Add SenseVoice-Small and Qwen3-ASR-0.6B

8a9b5fe verified 17 days ago

20.9 kB

[
 {
  "manifest": {
   "schema_version": 2,
   "serial": 6,
   "id": "tingmix-large-mlx-fp32-r75071a1",
   "family": "tingmix-large",
   "revision": "75071a1260fb21874f150df6ebd53716ac423f13",
   "backend": "mlx",
   "os": "macos",
   "arch": "aarch64",
   "precision": "fp32",
   "requirements": {
    "min_ram_bytes": 4294967296,
    "cpu_features": [],
    "apple_chip_min_generation": 1
   },
   "runtime": {
    "name": "mlx",
    "min_version": "0.32.4"
   },
   "base_model": {
    "id": "google/embeddinggemma-2",
    "revision": "914f7f89142e33e77833254d9c9b90c3cef7303b"
   },
   "adapter": {
    "id": "tc3oliver/TingMix-Large:encoder_delta.safetensors (merged into the tower weights, scale alpha/r = 2.0, fp32)",
    "revision": "75071a1260fb21874f150df6ebd53716ac423f13"
   },
   "tokenizer": {
    "id": "tc3oliver/TingMix-Large:units.json+tokenizer_bpe.model",
    "revision": "75071a1260fb21874f150df6ebd53716ac423f13"
   },
   "lm": {
    "id": "tc3oliver/TingMix-Large:lm/lmCN_cv_parl_ascend_ntuml.charlm.json.gz (alpha 0.35, beta 1.5, beam 16)",
    "revision": "75071a1260fb21874f150df6ebd53716ac423f13"
   },
   "files": [
    {
     "path": "tingmix-large-mlx-fp32.safetensors",
     "component": "model",
     "url": "https://huggingface.co/tc3oliver/TingMix-Desktop/resolve/a6d031854ac2fe234326ca1792dfb58fdc2821a4/tingmix-large-mlx-fp32-r75071a1/tingmix-large-mlx-fp32.safetensors",
     "size": 1308067132,
     "sha256": "9737153e4631bc773481ec458f34f242cc7800ccc6ae6ea1c509d895181fc7ec"
    },
    {
     "path": "config.json",
     "component": "meta",
     "url": "https://huggingface.co/tc3oliver/TingMix-Desktop/resolve/a6d031854ac2fe234326ca1792dfb58fdc2821a4/tingmix-large-mlx-fp32-r75071a1/config.json",
     "size": 660,
     "sha256": "eac5da8ce273202417906279698130b09c810111d75f82882ac323102ceb7431"
    },
    {
     "path": "tingmix.json",
     "component": "meta",
     "url": "https://huggingface.co/tc3oliver/TingMix-Desktop/resolve/a6d031854ac2fe234326ca1792dfb58fdc2821a4/tingmix-large-mlx-fp32-r75071a1/tingmix.json",
     "size": 412,
     "sha256": "5b818f0dcd7752539a449f49d834a684518e3feac006ceb929f8d6548d6beb6e"
    },
    {
     "path": "units.json",
     "component": "tokenizer",
     "url": "https://huggingface.co/tc3oliver/TingMix-Desktop/resolve/a6d031854ac2fe234326ca1792dfb58fdc2821a4/tingmix-large-mlx-fp32-r75071a1/units.json",
     "size": 35553,
     "sha256": "dbc1aba97abeeae49dc7556531a2f670d4dc5f39705b54ca28e84b8de9187756"
    },
    {
     "path": "tokenizer_bpe.model",
     "component": "tokenizer",
     "url": "https://huggingface.co/tc3oliver/TingMix-Desktop/resolve/a6d031854ac2fe234326ca1792dfb58fdc2821a4/tingmix-large-mlx-fp32-r75071a1/tokenizer_bpe.model",
     "size": 245401,
     "sha256": "959e7af76bb54482be709575e7ee3da49dce7f293f1e6e18cc99baaf2eeb0c7a"
    },
    {
     "path": "lm/lmCN_cv_parl_ascend_ntuml.charlm.json.gz",
     "component": "lm",
     "url": "https://huggingface.co/tc3oliver/TingMix-Desktop/resolve/a6d031854ac2fe234326ca1792dfb58fdc2821a4/tingmix-large-mlx-fp32-r75071a1/lm/lmCN_cv_parl_ascend_ntuml.charlm.json.gz",
     "size": 113932911,
     "sha256": "c8251c39fd003df0527f074287acd5094f9da476c3bcd0a5eb86a5d6d9cc5a84"
    }
   ],
   "licenses": [
    {
     "id": "tingmix-large-model",
     "name": "TingMix-Large model terms",
     "spdx": "Apache-2.0",
     "url": "https://huggingface.co/tc3oliver/TingMix-Large/blob/75071a1260fb21874f150df6ebd53716ac423f13/README.md",
     "applies_to": [
      "model",
      "meta",
      "tokenizer",
      "lm"
     ],
     "acceptance_required": true,
     "attribution": [
      "LibriSpeech (Panayotov et al. 2015)",
      "AMI (Carletta et al.)",
      "FLEURS (Conneau et al. 2022)",
      "People's Speech (Galvez et al. 2021)",
      "ASCEND (Lovenia et al., LREC 2022)",
      "VoxPopuli audio (c) European Union - Source: European Parliament",
      "Common Voice: speakers must not be identified"
     ],
     "notes": "Training data includes Taiwan Parliament speech (CC BY-NC 4.0), ASCEND (CC BY-SA 4.0) and the NTUML2021 train split (no stated licence). The model owner publishes these weights and the language model for use with TingMix Desktop and promises no commercial-use rights; see the data section of the model card."
    },
    {
     "id": "google-embeddinggemma-2",
     "name": "Google EmbeddingGemma 2 licence and prohibited use policy",
     "spdx": "Apache-2.0",
     "url": "https://ai.google.dev/gemma/prohibited_use_policy",
     "applies_to": [
      "model"
     ],
     "acceptance_required": true,
     "attribution": [
      "EmbeddingGemma 2, Google DeepMind, Apache-2.0 (https://huggingface.co/google/embeddinggemma-2)"
     ],
     "notes": "This package contains a modified copy of the google/embeddinggemma-2 audio tower (Apache-2.0): LoRA merged, layouts converted; see Provenance in the model card. Deployments must adhere to the Gemma Prohibited Use Policy: https://ai.google.dev/gemma/prohibited_use_policy"
    }
   ],
   "compat": {
    "status": "EXPERIMENTAL",
    "rtf_x1000": 15,
    "notes": "EXPERIMENTAL: greedy and beam+LM output equal the reference implementation on 8 synthetic speech clips. Measured on one machine (Apple M4 Max, MLX 0.32.4): warm decode 64 ms for 7.5 s of audio, peak worker memory 2.22 GiB. Real-speech accuracy is limited; see Limitations in the model card."
