254, + 482 + ], + "score": 1.0, + "content": "also consider self-training (wav2vec-", + "type": "text" + }, + { + "bbox": [ + 254, + 470, + 285, + 481 + ], + "score": 0.47, + "content": "\\mathbf { \\partial } . \\mathbf { U } + \\mathbf { S } \\mathbf { T }", + "type": "inline_equation" + }, + { + "bbox": [ + 286, + 469, + 291, + 482 + ], + "score": 1.0, + "content": ").", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 15, + "bbox_fs": [ + 105, + 426, + 506, + 482 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 486, + 505, + 541 + ], + "lines": [ + { + "bbox": [ + 105, + 485, + 505, + 498 + ], + "spans": [ + { + "bbox": [ + 105, + 485, + 505, + 498 + ], + "score": 1.0, + "content": "Table 2 shows that wav2vec-U outperforms prior unsupervised work in both the matched and", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 497, + 505, + 509 + ], + "spans": [ + { + "bbox": [ + 105, + 498, + 379, + 509 + ], + "score": 1.0, + "content": "unmatched settings, reducing PER on all-test in the matched setup by", + "type": "text" + }, + { + "bbox": [ + 379, + 497, + 399, + 508 + ], + "score": 0.87, + "content": "57 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 399, + 498, + 505, + 509 + ], + "score": 1.0, + "content": "relative compared to Chen", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 106, + 508, + 505, + 520 + ], + "spans": [ + { + "bbox": [ + 106, + 508, + 505, + 520 + ], + "score": 1.0, + "content": "et al. [2019]. Our method has lower performance than the best supervised methods but it performs", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 519, + 505, + 532 + ], + "spans": [ + { + "bbox": [ + 105, + 519, + 505, + 532 + ], + "score": 1.0, + "content": "still very well at PER 12 on core-test in the matched setup compared to PER 8.3 for the state of the", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 529, + 214, + 543 + ], + "spans": [ + { + "bbox": [ + 105, + 529, + 214, + 543 + ], + "score": 1.0, + "content": "art [Baevski et al., 2020c].", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 20, + "bbox_fs": [ + 105, + 485, + 505, + 543 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 555, + 297, + 567 + ], + "lines": [ + { + "bbox": [ + 105, + 553, + 298, + 570 + ], + "spans": [ + { + "bbox": [ + 105, + 553, + 298, + 570 + ], + "score": 1.0, + "content": "4.3 Performance on non-English languages", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 23 + }, + { + "type": "text", + "bbox": [ + 107, + 575, + 505, + 663 + ], + "lines": [ + { + "bbox": [ + 105, + 574, + 506, + 589 + ], + "spans": [ + { + "bbox": [ + 105, + 574, + 506, + 589 + ], + "score": 1.0, + "content": "To get a sense of how well the method works on non-English data, we experiment on six languages", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 586, + 506, + 599 + ], + "spans": [ + { + "bbox": [ + 105, + 586, + 506, + 599 + ], + "score": 1.0, + "content": "of the multilingual Librispeech corpus (MLS; Pratap et al. 2020). As baseline we consider the", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 597, + 506, + 610 + ], + "spans": [ + { + "bbox": [ + 105, + 597, + 364, + 610 + ], + "score": 1.0, + "content": "supervised systems of Pratap et al. [2020] trained on between", + "type": "text" + }, + { + "bbox": [ + 365, + 598, + 377, + 608 + ], + "score": 0.31, + "content": "2 \\mathrm { k }", + "type": "inline_equation" + }, + { + "bbox": [ + 377, + 597, + 506, + 610 + ], + "score": 1.0, + "content": "and 161 hours of labeled data,", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 608, + 506, + 621 + ], + "spans": [ + { + "bbox": [ + 106, + 608, + 506, + 621 + ], + "score": 1.0, + "content": "depending on the language. For adversarial learning we use 100 hours of unlabeled audio data from", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 618, + 506, + 632 + ], + "spans": [ + { + "bbox": [ + 105, + 618, + 506, + 632 + ], + "score": 1.0, + "content": "MLS for every language as well as the MLS language modeling data. As input to wav2vec-U we use", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 630, + 505, + 643 + ], + "spans": [ + { + "bbox": [ + 105, + 630, + 505, + 643 + ], + "score": 1.0, + "content": "the representations from XLSR-53 [Conneau et al., 2020], a wav2vec 2.0 model pre-trained on 53", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 106, + 641, + 506, + 653 + ], + "spans": [ + { + "bbox": [ + 106, + 641, + 506, + 653 + ], + "score": 1.0, + "content": "languages. Table 3 shows that wav2vec-U generalizes across a range of languages. Performance is", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 651, + 412, + 666 + ], + "spans": [ + { + "bbox": [ + 105, + 651, + 412, + 666 + ], + "score": 1.0, + "content": "lower than supervised systems but it shows the viability for other languages.", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 27.5, + "bbox_fs": [ + 105, + 574, + 506, + 666 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 668, + 506, + 702 + ], + "lines": [ + { + "bbox": [ + 105, + 667, + 506, + 681 + ], + "spans": [ + { + "bbox": [ + 105, + 667, + 506, + 681 + ], + "score": 1.0, + "content": "Next, we turn to three low-resource languages, Swahili, Kyrgyz, and Tatar. Swahili is an African lan-", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 679, + 505, + 691 + ], + "spans": [ + { + "bbox": [ + 105, + 679, + 351, + 691 + ], + "score": 1.0, + "content": "guage, Kyrgyz and Tatar are Turkic languages with only about", + "type": "text" + }, + { + "bbox": [ + 351, + 680, + 372, + 690 + ], + "score": 0.58, + "content": "4 . 3 \\mathrm { m }", + "type": "inline_equation" + }, + { + "bbox": [ + 373, + 679, + 389, + 691 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 389, + 680, + 411, + 690 + ], + "score": 0.68, + "content": "5 . 2 \\mathrm { m }", + "type": "inline_equation" + }, + { + "bbox": [ + 411, + 679, + 505, + 691 + ], + "score": 1.0, + "content": "speakers, respectively.5", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 689, + 506, + 703 + ], + "spans": [ + { + "bbox": [ + 105, + 689, + 480, + 703 + ], + "score": 1.0, + "content": "We use between 1.8 hours (Kyrgyz) and 9.2 hours of unlabeled audio (Swahili), see Appendix", + "type": "text" + }, + { + "bbox": [ + 481, + 690, + 503, + 702 + ], + "score": 0.83, + "content": "\\ S \\operatorname { E } . 1", + "type": "inline_equation" + }, + { + "bbox": [ + 503, + 689, + 506, + 703 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 34 + } + ], + "index": 33, + "bbox_fs": [ + 105, + 667, + 506, + 703 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "table", + "bbox": [ + 147, + 121, + 464, + 363 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 106, + 70, + 506, + 115 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 70, + 506, + 83 + ], + "spans": [ + { + "bbox": [ + 105, + 70, + 506, + 83 + ], + "score": 1.0, + "content": "Table 2: TIMIT Phoneme Error Rate (PER) in comparison to previous work for the matched and", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 82, + 505, + 93 + ], + "spans": [ + { + "bbox": [ + 105, + 82, + 281, + 93 + ], + "score": 1.0, + "content": "unmatched training data setups (Appendix", + "type": "text" + }, + { + "bbox": [ + 281, + 82, + 304, + 93 + ], + "score": 0.77, + "content": "\\ S \\operatorname { E . 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 304, + 82, + 505, + 93 + ], + "score": 1.0, + "content": "). PER is measured on the Kaldi dev and test sets", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 106, + 93, + 505, + 104 + ], + "spans": [ + { + "bbox": [ + 106, + 93, + 505, + 104 + ], + "score": 1.0, + "content": "(core-dev/core-test) as well as a slightly larger version of the test set (all-test) as used by some of the", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 103, + 491, + 116 + ], + "spans": [ + { + "bbox": [ + 105, + 103, + 154, + 116 + ], + "score": 1.0, + "content": "prior work.", + "type": "text" + }, + { + "bbox": [ + 155, + 103, + 168, + 115 + ], + "score": 0.76, + "content": "( ^ { * } )", + "type": "inline_equation" + }, + { + "bbox": [ + 168, + 103, + 491, + 116 + ], + "score": 1.0, + "content": "indicates experiments that do not use the standard split excluding SA utterances.", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 1.5 + }, + { + "type": "table_body", + "bbox": [ + 147, + 121, + 464, + 363 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 147, + 121, + 464, + 363 + ], + "spans": [ + { + "bbox": [ + 147, + 121, + 464, + 363 + ], + "score": 0.981, + "html": "
ModelLMcore-devcore-testall-test
Supervised learning
LiGRU [Ravanelli et al., 2018]14.9
LiGRU [Ravanelli et al., 2019]14.2
Self and semi-supervised learning
vq-wav2vec [Baevski et al.,2020b]9.611.6
wav2vec 2.0 [Baevski et al.,2020c]7.48.3
Unsupervised learning - matched setup
EODM[Yeh et al.,2019]5-gram36.5=
GAN*[Chen et al., 2019]9-gram=48.6
GAN + HMM* [Chen et al.,2019]9-gram■-26.1
wav2vec-U4-gram17.017.816.6
wav2vec-U + ST4-gram11.312.011.3
Unsupervised learning - unmatched setup
EODM[Yeh et al., 2019]5-gram41.6■
GAN* [Chen et al.,2019]9-gram=50.0
GAN + HMM* [Chen et al., 2019]9-gram=-33.1
wav2vec-U*4-gram21.322.324.4
wav2vec-U + ST*4-gram13.815.018.6
", + "type": "table", + "image_path": "77fd4860fa4ec19b6701be89b6868bc4fe53c80cb80e10678edaf6050869d0ac.jpg" + } + ] + } + ], + "index": 5, + "virtual_lines": [ + { + "bbox": [ + 147, + 121, + 464, + 201.66666666666669 + ], + "spans": [], + "index": 4 + }, + { + "bbox": [ + 147, + 201.66666666666669, + 464, + 282.33333333333337 + ], + "spans": [], + "index": 5 + }, + { + "bbox": [ + 147, + 282.33333333333337, + 464, + 363.00000000000006 + ], + "spans": [], + "index": 6 + } + ] + } + ], + "index": 3.25 + }, + { + "type": "table", + "bbox": [ + 108, + 409, + 503, + 521 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 107, + 369, + 505, + 403 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 105, + 368, + 505, + 382 + ], + "spans": [ + { + "bbox": [ + 105, + 368, + 505, + 382 + ], + "score": 1.0, + "content": "Table 3: WER on the Multilingual Librispeech (MLS) dataset using representations from the wav2vec", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 379, + 506, + 392 + ], + "spans": [ + { + "bbox": [ + 105, + 379, + 506, + 392 + ], + "score": 1.0, + "content": "2.0 XLSR-53 model. We consider German (de), Dutch (nl), French (fr), Spanish (es), Italian (it),", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 389, + 172, + 405 + ], + "spans": [ + { + "bbox": [ + 105, + 389, + 172, + 405 + ], + "score": 1.0, + "content": "Portuguese (pt).", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 8 + }, + { + "type": "table_body", + "bbox": [ + 108, + 409, + 503, + 521 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 108, + 409, + 503, + 521 + ], + "spans": [ + { + "bbox": [ + 108, + 409, + 503, + 521 + ], + "score": 0.983, + "html": "
ModelLabeled data usedLMdenlfresitptAvg
Labeled training hours (full)2k1.6k1.1k918247161
Supervised learning
Pratap et al. [2020]full 5-gram6.4912.025.586.0710.5419.4910.0
Unsupervised learning0h33.3
wav2vec-U 4-gram32.540.239.858.159.843.9
wav2vec-U + ST0h4-gram11.821.414.711.326.326.318.6
", + "type": "table", + "image_path": "b406f6c833d0575a5f43e3163dc4e9bbbbefc546de30a4f3e6945916da6d56ac.jpg" + } + ] + } + ], + "index": 11, + "virtual_lines": [ + { + "bbox": [ + 108, + 409, + 503, + 446.3333333333333 + ], + "spans": [], + "index": 10 + }, + { + "bbox": [ + 108, + 446.3333333333333, + 503, + 483.66666666666663 + ], + "spans": [], + "index": 11 + }, + { + "bbox": [ + 108, + 483.66666666666663, + 503, + 521.0 + ], + "spans": [], + "index": 12 + } + ] + } + ], + "index": 9.5 + }, + { + "type": "text", + "bbox": [ + 106, + 535, + 505, + 613 + ], + "lines": [ + { + "bbox": [ + 106, + 536, + 505, + 549 + ], + "spans": [ + { + "bbox": [ + 106, + 536, + 505, + 549 + ], + "score": 1.0, + "content": "To compare to prior work, we measure WER for Swahili and PER for Kyrgyz and Tatar. For Tatar", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 106, + 547, + 505, + 560 + ], + "spans": [ + { + "bbox": [ + 106, + 547, + 505, + 560 + ], + "score": 1.0, + "content": "and Kyrgyz we opted to use a reduced self-training regime for faster experimental turn-around where", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 104, + 557, + 505, + 572 + ], + "spans": [ + { + "bbox": [ + 104, + 557, + 505, + 572 + ], + "score": 1.0, + "content": "we only perform HMM self-training and we expect better performance with the full self-training", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 569, + 506, + 581 + ], + "spans": [ + { + "bbox": [ + 105, + 569, + 506, + 581 + ], + "score": 1.0, + "content": "setup (Appendix F). Table 4 and Table 5 show that wav2vec-U achieves good performance on these", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 579, + 506, + 592 + ], + "spans": [ + { + "bbox": [ + 105, + 579, + 506, + 592 + ], + "score": 1.0, + "content": "low-resource languages compared to previous work that utilized labeled data. We note that for Tatar", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 591, + 505, + 603 + ], + "spans": [ + { + "bbox": [ + 105, + 591, + 505, + 603 + ], + "score": 1.0, + "content": "and Kyrgyz we use a much smaller amount of speech audio than prior work: compared to XLSR-53", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 601, + 460, + 614 + ], + "spans": [ + { + "bbox": [ + 105, + 601, + 136, + 614 + ], + "score": 1.0, + "content": "we use", + "type": "text" + }, + { + "bbox": [ + 136, + 602, + 155, + 612 + ], + "score": 0.26, + "content": "1 . 8 \\mathrm { h }", + "type": "inline_equation" + }, + { + "bbox": [ + 155, + 601, + 460, + 614 + ], + "score": 1.0, + "content": "unlabeled data vs 17h of labeled data for Kyrgyz and 4.6h vs. 17h for Tatar.", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 16 + }, + { + "type": "title", + "bbox": [ + 107, + 630, + 197, + 643 + ], + "lines": [ + { + "bbox": [ + 104, + 629, + 198, + 644 + ], + "spans": [ + { + "bbox": [ + 104, + 629, + 198, + 644 + ], + "score": 1.0, + "content": "5 Related Work", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 20 + }, + { + "type": "text", + "bbox": [ + 107, + 655, + 505, + 722 + ], + "lines": [ + { + "bbox": [ + 106, + 656, + 505, + 668 + ], + "spans": [ + { + "bbox": [ + 106, + 656, + 505, + 668 + ], + "score": 1.0, + "content": "This paper builds on a large body of prior work which includes semi-supervised speech recognition", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 667, + 506, + 679 + ], + "spans": [ + { + "bbox": [ + 105, + 667, + 506, + 679 + ], + "score": 1.0, + "content": "such as self-training [Kahn et al., 2020a, Xu et al., 2020b, Park et al., 2020]. Some of the earliest", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 677, + 507, + 690 + ], + "spans": [ + { + "bbox": [ + 105, + 677, + 507, + 690 + ], + "score": 1.0, + "content": "work in self-supervised learning of speech representations was was done by van den Oord et al.", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 689, + 506, + 701 + ], + "spans": [ + { + "bbox": [ + 105, + 689, + 506, + 701 + ], + "score": 1.0, + "content": "[2018] for phoneme recognition which was simplified in Schneider et al. [2019] who applied it to full", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 699, + 506, + 713 + ], + "spans": [ + { + "bbox": [ + 105, + 699, + 506, + 713 + ], + "score": 1.0, + "content": "speech recognition. Other work includes language model-style pre-training [Chung et al., 2019a] and", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 711, + 506, + 723 + ], + "spans": [ + { + "bbox": [ + 105, + 711, + 506, + 723 + ], + "score": 1.0, + "content": "learning fixed size representations of audio segments [Chung and Glass, 2018]. There is also work", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 23.5 + } + ], + "page_idx": 7, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 302, + 742, + 308, + 750 + ], + "lines": [ + { + "bbox": [ + 301, + 740, + 310, + 752 + ], + "spans": [ + { + "bbox": [ + 301, + 740, + 310, + 752 + ], + "score": 1.0, + "content": "", + "type": "text", + "height": 12, + "width": 9 + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "table", + "bbox": [ + 147, + 121, + 464, + 363 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 106, + 70, + 506, + 115 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 70, + 506, + 83 + ], + "spans": [ + { + "bbox": [ + 105, + 70, + 506, + 83 + ], + "score": 1.0, + "content": "Table 2: TIMIT Phoneme Error Rate (PER) in comparison to previous work for the matched and", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 82, + 505, + 93 + ], + "spans": [ + { + "bbox": [ + 105, + 82, + 281, + 93 + ], + "score": 1.0, + "content": "unmatched training data setups (Appendix", + "type": "text" + }, + { + "bbox": [ + 281, + 82, + 304, + 93 + ], + "score": 0.77, + "content": "\\ S \\operatorname { E . 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 304, + 82, + 505, + 93 + ], + "score": 1.0, + "content": "). PER is measured on the Kaldi dev and test sets", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 106, + 93, + 505, + 104 + ], + "spans": [ + { + "bbox": [ + 106, + 93, + 505, + 104 + ], + "score": 1.0, + "content": "(core-dev/core-test) as well as a slightly larger version of the test set (all-test) as used by some of the", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 103, + 491, + 116 + ], + "spans": [ + { + "bbox": [ + 105, + 103, + 154, + 116 + ], + "score": 1.0, + "content": "prior work.", + "type": "text" + }, + { + "bbox": [ + 155, + 103, + 168, + 115 + ], + "score": 0.76, + "content": "( ^ { * } )", + "type": "inline_equation" + }, + { + "bbox": [ + 168, + 103, + 491, + 116 + ], + "score": 1.0, + "content": "indicates experiments that do not use the standard split excluding SA utterances.", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 1.5 + }, + { + "type": "table_body", + "bbox": [ + 147, + 121, + 464, + 363 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 147, + 121, + 464, + 363 + ], + "spans": [ + { + "bbox": [ + 147, + 121, + 464, + 363 + ], + "score": 0.981, + "html": "
ModelLMcore-devcore-testall-test
Supervised learning
LiGRU [Ravanelli et al., 2018]14.9
LiGRU [Ravanelli et al., 2019]14.2
Self and semi-supervised learning
vq-wav2vec [Baevski et al.,2020b]9.611.6
wav2vec 2.0 [Baevski et al.,2020c]7.48.3
Unsupervised learning - matched setup
EODM[Yeh et al.,2019]5-gram36.5=
GAN*[Chen et al., 2019]9-gram=48.6
GAN + HMM* [Chen et al.,2019]9-gram■-26.1
wav2vec-U4-gram17.017.816.6
wav2vec-U + ST4-gram11.312.011.3
Unsupervised learning - unmatched setup
EODM[Yeh et al., 2019]5-gram41.6■
GAN* [Chen et al.,2019]9-gram=50.0
GAN + HMM* [Chen et al., 2019]9-gram=-33.1
wav2vec-U*4-gram21.322.324.4
wav2vec-U + ST*4-gram13.815.018.6
", + "type": "table", + "image_path": "77fd4860fa4ec19b6701be89b6868bc4fe53c80cb80e10678edaf6050869d0ac.jpg" + } + ] + } + ], + "index": 5, + "virtual_lines": [ + { + "bbox": [ + 147, + 121, + 464, + 201.66666666666669 + ], + "spans": [], + "index": 4 + }, + { + "bbox": [ + 147, + 201.66666666666669, + 464, + 282.33333333333337 + ], + "spans": [], + "index": 5 + }, + { + "bbox": [ + 147, + 282.33333333333337, + 464, + 363.00000000000006 + ], + "spans": [], + "index": 6 + } + ] + } + ], + "index": 3.25 + }, + { + "type": "table", + "bbox": [ + 108, + 409, + 503, + 521 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 107, + 369, + 505, + 403 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 105, + 368, + 505, + 382 + ], + "spans": [ + { + "bbox": [ + 105, + 368, + 505, + 382 + ], + "score": 1.0, + "content": "Table 3: WER on the Multilingual Librispeech (MLS) dataset using representations from the wav2vec", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 379, + 506, + 392 + ], + "spans": [ + { + "bbox": [ + 105, + 379, + 506, + 392 + ], + "score": 1.0, + "content": "2.0 XLSR-53 model. We consider German (de), Dutch (nl), French (fr), Spanish (es), Italian (it),", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 389, + 172, + 405 + ], + "spans": [ + { + "bbox": [ + 105, + 389, + 172, + 405 + ], + "score": 1.0, + "content": "Portuguese (pt).", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 8 + }, + { + "type": "table_body", + "bbox": [ + 108, + 409, + 503, + 521 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 108, + 409, + 503, + 521 + ], + "spans": [ + { + "bbox": [ + 108, + 409, + 503, + 521 + ], + "score": 0.983, + "html": "
ModelLabeled data usedLMdenlfresitptAvg
Labeled training hours (full)2k1.6k1.1k918247161
Supervised learning
Pratap et al. [2020]full 5-gram6.4912.025.586.0710.5419.4910.0
Unsupervised learning0h33.3
wav2vec-U 4-gram32.540.239.858.159.843.9
wav2vec-U + ST0h4-gram11.821.414.711.326.326.318.6
