Model distribution mirror for video-understanding-skill. The pipeline downloads from the official sherpa-onnx releases first and falls back to this mirror. All model files are redistributed unmodified (except int8-only subsetting for the streaming package); see THIRD-PARTY-NOTICES.md for licenses and attribution. Repository docs only: MIT. Model weights keep their upstream licenses.
| File | Size |
|---|---|
README.md |
upstream readme (source attribution) |
model.int8.onnx |
227 MB |
tokens.txt |
309 KB |
- Source: sherpa-onnx release
asr-models—sherpa-onnx-sense-voice-zh-en-ja-ko-yue-int8-2025-09-09.tar.bz2(files unmodified, test wavs dropped). - Upstream model: SenseVoice-Small, converted by k2-fsa from ASLP-lab/WSYue-ASR
sensevoice_small_yue; base weights by Alibaba FunAudioLLM. - License: FunASR Model Open Source License Agreement v1.1 (see notices; attribution and model-name retention required — kept in package name and readme).
| File | Size |
|---|---|
README.md |
upstream readme (license: apache-2.0) |
encoder-epoch-99-avg-1.int8.onnx |
174 MB |
decoder-epoch-99-avg-1.int8.onnx |
13 MB |
joiner-epoch-99-avg-1.int8.onnx |
3.1 MB |
tokens.txt |
55 KB |
- Source: sherpa-onnx release
asr-models—sherpa-onnx-streaming-zipformer-bilingual-zh-en-2023-02-20.tar.bz2(int8 files only; fp32 variants and test wavs dropped). - Upstream model: bilingual zh-en streaming zipformer, torchscript from pfluo/k2fsa-zipformer-chinese-english-mixed, trained with k2-fsa/icefall.
- License: Apache-2.0.
cefa41698d842d5aed1f1dd6379372f4fd20c333eeff89b4e5fc100bbaf586e6 sense-voice-int8.tar.bz2
fb3b2898f5fb14bb2b99b44ebab12ad94b97416094915977a103e3eb6c728af9 streaming-zipformer-int8.tar.bz2
video-understanding-skill consumes these automatically as the fallback download chain in vus/model_setup.py. Manual download:
curl -L -O https://github.com/CommitStrip/vus-models/releases/download/v1.0/sense-voice-int8.tar.bz2
curl -L -O https://github.com/CommitStrip/vus-models/releases/download/v1.0/streaming-zipformer-int8.tar.bz2Not affiliated with k2-fsa, Alibaba, or ASLP-lab. Models are provided as-is under their upstream licenses.