   },
   "provenance": {
    "modified": true,
    "sources": [
     {
      "id": "google/embeddinggemma-2",
      "revision": "914f7f89142e33e77833254d9c9b90c3cef7303b"
     },
     {
      "id": "tc3oliver/TingMix-Large",
      "revision": "75071a1260fb21874f150df6ebd53716ac423f13"
     }
    ],
    "converter": "experiments/macos/mlx/convert.py",
    "converter_commit": "0b32104e54dfed0ba6ce4a6f70bf0fa8d78b87c5",
    "notes": "audio_tower of the Google base (text/vision towers dropped) with the 96 LoRA pairs (r16, alpha32) merged in fp32 as W + 2.0*B@A, conv weights transposed to MLX layouts (OIHW->OHWI, depthwise CK->KC), Reader GRU repacked to MLX gate layout, all fp32 in one safetensors file (tensor metadata format tingmix-large-mlx). Environment: mlx 0.32.4, mlx-vlm 0.7.8, transformers 5.19.0 (experiments/macos/mlx/uv.lock)."
   },
   "created": "2026-10-17",
   "min_app_version": "0.1.0"
  },
  "key_id": "tingmix-release-1",
  "signature": "c2c528496926019bcbafce38ec05205fb957b485b5ce61e88371ec1263d027c4ecfc69e88dcd2d96edef507f807c3a4bd0bbc741cce9d8f625620c49ece6df09"
 },
 {
  "manifest": {
   "schema_version": 2,
   "serial": 3,
   "id": "tingmix-large-onnx-fp32-r75071a1",
   "family": "tingmix-large",
   "revision": "75071a1260fb21874f150df6ebd53716ac423f13",
   "backend": "onnx-cpu",
   "os": "windows",
   "arch": "x86_64",
   "precision": "fp32",
   "requirements": {
    "min_ram_bytes": 6442450944,
    "cpu_features": [],
    "apple_chip_min_generation": null
   },
   "runtime": {
    "name": "onnxruntime",
    "min_version": "1.31.0"
   },
   "base_model": {
    "id": "google/embeddinggemma-2",
    "revision": "914f7f89142e33e77833254d9c9b90c3cef7303b"
   },
   "adapter": {
    "id": "tc3oliver/TingMix-Large:encoder_delta.safetensors (merged into the tower weights, scale alpha/r = 2.0, fp32)",
    "revision": "75071a1260fb21874f150df6ebd53716ac423f13"
   },
   "tokenizer": {
    "id": "tc3oliver/TingMix-Large:units.json+tokenizer_bpe.model",
    "revision": "75071a1260fb21874f150df6ebd53716ac423f13"
   },
   "lm": {
    "id": "tc3oliver/TingMix-Large:lm/lmCN_cv_parl_ascend_ntuml.charlm.json.gz (alpha 0.35, beta 1.5, beam 16)",
    "revision": "75071a1260fb21874f150df6ebd53716ac423f13"
   },
   "files": [
    {
     "path": "tingmix-large.onnx",
     "component": "model",
     "url": "https://huggingface.co/tc3oliver/TingMix-Desktop/resolve/e667c3db29bb08deb41623a3a27cbcce5722d5b6/tingmix-large-onnx-fp32-r75071a1/tingmix-large.onnx",
     "size": 1310190728,
     "sha256": "0b9f89c8ca521263cc6e1ddd52985d52f48591cfd1be2c1441d5cf95e618bc47"
    },
    {
     "path": "config.json",
     "component": "meta",
     "url": "https://huggingface.co/tc3oliver/TingMix-Desktop/resolve/a6d031854ac2fe234326ca1792dfb58fdc2821a4/tingmix-large-mlx-fp32-r75071a1/config.json",
     "size": 660,
     "sha256": "eac5da8ce273202417906279698130b09c810111d75f82882ac323102ceb7431"
    },
    {
     "path": "tingmix.json",
     "component": "meta",
     "url": "https://huggingface.co/tc3oliver/TingMix-Desktop/resolve/a6d031854ac2fe234326ca1792dfb58fdc2821a4/tingmix-large-mlx-fp32-r75071a1/tingmix.json",
     "size": 412,
     "sha256": "5b818f0dcd7752539a449f49d834a684518e3feac006ceb929f8d6548d6beb6e"
    },
    {
     "path": "units.json",
     "component": "tokenizer",
     "url": "https://huggingface.co/tc3oliver/TingMix-Desktop/resolve/a6d031854ac2fe234326ca1792dfb58fdc2821a4/tingmix-large-mlx-fp32-r75071a1/units.json",
     "size": 35553,
     "sha256": "dbc1aba97abeeae49dc7556531a2f670d4dc5f39705b54ca28e84b8de9187756"
    },
    {
     "path": "tokenizer_bpe.model",
     "component": "tokenizer",
     "url": "https://huggingface.co/tc3oliver/TingMix-Desktop/resolve/a6d031854ac2fe234326ca1792dfb58fdc2821a4/tingmix-large-mlx-fp32-r75071a1/tokenizer_bpe.model",
     "size": 245401,
     "sha256": "959e7af76bb54482be709575e7ee3da49dce7f293f1e6e18cc99baaf2eeb0c7a"
    },
    {
     "path": "lm/lmCN_cv_parl_ascend_ntuml.charlm.json.gz",
     "component": "lm",
     "url": "https://huggingface.co/tc3oliver/TingMix-Desktop/resolve/a6d031854ac2fe234326ca1792dfb58fdc2821a4/tingmix-large-mlx-fp32-r75071a1/lm/lmCN_cv_parl_ascend_ntuml.charlm.json.gz",
     "size": 113932911,
     "sha256": "c8251c39fd003df0527f074287acd5094f9da476c3bcd0a5eb86a5d6d9cc5a84"
    }
   ],
   "licenses": [
    {
     "id": "tingmix-large-model",
     "name": "TingMix-Large model terms",
     "spdx": "Apache-2.0",
     "url": "https://huggingface.co/tc3oliver/TingMix-Large/blob/75071a1260fb21874f150df6ebd53716ac423f13/README.md",
     "applies_to": [
      "model",
      "meta",
      "tokenizer",
      "lm"
     ],
     "acceptance_required": true,
     "attribution": [
      "LibriSpeech (Panayotov et al. 2015)",
      "AMI (Carletta et al.)",
      "FLEURS (Conneau et al. 2022)",
      "People's Speech (Galvez et al. 2021)",
      "ASCEND (Lovenia et al., LREC 2022)",
      "VoxPopuli audio (c) European Union - Source: European Parliament",
      "Common Voice: speakers must not be identified"
     ],
     "notes": "Training data includes Taiwan Parliament speech (CC BY-NC 4.0), ASCEND (CC BY-SA 4.0) and the NTUML2021 train split (no stated licence). The model owner publishes these weights and the language model for use with TingMix Desktop and promises no commercial-use rights; see the data section of the model card."