", + "type": "table", + "image_path": "b406f6c833d0575a5f43e3163dc4e9bbbbefc546de30a4f3e6945916da6d56ac.jpg" + } + ] + } + ], + "index": 11, + "virtual_lines": [ + { + "bbox": [ + 108, + 409, + 503, + 446.3333333333333 + ], + "spans": [], + "index": 10 + }, + { + "bbox": [ + 108, + 446.3333333333333, + 503, + 483.66666666666663 + ], + "spans": [], + "index": 11 + }, + { + "bbox": [ + 108, + 483.66666666666663, + 503, + 521.0 + ], + "spans": [], + "index": 12 + } + ] + } + ], + "index": 9.5 + }, + { + "type": "text", + "bbox": [ + 106, + 535, + 505, + 613 + ], + "lines": [ + { + "bbox": [ + 106, + 536, + 505, + 549 + ], + "spans": [ + { + "bbox": [ + 106, + 536, + 505, + 549 + ], + "score": 1.0, + "content": "To compare to prior work, we measure WER for Swahili and PER for Kyrgyz and Tatar. For Tatar", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 106, + 547, + 505, + 560 + ], + "spans": [ + { + "bbox": [ + 106, + 547, + 505, + 560 + ], + "score": 1.0, + "content": "and Kyrgyz we opted to use a reduced self-training regime for faster experimental turn-around where", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 104, + 557, + 505, + 572 + ], + "spans": [ + { + "bbox": [ + 104, + 557, + 505, + 572 + ], + "score": 1.0, + "content": "we only perform HMM self-training and we expect better performance with the full self-training", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 569, + 506, + 581 + ], + "spans": [ + { + "bbox": [ + 105, + 569, + 506, + 581 + ], + "score": 1.0, + "content": "setup (Appendix F). Table 4 and Table 5 show that wav2vec-U achieves good performance on these", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 579, + 506, + 592 + ], + "spans": [ + { + "bbox": [ + 105, + 579, + 506, + 592 + ], + "score": 1.0, + "content": "low-resource languages compared to previous work that utilized labeled data. We note that for Tatar", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 591, + 505, + 603 + ], + "spans": [ + { + "bbox": [ + 105, + 591, + 505, + 603 + ], + "score": 1.0, + "content": "and Kyrgyz we use a much smaller amount of speech audio than prior work: compared to XLSR-53", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 601, + 460, + 614 + ], + "spans": [ + { + "bbox": [ + 105, + 601, + 136, + 614 + ], + "score": 1.0, + "content": "we use", + "type": "text" + }, + { + "bbox": [ + 136, + 602, + 155, + 612 + ], + "score": 0.26, + "content": "1 . 8 \\mathrm { h }", + "type": "inline_equation" + }, + { + "bbox": [ + 155, + 601, + 460, + 614 + ], + "score": 1.0, + "content": "unlabeled data vs 17h of labeled data for Kyrgyz and 4.6h vs. 17h for Tatar.", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 16, + "bbox_fs": [ + 104, + 536, + 506, + 614 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 630, + 197, + 643 + ], + "lines": [ + { + "bbox": [ + 104, + 629, + 198, + 644 + ], + "spans": [ + { + "bbox": [ + 104, + 629, + 198, + 644 + ], + "score": 1.0, + "content": "5 Related Work", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 20 + }, + { + "type": "text", + "bbox": [ + 107, + 655, + 505, + 722 + ], + "lines": [ + { + "bbox": [ + 106, + 656, + 505, + 668 + ], + "spans": [ + { + "bbox": [ + 106, + 656, + 505, + 668 + ], + "score": 1.0, + "content": "This paper builds on a large body of prior work which includes semi-supervised speech recognition", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 667, + 506, + 679 + ], + "spans": [ + { + "bbox": [ + 105, + 667, + 506, + 679 + ], + "score": 1.0, + "content": "such as self-training [Kahn et al., 2020a, Xu et al., 2020b, Park et al., 2020]. Some of the earliest", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 677, + 507, + 690 + ], + "spans": [ + { + "bbox": [ + 105, + 677, + 507, + 690 + ], + "score": 1.0, + "content": "work in self-supervised learning of speech representations was was done by van den Oord et al.", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 689, + 506, + 701 + ], + "spans": [ + { + "bbox": [ + 105, + 689, + 506, + 701 + ], + "score": 1.0, + "content": "[2018] for phoneme recognition which was simplified in Schneider et al. [2019] who applied it to full", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 699, + 506, + 713 + ], + "spans": [ + { + "bbox": [ + 105, + 699, + 506, + 713 + ], + "score": 1.0, + "content": "speech recognition. Other work includes language model-style pre-training [Chung et al., 2019a] and", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 711, + 506, + 723 + ], + "spans": [ + { + "bbox": [ + 105, + 711, + 506, + 723 + ], + "score": 1.0, + "content": "learning fixed size representations of audio segments [Chung and Glass, 2018]. There is also work", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 261, + 505, + 273 + ], + "spans": [ + { + "bbox": [ + 106, + 261, + 505, + 273 + ], + "score": 1.0, + "content": "on quantization of the continuous speech data [Baevski et al., 2020b,a, Liu et al., 2019, van Niekerk", + "type": "text", + "cross_page": true + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 272, + 506, + 284 + ], + "spans": [ + { + "bbox": [ + 106, + 272, + 506, + 284 + ], + "score": 1.0, + "content": "et al., 2020, Baevski et al., 2020c, Hsu et al., 2021b] and on robustness to domain shift [Hsu et al.,", + "type": "text", + "cross_page": true + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 282, + 505, + 296 + ], + "spans": [ + { + "bbox": [ + 105, + 282, + 505, + 296 + ], + "score": 1.0, + "content": "2021a], multilingual pre-training [Kawakami et al., 2020, Conneau et al., 2020] as well as combining", + "type": "text", + "cross_page": true + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 294, + 272, + 306 + ], + "spans": [ + { + "bbox": [ + 105, + 294, + 272, + 306 + ], + "score": 1.0, + "content": "speech and vision [Harwath et al., 2020].", + "type": "text", + "cross_page": true + } + ], + "index": 19 + } + ], + "index": 23.5, + "bbox_fs": [ + 105, + 656, + 507, + 723 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "table", + "bbox": [ + 108, + 109, + 307, + 226 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 106, + 86, + 291, + 108 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 85, + 292, + 98 + ], + "spans": [ + { + "bbox": [ + 105, + 85, + 292, + 98 + ], + "score": 1.0, + "content": "Table 4: PER for low-resource languages,", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 96, + 214, + 110 + ], + "spans": [ + { + "bbox": [ + 105, + 96, + 214, + 110 + ], + "score": 1.0, + "content": "Tatar (tt) and Kyrgyz (ky).", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5 + }, + { + "type": "table_body", + "bbox": [ + 108, + 109, + 307, + 226 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 108, + 109, + 307, + 226 + ], + "spans": [ + { + "bbox": [ + 108, + 109, + 307, + 226 + ], + "score": 0.979, + "html": "
Modelttky
Supervised learning
Fer et al. [2017]42.538.7
m-CPC [Riviere et al., 2020]42.041.2
XLSR-53 [Conneau et al.,2020]5.16.1
Unsupervised learning
wav2vec-U25.724.1
wav2vec-U + HMM13.714.9
", + "type": "table", + "image_path": "2681bd270ee947c4735ec6105ffa578b603c0a8f14a89fb155bf16d2679be651.jpg" + } + ] + } + ], + "index": 6, + "virtual_lines": [ + { + "bbox": [ + 108, + 109, + 307, + 122.0 + ], + "spans": [], + "index": 2 + }, + { + "bbox": [ + 108, + 122.0, + 307, + 135.0 + ], + "spans": [], + "index": 3 + }, + { + "bbox": [ + 108, + 135.0, + 307, + 148.0 + ], + "spans": [], + "index": 4 + }, + { + "bbox": [ + 108, + 148.0, + 307, + 161.0 + ], + "spans": [], + "index": 5 + }, + { + "bbox": [ + 108, + 161.0, + 307, + 174.0 + ], + "spans": [], + "index": 6 + }, + { + "bbox": [ + 108, + 174.0, + 307, + 187.0 + ], + "spans": [], + "index": 7 + }, + { + "bbox": [ + 108, + 187.0, + 307, + 200.0 + ], + "spans": [], + "index": 8 + }, + { + "bbox": [ + 108, + 200.0, + 307, + 213.0 + ], + "spans": [], + "index": 9 + }, + { + "bbox": [ + 108, + 213.0, + 307, + 226.0 + ], + "spans": [], + "index": 10 + } + ] + } + ], + "index": 3.25 + }, + { + "type": "table", + "bbox": [ + 346, + 120, + 477, + 217 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 319, + 87, + 504, + 119 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 320, + 86, + 505, + 98 + ], + "spans": [ + { + "bbox": [ + 320, + 86, + 505, + 98 + ], + "score": 1.0, + "content": "Table 5: WER for Swahili from the ALFFA", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 320, + 98, + 505, + 110 + ], + "spans": [ + { + "bbox": [ + 320, + 98, + 505, + 110 + ], + "score": 1.0, + "content": "corpus. We compare to the supervised baseline", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 320, + 108, + 412, + 121 + ], + "spans": [ + { + "bbox": [ + 320, + 108, + 412, + 121 + ], + "score": 1.0, + "content": "of the ALFFA project.", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 12 + }, + { + "type": "table_body", + "bbox": [ + 346, + 120, + 477, + 217 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 346, + 120, + 477, + 217 + ], + "spans": [ + { + "bbox": [ + 346, + 120, + 477, + 217 + ], + "score": 0.974, + "html": "
Modelsw
Supervised learning
Besacier et al. [2015]27.36
Unsupervised learning
wav2vec-U52.6
wav2vec-U + ST32.2
", + "type": "table", + "image_path": "75c82e8aef7073d98bca3ee508ceb7960d3a4e7ae6963489d1130fb03b1695b7.jpg" + } + ] + } + ], + "index": 14.5, + "virtual_lines": [ + { + "bbox": [ + 346, + 120, + 477, + 168.5 + ], + "spans": [], + "index": 14 + }, + { + "bbox": [ + 346, + 168.5, + 477, + 217.0 + ], + "spans": [], + "index": 15 + } + ] + } + ], + "index": 13.25 + }, + { + "type": "text", + "bbox": [ + 107, + 261, + 505, + 305 + ], + "lines": [ + { + "bbox": [ + 106, + 261, + 505, + 273 + ], + "spans": [ + { + "bbox": [ + 106, + 261, + 505, + 273 + ], + "score": 1.0, + "content": "on quantization of the continuous speech data [Baevski et al., 2020b,a, Liu et al., 2019, van Niekerk", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 272, + 506, + 284 + ], + "spans": [ + { + "bbox": [ + 106, + 272, + 506, + 284 + ], + "score": 1.0, + "content": "et al., 2020, Baevski et al., 2020c, Hsu et al., 2021b] and on robustness to domain shift [Hsu et al.,", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 282, + 505, + 296 + ], + "spans": [ + { + "bbox": [ + 105, + 282, + 505, + 296 + ], + "score": 1.0, + "content": "2021a], multilingual pre-training [Kawakami et al., 2020, Conneau et al., 2020] as well as combining", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 294, + 272, + 306 + ], + "spans": [ + { + "bbox": [ + 105, + 294, + 272, + 306 + ], + "score": 1.0, + "content": "speech and vision [Harwath et al., 2020].", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 17.5 + }, + { + "type": "text", + "bbox": [ + 106, + 310, + 505, + 409 + ], + "lines": [ + { + "bbox": [ + 106, + 311, + 505, + 322 + ], + "spans": [ + { + "bbox": [ + 106, + 311, + 505, + 322 + ], + "score": 1.0, + "content": "Learning to map speech to phonemes without supervision using adversarial learning has been explored", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 321, + 506, + 334 + ], + "spans": [ + { + "bbox": [ + 105, + 321, + 506, + 334 + ], + "score": 1.0, + "content": "by Liu et al. [2018] who learn a mapping matrix between segment identifiers and phonemes. However,", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 106, + 332, + 506, + 344 + ], + "spans": [ + { + "bbox": [ + 106, + 332, + 506, + 344 + ], + "score": 1.0, + "content": "their work still relied on data segmented into phonemes by human annotators. This has been later", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 343, + 505, + 355 + ], + "spans": [ + { + "bbox": [ + 105, + 343, + 505, + 355 + ], + "score": 1.0, + "content": "extended to use an automatic segmentation [Chen et al., 2019] which is iteratively refined with", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 353, + 505, + 367 + ], + "spans": [ + { + "bbox": [ + 105, + 353, + 505, + 367 + ], + "score": 1.0, + "content": "HMMs. However, cross validation is still performed using labeled data (personal communication with", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 106, + 364, + 506, + 378 + ], + "spans": [ + { + "bbox": [ + 106, + 364, + 506, + 378 + ], + "score": 1.0, + "content": "authors). We also explored HMMs to refine segmentation boundaries (Table A2) but did not find it as", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 375, + 505, + 389 + ], + "spans": [ + { + "bbox": [ + 105, + 375, + 505, + 389 + ], + "score": 1.0, + "content": "effective as self-training. Our work is in part inspired by aligning word embedding spaces of different", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 386, + 506, + 399 + ], + "spans": [ + { + "bbox": [ + 106, + 386, + 506, + 399 + ], + "score": 1.0, + "content": "languages [Mikolov et al., 2013, Artetxe et al., 2017, Conneau et al., 2018] and full unsupervised", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 106, + 398, + 472, + 410 + ], + "spans": [ + { + "bbox": [ + 106, + 398, + 472, + 410 + ], + "score": 1.0, + "content": "machine translation [Lample et al., 2018, Artetxe et al., 2018, Conneau and Lample, 2019].", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 24 + }, + { + "type": "title", + "bbox": [ + 107, + 441, + 275, + 455 + ], + "lines": [ + { + "bbox": [ + 105, + 440, + 277, + 456 + ], + "spans": [ + { + "bbox": [ + 105, + 440, + 277, + 456 + ], + "score": 1.0, + "content": "6 Conclusion and Future Work", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 29 + }, + { + "type": "text", + "bbox": [ + 107, + 475, + 505, + 542 + ], + "lines": [ + { + "bbox": [ + 106, + 476, + 506, + 488 + ], + "spans": [ + { + "bbox": [ + 106, + 476, + 506, + 488 + ], + "score": 1.0, + "content": "wav2vec-U is a framework which enables building speech recognition models without labeled data.", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 488, + 506, + 499 + ], + "spans": [ + { + "bbox": [ + 105, + 488, + 506, + 499 + ], + "score": 1.0, + "content": "It embeds and segments the speech audio with self-supervised representations from wav2vec 2.0,", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 498, + 506, + 511 + ], + "spans": [ + { + "bbox": [ + 105, + 498, + 506, + 511 + ], + "score": 1.0, + "content": "learns a mapping to phonemes with adversarial learning, and cross-validates hyper-parameter choices", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 508, + 505, + 521 + ], + "spans": [ + { + "bbox": [ + 105, + 508, + 505, + 521 + ], + "score": 1.0, + "content": "as well as early stopping with an unsupervised metric. Experiments on the standard Librispeech", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 518, + 505, + 533 + ], + "spans": [ + { + "bbox": [ + 105, + 518, + 505, + 533 + ], + "score": 1.0, + "content": "benchmark show performance close to the state of the art models from only a few years ago, even", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 106, + 531, + 369, + 543 + ], + "spans": [ + { + "bbox": [ + 106, + 531, + 369, + 543 + ], + "score": 1.0, + "content": "though these models relied on nearly 1,000 hours of labeled data.", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 32.5 + }, + { + "type": "text", + "bbox": [ + 107, + 547, + 504, + 602 + ], + "lines": [ + { + "bbox": [ + 106, + 546, + 506, + 560 + ], + "spans": [ + { + "bbox": [ + 106, + 546, + 506, + 560 + ], + "score": 1.0, + "content": "Compared to the previous best unsupervised speech recognition approach, wav2vec-U reduces TIMIT", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 106, + 558, + 505, + 570 + ], + "spans": [ + { + "bbox": [ + 106, + 558, + 505, + 570 + ], + "score": 1.0, + "content": "phone error rate from 26.1 to 11.3. We also demonstrate the viability of our approach on several", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 569, + 506, + 582 + ], + "spans": [ + { + "bbox": [ + 105, + 569, + 506, + 582 + ], + "score": 1.0, + "content": "languages other than English, some of which are low-resource. The ability to build speech recognition", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 579, + 506, + 592 + ], + "spans": [ + { + "bbox": [ + 105, + 579, + 506, + 592 + ], + "score": 1.0, + "content": "models solely from unlabeled speech audio and unlabeled text drastically lowers the effort to build", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 591, + 340, + 603 + ], + "spans": [ + { + "bbox": [ + 105, + 591, + 340, + 603 + ], + "score": 1.0, + "content": "speech technology for many more languages of the world.", + "type": "text" + } + ], + "index": 40 + } + ], + "index": 38 + }, + { + "type": "text", + "bbox": [ + 107, + 607, + 505, + 651 + ], + "lines": [ + { + "bbox": [ + 106, + 607, + 505, + 619 + ], + "spans": [ + { + "bbox": [ + 106, + 607, + 505, + 619 + ], + "score": 1.0, + "content": "Our approach requires phonemization of the text for the language of interest. Moreover, phonemizers", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 617, + 505, + 631 + ], + "spans": [ + { + "bbox": [ + 105, + 617, + 505, + 631 + ], + "score": 1.0, + "content": "are not available for all languages and this presents a bottleneck. To address this, future work may", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 628, + 506, + 642 + ], + "spans": [ + { + "bbox": [ + 105, + 628, + 506, + 642 + ], + "score": 1.0, + "content": "develop phonemizers for more languages, explore phonemization approaches that generalize across", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 640, + 415, + 652 + ], + "spans": [ + { + "bbox": [ + 105, + 640, + 415, + 652 + ], + "score": 1.0, + "content": "languages, or unsupervised training with graphemic text units such as letters.", + "type": "text" + } + ], + "index": 44 + } + ], + "index": 42.5 + }, + { + "type": "text", + "bbox": [ + 107, + 656, + 505, + 722 + ], + "lines": [ + { + "bbox": [ + 105, + 655, + 505, + 668 + ], + "spans": [ + { + "bbox": [ + 105, + 655, + 505, + 668 + ], + "score": 1.0, + "content": "We explored a simple segmentation technique based on self-supervised representations, however, there", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 105, + 667, + 506, + 680 + ], + "spans": [ + { + "bbox": [ + 105, + 667, + 506, + 680 + ], + "score": 1.0, + "content": "is a large body of research on segmentation and some of these techniques may lead to improvements", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 105, + 677, + 505, + 690 + ], + "spans": [ + { + "bbox": [ + 105, + 677, + 505, + 690 + ], + "score": 1.0, + "content": "over our simple approach [Varadarajan et al., 2008, Zhang and Glass, 2009, Gish et al., 2009, Lee and", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 105, + 687, + 506, + 702 + ], + "spans": [ + { + "bbox": [ + 105, + 687, + 506, + 702 + ], + "score": 1.0, + "content": "Glass, 2012, Lee et al., 2015, Ondel et al., 2016, Kamper et al., 2017a,b, Kreuk et al., 2020]. Also,", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 105, + 699, + 505, + 713 + ], + "spans": [ + { + "bbox": [ + 105, + 699, + 505, + 713 + ], + "score": 1.0, + "content": "wav2vec 2.0 learns representations for fixed size units with a fixed stride, however, phonemic units", + "type": "text" + } + ], + "index": 49 + }, + { + "bbox": [ + 105, + 709, + 503, + 725 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 503, + 725 + ], + "score": 1.0, + "content": "are of variable size. Another direction is to learn variable sized representations during pre-training.", + "type": "text" + } + ], + "index": 50 + } + ], + "index": 47.5 + } + ], + "page_idx": 8, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 302, + 741, + 309, + 750 + ], + "lines": [ + { + "bbox": [ + 302, + 741, + 309, + 752 + ], + "spans": [ + { + "bbox": [ + 302, + 741, + 309, + 752 + ], + "score": 1.0, + "content": "9", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "table", + "bbox": [ + 108, + 109, + 307, + 226 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 106, + 86, + 291, + 108 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 85, + 292, + 98 + ], + "spans": [ + { + "bbox": [ + 105, + 85, + 292, + 98 + ], + "score": 1.0, + "content": "Table 4: PER for low-resource languages,", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 96, + 214, + 110 + ], + "spans": [ + { + "bbox": [ + 105, + 96, + 214, + 110 + ], + "score": 1.0, + "content": "Tatar (tt) and Kyrgyz (ky).", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5 + }, + { + "type": "table_body", + "bbox": [ + 108, + 109, + 307, + 226 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 108, + 109, + 307, + 226 + ], + "spans": [ + { + "bbox": [ + 108, + 109, + 307, + 226 + ], + "score": 0.979, + "html": "
Modelttky
Supervised learning
Fer et al. [2017]42.538.7
m-CPC [Riviere et al., 2020]42.041.2
XLSR-53 [Conneau et al.,2020]5.16.1
Unsupervised learning
wav2vec-U25.724.1
wav2vec-U + HMM13.714.9
