FunASR
VerifiedIndustrial speech recognition toolkit with 170x realtime speed.
What is the FunASR MCP server?
FunASR delivers industrial-grade ASR performance that is significantly faster than Whisper while running on both CPU and GPU. It includes models such as SenseVoice-Small and Paraformer-Large that automatically handle punctuation, timestamps, and speaker labels in a single call.
The toolkit supports LLM-enhanced ASR via Fun-ASR-Nano and offers deployment options including vLLM acceleration and a local MCP server for Claude and Cursor integration.
Install & connect
Add this to your MCP client config. Pick your client below and copy.
{
"mcpServers": {
"FunASR": {
"command": "uvx",
"args": [
"torch"
]
}
}
}Package: torch (pypi)
Other ways to install
pip
pip install torchExample prompts
Once connected, try asking your AI client:
Security & permissions
Runs locally via stdio with no remote API keys required. Accesses local audio files and GPU/CPU resources on the host machine.
What you can do with FunASR
Meeting transcription
Process long audio recordings to produce timestamped transcripts with speaker identification and emotion labels.
Multilingual subtitle generation
Transcribe videos or podcasts in 50+ languages with automatic punctuation and high accuracy.
Agent-powered voice interfaces
Connect FunASR via MCP to let AI agents analyze recorded audio and extract structured insights.
How to use FunASR
- 1Install funasr via pip along with torch and torchaudio
- 2Clone the FunASR repository and navigate to examples/mcp_server
- 3Start the MCP server using the provided launch script
- 4Configure your AI client (Claude/Cursor) to connect via stdio
- 5Test with a sample audio file to verify transcription output
FunASR: pros & cons
Pros
- +Extremely fast inference (up to 170x realtime on GPU)
- +Built-in support for diarization, emotion, and punctuation
- +Self-hosted and free with MIT license
- +OpenAI-compatible API and MCP integration available
Cons
- –Requires local GPU or CPU resources for best performance
- –Model downloads can be large
- –MCP server setup needs manual configuration from examples
Frequently asked questions
No, once models are downloaded it runs completely offline.
User reviews
Verified reviews from the community shape this listing's rating.
Loading reviews…
Similar MCP servers
Other ai & knowledge connectors worth a look.