    },
    {
     "id": "google-embeddinggemma-2",
     "name": "Google EmbeddingGemma 2 licence and prohibited use policy",
     "spdx": "Apache-2.0",
     "url": "https://ai.google.dev/gemma/prohibited_use_policy",
     "applies_to": [
      "model"
     ],
     "acceptance_required": true,
     "attribution": [
      "EmbeddingGemma 2, Google DeepMind, Apache-2.0 (https://huggingface.co/google/embeddinggemma-2)"
     ],
     "notes": "This package contains a modified copy of the google/embeddinggemma-2 audio tower (Apache-2.0): LoRA merged, layouts converted; see Provenance in the model card. Deployments must adhere to the Gemma Prohibited Use Policy: https://ai.google.dev/gemma/prohibited_use_policy"
    }
   ],
   "compat": {
    "status": "EXPERIMENTAL",
    "rtf_x1000": 28,
    "notes": "EXPERIMENTAL: greedy and beam+LM output equal the reference implementation on 8 synthetic speech clips. Measured on one machine (Windows 11, AMD Ryzen 7 9700X, ONNX Runtime 1.31.0 CPU): about 0.8 s for a 29.5 s window. Real-speech accuracy is limited; see Limitations in the model card."
   },
   "provenance": {
    "modified": true,
    "sources": [
     {
      "id": "google/embeddinggemma-2",
      "revision": "914f7f89142e33e77833254d9c9b90c3cef7303b"
     },
     {
      "id": "tc3oliver/TingMix-Large",
      "revision": "75071a1260fb21874f150df6ebd53716ac423f13"
     }
    ],
    "converter": "experiments/windows/large-onnx/export.py",
    "converter_commit": "e322b4c9b7f8293beab1cee7ab85c9a3f781cda2",
    "notes": "Audio tower of google/embeddinggemma-2 (text and vision towers dropped) with the 96 LoRA pairs of TingMix-Large merged in fp32 (W + 2.0*B@A), followed by the GRU-CTC reader, exported to ONNX (opset 17) with a dynamic frame axis. Numerically equivalent rewrites: local attention without unfold, in-graph attention mask, RMSNorm prescale, clamp in place of ReLU."
   },
   "created": "2026-10-17",
   "min_app_version": "0.1.0"
  },
  "key_id": "tingmix-release-1",
  "signature": "20c8b1906be85c414e26343cf95a848e01dfbac3bb5b75e5a62a5016f47f9512dcf6b06ac619d948a12aa39249adfdae84741a7ffa976fa9f94a5a098cfcbd02"
 },
 {
  "manifest": {
   "schema_version": 2,
   "serial": 1,
   "id": "qwen3-asr-0.6b-onnx-int4-r4fc24a1",
   "family": "qwen3-asr",
   "revision": "4fc24a1402e74db89c4d2ef256875e71680128c4",
   "backend": "onnx-cpu",
   "os": "any",
   "arch": "any",
   "precision": "int4",
   "requirements": {
    "min_ram_bytes": 8589934592,
    "cpu_features": [],
    "apple_chip_min_generation": null
   },
   "runtime": {
    "name": "onnxruntime",
    "min_version": "1.31.0"
   },
   "base_model": {
    "id": "Qwen/Qwen3-ASR-0.6B",
    "revision": "5eb144179a02acc5e5ba31e748d22b0cf3e303b0"
   },
   "adapter": null,
   "tokenizer": {
    "id": "andrewleech/qwen3-asr-0.6b-onnx:vocab.json+added_tokens.json",
    "revision": "4fc24a1402e74db89c4d2ef256875e71680128c4"
   },
   "lm": null,
   "files": [
    {
     "path": "encoder.int4.onnx",
     "component": "model",
     "url": "https://huggingface.co/andrewleech/qwen3-asr-0.6b-onnx/resolve/4fc24a1402e74db89c4d2ef256875e71680128c4/encoder.int4.onnx",
     "size": 745762694,
     "sha256": "3c027f880f677615de85e1f6934906e1d5d77624b724096d33accef99c753eed"
    },
    {
     "path": "decoder_init.int4.onnx",
     "component": "model",
     "url": "https://huggingface.co/andrewleech/qwen3-asr-0.6b-onnx/resolve/4fc24a1402e74db89c4d2ef256875e71680128c4/decoder_init.int4.onnx",
     "size": 355390,
     "sha256": "6d54633bbb9e6b1ecf372f218d04adb3bf6f2bfc442ce55c988526d532896faa"
    },
    {
     "path": "decoder_step.int4.onnx",
     "component": "model",
     "url": "https://huggingface.co/andrewleech/qwen3-asr-0.6b-onnx/resolve/4fc24a1402e74db89c4d2ef256875e71680128c4/decoder_step.int4.onnx",
     "size": 354888,
     "sha256": "7eaa4f31d5eb6ae2937ff45222c32f692f2d5ac29b46cbae9046da332a9f317d"
    },
    {
     "path": "decoder_weights.int4.data",
     "component": "model",
     "url": "https://huggingface.co/andrewleech/qwen3-asr-0.6b-onnx/resolve/4fc24a1402e74db89c4d2ef256875e71680128c4/decoder_weights.int4.data",
     "size": 962460672,
     "sha256": "d68cd1c0695a7ba42651d06b1bc1158e2e58af5f9adbb6dba874fbbb8a4f22cf"
    },
    {
     "path": "embed_tokens.bin",
     "component": "model",
     "url": "https://huggingface.co/andrewleech/qwen3-asr-0.6b-onnx/resolve/4fc24a1402e74db89c4d2ef256875e71680128c4/embed_tokens.bin",
     "size": 311164928,
     "sha256": "e80150119fa5f7e56e85aed64c3a02d5c78eb7a37cfdcb973d0987316f15bee2"
    },
    {
     "path": "vocab.json",
     "component": "tokenizer",
     "url": "https://huggingface.co/andrewleech/qwen3-asr-0.6b-onnx/resolve/4fc24a1402e74db89c4d2ef256875e71680128c4/vocab.json",
     "size": 2776833,
     "sha256": "ca10d7e9fb3ed18575dd1e277a2579c16d108e32f27439684afa0e10b1440910"
    },
    {
     "path": "added_tokens.json",
     "component": "tokenizer",
     "url": "https://huggingface.co/andrewleech/qwen3-asr-0.6b-onnx/resolve/4fc24a1402e74db89c4d2ef256875e71680128c4/added_tokens.json",
     "size": 1566,
     "sha256": "de40784677cbd1843cabe5fbee078c7e042cd0b62155f0810af5a13842e5722a"
    },
    {
     "path": "config.json",
     "component": "meta",
     "url": "https://huggingface.co/andrewleech/qwen3-asr-0.6b-onnx/resolve/4fc24a1402e74db89c4d2ef256875e71680128c4/config.json",
     "size": 1250,
     "sha256": "df31c4689abe9d782366fffd2454b546291d0205d082b3fd01b99fb76a45b11f"
    }
   ],
   "licenses": [
    {
     "id": "qwen3-asr-apache-2.0",
     "name": "Qwen3-ASR-0.6B (Apache-2.0)",
     "spdx": "Apache-2.0",
     "url": "https://huggingface.co/Qwen/Qwen3-ASR-0.6B",
     "applies_to": [
      "model",
      "tokenizer",
      "meta"
     ],
     "acceptance_required": false,
     "attribution": [
      "Qwen3-ASR, Qwen Team, Alibaba Group (Qwen/Qwen3-ASR-0.6B, Apache-2.0)",
      "ONNX export and int4 quantisation by Andrew Leech (andrewleech/qwen3-asr-onnx, Apache-2.0)"
     ],
     "notes": "The ONNX files are a modified form of the Qwen weights (graph export, int4 RTN quantisation of the decoder, FP16 embedding table) made by andrewleech/qwen3-asr-onnx; they are downloaded from that repository unchanged."
