FunASR
ActiveDescription
FunASR is an industrial-grade open-source speech recognition toolkit from Alibaba DAMO Academy, 170x faster than Whisper. Supports 50+ languages, speaker diarization, emotion recognition, and an OpenAI-compatible API with MCP server.
Key Features
- Ultra-fast recognition — SenseVoice-Small runs at 170× realtime on GPU and 17× realtime on CPU
- Speaker diarization + emotion recognition — Automatically annotate speaker IDs and emotion labels from a single audio clip
- Streaming recognition — WebSocket-based real-time streaming ASR for call QA and live captioning use cases
- OpenAI-compatible API + MCP — /v1/audio/transcriptions endpoint and an MCP server for AI agents
Use Cases
💡 Auto-transcribe meeting recordings and separate speakers, outputting structured subtitle files
💡 Real-time call transcription in a customer-service QA system with automatic flagging of abnormal conversations
💡 Integrate speech recognition into Claude / Cursor and other AI agent toolchains via the MCP server
Strengths & Limitations
✅ Strengths
- • Actively maintained, recent updates
- • High community interest (20.2k stars)
- • Permissive open-source license (MIT)
- • Established track record (3 years in production)
Categories
Quick Start
# One-liner install
pip install funasr
# Basic ASR with VAD + speaker diarization + timestamps
from funasr import AutoModel
model = AutoModel(model="iic/SenseVoiceSmall",
vad_model="fsmn-vad", spk_model="cam++", device="cuda")
result = model.generate(input="audio.wav")
# Start OpenAI-compatible API server
funasr-server --device cuda
# POST /v1/audio/transcriptions