", + "type": "table", + "image_path": "2681bd270ee947c4735ec6105ffa578b603c0a8f14a89fb155bf16d2679be651.jpg" + } + ] + } + ], + "index": 6, + "virtual_lines": [ + { + "bbox": [ + 108, + 109, + 307, + 122.0 + ], + "spans": [], + "index": 2 + }, + { + "bbox": [ + 108, + 122.0, + 307, + 135.0 + ], + "spans": [], + "index": 3 + }, + { + "bbox": [ + 108, + 135.0, + 307, + 148.0 + ], + "spans": [], + "index": 4 + }, + { + "bbox": [ + 108, + 148.0, + 307, + 161.0 + ], + "spans": [], + "index": 5 + }, + { + "bbox": [ + 108, + 161.0, + 307, + 174.0 + ], + "spans": [], + "index": 6 + }, + { + "bbox": [ + 108, + 174.0, + 307, + 187.0 + ], + "spans": [], + "index": 7 + }, + { + "bbox": [ + 108, + 187.0, + 307, + 200.0 + ], + "spans": [], + "index": 8 + }, + { + "bbox": [ + 108, + 200.0, + 307, + 213.0 + ], + "spans": [], + "index": 9 + }, + { + "bbox": [ + 108, + 213.0, + 307, + 226.0 + ], + "spans": [], + "index": 10 + } + ] + } + ], + "index": 3.25 + }, + { + "type": "table", + "bbox": [ + 346, + 120, + 477, + 217 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 319, + 87, + 504, + 119 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 320, + 86, + 505, + 98 + ], + "spans": [ + { + "bbox": [ + 320, + 86, + 505, + 98 + ], + "score": 1.0, + "content": "Table 5: WER for Swahili from the ALFFA", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 320, + 98, + 505, + 110 + ], + "spans": [ + { + "bbox": [ + 320, + 98, + 505, + 110 + ], + "score": 1.0, + "content": "corpus. We compare to the supervised baseline", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 320, + 108, + 412, + 121 + ], + "spans": [ + { + "bbox": [ + 320, + 108, + 412, + 121 + ], + "score": 1.0, + "content": "of the ALFFA project.", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 12 + }, + { + "type": "table_body", + "bbox": [ + 346, + 120, + 477, + 217 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 346, + 120, + 477, + 217 + ], + "spans": [ + { + "bbox": [ + 346, + 120, + 477, + 217 + ], + "score": 0.974, + "html": "
Modelsw
Supervised learning
Besacier et al. [2015]27.36
Unsupervised learning
wav2vec-U52.6
wav2vec-U + ST32.2
", + "type": "table", + "image_path": "75c82e8aef7073d98bca3ee508ceb7960d3a4e7ae6963489d1130fb03b1695b7.jpg" + } + ] + } + ], + "index": 14.5, + "virtual_lines": [ + { + "bbox": [ + 346, + 120, + 477, + 168.5 + ], + "spans": [], + "index": 14 + }, + { + "bbox": [ + 346, + 168.5, + 477, + 217.0 + ], + "spans": [], + "index": 15 + } + ] + } + ], + "index": 13.25 + }, + { + "type": "text", + "bbox": [ + 107, + 261, + 505, + 305 + ], + "lines": [], + "index": 17.5, + "bbox_fs": [ + 105, + 261, + 506, + 306 + ], + "lines_deleted": true + }, + { + "type": "text", + "bbox": [ + 106, + 310, + 505, + 409 + ], + "lines": [ + { + "bbox": [ + 106, + 311, + 505, + 322 + ], + "spans": [ + { + "bbox": [ + 106, + 311, + 505, + 322 + ], + "score": 1.0, + "content": "Learning to map speech to phonemes without supervision using adversarial learning has been explored", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 321, + 506, + 334 + ], + "spans": [ + { + "bbox": [ + 105, + 321, + 506, + 334 + ], + "score": 1.0, + "content": "by Liu et al. [2018] who learn a mapping matrix between segment identifiers and phonemes. However,", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 106, + 332, + 506, + 344 + ], + "spans": [ + { + "bbox": [ + 106, + 332, + 506, + 344 + ], + "score": 1.0, + "content": "their work still relied on data segmented into phonemes by human annotators. This has been later", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 343, + 505, + 355 + ], + "spans": [ + { + "bbox": [ + 105, + 343, + 505, + 355 + ], + "score": 1.0, + "content": "extended to use an automatic segmentation [Chen et al., 2019] which is iteratively refined with", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 353, + 505, + 367 + ], + "spans": [ + { + "bbox": [ + 105, + 353, + 505, + 367 + ], + "score": 1.0, + "content": "HMMs. However, cross validation is still performed using labeled data (personal communication with", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 106, + 364, + 506, + 378 + ], + "spans": [ + { + "bbox": [ + 106, + 364, + 506, + 378 + ], + "score": 1.0, + "content": "authors). We also explored HMMs to refine segmentation boundaries (Table A2) but did not find it as", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 375, + 505, + 389 + ], + "spans": [ + { + "bbox": [ + 105, + 375, + 505, + 389 + ], + "score": 1.0, + "content": "effective as self-training. Our work is in part inspired by aligning word embedding spaces of different", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 386, + 506, + 399 + ], + "spans": [ + { + "bbox": [ + 106, + 386, + 506, + 399 + ], + "score": 1.0, + "content": "languages [Mikolov et al., 2013, Artetxe et al., 2017, Conneau et al., 2018] and full unsupervised", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 106, + 398, + 472, + 410 + ], + "spans": [ + { + "bbox": [ + 106, + 398, + 472, + 410 + ], + "score": 1.0, + "content": "machine translation [Lample et al., 2018, Artetxe et al., 2018, Conneau and Lample, 2019].", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 24, + "bbox_fs": [ + 105, + 311, + 506, + 410 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 441, + 275, + 455 + ], + "lines": [ + { + "bbox": [ + 105, + 440, + 277, + 456 + ], + "spans": [ + { + "bbox": [ + 105, + 440, + 277, + 456 + ], + "score": 1.0, + "content": "6 Conclusion and Future Work", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 29 + }, + { + "type": "text", + "bbox": [ + 107, + 475, + 505, + 542 + ], + "lines": [ + { + "bbox": [ + 106, + 476, + 506, + 488 + ], + "spans": [ + { + "bbox": [ + 106, + 476, + 506, + 488 + ], + "score": 1.0, + "content": "wav2vec-U is a framework which enables building speech recognition models without labeled data.", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 488, + 506, + 499 + ], + "spans": [ + { + "bbox": [ + 105, + 488, + 506, + 499 + ], + "score": 1.0, + "content": "It embeds and segments the speech audio with self-supervised representations from wav2vec 2.0,", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 498, + 506, + 511 + ], + "spans": [ + { + "bbox": [ + 105, + 498, + 506, + 511 + ], + "score": 1.0, + "content": "learns a mapping to phonemes with adversarial learning, and cross-validates hyper-parameter choices", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 508, + 505, + 521 + ], + "spans": [ + { + "bbox": [ + 105, + 508, + 505, + 521 + ], + "score": 1.0, + "content": "as well as early stopping with an unsupervised metric. Experiments on the standard Librispeech", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 518, + 505, + 533 + ], + "spans": [ + { + "bbox": [ + 105, + 518, + 505, + 533 + ], + "score": 1.0, + "content": "benchmark show performance close to the state of the art models from only a few years ago, even", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 106, + 531, + 369, + 543 + ], + "spans": [ + { + "bbox": [ + 106, + 531, + 369, + 543 + ], + "score": 1.0, + "content": "though these models relied on nearly 1,000 hours of labeled data.", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 32.5, + "bbox_fs": [ + 105, + 476, + 506, + 543 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 547, + 504, + 602 + ], + "lines": [ + { + "bbox": [ + 106, + 546, + 506, + 560 + ], + "spans": [ + { + "bbox": [ + 106, + 546, + 506, + 560 + ], + "score": 1.0, + "content": "Compared to the previous best unsupervised speech recognition approach, wav2vec-U reduces TIMIT", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 106, + 558, + 505, + 570 + ], + "spans": [ + { + "bbox": [ + 106, + 558, + 505, + 570 + ], + "score": 1.0, + "content": "phone error rate from 26.1 to 11.3. We also demonstrate the viability of our approach on several", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 569, + 506, + 582 + ], + "spans": [ + { + "bbox": [ + 105, + 569, + 506, + 582 + ], + "score": 1.0, + "content": "languages other than English, some of which are low-resource. The ability to build speech recognition", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 579, + 506, + 592 + ], + "spans": [ + { + "bbox": [ + 105, + 579, + 506, + 592 + ], + "score": 1.0, + "content": "models solely from unlabeled speech audio and unlabeled text drastically lowers the effort to build", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 591, + 340, + 603 + ], + "spans": [ + { + "bbox": [ + 105, + 591, + 340, + 603 + ], + "score": 1.0, + "content": "speech technology for many more languages of the world.", + "type": "text" + } + ], + "index": 40 + } + ], + "index": 38, + "bbox_fs": [ + 105, + 546, + 506, + 603 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 607, + 505, + 651 + ], + "lines": [ + { + "bbox": [ + 106, + 607, + 505, + 619 + ], + "spans": [ + { + "bbox": [ + 106, + 607, + 505, + 619 + ], + "score": 1.0, + "content": "Our approach requires phonemization of the text for the language of interest. Moreover, phonemizers", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 617, + 505, + 631 + ], + "spans": [ + { + "bbox": [ + 105, + 617, + 505, + 631 + ], + "score": 1.0, + "content": "are not available for all languages and this presents a bottleneck. To address this, future work may", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 628, + 506, + 642 + ], + "spans": [ + { + "bbox": [ + 105, + 628, + 506, + 642 + ], + "score": 1.0, + "content": "develop phonemizers for more languages, explore phonemization approaches that generalize across", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 640, + 415, + 652 + ], + "spans": [ + { + "bbox": [ + 105, + 640, + 415, + 652 + ], + "score": 1.0, + "content": "languages, or unsupervised training with graphemic text units such as letters.", + "type": "text" + } + ], + "index": 44 + } + ], + "index": 42.5, + "bbox_fs": [ + 105, + 607, + 506, + 652 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 656, + 505, + 722 + ], + "lines": [ + { + "bbox": [ + 105, + 655, + 505, + 668 + ], + "spans": [ + { + "bbox": [ + 105, + 655, + 505, + 668 + ], + "score": 1.0, + "content": "We explored a simple segmentation technique based on self-supervised representations, however, there", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 105, + 667, + 506, + 680 + ], + "spans": [ + { + "bbox": [ + 105, + 667, + 506, + 680 + ], + "score": 1.0, + "content": "is a large body of research on segmentation and some of these techniques may lead to improvements", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 105, + 677, + 505, + 690 + ], + "spans": [ + { + "bbox": [ + 105, + 677, + 505, + 690 + ], + "score": 1.0, + "content": "over our simple approach [Varadarajan et al., 2008, Zhang and Glass, 2009, Gish et al., 2009, Lee and", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 105, + 687, + 506, + 702 + ], + "spans": [ + { + "bbox": [ + 105, + 687, + 506, + 702 + ], + "score": 1.0, + "content": "Glass, 2012, Lee et al., 2015, Ondel et al., 2016, Kamper et al., 2017a,b, Kreuk et al., 2020]. Also,", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 105, + 699, + 505, + 713 + ], + "spans": [ + { + "bbox": [ + 105, + 699, + 505, + 713 + ], + "score": 1.0, + "content": "wav2vec 2.0 learns representations for fixed size units with a fixed stride, however, phonemic units", + "type": "text" + } + ], + "index": 49 + }, + { + "bbox": [ + 105, + 709, + 503, + 725 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 503, + 725 + ], + "score": 1.0, + "content": "are of variable size. Another direction is to learn variable sized representations during pre-training.", + "type": "text" + } + ], + "index": 50 + } + ], + "index": 47.5, + "bbox_fs": [ + 105, + 655, + 506, + 725 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "title", + "bbox": [ + 107, + 70, + 339, + 85 + ], + "lines": [ + { + "bbox": [ + 105, + 69, + 341, + 88 + ], + "spans": [ + { + "bbox": [ + 105, + 69, + 341, + 88 + ], + "score": 1.0, + "content": "Acknowledgments and Disclosure of Funding", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "text", + "bbox": [ + 107, + 96, + 506, + 140 + ], + "lines": [ + { + "bbox": [ + 105, + 95, + 506, + 108 + ], + "spans": [ + { + "bbox": [ + 105, + 95, + 506, + 108 + ], + "score": 1.0, + "content": "We thank Zhouhan Lin for helping with initial explorations in this project, Tatiana Likhomanenko", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 106, + 506, + 120 + ], + "spans": [ + { + "bbox": [ + 105, + 106, + 506, + 120 + ], + "score": 1.0, + "content": "for helpful discussions about self-training, Da-Rong Liu for sharing details to reproduce the setup", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 117, + 507, + 131 + ], + "spans": [ + { + "bbox": [ + 105, + 117, + 507, + 131 + ], + "score": 1.0, + "content": "of Chen et al. [2019], Marc’Aurelio Ranzato for general helpful discussions, and Ruth Kipng’eno,", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 128, + 434, + 140 + ], + "spans": [ + { + "bbox": [ + 105, + 128, + 434, + 140 + ], + "score": 1.0, + "content": "Ruth Ndila Ndeto as well as Mark Mutitu for error analysis of our Swahili model.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 2.5 + }, + { + "type": "title", + "bbox": [ + 107, + 155, + 163, + 168 + ], + "lines": [ + { + "bbox": [ + 106, + 154, + 165, + 170 + ], + "spans": [ + { + "bbox": [ + 106, + 154, + 165, + 170 + ], + "score": 1.0, + "content": "References", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 5 + }, + { + "type": "text", + "bbox": [ + 105, + 172, + 507, + 718 + ], + "lines": [ + { + "bbox": [ + 104, + 172, + 506, + 189 + ], + "spans": [ + { + "bbox": [ + 104, + 172, + 506, + 189 + ], + "score": 1.0, + "content": "S. T. Abate, W. Menzel, and B. Tafila. An amharic speech corpus for large vocabulary continuous", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 114, + 185, + 320, + 199 + ], + "spans": [ + { + "bbox": [ + 114, + 185, + 320, + 199 + ], + "score": 1.0, + "content": "speech recognition. In Proc. of Interspeech, 2005.", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 203, + 506, + 217 + ], + "spans": [ + { + "bbox": [ + 105, + 203, + 506, + 217 + ], + "score": 1.0, + "content": "D. Amodei, S. Ananthanarayanan, R. Anubhai, J. Bai, E. Battenberg, C. Case, J. Casper, B. Catanzaro,", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 114, + 214, + 507, + 230 + ], + "spans": [ + { + "bbox": [ + 114, + 214, + 507, + 230 + ], + "score": 1.0, + "content": "Q. Cheng, G. Chen, et al. Deep speech 2: End-to-end speech recognition in english and mandarin.", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 115, + 226, + 216, + 239 + ], + "spans": [ + { + "bbox": [ + 115, + 226, + 216, + 239 + ], + "score": 1.0, + "content": "In Proc. of ICML, 2016.", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 244, + 506, + 257 + ], + "spans": [ + { + "bbox": [ + 105, + 244, + 506, + 257 + ], + "score": 1.0, + "content": "R. Ardila, M. Branson, K. Davis, M. Henretty, M. Kohler, J. Meyer, R. Morais, L. Saunders, F. M.", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 115, + 255, + 506, + 269 + ], + "spans": [ + { + "bbox": [ + 115, + 255, + 506, + 269 + ], + "score": 1.0, + "content": "Tyers, and G. Weber. Common voice: A massively-multilingual speech corpus. Proc. of LREC,", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 115, + 265, + 143, + 280 + ], + "spans": [ + { + "bbox": [ + 115, + 265, + 143, + 280 + ], + "score": 1.0, + "content": "2020.", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 286, + 430, + 297 + ], + "spans": [ + { + "bbox": [ + 105, + 286, + 430, + 297 + ], + "score": 1.0, + "content": "M. Arjovsky, S. Chintala, and L. Bottou. Wasserstein gan. Proc. of ICML, 2017.", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 304, + 505, + 317 + ], + "spans": [ + { + "bbox": [ + 105, + 304, + 505, + 317 + ], + "score": 1.0, + "content": "M. Artetxe, G. Labaka, and E. Agirre. Learning bilingual word embeddings with (almost) no bilingual", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 115, + 315, + 233, + 328 + ], + "spans": [ + { + "bbox": [ + 115, + 315, + 233, + 328 + ], + "score": 1.0, + "content": "data. In Proc. of ACL, 2017.", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 104, + 333, + 506, + 347 + ], + "spans": [ + { + "bbox": [ + 104, + 333, + 506, + 347 + ], + "score": 1.0, + "content": "M. Artetxe, G. Labaka, E. Agirre, and K. Cho. Unsupervised neural machine translation. Proc. of", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 115, + 345, + 168, + 356 + ], + "spans": [ + { + "bbox": [ + 115, + 345, + 168, + 356 + ], + "score": 1.0, + "content": "ICLR, 2018.", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 104, + 361, + 507, + 378 + ], + "spans": [ + { + "bbox": [ + 104, + 361, + 507, + 378 + ], + "score": 1.0, + "content": "A. Baevski and M. Auli. Adaptive input representations for neural language modeling. In Proc. of", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 115, + 374, + 168, + 386 + ], + "spans": [ + { + "bbox": [ + 115, + 374, + 168, + 386 + ], + "score": 1.0, + "content": "ICLR, 2018.", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 393, + 505, + 406 + ], + "spans": [ + { + "bbox": [ + 105, + 393, + 505, + 406 + ], + "score": 1.0, + "content": "A. Baevski, M. Auli, and A. Mohamed. Effectiveness of self-supervised pre-training for speech", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 115, + 404, + 269, + 417 + ], + "spans": [ + { + "bbox": [ + 115, + 404, + 269, + 417 + ], + "score": 1.0, + "content": "recognition. Proc. of ICASSP, 2020a.", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 421, + 506, + 437 + ], + "spans": [ + { + "bbox": [ + 105, + 421, + 506, + 437 + ], + "score": 1.0, + "content": "A. Baevski, S. Schneider, and M. Auli. vq-wav2vec: Self-supervised learning of discrete speech", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 114, + 433, + 284, + 447 + ], + "spans": [ + { + "bbox": [ + 114, + 433, + 284, + 447 + ], + "score": 1.0, + "content": "representations. In Proc. of ICLR, 2020b.", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 452, + 506, + 465 + ], + "spans": [ + { + "bbox": [ + 105, + 452, + 506, + 465 + ], + "score": 1.0, + "content": "A. Baevski, Y. Zhou, A. Mohamed, and M. Auli. wav2vec 2.0: A framework for self-supervised", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 115, + 463, + 372, + 476 + ], + "spans": [ + { + "bbox": [ + 115, + 463, + 372, + 476 + ], + "score": 1.0, + "content": "learning of speech representations. In Proc. of NeurIPS, 2020c.", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 104, + 482, + 506, + 495 + ], + "spans": [ + { + "bbox": [ + 104, + 482, + 506, + 495 + ], + "score": 1.0, + "content": "L. Besacier, E. Gauthier, M. Mangeot, P. Bretier, P. Bagshaw, O. Rosec, T. Moudenc, F. Pellegrino,", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 115, + 493, + 506, + 506 + ], + "spans": [ + { + "bbox": [ + 115, + 493, + 506, + 506 + ], + "score": 1.0, + "content": "S. Voisin, E. Marsico, and P. Nocera. Speech technologies for african languages: example of a", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 115, + 504, + 389, + 517 + ], + "spans": [ + { + "bbox": [ + 115, + 504, + 389, + 517 + ], + "score": 1.0, + "content": "multilingual calculator for education. In Proc. of Interspeech, 2015.", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 104, + 520, + 507, + 536 + ], + "spans": [ + { + "bbox": [ + 104, + 520, + 507, + 536 + ], + "score": 1.0, + "content": "D. Can, V. R. Martinez, P. Papadopoulos, and S. S. Narayanan. Pykaldi: A python wrapper for kaldi.", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 115, + 533, + 223, + 546 + ], + "spans": [ + { + "bbox": [ + 115, + 533, + 223, + 546 + ], + "score": 1.0, + "content": "In Proc. of ICASSP, 2018.", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 551, + 506, + 565 + ], + "spans": [ + { + "bbox": [ + 105, + 551, + 506, + 565 + ], + "score": 1.0, + "content": "K.-Y. Chen, C.-P. Tsai, D.-R. Liu, H.-Y. Lee, and L. shan Lee. Completely unsupervised speech", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 115, + 563, + 506, + 576 + ], + "spans": [ + { + "bbox": [ + 115, + 563, + 506, + 576 + ], + "score": 1.0, + "content": "recognition by a generative adversarial network harmonized with iteratively refined hidden markov", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 115, + 574, + 273, + 587 + ], + "spans": [ + { + "bbox": [ + 115, + 574, + 273, + 587 + ], + "score": 1.0, + "content": "models. In Proc. of Interspeech, 2019.", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 104, + 591, + 506, + 606 + ], + "spans": [ + { + "bbox": [ + 104, + 591, + 506, + 606 + ], + "score": 1.0, + "content": "Y. Chung, W. Weng, S. Tong, and J. R. Glass. Unsupervised cross-modal alignment of speech and", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 115, + 604, + 298, + 617 + ], + "spans": [ + { + "bbox": [ + 115, + 604, + 298, + 617 + ], + "score": 1.0, + "content": "text embedding spaces. Proc. of NIPS, 2018.", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 622, + 506, + 635 + ], + "spans": [ + { + "bbox": [ + 105, + 622, + 506, + 635 + ], + "score": 1.0, + "content": "Y. Chung, W. Hsu, H. Tang, and J. R. Glass. An unsupervised autoregressive model for speech", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 114, + 633, + 329, + 646 + ], + "spans": [ + { + "bbox": [ + 114, + 633, + 329, + 646 + ], + "score": 1.0, + "content": "representation learning. Proc. of Interspeech, 2019a.", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 650, + 506, + 666 + ], + "spans": [ + { + "bbox": [ + 105, + 650, + 506, + 666 + ], + "score": 1.0, + "content": "Y.