    }
   ],
   "compat": {
    "status": "EXPERIMENTAL",
    "rtf_x1000": 153,
    "notes": "EXPERIMENTAL: token ids equal the export's own Python ONNX path on 7 clips (macOS and Windows), with the first token fixed to 'language'. On 55 real zh-TW clips MER 0.094 against provisional references (SenseVoice-Small 0.102, TingMix-Large 0.212) and 0.091 on the live captioning path (0.103, 0.244); English word error 0.22 live. About 0.15 x real time on 4 CPU threads of an Apple M4 Max, 3.4 GB peak memory (benchmarks/qwen3-asr). Output is converted from Simplified to Traditional characters (OpenCC s2tw)."
   },
   "provenance": null,
   "created": "2026-10-20",
   "min_app_version": "0.2.0"
  },
  "key_id": "tingmix-release-1",
  "signature": "56237ef6c97a321ef12ced0ebee203b89dafe3b609d8103fd9a9704d5f19a1729b49f1a38f8ab0f3318311593b0a529be2a6fac7c7473b997e9626e437a3bd09"
 },
 {
  "manifest": {
   "schema_version": 2,
   "serial": 1,
   "id": "sensevoice-small-onnx-int8-r2365bae",
   "family": "sensevoice-small",
   "revision": "2365baeacb507f821a0c8120fcee3d484dba7a07",
   "backend": "onnx-cpu",
   "os": "any",
   "arch": "any",
   "precision": "int8",
   "requirements": {
    "min_ram_bytes": 2147483648,
    "cpu_features": [],
    "apple_chip_min_generation": null
   },
   "runtime": {
    "name": "onnxruntime",
    "min_version": "1.31.0"
   },
   "base_model": null,
   "adapter": null,
   "tokenizer": {
    "id": "csukuangfj/sherpa-onnx-sense-voice-zh-en-ja-ko-yue-2024-07-17:tokens.txt",
    "revision": "2365baeacb507f821a0c8120fcee3d484dba7a07"
   },
   "lm": null,
   "files": [
    {
     "path": "model.int8.onnx",
     "component": "model",
     "url": "https://huggingface.co/csukuangfj/sherpa-onnx-sense-voice-zh-en-ja-ko-yue-2024-07-17/resolve/2365baeacb507f821a0c8120fcee3d484dba7a07/model.int8.onnx",
     "size": 239233841,
     "sha256": "c71f0ce00bec95b07744e116345e33d8cbbe08cef896382cf907bf4b51a2cd51"
    },
    {
     "path": "tokens.txt",
     "component": "tokenizer",
     "url": "https://huggingface.co/csukuangfj/sherpa-onnx-sense-voice-zh-en-ja-ko-yue-2024-07-17/resolve/2365baeacb507f821a0c8120fcee3d484dba7a07/tokens.txt",
     "size": 315894,
     "sha256": "f449eb28dc567533d7fa59be34e2abca8784f771850c78a47fb731a31429a1dc"
    }
   ],
   "licenses": [
    {
     "id": "funasr-model-license-1.1",
     "name": "FunASR Model Open Source License Agreement 1.1 (SenseVoice-Small)",
     "spdx": null,
     "url": "https://github.com/modelscope/FunASR/blob/main/MODEL_LICENSE",
     "applies_to": [
      "model",
      "tokenizer"
     ],
     "acceptance_required": true,
     "attribution": [
      "SenseVoice-Small, Tongyi Lab, Alibaba Group (FunAudioLLM/SenseVoiceSmall)",
      "ONNX export by k2-fsa / sherpa-onnx (Xiaomi Corp.)"
     ],
     "notes": "Free to use, copy, modify and share with attribution and the model names kept. Provided for reference and learning purposes without warranty; the terms may be revised by the publisher. The ONNX export was made by sherpa-onnx from iic/SenseVoiceSmall on ModelScope (export-onnx.py in the export repository); the exact upstream revision is not recorded there."
    }
   ],
   "compat": {
    "status": "EXPERIMENTAL",
    "rtf_x1000": 11,
    "notes": "EXPERIMENTAL: text equals sherpa-onnx 1.13.8 on 10 clips (one differs by one comma). On 55 real zh-TW clips MER 0.102 against provisional references (TingMix-Large 0.212), 0.116 against one video's subtitles (0.277). About 0.011 x real time on 4 CPU threads of an Apple M4 Max, 0.5 GB peak memory (benchmarks/sensevoice). Output is converted from Simplified to Traditional characters (OpenCC s2tw)."
