FunASR MCP Server

What it does

The FunASR MCP Server, from the ModelScope FunASR speech-recognition project, gives an assistant one tool: transcribe a local audio file (wav, mp3, flac, m4a or ogg) to text. It recognises Mandarin, Cantonese, English, Japanese and Korean, detecting the language automatically or taking a hint, splits longer recordings with voice-activity detection, and returns per-segment timestamps when the model provides them.

Inference runs locally with the SenseVoiceSmall model by default, on CPU, CUDA or Apple MPS, and no API key is needed; the model can be swapped with FUNASR_MODEL. The model is fetched from ModelScope on first use and cached.

The published way to run it is the container image ghcr.io/modelscope/funasr-mcp (listed in the official MCP registry), with a host folder of audio mounted read-only at /audio and paths passed as /audio/... ; the docs also show running the script funasr_mcp.py with Python after pip install funasr.