-A. Chung and J. Glass. Speech2vec: A sequence-to-sequence framework for learning word", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 114, + 662, + 333, + 676 + ], + "spans": [ + { + "bbox": [ + 114, + 662, + 333, + 676 + ], + "score": 1.0, + "content": "embeddings from speech. Proc. of Interspeech, 2018.", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 679, + 506, + 695 + ], + "spans": [ + { + "bbox": [ + 105, + 679, + 506, + 695 + ], + "score": 1.0, + "content": "Y.-A. Chung, W.-N. Hsu, H. Tang, and J. Glass. An unsupervised autoregressive model for speech", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 115, + 692, + 330, + 705 + ], + "spans": [ + { + "bbox": [ + 115, + 692, + 330, + 705 + ], + "score": 1.0, + "content": "representation learning. Proc. of Interspeech, 2019b.", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 710, + 491, + 723 + ], + "spans": [ + { + "bbox": [ + 105, + 710, + 491, + 723 + ], + "score": 1.0, + "content": "A. Conneau and G. Lample. Cross-lingual language model pretraining. Proc. of NeurIPS, 2019.", + "type": "text" + } + ], + "index": 43 + } + ], + "index": 24.5 + } + ], + "page_idx": 9, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 300, + 741, + 311, + 750 + ], + "lines": [ + { + "bbox": [ + 299, + 740, + 313, + 754 + ], + "spans": [ + { + "bbox": [ + 299, + 740, + 313, + 754 + ], + "score": 1.0, + "content": "10", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "title", + "bbox": [ + 107, + 70, + 339, + 85 + ], + "lines": [ + { + "bbox": [ + 105, + 69, + 341, + 88 + ], + "spans": [ + { + "bbox": [ + 105, + 69, + 341, + 88 + ], + "score": 1.0, + "content": "Acknowledgments and Disclosure of Funding", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "text", + "bbox": [ + 107, + 96, + 506, + 140 + ], + "lines": [ + { + "bbox": [ + 105, + 95, + 506, + 108 + ], + "spans": [ + { + "bbox": [ + 105, + 95, + 506, + 108 + ], + "score": 1.0, + "content": "We thank Zhouhan Lin for helping with initial explorations in this project, Tatiana Likhomanenko", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 106, + 506, + 120 + ], + "spans": [ + { + "bbox": [ + 105, + 106, + 506, + 120 + ], + "score": 1.0, + "content": "for helpful discussions about self-training, Da-Rong Liu for sharing details to reproduce the setup", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 117, + 507, + 131 + ], + "spans": [ + { + "bbox": [ + 105, + 117, + 507, + 131 + ], + "score": 1.0, + "content": "of Chen et al. [2019], Marc’Aurelio Ranzato for general helpful discussions, and Ruth Kipng’eno,", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 128, + 434, + 140 + ], + "spans": [ + { + "bbox": [ + 105, + 128, + 434, + 140 + ], + "score": 1.0, + "content": "Ruth Ndila Ndeto as well as Mark Mutitu for error analysis of our Swahili model.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 2.5, + "bbox_fs": [ + 105, + 95, + 507, + 140 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 155, + 163, + 168 + ], + "lines": [ + { + "bbox": [ + 106, + 154, + 165, + 170 + ], + "spans": [ + { + "bbox": [ + 106, + 154, + 165, + 170 + ], + "score": 1.0, + "content": "References", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 5 + }, + { + "type": "list", + "bbox": [ + 105, + 172, + 507, + 718 + ], + "lines": [ + { + "bbox": [ + 104, + 172, + 506, + 189 + ], + "spans": [ + { + "bbox": [ + 104, + 172, + 506, + 189 + ], + "score": 1.0, + "content": "S. T. Abate, W. Menzel, and B. Tafila. An amharic speech corpus for large vocabulary continuous", + "type": "text" + } + ], + "index": 6, + "is_list_start_line": true + }, + { + "bbox": [ + 114, + 185, + 320, + 199 + ], + "spans": [ + { + "bbox": [ + 114, + 185, + 320, + 199 + ], + "score": 1.0, + "content": "speech recognition. In Proc. of Interspeech, 2005.", + "type": "text" + } + ], + "index": 7, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 203, + 506, + 217 + ], + "spans": [ + { + "bbox": [ + 105, + 203, + 506, + 217 + ], + "score": 1.0, + "content": "D. Amodei, S. Ananthanarayanan, R. Anubhai, J. Bai, E. Battenberg, C. Case, J. Casper, B. Catanzaro,", + "type": "text" + } + ], + "index": 8, + "is_list_start_line": true + }, + { + "bbox": [ + 114, + 214, + 507, + 230 + ], + "spans": [ + { + "bbox": [ + 114, + 214, + 507, + 230 + ], + "score": 1.0, + "content": "Q. Cheng, G. Chen, et al. Deep speech 2: End-to-end speech recognition in english and mandarin.", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 115, + 226, + 216, + 239 + ], + "spans": [ + { + "bbox": [ + 115, + 226, + 216, + 239 + ], + "score": 1.0, + "content": "In Proc. of ICML, 2016.", + "type": "text" + } + ], + "index": 10, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 244, + 506, + 257 + ], + "spans": [ + { + "bbox": [ + 105, + 244, + 506, + 257 + ], + "score": 1.0, + "content": "R. Ardila, M. Branson, K. Davis, M. Henretty, M. Kohler, J. Meyer, R. Morais, L. Saunders, F. M.", + "type": "text" + } + ], + "index": 11, + "is_list_start_line": true + }, + { + "bbox": [ + 115, + 255, + 506, + 269 + ], + "spans": [ + { + "bbox": [ + 115, + 255, + 506, + 269 + ], + "score": 1.0, + "content": "Tyers, and G. Weber. Common voice: A massively-multilingual speech corpus. Proc. of LREC,", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 115, + 265, + 143, + 280 + ], + "spans": [ + { + "bbox": [ + 115, + 265, + 143, + 280 + ], + "score": 1.0, + "content": "2020.", + "type": "text" + } + ], + "index": 13, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 286, + 430, + 297 + ], + "spans": [ + { + "bbox": [ + 105, + 286, + 430, + 297 + ], + "score": 1.0, + "content": "M. Arjovsky, S. Chintala, and L. Bottou. Wasserstein gan. Proc. of ICML, 2017.", + "type": "text" + } + ], + "index": 14, + "is_list_start_line": true, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 304, + 505, + 317 + ], + "spans": [ + { + "bbox": [ + 105, + 304, + 505, + 317 + ], + "score": 1.0, + "content": "M. Artetxe, G. Labaka, and E. Agirre. Learning bilingual word embeddings with (almost) no bilingual", + "type": "text" + } + ], + "index": 15, + "is_list_start_line": true + }, + { + "bbox": [ + 115, + 315, + 233, + 328 + ], + "spans": [ + { + "bbox": [ + 115, + 315, + 233, + 328 + ], + "score": 1.0, + "content": "data. In Proc. of ACL, 2017.", + "type": "text" + } + ], + "index": 16, + "is_list_end_line": true + }, + { + "bbox": [ + 104, + 333, + 506, + 347 + ], + "spans": [ + { + "bbox": [ + 104, + 333, + 506, + 347 + ], + "score": 1.0, + "content": "M. Artetxe, G. Labaka, E. Agirre, and K. Cho. Unsupervised neural machine translation. Proc. of", + "type": "text" + } + ], + "index": 17, + "is_list_start_line": true + }, + { + "bbox": [ + 115, + 345, + 168, + 356 + ], + "spans": [ + { + "bbox": [ + 115, + 345, + 168, + 356 + ], + "score": 1.0, + "content": "ICLR, 2018.", + "type": "text" + } + ], + "index": 18, + "is_list_end_line": true + }, + { + "bbox": [ + 104, + 361, + 507, + 378 + ], + "spans": [ + { + "bbox": [ + 104, + 361, + 507, + 378 + ], + "score": 1.0, + "content": "A. Baevski and M. Auli. Adaptive input representations for neural language modeling. In Proc. of", + "type": "text" + } + ], + "index": 19, + "is_list_start_line": true + }, + { + "bbox": [ + 115, + 374, + 168, + 386 + ], + "spans": [ + { + "bbox": [ + 115, + 374, + 168, + 386 + ], + "score": 1.0, + "content": "ICLR, 2018.", + "type": "text" + } + ], + "index": 20, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 393, + 505, + 406 + ], + "spans": [ + { + "bbox": [ + 105, + 393, + 505, + 406 + ], + "score": 1.0, + "content": "A. Baevski, M. Auli, and A. Mohamed. Effectiveness of self-supervised pre-training for speech", + "type": "text" + } + ], + "index": 21, + "is_list_start_line": true + }, + { + "bbox": [ + 115, + 404, + 269, + 417 + ], + "spans": [ + { + "bbox": [ + 115, + 404, + 269, + 417 + ], + "score": 1.0, + "content": "recognition. Proc. of ICASSP, 2020a.", + "type": "text" + } + ], + "index": 22, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 421, + 506, + 437 + ], + "spans": [ + { + "bbox": [ + 105, + 421, + 506, + 437 + ], + "score": 1.0, + "content": "A. Baevski, S. Schneider, and M. Auli. vq-wav2vec: Self-supervised learning of discrete speech", + "type": "text" + } + ], + "index": 23, + "is_list_start_line": true + }, + { + "bbox": [ + 114, + 433, + 284, + 447 + ], + "spans": [ + { + "bbox": [ + 114, + 433, + 284, + 447 + ], + "score": 1.0, + "content": "representations. In Proc. of ICLR, 2020b.", + "type": "text" + } + ], + "index": 24, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 452, + 506, + 465 + ], + "spans": [ + { + "bbox": [ + 105, + 452, + 506, + 465 + ], + "score": 1.0, + "content": "A. Baevski, Y. Zhou, A. Mohamed, and M. Auli. wav2vec 2.0: A framework for self-supervised", + "type": "text" + } + ], + "index": 25, + "is_list_start_line": true + }, + { + "bbox": [ + 115, + 463, + 372, + 476 + ], + "spans": [ + { + "bbox": [ + 115, + 463, + 372, + 476 + ], + "score": 1.0, + "content": "learning of speech representations. In Proc. of NeurIPS, 2020c.", + "type": "text" + } + ], + "index": 26, + "is_list_end_line": true + }, + { + "bbox": [ + 104, + 482, + 506, + 495 + ], + "spans": [ + { + "bbox": [ + 104, + 482, + 506, + 495 + ], + "score": 1.0, + "content": "L. Besacier, E. Gauthier, M. Mangeot, P. Bretier, P. Bagshaw, O. Rosec, T. Moudenc, F. Pellegrino,", + "type": "text" + } + ], + "index": 27, + "is_list_start_line": true + }, + { + "bbox": [ + 115, + 493, + 506, + 506 + ], + "spans": [ + { + "bbox": [ + 115, + 493, + 506, + 506 + ], + "score": 1.0, + "content": "S. Voisin, E. Marsico, and P. Nocera. Speech technologies for african languages: example of a", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 115, + 504, + 389, + 517 + ], + "spans": [ + { + "bbox": [ + 115, + 504, + 389, + 517 + ], + "score": 1.0, + "content": "multilingual calculator for education. In Proc. of Interspeech, 2015.", + "type": "text" + } + ], + "index": 29, + "is_list_end_line": true + }, + { + "bbox": [ + 104, + 520, + 507, + 536 + ], + "spans": [ + { + "bbox": [ + 104, + 520, + 507, + 536 + ], + "score": 1.0, + "content": "D. Can, V. R. Martinez, P. Papadopoulos, and S. S. Narayanan. Pykaldi: A python wrapper for kaldi.", + "type": "text" + } + ], + "index": 30, + "is_list_start_line": true + }, + { + "bbox": [ + 115, + 533, + 223, + 546 + ], + "spans": [ + { + "bbox": [ + 115, + 533, + 223, + 546 + ], + "score": 1.0, + "content": "In Proc. of ICASSP, 2018.", + "type": "text" + } + ], + "index": 31, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 551, + 506, + 565 + ], + "spans": [ + { + "bbox": [ + 105, + 551, + 506, + 565 + ], + "score": 1.0, + "content": "K.-Y. Chen, C.-P. Tsai, D.-R. Liu, H.-Y. Lee, and L. shan Lee. Completely unsupervised speech", + "type": "text" + } + ], + "index": 32, + "is_list_start_line": true + }, + { + "bbox": [ + 115, + 563, + 506, + 576 + ], + "spans": [ + { + "bbox": [ + 115, + 563, + 506, + 576 + ], + "score": 1.0, + "content": "recognition by a generative adversarial network harmonized with iteratively refined hidden markov", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 115, + 574, + 273, + 587 + ], + "spans": [ + { + "bbox": [ + 115, + 574, + 273, + 587 + ], + "score": 1.0, + "content": "models. In Proc. of Interspeech, 2019.", + "type": "text" + } + ], + "index": 34, + "is_list_end_line": true + }, + { + "bbox": [ + 104, + 591, + 506, + 606 + ], + "spans": [ + { + "bbox": [ + 104, + 591, + 506, + 606 + ], + "score": 1.0, + "content": "Y. Chung, W. Weng, S. Tong, and J. R. Glass. Unsupervised cross-modal alignment of speech and", + "type": "text" + } + ], + "index": 35, + "is_list_start_line": true + }, + { + "bbox": [ + 115, + 604, + 298, + 617 + ], + "spans": [ + { + "bbox": [ + 115, + 604, + 298, + 617 + ], + "score": 1.0, + "content": "text embedding spaces. Proc. of NIPS, 2018.", + "type": "text" + } + ], + "index": 36, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 622, + 506, + 635 + ], + "spans": [ + { + "bbox": [ + 105, + 622, + 506, + 635 + ], + "score": 1.0, + "content": "Y. Chung, W. Hsu, H. Tang, and J. R. Glass. An unsupervised autoregressive model for speech", + "type": "text" + } + ], + "index": 37, + "is_list_start_line": true + }, + { + "bbox": [ + 114, + 633, + 329, + 646 + ], + "spans": [ + { + "bbox": [ + 114, + 633, + 329, + 646 + ], + "score": 1.0, + "content": "representation learning. Proc. of Interspeech, 2019a.", + "type": "text" + } + ], + "index": 38, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 650, + 506, + 666 + ], + "spans": [ + { + "bbox": [ + 105, + 650, + 506, + 666 + ], + "score": 1.0, + "content": "Y.-A. Chung and J. Glass. Speech2vec: A sequence-to-sequence framework for learning word", + "type": "text" + } + ], + "index": 39, + "is_list_start_line": true + }, + { + "bbox": [ + 114, + 662, + 333, + 676 + ], + "spans": [ + { + "bbox": [ + 114, + 662, + 333, + 676 + ], + "score": 1.0, + "content": "embeddings from speech. Proc. of Interspeech, 2018.", + "type": "text" + } + ], + "index": 40, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 679, + 506, + 695 + ], + "spans": [ + { + "bbox": [ + 105, + 679, + 506, + 695 + ], + "score": 1.0, + "content": "Y.-A. Chung, W.-N. Hsu, H. Tang, and J. Glass. An unsupervised autoregressive model for speech", + "type": "text" + } + ], + "index": 41, + "is_list_start_line": true + }, + { + "bbox": [ + 115, + 692, + 330, + 705 + ], + "spans": [ + { + "bbox": [ + 115, + 692, + 330, + 705 + ], + "score": 1.0, + "content": "representation learning. Proc. of Interspeech, 2019b.", + "type": "text" + } + ], + "index": 42, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 710, + 491, + 723 + ], + "spans": [ + { + "bbox": [ + 105, + 710, + 491, + 723 + ], + "score": 1.0, + "content": "A. Conneau and G. Lample. Cross-lingual language model pretraining. Proc. of NeurIPS, 2019.", + "type": "text" + } + ], + "index": 43, + "is_list_start_line": true + }, + { + "bbox": [ + 105, + 73, + 506, + 87 + ], + "spans": [ + { + "bbox": [ + 105, + 73, + 506, + 87 + ], + "score": 1.0, + "content": "A. Conneau, G. Lample, M. Ranzato, L. Denoyer, and H. Jégou. Word translation without parallel", + "type": "text", + "cross_page": true + } + ], + "index": 0, + "is_list_start_line": true + }, + { + "bbox": [ + 116, + 84, + 225, + 96 + ], + "spans": [ + { + "bbox": [ + 116, + 84, + 225, + 96 + ], + "score": 1.0, + "content": "data. Proc. of ICLR, 2018.", + "type": "text", + "cross_page": true + } + ], + "index": 1, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 102, + 506, + 117 + ], + "spans": [ + { + "bbox": [ + 105, + 102, + 506, + 117 + ], + "score": 1.0, + "content": "A. Conneau, A. Baevski, R. Collobert, A. Mohamed, and M. Auli. Unsupervised cross-lingual", + "type": "text", + "cross_page": true + } + ], + "index": 2, + "is_list_start_line": true + }, + { + "bbox": [ + 114, + 114, + 427, + 128 + ], + "spans": [ + { + "bbox": [ + 114, + 114, + 427, + 128 + ], + "score": 1.0, + "content": "representation learning for speech recognition. arXiv, abs/2006.13979, 2020.", + "type": "text", + "cross_page": true + } + ], + "index": 3, + "is_list_end_line": true + }, + { + "bbox": [ + 104, + 133, + 506, + 147 + ], + "spans": [ + { + "bbox": [ + 104, + 133, + 506, + 147 + ], + "score": 1.0, + "content": "J. Devlin, M.-W. Chang, K. Lee, and K. Toutanova. Bert: Pre-training of deep bidirectional", + "type": "text", + "cross_page": true + } + ], + "index": 4, + "is_list_start_line": true + }, + { + "bbox": [ + 114, + 144, + 379, + 158 + ], + "spans": [ + { + "bbox": [ + 114, + 144, + 379, + 158 + ], + "score": 1.0, + "content": "transformers for language understanding. Proc. of NAACL, 2019.", + "type": "text", + "cross_page": true + } + ], + "index": 5, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 163, + 506, + 177 + ], + "spans": [ + { + "bbox": [ + 105, + 163, + 506, + 177 + ], + "score": 1.0, + "content": "L. Dong, S. Xu, and B. Xu. Speech-transformer: A no-recurrence sequence-to-sequence model for", + "type": "text", + "cross_page": true + } + ], + "index": 6, + "is_list_start_line": true + }, + { + "bbox": [ + 114, + 173, + 306, + 188 + ], + "spans": [ + { + "bbox": [ + 114, + 173, + 306, + 188 + ], + "score": 1.0, + "content": "speech recognition. In Proc. of ICASSP, 2018.", + "type": "text", + "cross_page": true + } + ], + "index": 7, + "is_list_end_line": true + }, + { + "bbox": [ + 103, + 190, + 506, + 209 + ], + "spans": [ + { + "bbox": [ + 103, + 190, + 506, + 209 + ], + "score": 1.0, + "content": "Z. Fan, M. Li, S. Zhou, and B. Xu. Exploring wav2vec 2.0 on speaker verification and language", + "type": "text", + "cross_page": true + } + ], + "index": 8, + "is_list_start_line": true + }, + { + "bbox": [ + 116, + 205, + 227, + 217 + ], + "spans": [ + { + "bbox": [ + 116, + 205, + 227, + 217 + ], + "score": 1.0, + "content": "identification. arXiv, 2021.", + "type": "text", + "cross_page": true + } + ], + "index": 9, + "is_list_end_line": true + }, + { + "bbox": [ + 104, + 223, + 506, + 237 + ], + "spans": [ + { + "bbox": [ + 104, + 223, + 506, + 237 + ], + "score": 1.0, + "content": "R. Fer, P. Matejka, F. Grézl, O. Plchot, K. Vesel ˇ y, and J. H. ` Cernock ˇ y. Multilingually trained `", + "type": "text", + "cross_page": true + } + ], + "index": 10, + "is_list_start_line": true + }, + { + "bbox": [ + 114, + 234, + 493, + 248 + ], + "spans": [ + { + "bbox": [ + 114, + 234, + 493, + 248 + ], + "score": 1.0, + "content": "bottleneck features in spoken language recognition. Computer Speech & Language, 46, 2017.", + "type": "text", + "cross_page": true + } + ], + "index": 11 + }, + { + "bbox": [ + 104, + 253, + 506, + 267 + ], + "spans": [ + { + "bbox": [ + 104, + 253, + 506, + 267 + ], + "score": 1.0, + "content": "J. S. Garofolo, L. F. Lamel, W. M. Fisher, J. G. Fiscus, D. S. Pallett, and N. L. Dahlgren. The DARPA", + "type": "text", + "cross_page": true + } + ], + "index": 12, + "is_list_start_line": true + }, + { + "bbox": [ + 113, + 263, + 507, + 279 + ], + "spans": [ + { + "bbox": [ + 113, + 263, + 507, + 279 + ], + "score": 1.0, + "content": "TIMIT Acoustic-Phonetic Continuous Speech Corpus CDROM. Linguistic Data Consortium,", + "type": "text", + "cross_page": true + } + ], + "index": 13 + }, + { + "bbox": [ + 115, + 276, + 142, + 289 + ], + "spans": [ + { + "bbox": [ + 115, + 276, + 142, + 289 + ], + "score": 1.0, + "content": "1993.", + "type": "text", + "cross_page": true + } + ], + "index": 14, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 294, + 506, + 308 + ], + "spans": [ + { + "bbox": [ + 105, + 294, + 506, + 308 + ], + "score": 1.0, + "content": "H. Gelas, L. Besacier, and F. Pellegrino. Developments of Swahili resources for an automatic speech", + "type": "text", + "cross_page": true + } + ], + "index": 15, + "is_list_start_line": true + }, + { + "bbox": [ + 113, + 306, + 297, + 319 + ], + "spans": [ + { + "bbox": [ + 113, + 306, + 297, + 319 + ], + "score": 1.0, + "content": "recognition system. In Proc. of SLTU, 2012.", + "type": "text", + "cross_page": true + } + ], + "index": 16, + "is_list_end_line": true + }, + { + "bbox": [ + 104, + 325, + 506, + 339 + ], + "spans": [ + { + "bbox": [ + 104, + 325, + 506, + 339 + ], + "score": 1.0, + "content": "H. Gish, M. Siu, A. Chan, and W. Belfield. Unsupervised training of an hmm-based speech recognizer", + "type": "text", + "cross_page": true + } + ], + "index": 17, + "is_list_start_line": true + }, + { + "bbox": [ + 114, + 335, + 332, + 349 + ], + "spans": [ + { + "bbox": [ + 114, + 335, + 332, + 349 + ], + "score": 1.0, + "content": "for topic classification. In Proc. of Interspeech, 2009.", + "type": "text", + "cross_page": true + } + ], + "index": 18, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 356, + 506, + 367 + ], + "spans": [ + { + "bbox": [ + 105, + 356, + 506, + 367 + ], + "score": 1.0, + "content": "I. J. Goodfellow, J. Pouget-Abadie, M. Mirza, B. Xu, D. Warde-Farley, S. Ozair, A. Courville, and", + "type": "text", + "cross_page": true + } + ], + "index": 19, + "is_list_start_line": true + }, + { + "bbox": [ + 115, + 367, + 381, + 379 + ], + "spans": [ + { + "bbox": [ + 115, + 367, + 381, + 379 + ], + "score": 1.0, + "content": "Y. Bengio. Generative adversarial networks. Proc. of NIPS, 2014.", + "type": "text", + "cross_page": true + } + ], + "index": 20, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 385, + 506, + 399 + ], + "spans": [ + { + "bbox": [ + 105, + 385, + 506, + 399 + ], + "score": 1.0, + "content": "Google. Google cloud: Speech-to-text. https://cloud.google.com/speech-to-text, 2021.", + "type": "text", + "cross_page": true + } + ], + "index": 21, + "is_list_start_line": true + }, + { + "bbox": [ + 115, + 397, + 211, + 408 + ], + "spans": [ + { + "bbox": [ + 115, + 397, + 211, + 408 + ], + "score": 1.0, + "content": "Accessed: 2021-05-13.", + "type": "text", + "cross_page": true + } + ], + "index": 22, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 415, + 506, + 429 + ], + "spans": [ + { + "bbox": [ + 105, + 415, + 506, + 429 + ], + "score": 1.0, + "content": "A. Graves, S. Fernández, and F. Gomez. Connectionist temporal classification: Labelling unsegmented", + "type": "text", + "cross_page": true + } + ], + "index": 23, + "is_list_start_line": true + }, + { + "bbox": [ + 114, + 426, + 403, + 440 + ], + "spans": [ + { + "bbox": [ + 114, + 426, + 403, + 440 + ], + "score": 1.0, + "content": "sequence data with recurrent neural networks. In Proc. of ICML, 2006.", + "type": "text", + "cross_page": true + } + ], + "index": 24, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 445, + 507, + 459 + ], + "spans": [ + { + "bbox": [ + 105, + 445, + 507, + 459 + ], + "score": 1.0, + "content": "A. Gulati, J. Qin, C.