   },
   "provenance": null,
   "created": "2026-10-19",
   "min_app_version": "0.2.0"
  },
  "key_id": "tingmix-release-1",
  "signature": "d6f04ac4047edd94965578f8d24c5edaee360d5b1b55de844b04931045eb7d0489a0a7b57a432b4b119a68e9015c8c727ee331787fffc75b374f9adabe62ea08"
 }
]

	[
	{
	"manifest": {
	"schema_version": 2,
	"serial": 6,
	"id": "tingmix-large-mlx-fp32-r75071a1",
	"family": "tingmix-large",
	"revision": "75071a1260fb21874f150df6ebd53716ac423f13",
	"backend": "mlx",
	"os": "macos",
	"arch": "aarch64",
	"precision": "fp32",
	"requirements": {
	"min_ram_bytes": 4294967296,
	"cpu_features": [],
	"apple_chip_min_generation": 1
	},
	"runtime": {
	"name": "mlx",
	"min_version": "0.32.4"
	},
	"base_model": {
	"id": "google/embeddinggemma-2",
	"revision": "914f7f89142e33e77833254d9c9b90c3cef7303b"
	},
	"adapter": {
	"id": "tc3oliver/TingMix-Large:encoder_delta.safetensors (merged into the tower weights, scale alpha/r = 2.0, fp32)",
	"revision": "75071a1260fb21874f150df6ebd53716ac423f13"
	},
	"tokenizer": {
	"id": "tc3oliver/TingMix-Large:units.json+tokenizer_bpe.model",
	"revision": "75071a1260fb21874f150df6ebd53716ac423f13"
	},
	"lm": {
	"id": "tc3oliver/TingMix-Large:lm/lmCN_cv_parl_ascend_ntuml.charlm.json.gz (alpha 0.35, beta 1.5, beam 16)",
	"revision": "75071a1260fb21874f150df6ebd53716ac423f13"
	},
	"files": [
	{
	"path": "tingmix-large-mlx-fp32.safetensors",
	"component": "model",
	"url": "https://huggingface.co/tc3oliver/TingMix-Desktop/resolve/a6d031854ac2fe234326ca1792dfb58fdc2821a4/tingmix-large-mlx-fp32-r75071a1/tingmix-large-mlx-fp32.safetensors",
	"size": 1308067132,
	"sha256": "9737153e4631bc773481ec458f34f242cc7800ccc6ae6ea1c509d895181fc7ec"
	},
	{
	"path": "config.json",
	"component": "meta",
	"url": "https://huggingface.co/tc3oliver/TingMix-Desktop/resolve/a6d031854ac2fe234326ca1792dfb58fdc2821a4/tingmix-large-mlx-fp32-r75071a1/config.json",
	"size": 660,
	"sha256": "eac5da8ce273202417906279698130b09c810111d75f82882ac323102ceb7431"
	},
	{
	"path": "tingmix.json",
	"component": "meta",
	"url": "https://huggingface.co/tc3oliver/TingMix-Desktop/resolve/a6d031854ac2fe234326ca1792dfb58fdc2821a4/tingmix-large-mlx-fp32-r75071a1/tingmix.json",
	"size": 412,
	"sha256": "5b818f0dcd7752539a449f49d834a684518e3feac006ceb929f8d6548d6beb6e"
	},
	{
	"path": "units.json",
	"component": "tokenizer",
	"url": "https://huggingface.co/tc3oliver/TingMix-Desktop/resolve/a6d031854ac2fe234326ca1792dfb58fdc2821a4/tingmix-large-mlx-fp32-r75071a1/units.json",
	"size": 35553,
	"sha256": "dbc1aba97abeeae49dc7556531a2f670d4dc5f39705b54ca28e84b8de9187756"
	},
	{
	"path": "tokenizer_bpe.model",
	"component": "tokenizer",
	"url": "https://huggingface.co/tc3oliver/TingMix-Desktop/resolve/a6d031854ac2fe234326ca1792dfb58fdc2821a4/tingmix-large-mlx-fp32-r75071a1/tokenizer_bpe.model",
	"size": 245401,
	"sha256": "959e7af76bb54482be709575e7ee3da49dce7f293f1e6e18cc99baaf2eeb0c7a"
	},
	{
	"path": "lm/lmCN_cv_parl_ascend_ntuml.charlm.json.gz",
	"component": "lm",
	"url": "https://huggingface.co/tc3oliver/TingMix-Desktop/resolve/a6d031854ac2fe234326ca1792dfb58fdc2821a4/tingmix-large-mlx-fp32-r75071a1/lm/lmCN_cv_parl_ascend_ntuml.charlm.json.gz",
	"size": 113932911,
	"sha256": "c8251c39fd003df0527f074287acd5094f9da476c3bcd0a5eb86a5d6d9cc5a84"
	}
	],
	"licenses": [
	{
	"id": "tingmix-large-model",
	"name": "TingMix-Large model terms",
	"spdx": "Apache-2.0",
	"url": "https://huggingface.co/tc3oliver/TingMix-Large/blob/75071a1260fb21874f150df6ebd53716ac423f13/README.md",
	"applies_to": [
	"model",
	"meta",
	"tokenizer",
	"lm"
	],
	"acceptance_required": true,
	"attribution": [
	"LibriSpeech (Panayotov et al. 2015)",
	"AMI (Carletta et al.)",
	"FLEURS (Conneau et al. 2022)",
	"People's Speech (Galvez et al. 2021)",
	"ASCEND (Lovenia et al., LREC 2022)",
	"VoxPopuli audio (c) European Union - Source: European Parliament",
	"Common Voice: speakers must not be identified"
	],
	"notes": "Training data includes Taiwan Parliament speech (CC BY-NC 4.0), ASCEND (CC BY-SA 4.0) and the NTUML2021 train split (no stated licence). The model owner publishes these weights and the language model for use with TingMix Desktop and promises no commercial-use rights; see the data section of the model card."
	},
	{
	"id": "google-embeddinggemma-2",
	"name": "Google EmbeddingGemma 2 licence and prohibited use policy",
	"spdx": "Apache-2.0",
	"url": "https://ai.google.dev/gemma/prohibited_use_policy",
	"applies_to": [
	"model"
	],
	"acceptance_required": true,
	"attribution": [
	"EmbeddingGemma 2, Google DeepMind, Apache-2.0 (https://huggingface.co/google/embeddinggemma-2)"
	],
	"notes": "This package contains a modified copy of the google/embeddinggemma-2 audio tower (Apache-2.0): LoRA merged, layouts converted; see Provenance in the model card. Deployments must adhere to the Gemma Prohibited Use Policy: https://ai.google.dev/gemma/prohibited_use_policy"
	}
	],
	"compat": {
	"status": "EXPERIMENTAL",
	"rtf_x1000": 15,
	"notes": "EXPERIMENTAL: greedy and beam+LM output equal the reference implementation on 8 synthetic speech clips. Measured on one machine (Apple M4 Max, MLX 0.32.4): warm decode 64 ms for 7.5 s of audio, peak worker memory 2.22 GiB. Real-speech accuracy is limited; see Limitations in the model card."
