FunASR — Industrial Speech Recognition Toolkit

• URL: https://github.com/modelscope/FunASR
• Date Saved: 2026-09-01
• Source: GitHub
• Tags: asr, ai-tools, dev-tools
• Repo: https://github.com/modelscope/FunASR (20.1k stars, MIT license)

Summary

Open-source speech recognition toolkit from ModelScope/Alibaba DAMO Academy. Not a single model but a full toolkit — pick the right model per job, compose pipelines (ASR + VAD + punctuation + speaker diarization), and deploy via OpenAI-compatible API or edge binary.

Key Models

• Fun-ASR-Nano (800M) — flagship LLM-ASR for zh/en/ja + Chinese dialects, GPU required
• Fun-ASR-MLT-Nano (800M) — 31 languages
• SenseVoiceSmall (234M) — 5-language ASR + emotion recognition + audio events, CPU-viable (17x realtime)
• Paraformer-zh (220M) — low-latency streaming ASR for zh/en
• Qwen3-ASR (1.7B) — 52 languages (the base model that AmphionASR builds on)
• Also wraps Whisper-large-v3, GLM-ASR-Nano, emotion2vec

Why It Matters

• 340x realtime with Fun-ASR-Nano + vLLM batch processing
• Full pipeline composition: VAD (fsmn-vad) + ASR + punctuation (ct-punc) + speaker diarization (CAM++) in one AutoModel call
• CPU/edge deployment via GGUF format (like whisper.cpp but ~3x lower CER on Chinese)
• OpenAI-compatible API server: funasr-server --device cuda
• MCP server integration for AI agents (Claude/Cursor)
• Streaming WebSocket support via Paraformer
• CLI tool: funasr audio.wav --spk --timestamps -f json

Deployment Options

• Python pip package (pip install funasr)
• vLLM acceleration for batch processing
• Docker containers for streaming service
• GGUF binary for CPU/edge (no Python needed, supports Vulkan GPU)
• OpenAI-compatible REST API

Context

• Very actively maintained (v1.4.11 released 2026-08-30, commits daily)
• This is the FunASR-Realtime that came within 0.52 points of AmphionASR on Chinese entity recognition — without any hotword input
• Practical choice for self-hosted ASR: mature toolkit, multiple model options, production-ready serving