-C. Chiu, N. Parmar, Y. Zhang, J. Yu, W. Han, S. Wang, Z. Zhang, Y. Wu,", + "type": "text", + "cross_page": true + } + ], + "index": 25, + "is_list_start_line": true + }, + { + "bbox": [ + 114, + 456, + 506, + 470 + ], + "spans": [ + { + "bbox": [ + 114, + 456, + 506, + 470 + ], + "score": 1.0, + "content": "and R. Pang. Conformer: Convolution-augmented transformer for speech recognition. Proc. of", + "type": "text", + "cross_page": true + } + ], + "index": 26 + }, + { + "bbox": [ + 114, + 468, + 192, + 480 + ], + "spans": [ + { + "bbox": [ + 114, + 468, + 192, + 480 + ], + "score": 1.0, + "content": "Interspeech, 2020.", + "type": "text", + "cross_page": true + } + ], + "index": 27, + "is_list_end_line": true + }, + { + "bbox": [ + 104, + 486, + 506, + 500 + ], + "spans": [ + { + "bbox": [ + 104, + 486, + 506, + 500 + ], + "score": 1.0, + "content": "I. Gulrajani, F. Ahmed, M. Arjovsky, V. Dumoulin, and A. Courville. Improved training of wasserstein", + "type": "text", + "cross_page": true + } + ], + "index": 28, + "is_list_start_line": true + }, + { + "bbox": [ + 114, + 497, + 227, + 510 + ], + "spans": [ + { + "bbox": [ + 114, + 497, + 227, + 510 + ], + "score": 1.0, + "content": "gans. Proc. of NIPS, 2017.", + "type": "text", + "cross_page": true + } + ], + "index": 29, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 516, + 507, + 530 + ], + "spans": [ + { + "bbox": [ + 105, + 516, + 507, + 530 + ], + "score": 1.0, + "content": "W. Han, Z. Zhang, Y. Zhang, J. Yu, C.-C. Chiu, J. Qin, A. Gulati, R. Pang, and Y. Wu. Contextnet:", + "type": "text", + "cross_page": true + } + ], + "index": 30, + "is_list_start_line": true + }, + { + "bbox": [ + 115, + 527, + 506, + 541 + ], + "spans": [ + { + "bbox": [ + 115, + 527, + 506, + 541 + ], + "score": 1.0, + "content": "Improving convolutional neural networks for automatic speech recognition with global context.", + "type": "text", + "cross_page": true + } + ], + "index": 31 + }, + { + "bbox": [ + 116, + 539, + 226, + 551 + ], + "spans": [ + { + "bbox": [ + 116, + 539, + 226, + 551 + ], + "score": 1.0, + "content": "Proc. of Interspeech, 2020.", + "type": "text", + "cross_page": true + } + ], + "index": 32, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 556, + 506, + 570 + ], + "spans": [ + { + "bbox": [ + 105, + 556, + 506, + 570 + ], + "score": 1.0, + "content": "D. Harwath and J. Glass. Towards visually grounded sub-word speech unit discovery. In Proc. of", + "type": "text", + "cross_page": true + } + ], + "index": 33, + "is_list_start_line": true + }, + { + "bbox": [ + 113, + 568, + 281, + 582 + ], + "spans": [ + { + "bbox": [ + 113, + 568, + 281, + 582 + ], + "score": 1.0, + "content": "ICASSP, pages 3017–3021. IEEE, 2019.", + "type": "text", + "cross_page": true + } + ], + "index": 34, + "is_list_end_line": true + }, + { + "bbox": [ + 104, + 586, + 506, + 602 + ], + "spans": [ + { + "bbox": [ + 104, + 586, + 506, + 602 + ], + "score": 1.0, + "content": "D. Harwath, W.-N. Hsu, and J. Glass. Learning hierarchical discrete linguistic units from visually-", + "type": "text", + "cross_page": true + } + ], + "index": 35, + "is_list_start_line": true + }, + { + "bbox": [ + 113, + 598, + 288, + 612 + ], + "spans": [ + { + "bbox": [ + 113, + 598, + 288, + 612 + ], + "score": 1.0, + "content": "grounded speech. In Proc. of ICLR, 2020.", + "type": "text", + "cross_page": true + } + ], + "index": 36, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 618, + 506, + 632 + ], + "spans": [ + { + "bbox": [ + 105, + 618, + 506, + 632 + ], + "score": 1.0, + "content": "K. Heafield. KenLM: Faster and smaller language model queries. In Proceedings of the Sixth", + "type": "text", + "cross_page": true + } + ], + "index": 37, + "is_list_start_line": true + }, + { + "bbox": [ + 115, + 629, + 507, + 643 + ], + "spans": [ + { + "bbox": [ + 115, + 629, + 507, + 643 + ], + "score": 1.0, + "content": "Workshop on Statistical Machine Translation, pages 187–197, Edinburgh, Scotland, July 2011.", + "type": "text", + "cross_page": true + } + ], + "index": 38 + }, + { + "bbox": [ + 113, + 639, + 293, + 653 + ], + "spans": [ + { + "bbox": [ + 113, + 639, + 293, + 653 + ], + "score": 1.0, + "content": "Association for Computational Linguistics.", + "type": "text", + "cross_page": true + } + ], + "index": 39, + "is_list_end_line": true + }, + { + "bbox": [ + 104, + 657, + 507, + 673 + ], + "spans": [ + { + "bbox": [ + 104, + 657, + 507, + 673 + ], + "score": 1.0, + "content": "K. Hirsh-Pasek, D. G. Kemler Nelson, P. W. Jusczyk, K. W. Cassidy, B. Druss, and L. Kennedy.", + "type": "text", + "cross_page": true + } + ], + "index": 40, + "is_list_start_line": true + }, + { + "bbox": [ + 115, + 670, + 439, + 684 + ], + "spans": [ + { + "bbox": [ + 115, + 670, + 439, + 684 + ], + "score": 1.0, + "content": "Clauses are perceptual units for young infants. Cognition, 26(3):269–286, 1987.", + "type": "text", + "cross_page": true + } + ], + "index": 41, + "is_list_end_line": true + }, + { + "bbox": [ + 106, + 689, + 507, + 703 + ], + "spans": [ + { + "bbox": [ + 106, + 689, + 507, + 703 + ], + "score": 1.0, + "content": "W.-N. Hsu, A. Sriram, A. Baevski, T. Likhomanenko, Q. Xu, V. Pratap, J. Kahn, A. Lee, R. Collobert,", + "type": "text", + "cross_page": true + } + ], + "index": 42, + "is_list_start_line": true + }, + { + "bbox": [ + 113, + 696, + 507, + 716 + ], + "spans": [ + { + "bbox": [ + 113, + 696, + 507, + 716 + ], + "score": 1.0, + "content": "G. Synnaeve, et al. Robust wav2vec 2.0: Analyzing domain shift in self-supervised pre-training.", + "type": "text", + "cross_page": true + } + ], + "index": 43 + }, + { + "bbox": [ + 115, + 711, + 284, + 725 + ], + "spans": [ + { + "bbox": [ + 115, + 711, + 284, + 725 + ], + "score": 1.0, + "content": "arXiv preprint arXiv:2104.01027, 2021a.", + "type": "text", + "cross_page": true + } + ], + "index": 44, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 71, + 506, + 86 + ], + "spans": [ + { + "bbox": [ + 105, + 71, + 506, + 86 + ], + "score": 1.0, + "content": "W.-N. Hsu, Y.-H. H. Tsai, B. Bolte, R. Salakhutdinov, and A. Mohamed. Hubert: How much can a", + "type": "text", + "cross_page": true + } + ], + "index": 0, + "is_list_start_line": true + }, + { + "bbox": [ + 115, + 84, + 383, + 96 + ], + "spans": [ + { + "bbox": [ + 115, + 84, + 383, + 96 + ], + "score": 1.0, + "content": "bad teacher benefit ASR pre-training? In Proc. of ICASSP, 2021b.", + "type": "text", + "cross_page": true + } + ], + "index": 1, + "is_list_end_line": true + }, + { + "bbox": [ + 104, + 102, + 506, + 116 + ], + "spans": [ + { + "bbox": [ + 104, + 102, + 506, + 116 + ], + "score": 1.0, + "content": "E. Jang, S. Gu, and B. Poole. Categorical reparameterization with gumbel-softmax. Proc. of ICLR,", + "type": "text", + "cross_page": true + } + ], + "index": 2, + "is_list_start_line": true + }, + { + "bbox": [ + 114, + 113, + 144, + 127 + ], + "spans": [ + { + "bbox": [ + 114, + 113, + 144, + 127 + ], + "score": 1.0, + "content": "2016.", + "type": "text", + "cross_page": true + } + ], + "index": 3, + "is_list_end_line": true + }, + { + "bbox": [ + 104, + 131, + 507, + 147 + ], + "spans": [ + { + "bbox": [ + 104, + 131, + 507, + 147 + ], + "score": 1.0, + "content": "H. Jegou, M. Douze, and C. Schmid. Product quantization for nearest neighbor search. IEEE Trans.", + "type": "text", + "cross_page": true + } + ], + "index": 4, + "is_list_start_line": true + }, + { + "bbox": [ + 115, + 143, + 335, + 157 + ], + "spans": [ + { + "bbox": [ + 115, + 143, + 335, + 157 + ], + "score": 1.0, + "content": "Pattern Anal. Mach. Intell., 33(1):117–128, Jan. 2011.", + "type": "text", + "cross_page": true + } + ], + "index": 5, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 162, + 506, + 176 + ], + "spans": [ + { + "bbox": [ + 105, + 162, + 506, + 176 + ], + "score": 1.0, + "content": "E. K. Johnson and P. W. Jusczyk. Word segmentation by 8-month-olds: When speech cues count", + "type": "text", + "cross_page": true + } + ], + "index": 6, + "is_list_start_line": true + }, + { + "bbox": [ + 115, + 173, + 429, + 187 + ], + "spans": [ + { + "bbox": [ + 115, + 173, + 429, + 187 + ], + "score": 1.0, + "content": "more than statistics. Journal of Memory and Language, 44(4):548–567, 2001.", + "type": "text", + "cross_page": true + } + ], + "index": 7, + "is_list_end_line": true + }, + { + "bbox": [ + 104, + 191, + 506, + 206 + ], + "spans": [ + { + "bbox": [ + 104, + 191, + 506, + 206 + ], + "score": 1.0, + "content": "J. Johnson, M. Douze, and H. Jégou. Billion-scale similarity search with gpus. IEEE Transactions on", + "type": "text", + "cross_page": true + } + ], + "index": 8, + "is_list_start_line": true + }, + { + "bbox": [ + 114, + 203, + 184, + 216 + ], + "spans": [ + { + "bbox": [ + 114, + 203, + 184, + 216 + ], + "score": 1.0, + "content": "Big Data, 2019.", + "type": "text", + "cross_page": true + } + ], + "index": 9, + "is_list_end_line": true + }, + { + "bbox": [ + 104, + 221, + 506, + 235 + ], + "spans": [ + { + "bbox": [ + 104, + 221, + 506, + 235 + ], + "score": 1.0, + "content": "P. W. Jusczyk, D. M. Houston, and M. Newsome. The beginnings of word segmentation in english-", + "type": "text", + "cross_page": true + } + ], + "index": 10, + "is_list_start_line": true + }, + { + "bbox": [ + 114, + 233, + 366, + 246 + ], + "spans": [ + { + "bbox": [ + 114, + 233, + 366, + 246 + ], + "score": 1.0, + "content": "learning infants. Cognitive Psychology, 39(3):159–207, 1999.", + "type": "text", + "cross_page": true + } + ], + "index": 11, + "is_list_end_line": true + }, + { + "bbox": [ + 104, + 252, + 507, + 266 + ], + "spans": [ + { + "bbox": [ + 104, + 252, + 507, + 266 + ], + "score": 1.0, + "content": "J. Kahn, A. Lee, and A. Hannun. Self-training for end-to-end speech recognition. In Proc. of ICASSP,", + "type": "text", + "cross_page": true + } + ], + "index": 12, + "is_list_start_line": true + }, + { + "bbox": [ + 114, + 262, + 148, + 277 + ], + "spans": [ + { + "bbox": [ + 114, + 262, + 148, + 277 + ], + "score": 1.0, + "content": "2020a.", + "type": "text", + "cross_page": true + } + ], + "index": 13, + "is_list_end_line": true + }, + { + "bbox": [ + 104, + 280, + 506, + 295 + ], + "spans": [ + { + "bbox": [ + 104, + 280, + 506, + 295 + ], + "score": 1.0, + "content": "J. Kahn et al. Libri-light: A benchmark for asr with limited or no supervision. In Proc. of ICASSP,", + "type": "text", + "cross_page": true + } + ], + "index": 14, + "is_list_start_line": true + }, + { + "bbox": [ + 115, + 293, + 147, + 305 + ], + "spans": [ + { + "bbox": [ + 115, + 293, + 147, + 305 + ], + "score": 1.0, + "content": "2020b.", + "type": "text", + "cross_page": true + } + ], + "index": 15, + "is_list_end_line": true + }, + { + "bbox": [ + 104, + 309, + 507, + 326 + ], + "spans": [ + { + "bbox": [ + 104, + 309, + 507, + 326 + ], + "score": 1.0, + "content": "H. Kamper, A. Jansen, and S. Goldwater. A segmental framework for fully-unsupervised large-", + "type": "text", + "cross_page": true + } + ], + "index": 16, + "is_list_start_line": true + }, + { + "bbox": [ + 114, + 323, + 419, + 336 + ], + "spans": [ + { + "bbox": [ + 114, + 323, + 419, + 336 + ], + "score": 1.0, + "content": "vocabulary speech recognition. Comput. Speech Lang., 46(C), Nov. 2017a.", + "type": "text", + "cross_page": true + } + ], + "index": 17, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 341, + 506, + 354 + ], + "spans": [ + { + "bbox": [ + 105, + 341, + 506, + 354 + ], + "score": 1.0, + "content": "H. Kamper, K. Livescu, and S. Goldwater. An embedded segmental k-means model for unsupervised", + "type": "text", + "cross_page": true + } + ], + "index": 18, + "is_list_start_line": true + }, + { + "bbox": [ + 114, + 352, + 368, + 367 + ], + "spans": [ + { + "bbox": [ + 114, + 352, + 368, + 367 + ], + "score": 1.0, + "content": "segmentation and clustering of speech. Proc. of ASRU, 2017b.", + "type": "text", + "cross_page": true + } + ], + "index": 19, + "is_list_end_line": true + }, + { + "bbox": [ + 104, + 369, + 506, + 386 + ], + "spans": [ + { + "bbox": [ + 104, + 369, + 506, + 386 + ], + "score": 1.0, + "content": "K. Kawakami, L. Wang, C. Dyer, P. Blunsom, and A. van den Oord. Learning robust and multilingual", + "type": "text", + "cross_page": true + } + ], + "index": 20, + "is_list_start_line": true + }, + { + "bbox": [ + 115, + 383, + 310, + 395 + ], + "spans": [ + { + "bbox": [ + 115, + 383, + 310, + 395 + ], + "score": 1.0, + "content": "speech representations. Proc. of EMNLP, 2020.", + "type": "text", + "cross_page": true + } + ], + "index": 21, + "is_list_end_line": true + }, + { + "bbox": [ + 104, + 401, + 492, + 415 + ], + "spans": [ + { + "bbox": [ + 104, + 401, + 492, + 415 + ], + "score": 1.0, + "content": "D. P. Kingma and J. Ba. Adam: A Method for Stochastic Optimization. In Proc. of ICLR, 2015.", + "type": "text", + "cross_page": true + } + ], + "index": 22, + "is_list_start_line": true, + "is_list_end_line": true + }, + { + "bbox": [ + 104, + 420, + 506, + 434 + ], + "spans": [ + { + "bbox": [ + 104, + 420, + 506, + 434 + ], + "score": 1.0, + "content": "F. Kreuk, J. Keshet, and Y. Adi. Self-supervised contrastive learning for unsupervised phoneme", + "type": "text", + "cross_page": true + } + ], + "index": 23, + "is_list_start_line": true + }, + { + "bbox": [ + 114, + 431, + 285, + 444 + ], + "spans": [ + { + "bbox": [ + 114, + 431, + 285, + 444 + ], + "score": 1.0, + "content": "segmentation. Proc. of Interspeech, 2020.", + "type": "text", + "cross_page": true + } + ], + "index": 24, + "is_list_end_line": true + }, + { + "bbox": [ + 104, + 448, + 506, + 465 + ], + "spans": [ + { + "bbox": [ + 104, + 448, + 506, + 465 + ], + "score": 1.0, + "content": "G. Lample, L. Denoyer, and M. Ranzato. Unsupervised machine translation using monolingual", + "type": "text", + "cross_page": true + } + ], + "index": 25, + "is_list_start_line": true + }, + { + "bbox": [ + 115, + 460, + 270, + 474 + ], + "spans": [ + { + "bbox": [ + 115, + 460, + 270, + 474 + ], + "score": 1.0, + "content": "corpora only. In Proc. of ICLR, 2018.", + "type": "text", + "cross_page": true + } + ], + "index": 26, + "is_list_end_line": true + }, + { + "bbox": [ + 104, + 478, + 508, + 495 + ], + "spans": [ + { + "bbox": [ + 104, + 478, + 508, + 495 + ], + "score": 1.0, + "content": "C. Lee and J. R. Glass. A nonparametric bayesian approach to acoustic model discovery. In Proc. of", + "type": "text", + "cross_page": true + } + ], + "index": 27, + "is_list_start_line": true + }, + { + "bbox": [ + 114, + 492, + 164, + 503 + ], + "spans": [ + { + "bbox": [ + 114, + 492, + 164, + 503 + ], + "score": 1.0, + "content": "ACL, 2012.", + "type": "text", + "cross_page": true + } + ], + "index": 28, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 510, + 506, + 523 + ], + "spans": [ + { + "bbox": [ + 105, + 510, + 506, + 523 + ], + "score": 1.0, + "content": "C. Lee, T. J. O’Donnell, and J. R. Glass. Unsupervised lexicon discovery from acoustic input. TACL,", + "type": "text", + "cross_page": true + } + ], + "index": 29, + "is_list_start_line": true + }, + { + "bbox": [ + 114, + 520, + 144, + 534 + ], + "spans": [ + { + "bbox": [ + 114, + 520, + 144, + 534 + ], + "score": 1.0, + "content": "2015.", + "type": "text", + "cross_page": true + } + ], + "index": 30, + "is_list_end_line": true + }, + { + "bbox": [ + 104, + 539, + 507, + 553 + ], + "spans": [ + { + "bbox": [ + 104, + 539, + 507, + 553 + ], + "score": 1.0, + "content": "M. P. Lewis, G. F. Simon, and C. D. Fennig. Ethnologue: Languages of the world, nineteenth edition.", + "type": "text", + "cross_page": true + } + ], + "index": 31, + "is_list_start_line": true + }, + { + "bbox": [ + 114, + 550, + 343, + 563 + ], + "spans": [ + { + "bbox": [ + 114, + 550, + 343, + 563 + ], + "score": 1.0, + "content": "Online version: http://www.ethnologue.com, 2016.", + "type": "text", + "cross_page": true + } + ], + "index": 32, + "is_list_end_line": true + }, + { + "bbox": [ + 104, + 568, + 506, + 583 + ], + "spans": [ + { + "bbox": [ + 104, + 568, + 506, + 583 + ], + "score": 1.0, + "content": "T. Likhomanenko, Q. Xu, J. Kahn, G. Synnaeve, and R. Collobert. slimipl: Language-model-free", + "type": "text", + "cross_page": true + } + ], + "index": 33, + "is_list_start_line": true + }, + { + "bbox": [ + 115, + 581, + 273, + 593 + ], + "spans": [ + { + "bbox": [ + 115, + 581, + 273, + 593 + ], + "score": 1.0, + "content": "iterative pseudo-labeling. arXiv, 2021.", + "type": "text", + "cross_page": true + } + ], + "index": 34, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 599, + 506, + 613 + ], + "spans": [ + { + "bbox": [ + 105, + 599, + 506, + 613 + ], + "score": 1.0, + "content": "A. H. Liu, T. Tu, H. yi Lee, and L. shan Lee. Towards unsupervised speech recognition and synthesis", + "type": "text", + "cross_page": true + } + ], + "index": 35, + "is_list_start_line": true + }, + { + "bbox": [ + 114, + 610, + 401, + 623 + ], + "spans": [ + { + "bbox": [ + 114, + 610, + 401, + 623 + ], + "score": 1.0, + "content": "with quantized speech representation learning. Proc. of ICASSP, 2019.", + "type": "text", + "cross_page": true + } + ], + "index": 36, + "is_list_end_line": true + }, + { + "bbox": [ + 104, + 628, + 506, + 642 + ], + "spans": [ + { + "bbox": [ + 104, + 628, + 506, + 642 + ], + "score": 1.0, + "content": "D.-R. Liu, K.-Y. Chen, H.-Y. Lee, and L. shan Lee. Completely unsupervised phoneme recognition", + "type": "text", + "cross_page": true + } + ], + "index": 37, + "is_list_start_line": true + }, + { + "bbox": [ + 114, + 640, + 507, + 654 + ], + "spans": [ + { + "bbox": [ + 114, + 640, + 507, + 654 + ], + "score": 1.0, + "content": "by adversarially learning mapping relationships from audio embeddings. Proc. of Interspeech,", + "type": "text", + "cross_page": true + } + ], + "index": 38 + }, + { + "bbox": [ + 114, + 650, + 144, + 666 + ], + "spans": [ + { + "bbox": [ + 114, + 650, + 144, + 666 + ], + "score": 1.0, + "content": "2018.", + "type": "text", + "cross_page": true + } + ], + "index": 39, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 670, + 506, + 684 + ], + "spans": [ + { + "bbox": [ + 105, + 670, + 506, + 684 + ], + "score": 1.0, + "content": "T. Mikolov, Q. V. Le, and I. Sutskever. Exploiting similarities among languages for machine", + "type": "text", + "cross_page": true + } + ], + "index": 40, + "is_list_start_line": true + }, + { + "bbox": [ + 115, + 681, + 323, + 695 + ], + "spans": [ + { + "bbox": [ + 115, + 681, + 323, + 695 + ], + "score": 1.0, + "content": "translation. arXiv preprint arXiv:1309.4168, 2013.", + "type": "text", + "cross_page": true + } + ], + "index": 41, + "is_list_end_line": true + }, + { + "bbox": [ + 104, + 699, + 506, + 714 + ], + "spans": [ + { + "bbox": [ + 104, + 699, + 506, + 714 + ], + "score": 1.0, + "content": "M. Mohri, F. Pereira, and M. Riley. Weighted finite-state transducers in speech recognition. Computer", + "type": "text", + "cross_page": true + } + ], + "index": 42, + "is_list_start_line": true + }, + { + "bbox": [ + 113, + 709, + 282, + 725 + ], + "spans": [ + { + "bbox": [ + 113, + 709, + 282, + 725 + ], + "score": 1.0, + "content": "Speech & Language, 16(1):69–88, 2002.", + "type": "text", + "cross_page": true + } + ], + "index": 43, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 72, + 506, + 86 + ], + "spans": [ + { + "bbox": [ + 105, + 72, + 506, + 86 + ], + "score": 1.0, + "content": "L. Ondel, L. Burget, and J. Cernocký. Variational inference for acoustic unit discovery. In Proc. of", + "type": "text", + "cross_page": true + } + ], + "index": 0, + "is_list_start_line": true + }, + { + "bbox": [ + 115, + 84, + 170, + 96 + ], + "spans": [ + { + "bbox": [ + 115, + 84, + 170, + 96 + ], + "score": 1.0, + "content": "SLTU, 2016.", + "type": "text", + "cross_page": true + } + ], + "index": 1, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 101, + 507, + 116 + ], + "spans": [ + { + "bbox": [ + 105, + 101, + 507, + 116 + ], + "score": 1.0, + "content": "M. Ott, S. Edunov, A. Baevski, A. Fan, S. Gross, N. Ng, D. Grangier, and M. Auli. fairseq: A fast,", + "type": "text", + "cross_page": true + } + ], + "index": 2, + "is_list_start_line": true + }, + { + "bbox": [ + 115, + 113, + 487, + 127 + ], + "spans": [ + { + "bbox": [ + 115, + 113, + 487, + 127 + ], + "score": 1.0, + "content": "extensible toolkit for sequence modeling. In Proc. of NAACL System Demonstrations, 2019.", + "type": "text", + "cross_page": true + } + ], + "index": 3, + "is_list_end_line": true + }, + { + "bbox": [ + 104, + 129, + 506, + 146 + ], + "spans": [ + { + "bbox": [ + 104, + 129, + 506, + 146 + ], + "score": 1.0, + "content": "V. Panayotov, G. Chen, D. Povey, and S. Khudanpur. Librispeech: an asr corpus based on public", + "type": "text", + "cross_page": true + } + ], + "index": 4, + "is_list_start_line": true + }, + { + "bbox": [ + 115, + 142, + 412, + 156 + ], + "spans": [ + { + "bbox": [ + 115, + 142, + 412, + 156 + ], + "score": 1.0, + "content": "domain audio books. In Proc. of ICASSP, pages 5206–5210. IEEE, 2015.", + "type": "text", + "cross_page": true + } + ], + "index": 5, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 160, + 506, + 174 + ], + "spans": [ + { + "bbox": [ + 105, + 160, + 506, + 174 + ], + "score": 1.0, + "content": "D. S. Park, W. Chan, Y. Zhang, C.