	},
	"provenance": {
	"modified": true,
	"sources": [
	{
	"id": "google/embeddinggemma-2",
	"revision": "914f7f89142e33e77833254d9c9b90c3cef7303b"
	},
	{
	"id": "tc3oliver/TingMix-Large",
	"revision": "75071a1260fb21874f150df6ebd53716ac423f13"
	}
	],
	"converter": "experiments/macos/mlx/convert.py",
	"converter_commit": "0b32104e54dfed0ba6ce4a6f70bf0fa8d78b87c5",
	"notes": "audio_tower of the Google base (text/vision towers dropped) with the 96 LoRA pairs (r16, alpha32) merged in fp32 as W + 2.0*B@A, conv weights transposed to MLX layouts (OIHW->OHWI, depthwise CK->KC), Reader GRU repacked to MLX gate layout, all fp32 in one safetensors file (tensor metadata format tingmix-large-mlx). Environment: mlx 0.32.4, mlx-vlm 0.7.8, transformers 5.19.0 (experiments/macos/mlx/uv.lock)."
	},
	"created": "2026-10-17",
	"min_app_version": "0.1.0"
	},
	"key_id": "tingmix-release-1",
	"signature": "c2c528496926019bcbafce38ec05205fb957b485b5ce61e88371ec1263d027c4ecfc69e88dcd2d96edef507f807c3a4bd0bbc741cce9d8f625620c49ece6df09"
	},
	{
	"manifest": {
	"schema_version": 2,
	"serial": 3,
	"id": "tingmix-large-onnx-fp32-r75071a1",
	"family": "tingmix-large",
	"revision": "75071a1260fb21874f150df6ebd53716ac423f13",
	"backend": "onnx-cpu",
	"os": "windows",
	"arch": "x86_64",
	"precision": "fp32",
	"requirements": {
	"min_ram_bytes": 6442450944,
	"cpu_features": [],
	"apple_chip_min_generation": null
	},
	"runtime": {
	"name": "onnxruntime",
	"min_version": "1.31.0"
	},
	"base_model": {
	"id": "google/embeddinggemma-2",
	"revision": "914f7f89142e33e77833254d9c9b90c3cef7303b"
	},
	"adapter": {
	"id": "tc3oliver/TingMix-Large:encoder_delta.safetensors (merged into the tower weights, scale alpha/r = 2.0, fp32)",
	"revision": "75071a1260fb21874f150df6ebd53716ac423f13"
	},
	"tokenizer": {
	"id": "tc3oliver/TingMix-Large:units.json+tokenizer_bpe.model",
	"revision": "75071a1260fb21874f150df6ebd53716ac423f13"
	},
	"lm": {
	"id": "tc3oliver/TingMix-Large:lm/lmCN_cv_parl_ascend_ntuml.charlm.json.gz (alpha 0.35, beta 1.5, beam 16)",
	"revision": "75071a1260fb21874f150df6ebd53716ac423f13"
	},
	"files": [
	{
	"path": "tingmix-large.onnx",
	"component": "model",
	"url": "https://huggingface.co/tc3oliver/TingMix-Desktop/resolve/e667c3db29bb08deb41623a3a27cbcce5722d5b6/tingmix-large-onnx-fp32-r75071a1/tingmix-large.onnx",
	"size": 1310190728,
	"sha256": "0b9f89c8ca521263cc6e1ddd52985d52f48591cfd1be2c1441d5cf95e618bc47"
	},
	{
	"path": "config.json",
	"component": "meta",
	"url": "https://huggingface.co/tc3oliver/TingMix-Desktop/resolve/a6d031854ac2fe234326ca1792dfb58fdc2821a4/tingmix-large-mlx-fp32-r75071a1/config.json",
	"size": 660,
	"sha256": "eac5da8ce273202417906279698130b09c810111d75f82882ac323102ceb7431"
	},
	{
	"path": "tingmix.json",
	"component": "meta",
	"url": "https://huggingface.co/tc3oliver/TingMix-Desktop/resolve/a6d031854ac2fe234326ca1792dfb58fdc2821a4/tingmix-large-mlx-fp32-r75071a1/tingmix.json",
	"size": 412,
	"sha256": "5b818f0dcd7752539a449f49d834a684518e3feac006ceb929f8d6548d6beb6e"
	},
	{
	"path": "units.json",
	"component": "tokenizer",
	"url": "https://huggingface.co/tc3oliver/TingMix-Desktop/resolve/a6d031854ac2fe234326ca1792dfb58fdc2821a4/tingmix-large-mlx-fp32-r75071a1/units.json",
	"size": 35553,
	"sha256": "dbc1aba97abeeae49dc7556531a2f670d4dc5f39705b54ca28e84b8de9187756"
	},
	{
	"path": "tokenizer_bpe.model",
	"component": "tokenizer",
	"url": "https://huggingface.co/tc3oliver/TingMix-Desktop/resolve/a6d031854ac2fe234326ca1792dfb58fdc2821a4/tingmix-large-mlx-fp32-r75071a1/tokenizer_bpe.model",
	"size": 245401,
	"sha256": "959e7af76bb54482be709575e7ee3da49dce7f293f1e6e18cc99baaf2eeb0c7a"
	},
	{
	"path": "lm/lmCN_cv_parl_ascend_ntuml.charlm.json.gz",
	"component": "lm",
	"url": "https://huggingface.co/tc3oliver/TingMix-Desktop/resolve/a6d031854ac2fe234326ca1792dfb58fdc2821a4/tingmix-large-mlx-fp32-r75071a1/lm/lmCN_cv_parl_ascend_ntuml.charlm.json.gz",
	"size": 113932911,
	"sha256": "c8251c39fd003df0527f074287acd5094f9da476c3bcd0a5eb86a5d6d9cc5a84"
	}
	],
	"licenses": [
	{
	"id": "tingmix-large-model",
	"name": "TingMix-Large model terms",
	"spdx": "Apache-2.0",
	"url": "https://huggingface.co/tc3oliver/TingMix-Large/blob/75071a1260fb21874f150df6ebd53716ac423f13/README.md",
	"applies_to": [
	"model",
	"meta",
	"tokenizer",
	"lm"
	],
	"acceptance_required": true,
	"attribution": [
	"LibriSpeech (Panayotov et al. 2015)",
	"AMI (Carletta et al.)",
	"FLEURS (Conneau et al. 2022)",
	"People's Speech (Galvez et al. 2021)",
	"ASCEND (Lovenia et al., LREC 2022)",
	"VoxPopuli audio (c) European Union - Source: European Parliament",
	"Common Voice: speakers must not be identified"
	],
	"notes": "Training data includes Taiwan Parliament speech (CC BY-NC 4.0), ASCEND (CC BY-SA 4.0) and the NTUML2021 train split (no stated licence). The model owner publishes these weights and the language model for use with TingMix Desktop and promises no commercial-use rights; see the data section of the model card."
