OpenASR/sensevoice-small
01.6k
1---2license: other3license_name: funasr-model-license-v1.14license_link: https://github.com/modelscope/FunASR/blob/main/MODEL_LICENSE5base_model: FunAudioLLM/SenseVoiceSmall6pipeline_tag: automatic-speech-recognition7library_name: openasr8tags:9 - automatic-speech-recognition10 - speech-to-text11 - openasr12 - oasr13 - sensevoice14---15 16<div align="center">17 18# SenseVoice Small · OpenASR19 20**Fast multilingual speech recognition from FunAudioLLM — non-autoregressive SenseVoice, tuned for Chinese, Cantonese, English, Japanese and Korean**21 22[](https://github.com/modelscope/FunASR/blob/main/MODEL_LICENSE)23[](https://github.com/QuintinShaw/openasr)24[](https://openasr.org)25[](https://huggingface.co/FunAudioLLM/SenseVoiceSmall)26 27Native speech-to-text in the **[OpenASR](https://github.com/QuintinShaw/openasr)** runtime —28engineered for peak performance on CPU & GPU, **no Python at inference time**.29 30</div>31 32---33 34## ✨ Highlights35 36- 🌏 **Multilingual, zh-first** — high-precision Mandarin, Cantonese, English, Japanese and Korean with automatic language detection37- ⚡ **Non-autoregressive speed** — an end-to-end architecture the upstream clocks at about 70 ms for 10 seconds of audio, 15 times faster than Whisper-Large38- 🀄 **Chinese benchmark strength** — trained on over 400,000 hours of speech; the upstream reports better Chinese and Cantonese accuracy than Whisper on AISHELL and WenetSpeech39- 🪶 **Compact and local** — a small checkpoint that transcribes fully offline, from a 130 MB q4_k build up to full-fidelity fp1640- 🦀 **Native in OpenASR** — `.oasr` packs run with no Python at inference, engineered for peak performance on CPU & GPU41 42## 🚀 Quickstart43 44```bash45# 1. Install the OpenASR CLI · https://openasr.org46# 2. Pull a build (pick a quant — see the table below)47openasr pull sensevoice-small:fp1648 49# 3. Transcribe50openasr transcribe audio.wav --model sensevoice-small51```52 53All builds for this model:54 55```bash56openasr pull sensevoice-small:fp1657openasr pull sensevoice-small:q858openasr pull sensevoice-small:q459```60 61## 📦 Available builds62 63| Quant | File (`.oasr`) | Size | RAM peak | RTF · M1 CPU | RTF · M1 GPU | JFK ΔWER vs fp16 |64|:------|:---------------|-----:|---------:|-------------:|-------------:|-----------------:|65| fp16 | `sensevoice-small-fp16.oasr` | 470 MB | 745 MB | 0.18× | 0.04× | 0.0% |66| q8_0 | `sensevoice-small-q8_0.oasr` | 252 MB | 514 MB | 0.18× | 0.04× | 0.0% |67| q4_k | `sensevoice-small-q4_k.oasr` | 136 MB | 395 MB | 0.23× | 0.05× | 0.0% |68 69<sub>RTF = real-time factor on the fixed 11s JFK clip (**lower is faster**); RAM peak measured per pack70in an isolated subprocess. JFK ΔWER compares each quantized build's JFK transcript to this model's71fp16 JFK transcript, so it measures quantization drift rather than absolute recognition accuracy.72**fp16** is the recommended default — near-reference quality at a fraction of the73footprint.</sub>74 75## 🧠 About SenseVoice Small76 77SenseVoice Small is the compact member of **SenseVoice**, the speech understanding model family78open-sourced by **FunAudioLLM** (Alibaba). Trained on more than **400,000 hours** of speech, it79delivers high-precision transcription with automatic language detection for **Mandarin Chinese,80Cantonese, English, Japanese and Korean**, and the upstream card reports Chinese and Cantonese81accuracy ahead of Whisper on open benchmarks such as AISHELL and WenetSpeech. Its82**non-autoregressive end-to-end** architecture makes inference exceptionally fast — the upstream83team clocks about 70 ms for 10 seconds of audio, 15x faster than Whisper-Large. The upstream model84also carries speech emotion recognition and audio event detection; the OpenASR packs currently85surface plain transcription only (emotion/event tags are not yet exposed). This OpenASR repo86repackages the original weights as `.oasr` packs that run natively in the OpenASR runtime — no87Python at inference time. The **fp16** build is the recommended default for maximum fidelity;88**q8_0** halves the footprint at near-reference quality and **q4_k** suits tight-memory devices.89 90## ⚙️ How these packs were made91 92Converted from [FunAudioLLM/SenseVoiceSmall](https://huggingface.co/FunAudioLLM/SenseVoiceSmall) with the OpenASR importer:93 94```bash95openasr model-pack import sensevoice <src> <out>.oasr \96 --package-id sensevoice-small --quantization {fp16,q8-0,q4-k}97```98 99The `.oasr` container is GGUF-backed; packs use zero-copy mmap weight binding and graph100buffer reuse to keep peak memory low.101 102## ⚖️ License103 104These packs **inherit the upstream model's license: FunASR Model License v1.1**105([source](https://github.com/modelscope/FunASR/blob/main/MODEL_LICENSE)). OpenASR packaging retains the upstream copyright and106NOTICE; the only modifications are format conversion and quantization.107 108## 🙏 Acknowledgements109 110This pack is a redistribution of **SenseVoice Small**, created and open-sourced by the111**FunAudioLLM team at Alibaba**112([FunAudioLLM/SenseVoiceSmall](https://huggingface.co/FunAudioLLM/SenseVoiceSmall)), built on the113**FunASR** open-source speech toolkit from Alibaba's ModelScope community. All credit for the114original architecture, training, and weights belongs to the FunAudioLLM and FunASR teams; the115license is inherited from and identical to the upstream model — the116**[FunASR Model License v1.1](https://github.com/modelscope/FunASR/blob/main/MODEL_LICENSE)**,117which permits commercial use and requires this attribution. Thank you to FunAudioLLM, the FunASR118team, and Alibaba for releasing their work openly. OpenASR only performs format conversion,119quantization, runtime verification, and local-inference adaptation.120 121## 🔗 Links122 123- 🦀 **OpenASR** — <https://github.com/QuintinShaw/openasr>124- 🌐 **Website** — <https://openasr.org>125- 🤗 **Upstream model** — [FunAudioLLM/SenseVoiceSmall](https://huggingface.co/FunAudioLLM/SenseVoiceSmall)126