-C. Chiu, B. Zoph, E. D. Cubuk, and Q. V. Le. Specaugment: A", + "type": "text", + "cross_page": true + } + ], + "index": 6, + "is_list_start_line": true + }, + { + "bbox": [ + 115, + 172, + 507, + 186 + ], + "spans": [ + { + "bbox": [ + 115, + 172, + 507, + 186 + ], + "score": 1.0, + "content": "simple data augmentation method for automatic speech recognition. In Proc. of Interspeech, 2019.", + "type": "text", + "cross_page": true + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 190, + 505, + 204 + ], + "spans": [ + { + "bbox": [ + 105, + 190, + 505, + 204 + ], + "score": 1.0, + "content": "D. S. Park, Y. Zhang, Y. Jia, W. Han, C.-C. Chiu, B. Li, Y. Wu, and Q. V. Le. Improved noisy student", + "type": "text", + "cross_page": true + } + ], + "index": 8, + "is_list_start_line": true + }, + { + "bbox": [ + 114, + 201, + 397, + 214 + ], + "spans": [ + { + "bbox": [ + 114, + 201, + 397, + 214 + ], + "score": 1.0, + "content": "training for automatic speech recognition. Proc. of Interspeech, 2020.", + "type": "text", + "cross_page": true + } + ], + "index": 9, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 219, + 406, + 233 + ], + "spans": [ + { + "bbox": [ + 105, + 219, + 406, + 233 + ], + "score": 1.0, + "content": "K. Park and J. Kim. g2pe. https://github.com/Kyubyong/g2p, 2019.", + "type": "text", + "cross_page": true + } + ], + "index": 10, + "is_list_start_line": true, + "is_list_end_line": true + }, + { + "bbox": [ + 103, + 235, + 507, + 253 + ], + "spans": [ + { + "bbox": [ + 103, + 235, + 507, + 253 + ], + "score": 1.0, + "content": "L. Pepino, P. Riera, and L. Ferrer. Emotion recognition from speech using wav2vec 2.0 embeddings.", + "type": "text", + "cross_page": true + } + ], + "index": 11, + "is_list_start_line": true + }, + { + "bbox": [ + 114, + 248, + 169, + 262 + ], + "spans": [ + { + "bbox": [ + 114, + 248, + 169, + 262 + ], + "score": 1.0, + "content": "arXiv, 2021.", + "type": "text", + "cross_page": true + } + ], + "index": 12, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 268, + 505, + 280 + ], + "spans": [ + { + "bbox": [ + 105, + 268, + 505, + 280 + ], + "score": 1.0, + "content": "L. Polka and J. F. Werker. Developmental changes in perception of nonnative vowel contrasts. Journal", + "type": "text", + "cross_page": true + } + ], + "index": 13, + "is_list_start_line": true + }, + { + "bbox": [ + 114, + 277, + 455, + 292 + ], + "spans": [ + { + "bbox": [ + 114, + 277, + 455, + 292 + ], + "score": 1.0, + "content": "of Experimental Psychology: Human perception and performance, 20(2):421, 1994.", + "type": "text", + "cross_page": true + } + ], + "index": 14, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 296, + 506, + 310 + ], + "spans": [ + { + "bbox": [ + 105, + 296, + 506, + 310 + ], + "score": 1.0, + "content": "D. Povey, A. Ghoshal, G. Boulianne, L. Burget, O. Glembek, N. Goel, M. Hannemann, P. Motlicek,", + "type": "text", + "cross_page": true + } + ], + "index": 15, + "is_list_start_line": true + }, + { + "bbox": [ + 113, + 306, + 507, + 322 + ], + "spans": [ + { + "bbox": [ + 113, + 306, + 507, + 322 + ], + "score": 1.0, + "content": "Y. Qian, P. Schwarz, J. Silovsky, G. Stemmer, and K. Vesely. The kaldi speech recognition toolkit.", + "type": "text", + "cross_page": true + } + ], + "index": 16 + }, + { + "bbox": [ + 115, + 318, + 217, + 332 + ], + "spans": [ + { + "bbox": [ + 115, + 318, + 217, + 332 + ], + "score": 1.0, + "content": "In Proc. of ASRU, 2011.", + "type": "text", + "cross_page": true + } + ], + "index": 17, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 336, + 506, + 351 + ], + "spans": [ + { + "bbox": [ + 105, + 336, + 506, + 351 + ], + "score": 1.0, + "content": "V. Pratap, A. Hannun, Q. Xu, J. Cai, J. Kahn, G. Synnaeve, V. Liptchinsky, and R. Collobert.", + "type": "text", + "cross_page": true + } + ], + "index": 18, + "is_list_start_line": true + }, + { + "bbox": [ + 114, + 347, + 473, + 363 + ], + "spans": [ + { + "bbox": [ + 114, + 347, + 473, + 363 + ], + "score": 1.0, + "content": "Wav2letter++: A fast open-source speech recognition system. In Proc. of ICASSP, 2019.", + "type": "text", + "cross_page": true + } + ], + "index": 19, + "is_list_end_line": true + }, + { + "bbox": [ + 104, + 363, + 507, + 381 + ], + "spans": [ + { + "bbox": [ + 104, + 363, + 507, + 381 + ], + "score": 1.0, + "content": "V. Pratap, Q. Xu, A. Sriram, G. Synnaeve, and R. Collobert. Mls: A large-scale multilingual dataset", + "type": "text", + "cross_page": true + } + ], + "index": 20, + "is_list_start_line": true + }, + { + "bbox": [ + 115, + 377, + 321, + 391 + ], + "spans": [ + { + "bbox": [ + 115, + 377, + 321, + 391 + ], + "score": 1.0, + "content": "for speech research. In Proc. of Interspeech, 2020.", + "type": "text", + "cross_page": true + } + ], + "index": 21, + "is_list_end_line": true + }, + { + "bbox": [ + 104, + 394, + 506, + 410 + ], + "spans": [ + { + "bbox": [ + 104, + 394, + 506, + 410 + ], + "score": 1.0, + "content": "O. Rasanen, G. Doyle, and M. C. Frank. Unsupervised word discovery from speech using automatic", + "type": "text", + "cross_page": true + } + ], + "index": 22, + "is_list_start_line": true + }, + { + "bbox": [ + 113, + 406, + 389, + 420 + ], + "spans": [ + { + "bbox": [ + 113, + 406, + 389, + 420 + ], + "score": 1.0, + "content": "segmentation into syllable-like units. In Proc. of Interspeech, 2015.", + "type": "text", + "cross_page": true + } + ], + "index": 23, + "is_list_end_line": true + }, + { + "bbox": [ + 104, + 423, + 507, + 439 + ], + "spans": [ + { + "bbox": [ + 104, + 423, + 507, + 439 + ], + "score": 1.0, + "content": "M. Ravanelli, P. Brakel, M. Omologo, and Y. Bengio. Light gated recurrent units for speech", + "type": "text", + "cross_page": true + } + ], + "index": 24, + "is_list_start_line": true + }, + { + "bbox": [ + 115, + 436, + 402, + 450 + ], + "spans": [ + { + "bbox": [ + 115, + 436, + 402, + 450 + ], + "score": 1.0, + "content": "recognition. IEEE Trans. on Emerging Topics in Comp. Intel., 2, 2018.", + "type": "text", + "cross_page": true + } + ], + "index": 25, + "is_list_end_line": true + }, + { + "bbox": [ + 103, + 452, + 507, + 469 + ], + "spans": [ + { + "bbox": [ + 103, + 452, + 507, + 469 + ], + "score": 1.0, + "content": "M. Ravanelli, T. Parcollet, and Y. Bengio. The pytorch-kaldi speech recognition toolkit. Proc. of", + "type": "text", + "cross_page": true + } + ], + "index": 26, + "is_list_start_line": true + }, + { + "bbox": [ + 115, + 466, + 178, + 478 + ], + "spans": [ + { + "bbox": [ + 115, + 466, + 178, + 478 + ], + "score": 1.0, + "content": "ICASSP, 2019.", + "type": "text", + "cross_page": true + } + ], + "index": 27, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 483, + 506, + 497 + ], + "spans": [ + { + "bbox": [ + 105, + 483, + 506, + 497 + ], + "score": 1.0, + "content": "M. Rivière, A. Joulin, P.-E. Mazaré, and E. Dupoux. Unsupervised pretraining transfers well across", + "type": "text", + "cross_page": true + } + ], + "index": 28, + "is_list_start_line": true + }, + { + "bbox": [ + 114, + 494, + 270, + 507 + ], + "spans": [ + { + "bbox": [ + 114, + 494, + 270, + 507 + ], + "score": 1.0, + "content": "languages. In Proc. of ICASSP, 2020.", + "type": "text", + "cross_page": true + } + ], + "index": 29, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 513, + 506, + 527 + ], + "spans": [ + { + "bbox": [ + 105, + 513, + 506, + 527 + ], + "score": 1.0, + "content": "S. Schneider, A. Baevski, R. Collobert, and M. Auli. wav2vec: Unsupervised pre-training for speech", + "type": "text", + "cross_page": true + } + ], + "index": 30, + "is_list_start_line": true + }, + { + "bbox": [ + 115, + 524, + 289, + 538 + ], + "spans": [ + { + "bbox": [ + 115, + 524, + 289, + 538 + ], + "score": 1.0, + "content": "recognition. In Proc. of Interspeech, 2019.", + "type": "text", + "cross_page": true + } + ], + "index": 31, + "is_list_end_line": true + }, + { + "bbox": [ + 103, + 540, + 506, + 558 + ], + "spans": [ + { + "bbox": [ + 103, + 540, + 506, + 558 + ], + "score": 1.0, + "content": "N. Srivastava, G. Hinton, A. Krizhevsky, I. Sutskever, and R. Salakhutdinov. Dropout: a simple way", + "type": "text", + "cross_page": true + } + ], + "index": 32, + "is_list_start_line": true + }, + { + "bbox": [ + 113, + 553, + 349, + 567 + ], + "spans": [ + { + "bbox": [ + 113, + 553, + 349, + 567 + ], + "score": 1.0, + "content": "to prevent neural networks from overfitting. JMLR, 2014.", + "type": "text", + "cross_page": true + } + ], + "index": 33, + "is_list_end_line": true + }, + { + "bbox": [ + 104, + 569, + 507, + 586 + ], + "spans": [ + { + "bbox": [ + 104, + 569, + 507, + 586 + ], + "score": 1.0, + "content": "G. Synnaeve, Q. Xu, J. Kahn, T. Likhomanenko, E. Grave, V. Pratap, A. Sriram, V. Liptchinsky,", + "type": "text", + "cross_page": true + } + ], + "index": 34, + "is_list_start_line": true + }, + { + "bbox": [ + 115, + 583, + 505, + 595 + ], + "spans": [ + { + "bbox": [ + 115, + 583, + 505, + 595 + ], + "score": 1.0, + "content": "and R. Collobert. End-to-end ASR: from Supervised to Semi-Supervised Learning with Modern", + "type": "text", + "cross_page": true + } + ], + "index": 35 + }, + { + "bbox": [ + 115, + 593, + 495, + 607 + ], + "spans": [ + { + "bbox": [ + 115, + 593, + 495, + 607 + ], + "score": 1.0, + "content": "Architectures. Proc. of ICML workshop on Self-supervision in Audio and Speech (SAS), 2020.", + "type": "text", + "cross_page": true + } + ], + "index": 36 + }, + { + "bbox": [ + 103, + 609, + 507, + 627 + ], + "spans": [ + { + "bbox": [ + 103, + 609, + 507, + 627 + ], + "score": 1.0, + "content": "M. Tachbelie, S. T. Abate, and L. Besacier. Using different acoustic, lexical and language modeling", + "type": "text", + "cross_page": true + } + ], + "index": 37, + "is_list_start_line": true + }, + { + "bbox": [ + 115, + 623, + 478, + 637 + ], + "spans": [ + { + "bbox": [ + 115, + 623, + 478, + 637 + ], + "score": 1.0, + "content": "units for asr of an under-resourced language - amharic. Speech Communication, 56, 2014.", + "type": "text", + "cross_page": true + } + ], + "index": 38, + "is_list_end_line": true + }, + { + "bbox": [ + 104, + 640, + 505, + 655 + ], + "spans": [ + { + "bbox": [ + 104, + 640, + 505, + 655 + ], + "score": 1.0, + "content": "Z. Tan, A. K. Sarkar, and N. Dehak. rvad: An unsupervised segment-based robust voice activity", + "type": "text", + "cross_page": true + } + ], + "index": 39, + "is_list_start_line": true + }, + { + "bbox": [ + 114, + 652, + 378, + 664 + ], + "spans": [ + { + "bbox": [ + 114, + 652, + 378, + 664 + ], + "score": 1.0, + "content": "detection method. Computer speech & language, 59:1–21, 2020.", + "type": "text", + "cross_page": true + } + ], + "index": 40, + "is_list_end_line": true + }, + { + "bbox": [ + 104, + 669, + 506, + 685 + ], + "spans": [ + { + "bbox": [ + 104, + 669, + 506, + 685 + ], + "score": 1.0, + "content": "A. van den Oord, Y. Li, and O. Vinyals. Representation learning with contrastive predictive coding.", + "type": "text", + "cross_page": true + } + ], + "index": 41, + "is_list_start_line": true + }, + { + "bbox": [ + 114, + 680, + 202, + 695 + ], + "spans": [ + { + "bbox": [ + 114, + 680, + 202, + 695 + ], + "score": 1.0, + "content": "Proc. of NIPS, 2018.", + "type": "text", + "cross_page": true + } + ], + "index": 42, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 699, + 506, + 713 + ], + "spans": [ + { + "bbox": [ + 105, + 699, + 506, + 713 + ], + "score": 1.0, + "content": "B. van Niekerk, L. Nortje, and H. Kamper. Vector-quantized neural networks for acoustic unit", + "type": "text", + "cross_page": true + } + ], + "index": 43, + "is_list_start_line": true + }, + { + "bbox": [ + 115, + 711, + 406, + 724 + ], + "spans": [ + { + "bbox": [ + 115, + 711, + 406, + 724 + ], + "score": 1.0, + "content": "discovery in the zerospeech 2020 challenge. Proc. of Interspeech, 2020.", + "type": "text", + "cross_page": true + } + ], + "index": 44, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 72, + 505, + 87 + ], + "spans": [ + { + "bbox": [ + 105, + 72, + 505, + 87 + ], + "score": 1.0, + "content": "B. Varadarajan, S. Khudanpur, and E. Dupoux. Unsupervised learning of acoustic sub-word units. In", + "type": "text", + "cross_page": true + } + ], + "index": 0, + "is_list_start_line": true + }, + { + "bbox": [ + 115, + 84, + 198, + 96 + ], + "spans": [ + { + "bbox": [ + 115, + 84, + 198, + 96 + ], + "score": 1.0, + "content": "Proc. of ACL, 2008.", + "type": "text", + "cross_page": true + } + ], + "index": 1, + "is_list_end_line": true + }, + { + "bbox": [ + 106, + 102, + 506, + 115 + ], + "spans": [ + { + "bbox": [ + 106, + 102, + 506, + 115 + ], + "score": 1.0, + "content": "A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez, L. Kaiser, and I. Polosukhin.", + "type": "text", + "cross_page": true + } + ], + "index": 2, + "is_list_start_line": true + }, + { + "bbox": [ + 116, + 114, + 316, + 127 + ], + "spans": [ + { + "bbox": [ + 116, + 114, + 316, + 127 + ], + "score": 1.0, + "content": "Attention is all you need. In Proc. of NIPS, 2017.", + "type": "text", + "cross_page": true + } + ], + "index": 3, + "is_list_end_line": true + }, + { + "bbox": [ + 106, + 132, + 505, + 145 + ], + "spans": [ + { + "bbox": [ + 106, + 132, + 505, + 145 + ], + "score": 1.0, + "content": "C. Wang, A. Wu, J. Pino, A. Baevski, M. Auli, and A. Conneau. Large-scale self- and semi-supervised", + "type": "text", + "cross_page": true + } + ], + "index": 4, + "is_list_start_line": true + }, + { + "bbox": [ + 115, + 143, + 295, + 155 + ], + "spans": [ + { + "bbox": [ + 115, + 143, + 295, + 155 + ], + "score": 1.0, + "content": "learning for speech translation. arXiv, 2021.", + "type": "text", + "cross_page": true + } + ], + "index": 5, + "is_list_end_line": true + }, + { + "bbox": [ + 104, + 161, + 505, + 176 + ], + "spans": [ + { + "bbox": [ + 104, + 161, + 505, + 176 + ], + "score": 1.0, + "content": "J. F. Werker and R. C. Tees. Cross-language speech perception: Evidence for perceptual reorganization", + "type": "text", + "cross_page": true + } + ], + "index": 6, + "is_list_start_line": true + }, + { + "bbox": [ + 116, + 173, + 439, + 186 + ], + "spans": [ + { + "bbox": [ + 116, + 173, + 439, + 186 + ], + "score": 1.0, + "content": "during the first year of life. Infant behavior and development, 7(1):49–63, 1984.", + "type": "text", + "cross_page": true + } + ], + "index": 7, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 192, + 506, + 205 + ], + "spans": [ + { + "bbox": [ + 105, + 192, + 506, + 205 + ], + "score": 1.0, + "content": "H. Xu, K. Li, Y. Wang, J. Wang, S. Kang, X. Chen, D. Povey, and S. Khudanpur. Neural network", + "type": "text", + "cross_page": true + } + ], + "index": 8, + "is_list_start_line": true + }, + { + "bbox": [ + 115, + 202, + 507, + 217 + ], + "spans": [ + { + "bbox": [ + 115, + 202, + 507, + 217 + ], + "score": 1.0, + "content": "language modeling with letter-based features and importance sampling. In Proc. of ICASSP, 2018.", + "type": "text", + "cross_page": true + } + ], + "index": 9 + }, + { + "bbox": [ + 106, + 221, + 506, + 234 + ], + "spans": [ + { + "bbox": [ + 106, + 221, + 506, + 234 + ], + "score": 1.0, + "content": "Q. Xu, A. Baevski, T. Likhomanenko, P. Tomasello, A. Conneau, R. Collobert, G. Synnaeve, and", + "type": "text", + "cross_page": true + } + ], + "index": 10, + "is_list_start_line": true + }, + { + "bbox": [ + 115, + 232, + 507, + 247 + ], + "spans": [ + { + "bbox": [ + 115, + 232, + 507, + 247 + ], + "score": 1.0, + "content": "M. Auli. Self-training and pre-training are complementary for speech recognition. In Proc. of", + "type": "text", + "cross_page": true + } + ], + "index": 11 + }, + { + "bbox": [ + 116, + 244, + 182, + 256 + ], + "spans": [ + { + "bbox": [ + 116, + 244, + 182, + 256 + ], + "score": 1.0, + "content": "ICASSP, 2020a.", + "type": "text", + "cross_page": true + } + ], + "index": 12, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 261, + 506, + 276 + ], + "spans": [ + { + "bbox": [ + 105, + 261, + 506, + 276 + ], + "score": 1.0, + "content": "Q. Xu, T. Likhomanenko, J. Kahn, A. Hannun, G. Synnaeve, and R. Collobert. Iterative pseudo-", + "type": "text", + "cross_page": true + } + ], + "index": 13, + "is_list_start_line": true + }, + { + "bbox": [ + 115, + 273, + 362, + 286 + ], + "spans": [ + { + "bbox": [ + 115, + 273, + 362, + 286 + ], + "score": 1.0, + "content": "labeling for speech recognition. Proc. of Interspeech, 2020b.", + "type": "text", + "cross_page": true + } + ], + "index": 14, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 291, + 505, + 306 + ], + "spans": [ + { + "bbox": [ + 105, + 291, + 505, + 306 + ], + "score": 1.0, + "content": "C.-K. Yeh, J. Chen, C. Yu, and D. Yu. Unsupervised speech recognition via segmental empirical", + "type": "text", + "cross_page": true + } + ], + "index": 15, + "is_list_start_line": true + }, + { + "bbox": [ + 115, + 303, + 333, + 316 + ], + "spans": [ + { + "bbox": [ + 115, + 303, + 333, + 316 + ], + "score": 1.0, + "content": "output distribution matching. In Proc. of ICLR, 2019.", + "type": "text", + "cross_page": true + } + ], + "index": 16, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 321, + 505, + 335 + ], + "spans": [ + { + "bbox": [ + 105, + 321, + 505, + 335 + ], + "score": 1.0, + "content": "N. Zeghidour, Q. Xu, V. Liptchinsky, N. Usunier, G. Synnaeve, and R. Collobert. Fully convolutional", + "type": "text", + "cross_page": true + } + ], + "index": 17, + "is_list_start_line": true + }, + { + "bbox": [ + 115, + 333, + 318, + 345 + ], + "spans": [ + { + "bbox": [ + 115, + 333, + 318, + 345 + ], + "score": 1.0, + "content": "speech recognition. arXiv, abs/1812.06864, 2018.", + "type": "text", + "cross_page": true + } + ], + "index": 18, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 351, + 505, + 366 + ], + "spans": [ + { + "bbox": [ + 105, + 351, + 505, + 366 + ], + "score": 1.0, + "content": "F. Zhang, Y. Wang, X. Zhang, C. Liu, Y. Saraf, and G. Zweig. Faster, simpler and more accurate", + "type": "text", + "cross_page": true + } + ], + "index": 19, + "is_list_start_line": true + }, + { + "bbox": [ + 115, + 362, + 384, + 376 + ], + "spans": [ + { + "bbox": [ + 115, + 362, + 384, + 376 + ], + "score": 1.0, + "content": "hybrid asr systems using wordpieces. Proc. of Interspeech, 2020a.", + "type": "text", + "cross_page": true + } + ], + "index": 20, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 380, + 505, + 395 + ], + "spans": [ + { + "bbox": [ + 105, + 380, + 505, + 395 + ], + "score": 1.0, + "content": "Y. Zhang and J. R. Glass. Unsupervised spoken keyword spotting via segmental dtw on gaussian", + "type": "text", + "cross_page": true + } + ], + "index": 21, + "is_list_start_line": true + }, + { + "bbox": [ + 115, + 392, + 485, + 406 + ], + "spans": [ + { + "bbox": [ + 115, + 392, + 485, + 406 + ], + "score": 1.0, + "content": "posteriorgrams. IEEE Workshop on Automatic Speech Recognition & Understanding, 2009.", + "type": "text", + "cross_page": true + } + ], + "index": 22, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 410, + 505, + 425 + ], + "spans": [ + { + "bbox": [ + 105, + 410, + 505, + 425 + ], + "score": 1.0, + "content": "Y. Zhang, J. Qin, D. S. Park, W. Han, C.