	},
	{
	"id": "google-embeddinggemma-2",
	"name": "Google EmbeddingGemma 2 licence and prohibited use policy",
	"spdx": "Apache-2.0",
	"url": "https://ai.google.dev/gemma/prohibited_use_policy",
	"applies_to": [
	"model"
	],
	"acceptance_required": true,
	"attribution": [
	"EmbeddingGemma 2, Google DeepMind, Apache-2.0 (https://huggingface.co/google/embeddinggemma-2)"
	],
	"notes": "This package contains a modified copy of the google/embeddinggemma-2 audio tower (Apache-2.0): LoRA merged, layouts converted; see Provenance in the model card. Deployments must adhere to the Gemma Prohibited Use Policy: https://ai.google.dev/gemma/prohibited_use_policy"
	}
	],
	"compat": {
	"status": "EXPERIMENTAL",
	"rtf_x1000": 28,
	"notes": "EXPERIMENTAL: greedy and beam+LM output equal the reference implementation on 8 synthetic speech clips. Measured on one machine (Windows 11, AMD Ryzen 7 9700X, ONNX Runtime 1.31.0 CPU): about 0.8 s for a 29.5 s window. Real-speech accuracy is limited; see Limitations in the model card."
	},
	"provenance": {
	"modified": true,
	"sources": [
	{
	"id": "google/embeddinggemma-2",
	"revision": "914f7f89142e33e77833254d9c9b90c3cef7303b"
	},
	{
	"id": "tc3oliver/TingMix-Large",
	"revision": "75071a1260fb21874f150df6ebd53716ac423f13"
	}
	],
	"converter": "experiments/windows/large-onnx/export.py",
	"converter_commit": "e322b4c9b7f8293beab1cee7ab85c9a3f781cda2",
	"notes": "Audio tower of google/embeddinggemma-2 (text and vision towers dropped) with the 96 LoRA pairs of TingMix-Large merged in fp32 (W + 2.0*B@A), followed by the GRU-CTC reader, exported to ONNX (opset 17) with a dynamic frame axis. Numerically equivalent rewrites: local attention without unfold, in-graph attention mask, RMSNorm prescale, clamp in place of ReLU."
	},
	"created": "2026-10-17",
	"min_app_version": "0.1.0"
	},
	"key_id": "tingmix-release-1",
	"signature": "20c8b1906be85c414e26343cf95a848e01dfbac3bb5b75e5a62a5016f47f9512dcf6b06ac619d948a12aa39249adfdae84741a7ffa976fa9f94a5a098cfcbd02"
	},
	{
	"manifest": {
	"schema_version": 2,
	"serial": 1,
	"id": "qwen3-asr-0.6b-onnx-int4-r4fc24a1",
	"family": "qwen3-asr",
	"revision": "4fc24a1402e74db89c4d2ef256875e71680128c4",
	"backend": "onnx-cpu",
	"os": "any",
	"arch": "any",
	"precision": "int4",
	"requirements": {
	"min_ram_bytes": 8589934592,
	"cpu_features": [],
	"apple_chip_min_generation": null
	},
	"runtime": {
	"name": "onnxruntime",
	"min_version": "1.31.0"
	},
	"base_model": {
	"id": "Qwen/Qwen3-ASR-0.6B",
	"revision": "5eb144179a02acc5e5ba31e748d22b0cf3e303b0"
	},
	"adapter": null,
	"tokenizer": {
	"id": "andrewleech/qwen3-asr-0.6b-onnx:vocab.json+added_tokens.json",
	"revision": "4fc24a1402e74db89c4d2ef256875e71680128c4"
	},
	"lm": null,
	"files": [
	{
	"path": "encoder.int4.onnx",
	"component": "model",
	"url": "https://huggingface.co/andrewleech/qwen3-asr-0.6b-onnx/resolve/4fc24a1402e74db89c4d2ef256875e71680128c4/encoder.int4.onnx",
	"size": 745762694,
	"sha256": "3c027f880f677615de85e1f6934906e1d5d77624b724096d33accef99c753eed"
	},
	{
	"path": "decoder_init.int4.onnx",
	"component": "model",
	"url": "https://huggingface.co/andrewleech/qwen3-asr-0.6b-onnx/resolve/4fc24a1402e74db89c4d2ef256875e71680128c4/decoder_init.int4.onnx",
	"size": 355390,
	"sha256": "6d54633bbb9e6b1ecf372f218d04adb3bf6f2bfc442ce55c988526d532896faa"
	},
	{
	"path": "decoder_step.int4.onnx",
	"component": "model",
	"url": "https://huggingface.co/andrewleech/qwen3-asr-0.6b-onnx/resolve/4fc24a1402e74db89c4d2ef256875e71680128c4/decoder_step.int4.onnx",
	"size": 354888,
	"sha256": "7eaa4f31d5eb6ae2937ff45222c32f692f2d5ac29b46cbae9046da332a9f317d"
	},
	{
	"path": "decoder_weights.int4.data",
	"component": "model",
	"url": "https://huggingface.co/andrewleech/qwen3-asr-0.6b-onnx/resolve/4fc24a1402e74db89c4d2ef256875e71680128c4/decoder_weights.int4.data",
	"size": 962460672,
	"sha256": "d68cd1c0695a7ba42651d06b1bc1158e2e58af5f9adbb6dba874fbbb8a4f22cf"
	},
	{
	"path": "embed_tokens.bin",
	"component": "model",
	"url": "https://huggingface.co/andrewleech/qwen3-asr-0.6b-onnx/resolve/4fc24a1402e74db89c4d2ef256875e71680128c4/embed_tokens.bin",
	"size": 311164928,
	"sha256": "e80150119fa5f7e56e85aed64c3a02d5c78eb7a37cfdcb973d0987316f15bee2"
	},
	{
	"path": "vocab.json",
	"component": "tokenizer",
	"url": "https://huggingface.co/andrewleech/qwen3-asr-0.6b-onnx/resolve/4fc24a1402e74db89c4d2ef256875e71680128c4/vocab.json",
	"size": 2776833,
	"sha256": "ca10d7e9fb3ed18575dd1e277a2579c16d108e32f27439684afa0e10b1440910"
	},
	{
	"path": "added_tokens.json",
	"component": "tokenizer",
	"url": "https://huggingface.co/andrewleech/qwen3-asr-0.6b-onnx/resolve/4fc24a1402e74db89c4d2ef256875e71680128c4/added_tokens.json",
	"size": 1566,
	"sha256": "de40784677cbd1843cabe5fbee078c7e042cd0b62155f0810af5a13842e5722a"
	},
	{
	"path": "config.json",
	"component": "meta",
	"url": "https://huggingface.co/andrewleech/qwen3-asr-0.6b-onnx/resolve/4fc24a1402e74db89c4d2ef256875e71680128c4/config.json",
	"size": 1250,
	"sha256": "df31c4689abe9d782366fffd2454b546291d0205d082b3fd01b99fb76a45b11f"
	}
	],
	"licenses": [
	{
	"id": "qwen3-asr-apache-2.0",
	"name": "Qwen3-ASR-0.6B (Apache-2.0)",
	"spdx": "Apache-2.0",
	"url": "https://huggingface.co/Qwen/Qwen3-ASR-0.6B",
	"applies_to": [
	"model",
	"tokenizer",
	"meta"
	],
	"acceptance_required": false,
	"attribution": [
	"Qwen3-ASR, Qwen Team, Alibaba Group (Qwen/Qwen3-ASR-0.6B, Apache-2.0)",
	"ONNX export and int4 quantisation by Andrew Leech (andrewleech/qwen3-asr-onnx, Apache-2.0)"
	],
	"notes": "The ONNX files are a modified form of the Qwen weights (graph export, int4 RTN quantisation of the decoder, FP16 embedding table) made by andrewleech/qwen3-asr-onnx; they are downloaded from that repository unchanged."
