FSMN-VAD ONNX — official Alibaba model mirror
Unmodified mirror of Alibaba / FunASR FSMN-Monophone VAD, published by iic as iic/speech_fsmn_vad_zh-cn-16k-common-onnx. Mirrored by neavo; model authorship remains with Alibaba / FunASR.
本仓库原样转存阿里官方发布的 16 kHz FSMN-VAD ONNX 量化模型，同时保留 am.mvn、config.yaml 和 configuration.json。本次未重新量化、导出或修改运行时文件。
Source and integrity
- Original repository: https://modelscope.cn/models/iic/speech_fsmn_vad_zh-cn-16k-common-onnx
- Pinned revision:
a158155ef9e81f405c052f40b6d1ad43b87e6215. - Source files: https://modelscope.cn/models/iic/speech_fsmn_vad_zh-cn-16k-common-onnx/tree/a158155ef9e81f405c052f40b6d1ad43b87e6215
| Runtime file | Bytes | SHA-256 |
|---|---|---|
am.mvn |
8,040 | 6820fef9687708c4fc3fab2530179c8fcea6262daa25514380056cd8f6eb1754 |
config.yaml |
1,215 | 2ef334f2d7776edd86ed696296774f87550206c256bfabc597872ad861831589 |
configuration.json |
598 | f62fcbcebbaade798714b34c521cad87d869ee6a327ceed948e01ff2336bbfc0 |
model_quant.onnx |
506,744 | 5289eb2aa3c9af2d7a4284bcfa7c3ceb81d360814ed4203239b6c5d0569da8a1 |
All runtime files were checked against the pinned upstream SHA-256 values. SHA256SUMS includes the accompanying files as well.
Usage
Use the model with the matching FunASR FSMN-VAD frontend and endpoint detection logic. am.mvn provides CMVN normalization statistics; config.yaml and configuration.json preserve the upstream model/frontend/runtime settings. The model is a voice activity detector, not a speech transcription model. Retain the upstream streaming cache and postprocessing semantics when integrating it.
Original documentation is preserved in UPSTREAM_README.md and quickstart.md. See FunASR for maintained runtime documentation.
This is a file mirror. No additional inference accuracy, GPU/WebGPU compatibility, or end-to-end pipeline tests were performed. An ONNX filename alone does not guarantee GPU support.
License and attribution
The pinned upstream model card and ModelScope repository declare Apache License 2.0. A copy is included as LICENSE. Preserve Alibaba / FunASR model attribution and the original FSMN-Monophone VAD model name. This mirror does not claim authorship of the weights.
- Downloads last month
- 13