-C. Chiu, R. Pang, Q. V. Le, and Y. Wu. Pushing the limits", + "type": "text", + "cross_page": true + } + ], + "index": 23, + "is_list_start_line": true + }, + { + "bbox": [ + 115, + 421, + 507, + 437 + ], + "spans": [ + { + "bbox": [ + 115, + 421, + 507, + 437 + ], + "score": 1.0, + "content": "of semi-supervised learning for automatic speech recognition. Proc. of NeurIPS SAS Workshop,", + "type": "text", + "cross_page": true + } + ], + "index": 24 + }, + { + "bbox": [ + 115, + 433, + 147, + 445 + ], + "spans": [ + { + "bbox": [ + 115, + 433, + 147, + 445 + ], + "score": 1.0, + "content": "2020b.", + "type": "text", + "cross_page": true + } + ], + "index": 25, + "is_list_end_line": true + } + ], + "index": 24.5, + "bbox_fs": [ + 104, + 172, + 507, + 723 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 104, + 43, + 507, + 724 + ], + "lines": [ + { + "bbox": [ + 105, + 73, + 506, + 87 + ], + "spans": [ + { + "bbox": [ + 105, + 73, + 506, + 87 + ], + "score": 1.0, + "content": "A. Conneau, G. Lample, M. Ranzato, L. Denoyer, and H. Jégou. Word translation without parallel", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 116, + 84, + 225, + 96 + ], + "spans": [ + { + "bbox": [ + 116, + 84, + 225, + 96 + ], + "score": 1.0, + "content": "data. Proc. of ICLR, 2018.", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 102, + 506, + 117 + ], + "spans": [ + { + "bbox": [ + 105, + 102, + 506, + 117 + ], + "score": 1.0, + "content": "A. Conneau, A. Baevski, R. Collobert, A. Mohamed, and M. Auli. Unsupervised cross-lingual", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 114, + 114, + 427, + 128 + ], + "spans": [ + { + "bbox": [ + 114, + 114, + 427, + 128 + ], + "score": 1.0, + "content": "representation learning for speech recognition. arXiv, abs/2006.13979, 2020.", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 104, + 133, + 506, + 147 + ], + "spans": [ + { + "bbox": [ + 104, + 133, + 506, + 147 + ], + "score": 1.0, + "content": "J. Devlin, M.-W. Chang, K. Lee, and K. Toutanova. Bert: Pre-training of deep bidirectional", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 114, + 144, + 379, + 158 + ], + "spans": [ + { + "bbox": [ + 114, + 144, + 379, + 158 + ], + "score": 1.0, + "content": "transformers for language understanding. Proc. of NAACL, 2019.", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 163, + 506, + 177 + ], + "spans": [ + { + "bbox": [ + 105, + 163, + 506, + 177 + ], + "score": 1.0, + "content": "L. Dong, S. Xu, and B. Xu. Speech-transformer: A no-recurrence sequence-to-sequence model for", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 114, + 173, + 306, + 188 + ], + "spans": [ + { + "bbox": [ + 114, + 173, + 306, + 188 + ], + "score": 1.0, + "content": "speech recognition. In Proc. of ICASSP, 2018.", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 103, + 190, + 506, + 209 + ], + "spans": [ + { + "bbox": [ + 103, + 190, + 506, + 209 + ], + "score": 1.0, + "content": "Z. Fan, M. Li, S. Zhou, and B. Xu. Exploring wav2vec 2.0 on speaker verification and language", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 116, + 205, + 227, + 217 + ], + "spans": [ + { + "bbox": [ + 116, + 205, + 227, + 217 + ], + "score": 1.0, + "content": "identification. arXiv, 2021.", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 104, + 223, + 506, + 237 + ], + "spans": [ + { + "bbox": [ + 104, + 223, + 506, + 237 + ], + "score": 1.0, + "content": "R. Fer, P. Matejka, F. Grézl, O. Plchot, K. Vesel ˇ y, and J. H. ` Cernock ˇ y. Multilingually trained `", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 114, + 234, + 493, + 248 + ], + "spans": [ + { + "bbox": [ + 114, + 234, + 493, + 248 + ], + "score": 1.0, + "content": "bottleneck features in spoken language recognition. Computer Speech & Language, 46, 2017.", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 104, + 253, + 506, + 267 + ], + "spans": [ + { + "bbox": [ + 104, + 253, + 506, + 267 + ], + "score": 1.0, + "content": "J. S. Garofolo, L. F. Lamel, W. M. Fisher, J. G. Fiscus, D. S. Pallett, and N. L. Dahlgren. The DARPA", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 113, + 263, + 507, + 279 + ], + "spans": [ + { + "bbox": [ + 113, + 263, + 507, + 279 + ], + "score": 1.0, + "content": "TIMIT Acoustic-Phonetic Continuous Speech Corpus CDROM. Linguistic Data Consortium,", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 115, + 276, + 142, + 289 + ], + "spans": [ + { + "bbox": [ + 115, + 276, + 142, + 289 + ], + "score": 1.0, + "content": "1993.", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 294, + 506, + 308 + ], + "spans": [ + { + "bbox": [ + 105, + 294, + 506, + 308 + ], + "score": 1.0, + "content": "H. Gelas, L. Besacier, and F. Pellegrino. Developments of Swahili resources for an automatic speech", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 113, + 306, + 297, + 319 + ], + "spans": [ + { + "bbox": [ + 113, + 306, + 297, + 319 + ], + "score": 1.0, + "content": "recognition system. In Proc. of SLTU, 2012.", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 104, + 325, + 506, + 339 + ], + "spans": [ + { + "bbox": [ + 104, + 325, + 506, + 339 + ], + "score": 1.0, + "content": "H. Gish, M. Siu, A. Chan, and W. Belfield. Unsupervised training of an hmm-based speech recognizer", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 114, + 335, + 332, + 349 + ], + "spans": [ + { + "bbox": [ + 114, + 335, + 332, + 349 + ], + "score": 1.0, + "content": "for topic classification. In Proc. of Interspeech, 2009.", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 356, + 506, + 367 + ], + "spans": [ + { + "bbox": [ + 105, + 356, + 506, + 367 + ], + "score": 1.0, + "content": "I. J. Goodfellow, J. Pouget-Abadie, M. Mirza, B. Xu, D. Warde-Farley, S. Ozair, A. Courville, and", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 115, + 367, + 381, + 379 + ], + "spans": [ + { + "bbox": [ + 115, + 367, + 381, + 379 + ], + "score": 1.0, + "content": "Y. Bengio. Generative adversarial networks. Proc. of NIPS, 2014.", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 385, + 506, + 399 + ], + "spans": [ + { + "bbox": [ + 105, + 385, + 506, + 399 + ], + "score": 1.0, + "content": "Google. Google cloud: Speech-to-text. https://cloud.google.com/speech-to-text, 2021.", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 115, + 397, + 211, + 408 + ], + "spans": [ + { + "bbox": [ + 115, + 397, + 211, + 408 + ], + "score": 1.0, + "content": "Accessed: 2021-05-13.", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 415, + 506, + 429 + ], + "spans": [ + { + "bbox": [ + 105, + 415, + 506, + 429 + ], + "score": 1.0, + "content": "A. Graves, S. Fernández, and F. Gomez. Connectionist temporal classification: Labelling unsegmented", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 114, + 426, + 403, + 440 + ], + "spans": [ + { + "bbox": [ + 114, + 426, + 403, + 440 + ], + "score": 1.0, + "content": "sequence data with recurrent neural networks. In Proc. of ICML, 2006.", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 445, + 507, + 459 + ], + "spans": [ + { + "bbox": [ + 105, + 445, + 507, + 459 + ], + "score": 1.0, + "content": "A. Gulati, J. Qin, C.-C. Chiu, N. Parmar, Y. Zhang, J. Yu, W. Han, S. Wang, Z. Zhang, Y. Wu,", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 114, + 456, + 506, + 470 + ], + "spans": [ + { + "bbox": [ + 114, + 456, + 506, + 470 + ], + "score": 1.0, + "content": "and R. Pang. Conformer: Convolution-augmented transformer for speech recognition. Proc. of", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 114, + 468, + 192, + 480 + ], + "spans": [ + { + "bbox": [ + 114, + 468, + 192, + 480 + ], + "score": 1.0, + "content": "Interspeech, 2020.", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 104, + 486, + 506, + 500 + ], + "spans": [ + { + "bbox": [ + 104, + 486, + 506, + 500 + ], + "score": 1.0, + "content": "I. Gulrajani, F. Ahmed, M. Arjovsky, V. Dumoulin, and A. Courville. Improved training of wasserstein", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 114, + 497, + 227, + 510 + ], + "spans": [ + { + "bbox": [ + 114, + 497, + 227, + 510 + ], + "score": 1.0, + "content": "gans. Proc. of NIPS, 2017.", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 516, + 507, + 530 + ], + "spans": [ + { + "bbox": [ + 105, + 516, + 507, + 530 + ], + "score": 1.0, + "content": "W. Han, Z. Zhang, Y. Zhang, J. Yu, C.-C. Chiu, J. Qin, A. Gulati, R. Pang, and Y. Wu. Contextnet:", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 115, + 527, + 506, + 541 + ], + "spans": [ + { + "bbox": [ + 115, + 527, + 506, + 541 + ], + "score": 1.0, + "content": "Improving convolutional neural networks for automatic speech recognition with global context.", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 116, + 539, + 226, + 551 + ], + "spans": [ + { + "bbox": [ + 116, + 539, + 226, + 551 + ], + "score": 1.0, + "content": "Proc. of Interspeech, 2020.", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 556, + 506, + 570 + ], + "spans": [ + { + "bbox": [ + 105, + 556, + 506, + 570 + ], + "score": 1.0, + "content": "D. Harwath and J. Glass. Towards visually grounded sub-word speech unit discovery. In Proc. of", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 113, + 568, + 281, + 582 + ], + "spans": [ + { + "bbox": [ + 113, + 568, + 281, + 582 + ], + "score": 1.0, + "content": "ICASSP, pages 3017–3021. IEEE, 2019.", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 104, + 586, + 506, + 602 + ], + "spans": [ + { + "bbox": [ + 104, + 586, + 506, + 602 + ], + "score": 1.0, + "content": "D. Harwath, W.-N. Hsu, and J. Glass. Learning hierarchical discrete linguistic units from visually-", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 113, + 598, + 288, + 612 + ], + "spans": [ + { + "bbox": [ + 113, + 598, + 288, + 612 + ], + "score": 1.0, + "content": "grounded speech. In Proc. of ICLR, 2020.", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 618, + 506, + 632 + ], + "spans": [ + { + "bbox": [ + 105, + 618, + 506, + 632 + ], + "score": 1.0, + "content": "K. Heafield. KenLM: Faster and smaller language model queries. In Proceedings of the Sixth", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 115, + 629, + 507, + 643 + ], + "spans": [ + { + "bbox": [ + 115, + 629, + 507, + 643 + ], + "score": 1.0, + "content": "Workshop on Statistical Machine Translation, pages 187–197, Edinburgh, Scotland, July 2011.", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 113, + 639, + 293, + 653 + ], + "spans": [ + { + "bbox": [ + 113, + 639, + 293, + 653 + ], + "score": 1.0, + "content": "Association for Computational Linguistics.", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 104, + 657, + 507, + 673 + ], + "spans": [ + { + "bbox": [ + 104, + 657, + 507, + 673 + ], + "score": 1.0, + "content": "K. Hirsh-Pasek, D. G. Kemler Nelson, P. W. Jusczyk, K. W. Cassidy, B. Druss, and L. Kennedy.", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 115, + 670, + 439, + 684 + ], + "spans": [ + { + "bbox": [ + 115, + 670, + 439, + 684 + ], + "score": 1.0, + "content": "Clauses are perceptual units for young infants. Cognition, 26(3):269–286, 1987.", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 106, + 689, + 507, + 703 + ], + "spans": [ + { + "bbox": [ + 106, + 689, + 507, + 703 + ], + "score": 1.0, + "content": "W.-N. Hsu, A. Sriram, A. Baevski, T. Likhomanenko, Q. Xu, V. Pratap, J. Kahn, A. Lee, R. Collobert,", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 113, + 696, + 507, + 716 + ], + "spans": [ + { + "bbox": [ + 113, + 696, + 507, + 716 + ], + "score": 1.0, + "content": "G. Synnaeve, et al. Robust wav2vec 2.0: Analyzing domain shift in self-supervised pre-training.", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 115, + 711, + 284, + 725 + ], + "spans": [ + { + "bbox": [ + 115, + 711, + 284, + 725 + ], + "score": 1.0, + "content": "arXiv preprint arXiv:2104.01027, 2021a.", + "type": "text" + } + ], + "index": 44 + } + ], + "index": 22 + } + ], + "page_idx": 10, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 300, + 741, + 311, + 751 + ], + "lines": [ + { + "bbox": [ + 299, + 740, + 312, + 754 + ], + "spans": [ + { + "bbox": [ + 299, + 740, + 312, + 754 + ], + "score": 1.0, + "content": "11", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "list", + "bbox": [ + 104, + 43, + 507, + 724 + ], + "lines": [], + "index": 22, + "bbox_fs": [ + 103, + 73, + 507, + 725 + ], + "lines_deleted": true + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 104, + 69, + 508, + 729 + ], + "lines": [ + { + "bbox": [ + 105, + 71, + 506, + 86 + ], + "spans": [ + { + "bbox": [ + 105, + 71, + 506, + 86 + ], + "score": 1.0, + "content": "W.-N. Hsu, Y.-H. H. Tsai, B. Bolte, R. Salakhutdinov, and A. Mohamed. Hubert: How much can a", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 115, + 84, + 383, + 96 + ], + "spans": [ + { + "bbox": [ + 115, + 84, + 383, + 96 + ], + "score": 1.0, + "content": "bad teacher benefit ASR pre-training? In Proc. of ICASSP, 2021b.", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 104, + 102, + 506, + 116 + ], + "spans": [ + { + "bbox": [ + 104, + 102, + 506, + 116 + ], + "score": 1.0, + "content": "E. Jang, S. Gu, and B. Poole. Categorical reparameterization with gumbel-softmax. Proc. of ICLR,", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 114, + 113, + 144, + 127 + ], + "spans": [ + { + "bbox": [ + 114, + 113, + 144, + 127 + ], + "score": 1.0, + "content": "2016.", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 104, + 131, + 507, + 147 + ], + "spans": [ + { + "bbox": [ + 104, + 131, + 507, + 147 + ], + "score": 1.0, + "content": "H. Jegou, M. Douze, and C. Schmid. Product quantization for nearest neighbor search. IEEE Trans.", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 115, + 143, + 335, + 157 + ], + "spans": [ + { + "bbox": [ + 115, + 143, + 335, + 157 + ], + "score": 1.0, + "content": "Pattern Anal. Mach. Intell., 33(1):117–128, Jan. 2011.", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 162, + 506, + 176 + ], + "spans": [ + { + "bbox": [ + 105, + 162, + 506, + 176 + ], + "score": 1.0, + "content": "E. K. Johnson and P. W. Jusczyk. Word segmentation by 8-month-olds: When speech cues count", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 115, + 173, + 429, + 187 + ], + "spans": [ + { + "bbox": [ + 115, + 173, + 429, + 187 + ], + "score": 1.0, + "content": "more than statistics. Journal of Memory and Language, 44(4):548–567, 2001.", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 104, + 191, + 506, + 206 + ], + "spans": [ + { + "bbox": [ + 104, + 191, + 506, + 206 + ], + "score": 1.0, + "content": "J. Johnson, M. Douze, and H. Jégou. Billion-scale similarity search with gpus. IEEE Transactions on", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 114, + 203, + 184, + 216 + ], + "spans": [ + { + "bbox": [ + 114, + 203, + 184, + 216 + ], + "score": 1.0, + "content": "Big Data, 2019.", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 104, + 221, + 506, + 235 + ], + "spans": [ + { + "bbox": [ + 104, + 221, + 506, + 235 + ], + "score": 1.0, + "content": "P. W. Jusczyk, D. M. Houston, and M. Newsome. The beginnings of word segmentation in english-", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 114, + 233, + 366, + 246 + ], + "spans": [ + { + "bbox": [ + 114, + 233, + 366, + 246 + ], + "score": 1.0, + "content": "learning infants. Cognitive Psychology, 39(3):159–207, 1999.", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 104, + 252, + 507, + 266 + ], + "spans": [ + { + "bbox": [ + 104, + 252, + 507, + 266 + ], + "score": 1.0, + "content": "J. Kahn, A. Lee, and A. Hannun. Self-training for end-to-end speech recognition. In Proc. of ICASSP,", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 114, + 262, + 148, + 277 + ], + "spans": [ + { + "bbox": [ + 114, + 262, + 148, + 277 + ], + "score": 1.0, + "content": "2020a.", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 104, + 280, + 506, + 295 + ], + "spans": [ + { + "bbox": [ + 104, + 280, + 506, + 295 + ], + "score": 1.0, + "content": "J. Kahn et al. Libri-light: A benchmark for asr with limited or no supervision. In Proc. of ICASSP,", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 115, + 293, + 147, + 305 + ], + "spans": [ + { + "bbox": [ + 115, + 293, + 147, + 305 + ], + "score": 1.0, + "content": "2020b.", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 104, + 309, + 507, + 326 + ], + "spans": [ + { + "bbox": [ + 104, + 309, + 507, + 326 + ], + "score": 1.0, + "content": "H. Kamper, A. Jansen, and S. Goldwater. A segmental framework for fully-unsupervised large-", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 114, + 323, + 419, + 336 + ], + "spans": [ + { + "bbox": [ + 114, + 323, + 419, + 336 + ], + "score": 1.0, + "content": "vocabulary speech recognition. Comput. Speech Lang., 46(C), Nov. 2017a.", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 341, + 506, + 354 + ], + "spans": [ + { + "bbox": [ + 105, + 341, + 506, + 354 + ], + "score": 1.0, + "content": "H. Kamper, K. Livescu, and S. Goldwater. An embedded segmental k-means model for unsupervised", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 114, + 352, + 368, + 367 + ], + "spans": [ + { + "bbox": [ + 114, + 352, + 368, + 367 + ], + "score": 1.0, + "content": "segmentation and clustering of speech. Proc. of ASRU, 2017b.", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 104, + 369, + 506, + 386 + ], + "spans": [ + { + "bbox": [ + 104, + 369, + 506, + 386 + ], + "score": 1.0, + "content": "K. Kawakami, L. Wang, C. Dyer, P. Blunsom, and A. van den Oord. Learning robust and multilingual", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 115, + 383, + 310, + 395 + ], + "spans": [ + { + "bbox": [ + 115, + 383, + 310, + 395 + ], + "score": 1.0, + "content": "speech representations. Proc. of EMNLP, 2020.", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 104, + 401, + 492, + 415 + ], + "spans": [ + { + "bbox": [ + 104, + 401, + 492, + 415 + ], + "score": 1.0, + "content": "D. P. Kingma and J. Ba. Adam: A Method for Stochastic Optimization. In Proc. of ICLR, 2015.", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 104, + 420, + 506, + 434 + ], + "spans": [ + { + "bbox": [ + 104, + 420, + 506, + 434 + ], + "score": 1.0, + "content": "F. Kreuk, J. Keshet, and Y. Adi. Self-supervised contrastive learning for unsupervised phoneme", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 114, + 431, + 285, + 444 + ], + "spans": [ + { + "bbox": [ + 114, + 431, + 285, + 444 + ], + "score": 1.0, + "content": "segmentation. Proc. of Interspeech, 2020.", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 104, + 448, + 506, + 465 + ], + "spans": [ + { + "bbox": [ + 104, + 448, + 506, + 465 + ], + "score": 1.0, + "content": "G. Lample, L. Denoyer, and M. Ranzato. Unsupervised machine translation using monolingual", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 115, + 460, + 270, + 474 + ], + "spans": [ + { + "bbox": [ + 115, + 460, + 270, + 474 + ], + "score": 1.0, + "content": "corpora only. In Proc. of ICLR, 2018.", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 104, + 478, + 508, + 495 + ], + "spans": [ + { + "bbox": [ + 104, + 478, + 508, + 495 + ], + "score": 1.0, + "content": "C. Lee and J. R. Glass. A nonparametric bayesian approach to acoustic model discovery. In Proc. of", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 114, + 492, + 164, + 503 + ], + "spans": [ + { + "bbox": [ + 114, + 492, + 164, + 503 + ], + "score": 1.0, + "content": "ACL, 2012.", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 510, + 506, + 523 + ], + "spans": [ + { + "bbox": [ + 105, + 510, + 506, + 523 + ], + "score": 1.0, + "content": "C. Lee, T. J. O’Donnell, and J. R. Glass. Unsupervised lexicon discovery from acoustic input. TACL,", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 114, + 520, + 144, + 534 + ], + "spans": [ + { + "bbox": [ + 114, + 520, + 144, + 534 + ], + "score": 1.0, + "content": "2015.", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 104, + 539, + 507, + 553 + ], + "spans": [ + { + "bbox": [ + 104, + 539, + 507, + 553 + ], + "score": 1.0, + "content": "M. P. Lewis, G. F. Simon, and C. D. Fennig. Ethnologue: Languages of the world, nineteenth edition.", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 114, + 550, + 343, + 563 + ], + "spans": [ + { + "bbox": [ + 114, + 550, + 343, + 563 + ], + "score": 1.0, + "content": "Online version: http://www.ethnologue.com, 2016.", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 104, + 568, + 506, + 583 + ], + "spans": [ + { + "bbox": [ + 104, + 568, + 506, + 583 + ], + "score": 1.0, + "content": "T. Likhomanenko, Q. Xu, J. Kahn, G. Synnaeve, and R. Collobert. slimipl: Language-model-free", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 115, + 581, + 273, + 593 + ], + "spans": [ + { + "bbox": [ + 115, + 581, + 273, + 593 + ], + "score": 1.0, + "content": "iterative pseudo-labeling. arXiv, 2021.", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 599, + 506, + 613 + ], + "spans": [ + { + "bbox": [ + 105, + 599, + 506, + 613 + ], + "score": 1.0, + "content": "A. H. Liu, T. Tu, H. yi Lee, and L. shan Lee. Towards unsupervised speech recognition and synthesis", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 114, + 610, + 401, + 623 + ], + "spans": [ + { + "bbox": [ + 114, + 610, + 401, + 623 + ], + "score": 1.0, + "content": "with quantized speech representation learning. Proc. of ICASSP, 2019.", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 104, + 628, + 506, + 642 + ], + "spans": [ + { + "bbox": [ + 104, + 628, + 506, + 642 + ], + "score": 1.0, + "content": "D.-R. Liu, K.-Y. Chen, H.