	}
	],
	"compat": {
	"status": "EXPERIMENTAL",
	"rtf_x1000": 153,
	"notes": "EXPERIMENTAL: token ids equal the export's own Python ONNX path on 7 clips (macOS and Windows), with the first token fixed to 'language'. On 55 real zh-TW clips MER 0.094 against provisional references (SenseVoice-Small 0.102, TingMix-Large 0.212) and 0.091 on the live captioning path (0.103, 0.244); English word error 0.22 live. About 0.15 x real time on 4 CPU threads of an Apple M4 Max, 3.4 GB peak memory (benchmarks/qwen3-asr). Output is converted from Simplified to Traditional characters (OpenCC s2tw)."
	},
	"provenance": null,
	"created": "2026-10-20",
	"min_app_version": "0.2.0"
	},
	"key_id": "tingmix-release-1",
	"signature": "56237ef6c97a321ef12ced0ebee203b89dafe3b609d8103fd9a9704d5f19a1729b49f1a38f8ab0f3318311593b0a529be2a6fac7c7473b997e9626e437a3bd09"
	},
	{
	"manifest": {
	"schema_version": 2,
	"serial": 1,
	"id": "sensevoice-small-onnx-int8-r2365bae",
	"family": "sensevoice-small",
	"revision": "2365baeacb507f821a0c8120fcee3d484dba7a07",
	"backend": "onnx-cpu",
	"os": "any",
	"arch": "any",
	"precision": "int8",
	"requirements": {
	"min_ram_bytes": 2147483648,
	"cpu_features": [],
	"apple_chip_min_generation": null
	},
	"runtime": {
	"name": "onnxruntime",
	"min_version": "1.31.0"
	},
	"base_model": null,
	"adapter": null,
	"tokenizer": {
	"id": "csukuangfj/sherpa-onnx-sense-voice-zh-en-ja-ko-yue-2024-07-17:tokens.txt",
	"revision": "2365baeacb507f821a0c8120fcee3d484dba7a07"
	},
	"lm": null,
	"files": [
	{
	"path": "model.int8.onnx",
	"component": "model",
	"url": "https://huggingface.co/csukuangfj/sherpa-onnx-sense-voice-zh-en-ja-ko-yue-2024-07-17/resolve/2365baeacb507f821a0c8120fcee3d484dba7a07/model.int8.onnx",
	"size": 239233841,
	"sha256": "c71f0ce00bec95b07744e116345e33d8cbbe08cef896382cf907bf4b51a2cd51"
	},
	{
	"path": "tokens.txt",
	"component": "tokenizer",
	"url": "https://huggingface.co/csukuangfj/sherpa-onnx-sense-voice-zh-en-ja-ko-yue-2024-07-17/resolve/2365baeacb507f821a0c8120fcee3d484dba7a07/tokens.txt",
	"size": 315894,
	"sha256": "f449eb28dc567533d7fa59be34e2abca8784f771850c78a47fb731a31429a1dc"
	}
	],
	"licenses": [
	{
	"id": "funasr-model-license-1.1",
	"name": "FunASR Model Open Source License Agreement 1.1 (SenseVoice-Small)",
	"spdx": null,
	"url": "https://github.com/modelscope/FunASR/blob/main/MODEL_LICENSE",
	"applies_to": [
	"model",
	"tokenizer"
	],
	"acceptance_required": true,
	"attribution": [
	"SenseVoice-Small, Tongyi Lab, Alibaba Group (FunAudioLLM/SenseVoiceSmall)",
	"ONNX export by k2-fsa / sherpa-onnx (Xiaomi Corp.)"
	],
	"notes": "Free to use, copy, modify and share with attribution and the model names kept. Provided for reference and learning purposes without warranty; the terms may be revised by the publisher. The ONNX export was made by sherpa-onnx from iic/SenseVoiceSmall on ModelScope (export-onnx.py in the export repository); the exact upstream revision is not recorded there."
	}
	],
	"compat": {
	"status": "EXPERIMENTAL",
	"rtf_x1000": 11,
	"notes": "EXPERIMENTAL: text equals sherpa-onnx 1.13.8 on 10 clips (one differs by one comma). On 55 real zh-TW clips MER 0.102 against provisional references (TingMix-Large 0.212), 0.116 against one video's subtitles (0.277). About 0.011 x real time on 4 CPU threads of an Apple M4 Max, 0.5 GB peak memory (benchmarks/sensevoice). Output is converted from Simplified to Traditional characters (OpenCC s2tw)."
	},
	"provenance": null,
	"created": "2026-10-19",
	"min_app_version": "0.2.0"
	},
	"key_id": "tingmix-release-1",
	"signature": "d6f04ac4047edd94965578f8d24c5edaee360d5b1b55de844b04931045eb7d0489a0a7b57a432b4b119a68e9015c8c727ee331787fffc75b374f9adabe62ea08"
	}
	]