Loading listing…
FunASR
Security scan passedTranscribes audio files on your own machine with FunASR and SenseVoice, in Mandarin, Cantonese, English, Japanese and Korean.
by modelscope· MediaUnclaimed · Yours? Claim it
Preview
4 images
What FunASR does
About
The FunASR MCP server is ModelScope's local speech-to-text server for agents. It runs FunASR, the open-source speech recognition toolkit, with the SenseVoiceSmall model on your own machine, so your agent can turn a recording into text without sending the audio to a cloud service and without an API key. What you can do with it - Transcribe a meeting recording, an interview or a voice memo from a file path - Recognise Mandarin, Cantonese, English, Japanese and Korean, detected automatically or set with a language hint - Get time-stamped segments along with the text when the model returns them - Split longer recordings with voice activity detection before recognition How it connects - A local server over stdio, published as the Docker image ghcr.io/modelscope/funasr-mcp - You mount one folder of audio files read-only at /audio, then ask about files by their /audio path, such as /audio/meeting.wav - A Docker volume keeps the downloaded model between runs - Without Docker, install funasr with pip and point your agent at funasr_mcp.py from the FunASR repository; FUNASR_DEVICE picks CPU, CUDA or Apple's MPS, and FUNASR_MODEL another model Good to know - One tool, transcribe_audio, which reads files and changes nothing - The first transcription downloads the model weights from ModelScope, so it takes longer than later ones - It takes WAV, MP3, FLAC and other common audio files; not URLs or live microphone audio - The Docker image runs on the CPU and is built for x86-64 machines - MIT licensed, maintained in the modelscope/FunASR repository
More from Media
Plugins people install alongside FunASR.
