README.md

---
license: other
license_name: funasr-model-license-v1.1
license_link: https://github.com/modelscope/FunASR/blob/main/MODEL_LICENSE
base_model: FunAudioLLM/SenseVoiceSmall
pipeline_tag: automatic-speech-recognition
library_name: openasr
tags:
  - automatic-speech-recognition
  - speech-to-text
  - openasr
  - oasr
  - sensevoice
---

<div align="center">

# SenseVoice Small · OpenASR

**Fast multilingual speech recognition from FunAudioLLM — non-autoregressive SenseVoice, tuned for Chinese, Cantonese, English, Japanese and Korean**

[![License](https://img.shields.io/badge/license-FunASR_Model_License_v1.1-2563eb.svg)](https://github.com/modelscope/FunASR/blob/main/MODEL_LICENSE)
[![Format](https://img.shields.io/badge/format-.oasr-7c3aed.svg)](https://github.com/QuintinShaw/openasr)
[![Runtime](https://img.shields.io/badge/runtime-OpenASR-111827.svg)](https://openasr.org)
[![Base model](https://img.shields.io/badge/base-SenseVoiceSmall-f59e0b.svg)](https://huggingface.co/FunAudioLLM/SenseVoiceSmall)

Native speech-to-text in the **[OpenASR](https://github.com/QuintinShaw/openasr)** runtime —
engineered for peak performance on CPU & GPU, **no Python at inference time**.

</div>

---

## ✨ Highlights

- 🌏 **Multilingual, zh-first** — high-precision Mandarin, Cantonese, English, Japanese and Korean with automatic language detection
- ⚡ **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-Large
- 🀄 **Chinese benchmark strength** — trained on over 400,000 hours of speech; the upstream reports better Chinese and Cantonese accuracy than Whisper on AISHELL and WenetSpeech
- 🪶 **Compact and local** — a small checkpoint that transcribes fully offline, from a 130 MB q4_k build up to full-fidelity fp16
- 🦀 **Native in OpenASR** — `.oasr` packs run with no Python at inference, engineered for peak performance on CPU & GPU

## 🚀 Quickstart

```bash
# 1. Install the OpenASR CLI  ·  https://openasr.org
# 2. Pull a build (pick a quant — see the table below)
openasr pull sensevoice-small:fp16

# 3. Transcribe
openasr transcribe audio.wav --model sensevoice-small
```

All builds for this model:

```bash
openasr pull sensevoice-small:fp16
openasr pull sensevoice-small:q8
openasr pull sensevoice-small:q4
```

## 📦 Available builds

| Quant | File (`.oasr`) | Size | RAM peak | RTF · M1 CPU | RTF · M1 GPU | JFK ΔWER vs fp16 |
|:------|:---------------|-----:|---------:|-------------:|-------------:|-----------------:|
| fp16 | `sensevoice-small-fp16.oasr` | 470 MB | 745 MB | 0.18× | 0.04× | 0.0% |
| q8_0 | `sensevoice-small-q8_0.oasr` | 252 MB | 514 MB | 0.18× | 0.04× | 0.0% |
| q4_k | `sensevoice-small-q4_k.oasr` | 136 MB | 395 MB | 0.23× | 0.05× | 0.0% |

<sub>RTF = real-time factor on the fixed 11s JFK clip (**lower is faster**); RAM peak measured per pack
in an isolated subprocess. JFK ΔWER compares each quantized build's JFK transcript to this model's
fp16 JFK transcript, so it measures quantization drift rather than absolute recognition accuracy.
**fp16** is the recommended default — near-reference quality at a fraction of the
footprint.</sub>

## 🧠 About SenseVoice Small

SenseVoice Small is the compact member of **SenseVoice**, the speech understanding model family
open-sourced by **FunAudioLLM** (Alibaba). Trained on more than **400,000 hours** of speech, it
delivers high-precision transcription with automatic language detection for **Mandarin Chinese,
Cantonese, English, Japanese and Korean**, and the upstream card reports Chinese and Cantonese
accuracy ahead of Whisper on open benchmarks such as AISHELL and WenetSpeech. Its
**non-autoregressive end-to-end** architecture makes inference exceptionally fast — the upstream
team clocks about 70 ms for 10 seconds of audio, 15x faster than Whisper-Large. The upstream model
also carries speech emotion recognition and audio event detection; the OpenASR packs currently
surface plain transcription only (emotion/event tags are not yet exposed). This OpenASR repo
repackages the original weights as `.oasr` packs that run natively in the OpenASR runtime — no
Python at inference time. The **fp16** build is the recommended default for maximum fidelity;
**q8_0** halves the footprint at near-reference quality and **q4_k** suits tight-memory devices.

## ⚙️ How these packs were made

Converted from [FunAudioLLM/SenseVoiceSmall](https://huggingface.co/FunAudioLLM/SenseVoiceSmall) with the OpenASR importer:

```bash
openasr model-pack import sensevoice <src> <out>.oasr \
  --package-id sensevoice-small --quantization {fp16,q8-0,q4-k}
```

The `.oasr` container is GGUF-backed; packs use zero-copy mmap weight binding and graph
buffer reuse to keep peak memory low.

## ⚖️ License

These packs **inherit the upstream model's license: FunASR Model License v1.1**
([source](https://github.com/modelscope/FunASR/blob/main/MODEL_LICENSE)). OpenASR packaging retains the upstream copyright and
NOTICE; the only modifications are format conversion and quantization.

## 🙏 Acknowledgements

This pack is a redistribution of **SenseVoice Small**, created and open-sourced by the
**FunAudioLLM team at Alibaba**
([FunAudioLLM/SenseVoiceSmall](https://huggingface.co/FunAudioLLM/SenseVoiceSmall)), built on the
**FunASR** open-source speech toolkit from Alibaba's ModelScope community. All credit for the
original architecture, training, and weights belongs to the FunAudioLLM and FunASR teams; the
license is inherited from and identical to the upstream model — the
**[FunASR Model License v1.1](https://github.com/modelscope/FunASR/blob/main/MODEL_LICENSE)**,
which permits commercial use and requires this attribution. Thank you to FunAudioLLM, the FunASR
team, and Alibaba for releasing their work openly. OpenASR only performs format conversion,
quantization, runtime verification, and local-inference adaptation.

## 🔗 Links

- 🦀 **OpenASR** — <https://github.com/QuintinShaw/openasr>
- 🌐 **Website** — <https://openasr.org>
- 🤗 **Upstream model** — [FunAudioLLM/SenseVoiceSmall](https://huggingface.co/FunAudioLLM/SenseVoiceSmall)