FunASR-Powered MCP Server (MCPServer)

What is FunASR-Powered MCP Server (MCPServer)?

MCPServer is a Python-based server that leverages Alibaba's FunASR library to provide speech processing services through the FastMCP framework. It offers tools for audio validation, asynchronous speech transcription with models like Paraformer, and voice activity detection (VAD). It is designed for developers integrating speech capabilities into MCP-compatible applications.

How to use FunASR-Powered MCP Server (MCPServer)?

Install dependencies with pip install -r requirements.txt, then run the server using uvicorn main:app --host 0.0.0.0 --port 9000. Interact via any MCP client; the server exposes tools at http://0.0.0.0:9000/sse. Optionally set MODELSCOPE_API_TOKEN for model downloads.

Key features of FunASR-Powered MCP Server (MCPServer)

• Audio file validation (format, duration, sample rate).
• Asynchronous speech-to-text transcription with task management.
• Detailed transcription results including segment and word timestamps.
• Voice activity detection returning precise speech segment timestamps.
• Dynamic loading and switching of ASR and VAD models.
• Configurable model parameters per request.

Use cases of FunASR-Powered MCP Server (MCPServer)

• Transcribing long audio recordings (meetings, lectures) asynchronously.
• Validating audio file integrity before further processing.
• Detecting speech segments in audio for speaker diarization or silence removal.
• Building custom voice‑based applications with flexible model selection.

FAQ from FunASR-Powered MCP Server (MCPServer)

What are the runtime requirements?

Python 3.8 or higher and pip. FunASR downloads models from ModelScope; a MODELSCOPE_API_TOKEN may be needed to avoid rate limits or access private models.

How does transcription work?

Transcription tasks are fully asynchronous – submit a job, poll its status, and retrieve the result when completed. Results include full text, segment timestamps, and word‑level timestamps where available.

Can I use different ASR or VAD models?

Yes. You can specify a model per request (e.g., via model_name parameter) or load a new default model using load_asr_model or load_vad_model tools.

Where are audio files processed?

Audio files are processed on the server from local file paths. The server does not store or stream audio externally; all processing occurs in‑memory.

What transport and authentication does the server use?

The server uses SSE (Server‑Sent Events) at the /sse endpoint. No authentication is described in the README.