-Y. Lee, and L. shan Lee. Completely unsupervised phoneme recognition", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 114, + 640, + 507, + 654 + ], + "spans": [ + { + "bbox": [ + 114, + 640, + 507, + 654 + ], + "score": 1.0, + "content": "by adversarially learning mapping relationships from audio embeddings. Proc. of Interspeech,", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 114, + 650, + 144, + 666 + ], + "spans": [ + { + "bbox": [ + 114, + 650, + 144, + 666 + ], + "score": 1.0, + "content": "2018.", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 670, + 506, + 684 + ], + "spans": [ + { + "bbox": [ + 105, + 670, + 506, + 684 + ], + "score": 1.0, + "content": "T. Mikolov, Q. V. Le, and I. Sutskever. Exploiting similarities among languages for machine", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 115, + 681, + 323, + 695 + ], + "spans": [ + { + "bbox": [ + 115, + 681, + 323, + 695 + ], + "score": 1.0, + "content": "translation. arXiv preprint arXiv:1309.4168, 2013.", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 104, + 699, + 506, + 714 + ], + "spans": [ + { + "bbox": [ + 104, + 699, + 506, + 714 + ], + "score": 1.0, + "content": "M. Mohri, F. Pereira, and M. Riley. Weighted finite-state transducers in speech recognition. Computer", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 113, + 709, + 282, + 725 + ], + "spans": [ + { + "bbox": [ + 113, + 709, + 282, + 725 + ], + "score": 1.0, + "content": "Speech & Language, 16(1):69–88, 2002.", + "type": "text" + } + ], + "index": 43 + } + ], + "index": 21.5 + } + ], + "page_idx": 11, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 300, + 741, + 311, + 751 + ], + "lines": [ + { + "bbox": [ + 299, + 740, + 312, + 754 + ], + "spans": [ + { + "bbox": [ + 299, + 740, + 312, + 754 + ], + "score": 1.0, + "content": "12", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "list", + "bbox": [ + 104, + 69, + 508, + 729 + ], + "lines": [], + "index": 21.5, + "bbox_fs": [ + 104, + 71, + 508, + 725 + ], + "lines_deleted": true + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 104, + 58, + 507, + 730 + ], + "lines": [ + { + "bbox": [ + 105, + 72, + 506, + 86 + ], + "spans": [ + { + "bbox": [ + 105, + 72, + 506, + 86 + ], + "score": 1.0, + "content": "L. Ondel, L. Burget, and J. Cernocký. Variational inference for acoustic unit discovery. In Proc. of", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 115, + 84, + 170, + 96 + ], + "spans": [ + { + "bbox": [ + 115, + 84, + 170, + 96 + ], + "score": 1.0, + "content": "SLTU, 2016.", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 101, + 507, + 116 + ], + "spans": [ + { + "bbox": [ + 105, + 101, + 507, + 116 + ], + "score": 1.0, + "content": "M. Ott, S. Edunov, A. Baevski, A. Fan, S. Gross, N. Ng, D. Grangier, and M. Auli. fairseq: A fast,", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 115, + 113, + 487, + 127 + ], + "spans": [ + { + "bbox": [ + 115, + 113, + 487, + 127 + ], + "score": 1.0, + "content": "extensible toolkit for sequence modeling. In Proc. of NAACL System Demonstrations, 2019.", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 104, + 129, + 506, + 146 + ], + "spans": [ + { + "bbox": [ + 104, + 129, + 506, + 146 + ], + "score": 1.0, + "content": "V. Panayotov, G. Chen, D. Povey, and S. Khudanpur. Librispeech: an asr corpus based on public", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 115, + 142, + 412, + 156 + ], + "spans": [ + { + "bbox": [ + 115, + 142, + 412, + 156 + ], + "score": 1.0, + "content": "domain audio books. In Proc. of ICASSP, pages 5206–5210. IEEE, 2015.", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 160, + 506, + 174 + ], + "spans": [ + { + "bbox": [ + 105, + 160, + 506, + 174 + ], + "score": 1.0, + "content": "D. S. Park, W. Chan, Y. Zhang, C.-C. Chiu, B. Zoph, E. D. Cubuk, and Q. V. Le. Specaugment: A", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 115, + 172, + 507, + 186 + ], + "spans": [ + { + "bbox": [ + 115, + 172, + 507, + 186 + ], + "score": 1.0, + "content": "simple data augmentation method for automatic speech recognition. In Proc. of Interspeech, 2019.", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 190, + 505, + 204 + ], + "spans": [ + { + "bbox": [ + 105, + 190, + 505, + 204 + ], + "score": 1.0, + "content": "D. S. Park, Y. Zhang, Y. Jia, W. Han, C.-C. Chiu, B. Li, Y. Wu, and Q. V. Le. Improved noisy student", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 114, + 201, + 397, + 214 + ], + "spans": [ + { + "bbox": [ + 114, + 201, + 397, + 214 + ], + "score": 1.0, + "content": "training for automatic speech recognition. Proc. of Interspeech, 2020.", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 219, + 406, + 233 + ], + "spans": [ + { + "bbox": [ + 105, + 219, + 406, + 233 + ], + "score": 1.0, + "content": "K. Park and J. Kim. g2pe. https://github.com/Kyubyong/g2p, 2019.", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 103, + 235, + 507, + 253 + ], + "spans": [ + { + "bbox": [ + 103, + 235, + 507, + 253 + ], + "score": 1.0, + "content": "L. Pepino, P. Riera, and L. Ferrer. Emotion recognition from speech using wav2vec 2.0 embeddings.", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 114, + 248, + 169, + 262 + ], + "spans": [ + { + "bbox": [ + 114, + 248, + 169, + 262 + ], + "score": 1.0, + "content": "arXiv, 2021.", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 268, + 505, + 280 + ], + "spans": [ + { + "bbox": [ + 105, + 268, + 505, + 280 + ], + "score": 1.0, + "content": "L. Polka and J. F. Werker. Developmental changes in perception of nonnative vowel contrasts. Journal", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 114, + 277, + 455, + 292 + ], + "spans": [ + { + "bbox": [ + 114, + 277, + 455, + 292 + ], + "score": 1.0, + "content": "of Experimental Psychology: Human perception and performance, 20(2):421, 1994.", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 296, + 506, + 310 + ], + "spans": [ + { + "bbox": [ + 105, + 296, + 506, + 310 + ], + "score": 1.0, + "content": "D. Povey, A. Ghoshal, G. Boulianne, L. Burget, O. Glembek, N. Goel, M. Hannemann, P. Motlicek,", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 113, + 306, + 507, + 322 + ], + "spans": [ + { + "bbox": [ + 113, + 306, + 507, + 322 + ], + "score": 1.0, + "content": "Y. Qian, P. Schwarz, J. Silovsky, G. Stemmer, and K. Vesely. The kaldi speech recognition toolkit.", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 115, + 318, + 217, + 332 + ], + "spans": [ + { + "bbox": [ + 115, + 318, + 217, + 332 + ], + "score": 1.0, + "content": "In Proc. of ASRU, 2011.", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 336, + 506, + 351 + ], + "spans": [ + { + "bbox": [ + 105, + 336, + 506, + 351 + ], + "score": 1.0, + "content": "V. Pratap, A. Hannun, Q. Xu, J. Cai, J. Kahn, G. Synnaeve, V. Liptchinsky, and R. Collobert.", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 114, + 347, + 473, + 363 + ], + "spans": [ + { + "bbox": [ + 114, + 347, + 473, + 363 + ], + "score": 1.0, + "content": "Wav2letter++: A fast open-source speech recognition system. In Proc. of ICASSP, 2019.", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 104, + 363, + 507, + 381 + ], + "spans": [ + { + "bbox": [ + 104, + 363, + 507, + 381 + ], + "score": 1.0, + "content": "V. Pratap, Q. Xu, A. Sriram, G. Synnaeve, and R. Collobert. Mls: A large-scale multilingual dataset", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 115, + 377, + 321, + 391 + ], + "spans": [ + { + "bbox": [ + 115, + 377, + 321, + 391 + ], + "score": 1.0, + "content": "for speech research. In Proc. of Interspeech, 2020.", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 104, + 394, + 506, + 410 + ], + "spans": [ + { + "bbox": [ + 104, + 394, + 506, + 410 + ], + "score": 1.0, + "content": "O. Rasanen, G. Doyle, and M. C. Frank. Unsupervised word discovery from speech using automatic", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 113, + 406, + 389, + 420 + ], + "spans": [ + { + "bbox": [ + 113, + 406, + 389, + 420 + ], + "score": 1.0, + "content": "segmentation into syllable-like units. In Proc. of Interspeech, 2015.", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 104, + 423, + 507, + 439 + ], + "spans": [ + { + "bbox": [ + 104, + 423, + 507, + 439 + ], + "score": 1.0, + "content": "M. Ravanelli, P. Brakel, M. Omologo, and Y. Bengio. Light gated recurrent units for speech", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 115, + 436, + 402, + 450 + ], + "spans": [ + { + "bbox": [ + 115, + 436, + 402, + 450 + ], + "score": 1.0, + "content": "recognition. IEEE Trans. on Emerging Topics in Comp. Intel., 2, 2018.", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 103, + 452, + 507, + 469 + ], + "spans": [ + { + "bbox": [ + 103, + 452, + 507, + 469 + ], + "score": 1.0, + "content": "M. Ravanelli, T. Parcollet, and Y. Bengio. The pytorch-kaldi speech recognition toolkit. Proc. of", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 115, + 466, + 178, + 478 + ], + "spans": [ + { + "bbox": [ + 115, + 466, + 178, + 478 + ], + "score": 1.0, + "content": "ICASSP, 2019.", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 483, + 506, + 497 + ], + "spans": [ + { + "bbox": [ + 105, + 483, + 506, + 497 + ], + "score": 1.0, + "content": "M. Rivière, A. Joulin, P.-E. Mazaré, and E. Dupoux. Unsupervised pretraining transfers well across", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 114, + 494, + 270, + 507 + ], + "spans": [ + { + "bbox": [ + 114, + 494, + 270, + 507 + ], + "score": 1.0, + "content": "languages. In Proc. of ICASSP, 2020.", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 513, + 506, + 527 + ], + "spans": [ + { + "bbox": [ + 105, + 513, + 506, + 527 + ], + "score": 1.0, + "content": "S. Schneider, A. Baevski, R. Collobert, and M. Auli. wav2vec: Unsupervised pre-training for speech", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 115, + 524, + 289, + 538 + ], + "spans": [ + { + "bbox": [ + 115, + 524, + 289, + 538 + ], + "score": 1.0, + "content": "recognition. In Proc. of Interspeech, 2019.", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 103, + 540, + 506, + 558 + ], + "spans": [ + { + "bbox": [ + 103, + 540, + 506, + 558 + ], + "score": 1.0, + "content": "N. Srivastava, G. Hinton, A. Krizhevsky, I. Sutskever, and R. Salakhutdinov. Dropout: a simple way", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 113, + 553, + 349, + 567 + ], + "spans": [ + { + "bbox": [ + 113, + 553, + 349, + 567 + ], + "score": 1.0, + "content": "to prevent neural networks from overfitting. JMLR, 2014.", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 104, + 569, + 507, + 586 + ], + "spans": [ + { + "bbox": [ + 104, + 569, + 507, + 586 + ], + "score": 1.0, + "content": "G. Synnaeve, Q. Xu, J. Kahn, T. Likhomanenko, E. Grave, V. Pratap, A. Sriram, V. Liptchinsky,", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 115, + 583, + 505, + 595 + ], + "spans": [ + { + "bbox": [ + 115, + 583, + 505, + 595 + ], + "score": 1.0, + "content": "and R. Collobert. End-to-end ASR: from Supervised to Semi-Supervised Learning with Modern", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 115, + 593, + 495, + 607 + ], + "spans": [ + { + "bbox": [ + 115, + 593, + 495, + 607 + ], + "score": 1.0, + "content": "Architectures. Proc. of ICML workshop on Self-supervision in Audio and Speech (SAS), 2020.", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 103, + 609, + 507, + 627 + ], + "spans": [ + { + "bbox": [ + 103, + 609, + 507, + 627 + ], + "score": 1.0, + "content": "M. Tachbelie, S. T. Abate, and L. Besacier. Using different acoustic, lexical and language modeling", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 115, + 623, + 478, + 637 + ], + "spans": [ + { + "bbox": [ + 115, + 623, + 478, + 637 + ], + "score": 1.0, + "content": "units for asr of an under-resourced language - amharic. Speech Communication, 56, 2014.", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 104, + 640, + 505, + 655 + ], + "spans": [ + { + "bbox": [ + 104, + 640, + 505, + 655 + ], + "score": 1.0, + "content": "Z. Tan, A. K. Sarkar, and N. Dehak. rvad: An unsupervised segment-based robust voice activity", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 114, + 652, + 378, + 664 + ], + "spans": [ + { + "bbox": [ + 114, + 652, + 378, + 664 + ], + "score": 1.0, + "content": "detection method. Computer speech & language, 59:1–21, 2020.", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 104, + 669, + 506, + 685 + ], + "spans": [ + { + "bbox": [ + 104, + 669, + 506, + 685 + ], + "score": 1.0, + "content": "A. van den Oord, Y. Li, and O. Vinyals. Representation learning with contrastive predictive coding.", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 114, + 680, + 202, + 695 + ], + "spans": [ + { + "bbox": [ + 114, + 680, + 202, + 695 + ], + "score": 1.0, + "content": "Proc. of NIPS, 2018.", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 699, + 506, + 713 + ], + "spans": [ + { + "bbox": [ + 105, + 699, + 506, + 713 + ], + "score": 1.0, + "content": "B. van Niekerk, L. Nortje, and H. Kamper. Vector-quantized neural networks for acoustic unit", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 115, + 711, + 406, + 724 + ], + "spans": [ + { + "bbox": [ + 115, + 711, + 406, + 724 + ], + "score": 1.0, + "content": "discovery in the zerospeech 2020 challenge. Proc. of Interspeech, 2020.", + "type": "text" + } + ], + "index": 44 + } + ], + "index": 22 + } + ], + "page_idx": 12, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 300, + 741, + 311, + 751 + ], + "lines": [ + { + "bbox": [ + 299, + 740, + 312, + 754 + ], + "spans": [ + { + "bbox": [ + 299, + 740, + 312, + 754 + ], + "score": 1.0, + "content": "13", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "list", + "bbox": [ + 104, + 58, + 507, + 730 + ], + "lines": [], + "index": 22, + "bbox_fs": [ + 103, + 72, + 507, + 724 + ], + "lines_deleted": true + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 104, + 68, + 507, + 448 + ], + "lines": [ + { + "bbox": [ + 105, + 72, + 505, + 87 + ], + "spans": [ + { + "bbox": [ + 105, + 72, + 505, + 87 + ], + "score": 1.0, + "content": "B. Varadarajan, S. Khudanpur, and E. Dupoux. Unsupervised learning of acoustic sub-word units. In", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 115, + 84, + 198, + 96 + ], + "spans": [ + { + "bbox": [ + 115, + 84, + 198, + 96 + ], + "score": 1.0, + "content": "Proc. of ACL, 2008.", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 106, + 102, + 506, + 115 + ], + "spans": [ + { + "bbox": [ + 106, + 102, + 506, + 115 + ], + "score": 1.0, + "content": "A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez, L. Kaiser, and I. Polosukhin.", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 116, + 114, + 316, + 127 + ], + "spans": [ + { + "bbox": [ + 116, + 114, + 316, + 127 + ], + "score": 1.0, + "content": "Attention is all you need. In Proc. of NIPS, 2017.", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 132, + 505, + 145 + ], + "spans": [ + { + "bbox": [ + 106, + 132, + 505, + 145 + ], + "score": 1.0, + "content": "C. Wang, A. Wu, J. Pino, A. Baevski, M. Auli, and A. Conneau. Large-scale self- and semi-supervised", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 115, + 143, + 295, + 155 + ], + "spans": [ + { + "bbox": [ + 115, + 143, + 295, + 155 + ], + "score": 1.0, + "content": "learning for speech translation. arXiv, 2021.", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 104, + 161, + 505, + 176 + ], + "spans": [ + { + "bbox": [ + 104, + 161, + 505, + 176 + ], + "score": 1.0, + "content": "J. F. Werker and R. C. Tees. Cross-language speech perception: Evidence for perceptual reorganization", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 116, + 173, + 439, + 186 + ], + "spans": [ + { + "bbox": [ + 116, + 173, + 439, + 186 + ], + "score": 1.0, + "content": "during the first year of life. Infant behavior and development, 7(1):49–63, 1984.", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 192, + 506, + 205 + ], + "spans": [ + { + "bbox": [ + 105, + 192, + 506, + 205 + ], + "score": 1.0, + "content": "H. Xu, K. Li, Y. Wang, J. Wang, S. Kang, X. Chen, D. Povey, and S. Khudanpur. Neural network", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 115, + 202, + 507, + 217 + ], + "spans": [ + { + "bbox": [ + 115, + 202, + 507, + 217 + ], + "score": 1.0, + "content": "language modeling with letter-based features and importance sampling. In Proc. of ICASSP, 2018.", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 106, + 221, + 506, + 234 + ], + "spans": [ + { + "bbox": [ + 106, + 221, + 506, + 234 + ], + "score": 1.0, + "content": "Q. Xu, A. Baevski, T. Likhomanenko, P. Tomasello, A. Conneau, R. Collobert, G. Synnaeve, and", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 115, + 232, + 507, + 247 + ], + "spans": [ + { + "bbox": [ + 115, + 232, + 507, + 247 + ], + "score": 1.0, + "content": "M. Auli. Self-training and pre-training are complementary for speech recognition. In Proc. of", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 116, + 244, + 182, + 256 + ], + "spans": [ + { + "bbox": [ + 116, + 244, + 182, + 256 + ], + "score": 1.0, + "content": "ICASSP, 2020a.", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 261, + 506, + 276 + ], + "spans": [ + { + "bbox": [ + 105, + 261, + 506, + 276 + ], + "score": 1.0, + "content": "Q. Xu, T. Likhomanenko, J. Kahn, A. Hannun, G. Synnaeve, and R. Collobert. Iterative pseudo-", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 115, + 273, + 362, + 286 + ], + "spans": [ + { + "bbox": [ + 115, + 273, + 362, + 286 + ], + "score": 1.0, + "content": "labeling for speech recognition. Proc. of Interspeech, 2020b.", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 291, + 505, + 306 + ], + "spans": [ + { + "bbox": [ + 105, + 291, + 505, + 306 + ], + "score": 1.0, + "content": "C.-K. Yeh, J. Chen, C. Yu, and D. Yu. Unsupervised speech recognition via segmental empirical", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 115, + 303, + 333, + 316 + ], + "spans": [ + { + "bbox": [ + 115, + 303, + 333, + 316 + ], + "score": 1.0, + "content": "output distribution matching. In Proc. of ICLR, 2019.", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 321, + 505, + 335 + ], + "spans": [ + { + "bbox": [ + 105, + 321, + 505, + 335 + ], + "score": 1.0, + "content": "N. Zeghidour, Q. Xu, V. Liptchinsky, N. Usunier, G. Synnaeve, and R. Collobert. Fully convolutional", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 115, + 333, + 318, + 345 + ], + "spans": [ + { + "bbox": [ + 115, + 333, + 318, + 345 + ], + "score": 1.0, + "content": "speech recognition. arXiv, abs/1812.06864, 2018.", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 351, + 505, + 366 + ], + "spans": [ + { + "bbox": [ + 105, + 351, + 505, + 366 + ], + "score": 1.0, + "content": "F. Zhang, Y. Wang, X. Zhang, C. Liu, Y. Saraf, and G. Zweig. Faster, simpler and more accurate", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 115, + 362, + 384, + 376 + ], + "spans": [ + { + "bbox": [ + 115, + 362, + 384, + 376 + ], + "score": 1.0, + "content": "hybrid asr systems using wordpieces. Proc. of Interspeech, 2020a.", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 380, + 505, + 395 + ], + "spans": [ + { + "bbox": [ + 105, + 380, + 505, + 395 + ], + "score": 1.0, + "content": "Y. Zhang and J. R. Glass. Unsupervised spoken keyword spotting via segmental dtw on gaussian", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 115, + 392, + 485, + 406 + ], + "spans": [ + { + "bbox": [ + 115, + 392, + 485, + 406 + ], + "score": 1.0, + "content": "posteriorgrams. IEEE Workshop on Automatic Speech Recognition & Understanding, 2009.", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 410, + 505, + 425 + ], + "spans": [ + { + "bbox": [ + 105, + 410, + 505, + 425 + ], + "score": 1.0, + "content": "Y. Zhang, J. Qin, D. S. Park, W. Han, C.-C. Chiu, R. Pang, Q. V. Le, and Y. Wu. Pushing the limits", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 115, + 421, + 507, + 437 + ], + "spans": [ + { + "bbox": [ + 115, + 421, + 507, + 437 + ], + "score": 1.0, + "content": "of semi-supervised learning for automatic speech recognition. Proc. of NeurIPS SAS Workshop,", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 115, + 433, + 147, + 445 + ], + "spans": [ + { + "bbox": [ + 115, + 433, + 147, + 445 + ], + "score": 1.0, + "content": "2020b.", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 12.5 + } + ], + "page_idx": 13, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 300, + 741, + 311, + 750 + ], + "lines": [ + { + "bbox": [ + 299, + 740, + 313, + 754 + ], + "spans": [ + { + "bbox": [ + 299, + 740, + 313, + 754 + ], + "score": 1.0, + "content": "14", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "list", + "bbox": [ + 104, + 68, + 507, + 448 + ], + "lines": [], + "index": 12.5, + "bbox_fs": [ + 104, + 72, + 507, + 445 + ], + "lines_deleted": true + } + ] + } + ], + "_backend": "pipeline", + "_version_name": "2.2.2" +} \ No newline at end of file