A Fundamental End-to-End Speech Recognition Toolkit

English | 简体中文 | Documentation | Paper

---

FunASR is an industrial-grade speech recognition toolkit developed by Tongyi Lab, Alibaba Group. It provides a unified Python interface for the complete speech understanding pipeline — from raw audio to speaker-attributed, punctuated transcripts — in a single API call.

from funasr import AutoModel

model = AutoModel(
    model="paraformer-zh",       # speech recognition
    vad_model="fsmn-vad",        # voice activity detection
    punc_model="ct-punc",        # punctuation restoration
    spk_model="cam++",           # speaker diarization
)
res = model.generate(input="meeting.wav", batch_size_s=300)

for s in res[0]["sentence_info"]:
    print(f"[Speaker {s['spk']}] {s['text']}")

Output:

[Speaker 0] 欢迎大家来体验达摩院推出的语音识别模型。
[Speaker 1] 非常感谢，今天我们主要讨论三个议题。
[Speaker 0] 好的，请开始吧。

---

Key Features

• Unified Pipeline — VAD, ASR, punctuation, speaker diarization, emotion detection composed in one AutoModel() call
• 50+ Languages — Fun-ASR-Nano (31 languages incl. Chinese dialects), Qwen3-ASR (52 languages with auto-detection)
• High Performance — Non-autoregressive inference; SenseVoice achieves 70ms RTF for 10s audio (15× faster than Whisper)
• Speaker Diarization — Per-sentence speaker labels, compatible with Paraformer, Fun-ASR-Nano, and SenseVoice
• Emotion & Audio Events — SenseVoice classifies emotions (happy/sad/angry/neutral) and detects BGM, laughter, applause
• Production Ready — Fine-tune with DeepSpeed, export to ONNX, deploy via Docker runtime or Python SDK

Installation

pip install -U funasr

# Or install from source for latest models
git clone https://github.com/modelscope/FunASR.git && pip install -e ./FunASR

Models auto-download from ModelScope (fast in China). Add hub="hf" for HuggingFace.

Model Zoo

Model	Type	Languages	Params	Inference Speed	Links
Fun-ASR-Nano	ASR	31	800M	中	MS · HF
Paraformer-zh	ASR	2	220M	快	MS · HF
SenseVoice	ASR+SER+AED	5	234M	极快 (70ms/10s)	MS · HF
Qwen3-ASR	ASR (LLM)	52	1.7B	慢	MS · HF
GLM-ASR-Nano	ASR (LLM)	17	1.5B	慢	MS · HF
Paraformer-Streaming	ASR	1	220M	实时	MS · HF
fsmn-vad	VAD	2	0.4M	—	MS · HF
ct-punc	Punctuation	2	290M	—	MS · HF
cam++	Speaker	—	7.2M	—	MS · HF
emotion2vec+	Emotion	—	300M	—	MS · HF

Full model list → model_zoo/  |  Detailed usage → Documentation

What's New

Date	Update
2026.05	Qwen3-ASR — 52 languages, LLM-based, auto language detection
2026.05	GLM-ASR-Nano — 17 languages, dialect & low-volume optimization
2026.05	Speaker diarization support added for Fun-ASR-Nano and SenseVoice
2025.12	Fun-ASR-Nano — 31-language end-to-end ASR, trained on tens of millions of hours
2024.07	SenseVoice — multi-task speech understanding (ASR + emotion + events)

View full changelog

See CHANGELOG.md for complete release history.

Learn More

Resource	Description
Tutorial	Install, choose a model, run ASR/VAD/speaker diarization
Training Guide	Fine-tune Paraformer, SenseVoice, Fun-ASR-Nano on custom data
Developer Guide	Add a new model, understand the registry, test & contribute
API Reference	Auto-generated class & method docs with source links
Runtime / Deployment	File transcription service, real-time streaming service (CPU/GPU)

Ecosystem

Project	Description
Fun-ASR-Nano	Multi-language ASR large model — 31 languages, timestamps, hotwords, speaker diarization
SenseVoice	Multi-task speech understanding — ASR, language ID, emotion, audio events
FunClip	AI video clipping powered by FunASR and LLM-assisted editing
CosyVoice	Natural speech generation with multi-language, timbre, and emotion control

Community

• Issues & feature requests → GitHub Issues
• Join our DingTalk discussion group:

Citation

@inproceedings{gao2023funasr,
  author    = {Zhifu Gao and Zerui Li and Jiaming Wang and Haoneng Luo and Xian Shi and Mengzhe Chen and Yabin Li and Lingyun Zuo and Zhihao Du and Zhangyu Xiao and Shiliang Zhang},
  title     = {FunASR: A Fundamental End-to-End Speech Recognition Toolkit},
  booktitle = {INTERSPEECH},
  year      = {2023},
}

License

Code: MIT License  ·  Model weights: FunASR Model License (commercial use permitted with attribution)