{
  "release_name": "Swarit Indic Talk",
  "release_version": "1.0.0",
  "artifact_format": "safetensors",
  "artifact_filename": "swarit-indic-talk-v1.safetensors",
  "hugging_face_repo_id": "Vectornotion/swarit-indic-talk",
  "artifact": {
    "bytes": 14344705651,
    "sha256": "374f556fd18c403b078bddaf6c6b2bb18fcac260b85d7f7a038cd0f947a281f6",
    "tensor_count": 5768,
    "verified_object_store_readback": true
  },
  "scope": "weights-only; combined namespaces for selected inference components",
  "selected_stack_manifest": "model/stack/selected_packages.json",
  "components": {
    "asr": {
      "selection": "Qwen3-ASR 1.7B; 345k continuation epoch 3 / cumulative step 5391; BF16",
      "license": "Apache-2.0",
      "upstream_model": "Qwen/Qwen3-ASR-1.7B"
    },
    "tts": {
      "selection": "Qwen3-TTS V2 epoch 12 / step 26264; BF16 talker/predictor, FP32 codec",
      "license": "Apache-2.0",
      "upstream_model": "Qwen/Qwen3-TTS-12Hz-0.6B-Base",
      "include": [
        "main/model.safetensors",
        "speech_tokenizer/model.safetensors",
        "voices/female.f32",
        "voices/male.f32"
      ],
      "exclude": [
        "speaker-embedding.f32",
        "reference audio",
        "speaker encoder"
      ]
    },
    "response": {
      "selection": "Gemma 4 E4B epoch 2; selected NF4 base linears, INT4 PLE and unmerged LoRA adapter",
      "license": "Apache-2.0",
      "upstream_model": "google/gemma-4-E4B-it"
    },
    "conversation_detector": {
      "selection": "wav2vec2 XLS-R 300M acoustic epoch 9 + policy head 29",
      "license": "Apache-2.0",
      "upstream_model": "facebook/wav2vec2-xls-r-300m"
    },
    "conversation_memory": {
      "selection": "XLM-R large XNLI pinned encoder plus selected adapter and state heads",
      "license": "MIT",
      "upstream_model": "joeddav/xlm-roberta-large-xnli",
      "upstream_revision": "b227ee8435ceadfa86dc1368a34254e2838bf242"
    }
  },
  "voice_attribution": {
    "source": "IISc SYSPIN / LIMMITS24",
    "license": "CC-BY-4.0",
    "copyright_notice": "© Indian Institute of Science (IISc), Bengaluru — SYSPIN / SPIRE Lab, LIMMITS corpus",
    "attribution": "SYSPIN / SPIRE Lab, Indian Institute of Science, Bengaluru — LIMMITS corpus",
    "modifications": "Selected recordings were edge-trimmed, resampled for speaker encoding, and converted to float32 speaker vectors; the female reference was also peak-normalized to -3 dBFS.",
    "source_disclaimer": "IISc does not endorse this release. Source material is provided without warranties; see CC BY 4.0 for the full warranty disclaimer and license terms.",
    "source_material_url": "https://vaani.iisc.ac.in/dataset/syspindataset",
    "license_url": "https://creativecommons.org/licenses/by/4.0/",
    "include_audio": false,
    "include_selected_embeddings": true,
    "female_embedding_sha256": "1511538fe1b0cdf9e3f1be4a45a3c48ebd4984b9c9c964e4ab8380c7c3b2f6bd",
    "male_embedding_sha256": "53bae3aaad0d17e0a61eab6c1b6de6917b79b43d486a5f97e74225cf7cb47074"
  },
  "excluded": [
    "inference and serving code",
    "continuous batching, scheduling, orchestration and deployment harness",
    "tokenizers and runtime configs",
    "training and evaluation datasets",
    "raw speaker reference audio",
    "optimizer state and training checkpoints",
    "legacy TTS speaker-embedding.f32 (not a currently selected voice)"
  ],
  "release_selection": "quantized-20261016"
}