lvmmai2/asrtools-models

ASRTools offline models

Model collection used by the ASRTools offline transcription flow: speech recognition, voice activity detection and per-character timestamp alignment, all running on CPU without a compiler, Python environment or GPU on the receiving machine.

These are the upstream weights redistributed byte-for-byte. Nothing here is retrained, quantized, fused or re-exported by this repository; every file is published exactly as it was received from the upstream projects credited below. This repository only repackages the files so the ASRTools application can fetch them separately from its application archive.

This repository does not contain application code, inference runtimes (PyTorch, FunASR, sherpa-onnx, Qt, FFmpeg) or any user audio. The application archive still carries those; the models here are data files it loads.

Contents

Path	Purpose	Size	Licence
sensevoice-small-int8/model.int8.onnx	Speech recognition (zh / en / ja / ko / yue)	239,233,841 B	FunASR Model License
sensevoice-small-int8/tokens.txt	Token map for the recogniser	315,894 B	FunASR Model License
silero-vad/silero_vad.onnx	Voice activity detection and segmentation	643,854 B	MIT
fa-zh/model.pt	Per-character timestamp prediction	158,469,618 B	FunASR Model License
fa-zh/config.yaml	fa-zh model configuration	2,455 B	FunASR Model License
fa-zh/configuration.json	fa-zh model configuration	448 B	FunASR Model License
fa-zh/tokens.json	fa-zh token map	93,676 B	FunASR Model License
fa-zh/seg_dict	fa-zh segmentation dictionary	8,287,834 B	FunASR Model License
fa-zh/am.mvn	fa-zh feature normalisation statistics	11,203 B	FunASR Model License

Nine model files, 407,058,823 bytes in total (388.20 MiB). offline-models.json records the SHA-256 of each one; SHA256SUMS.txt covers every file in this repository, including the documentation below.

Sources and attribution

Here	Upstream	Licence
sensevoice-small-int8/	FunAudioLLM / Alibaba SenseVoice Small, converted to INT8 ONNX by the k2-fsa sherpa-onnx project, archive sherpa-onnx-sense-voice-zh-en-ja-ko-yue-int8-2024-07-17	FunASR Model Open Source License Agreement v1.1
silero-vad/silero_vad.onnx	Silero Team VAD, redistributed from the same sherpa-onnx release	MIT
fa-zh/	Alibaba / FunASR fa-zh timestamp prediction model, revision d7701644d6e336f093501241beca010519c0986f	FunASR Model Open Source License Agreement v1.1

Upstream locations: FunAudioLLM/SenseVoice, sherpa-onnx pretrained models, snakers4/silero-vad, funasr/fa-zh.

The FunASR agreement requires attribution of source and author and retention of the model names; both are kept here, and the agreement text, the retained notices and snapshots of the upstream licence files are in licenses/, with licenses/MODEL-SOURCES.txt recording where every file came from. The original model cards ship with the weights (sensevoice-small-int8/README.md, fa-zh/README.md) and are not replaced by this file.

The license_name: model-license metadata above is the licence of the redistributed weights, not of the ASRTools application. sensevoice-small-int8/LICENSE is the 71-byte pointer the conversion archive shipped, kept as received.

The repositories also publish a code licence (Apache-2.0 for FunASR and sherpa-onnx, MIT for SenseVoice) that does not cover the weights; the Apache-2.0 text that sherpa-onnx releases ship next to the onnxruntime wheels is retained in licenses/sherpa-onnx-LICENSE.txt.

Download

Command-line, huggingface_hub installed:

hf download lvmmai2/asrtools-models --revision <full-commit-sha> --local-dir ./_runtime/models

Source checkouts use --local-dir ./models instead. Pin the full commit SHA, not main, so the bytes cannot change between runs.

The layout matches what the application loads: the three directories sit at the repository root, exactly as they must appear under models/. Without the Hub CLI, download the nine files from the file browser into that directory and verify against offline-models.json.

Not included, by design: export-onnx.py from the conversion archive (a conversion tool, not needed to run the ONNX file) and its test_wavs/ demo audio (not used by the offline flow, and its own audio licence was not reviewed).