Simple Asr

Minimal MCP server for local speech recognition using faster-whisper. Runs on CPU, no cloud required.

Русская версия

Installation

pipx install simple-asr-mcp

Or with pip:

pip install simple-asr-mcp

Requirements

• Python 3.10+
• ffmpeg (required by faster-whisper for audio decoding)

On Windows, install ffmpeg via winget:

winget install Gyan.FFmpeg

Quick Setup (for Claude Code agents)

Run these two commands — no user interaction required:

pipx install simple-asr-mcp
claude mcp add asr --scope user -- simple-asr-mcp

Then ask the user to restart Claude Code or run /mcp to reconnect.

Usage

CLI

# Transcribe an audio file
simple-asr-mcp transcribe recording.wav

# Specify language and model
simple-asr-mcp transcribe recording.wav --language ru --model medium

# List available models
simple-asr-mcp models

MCP Server (Claude Code)

MCP tools available after setup:

• transcribe_file — transcribe any audio file by path
• list_models — see available Whisper models

Configuration

Environment variables:

Variable	Default	Description
WHISPER_MODEL	small	Default Whisper model
WHISPER_DEVICE	cpu	Device: cpu, cuda, or auto
WHISPER_COMPUTE_TYPE	int8	Quantization type

Example with custom config:

claude mcp add asr --scope user -e WHISPER_MODEL=medium -e WHISPER_DEVICE=cuda -- simple-asr-mcp

Available Models

Model	Size	RAM (est.)
tiny	75 MB	~1 GB
base	142 MB	~1 GB
small	466 MB	~2 GB
medium	1.5 GB	~5 GB
large-v3	3.1 GB	~10 GB

The model is downloaded automatically on first use and cached locally. It stays in memory until the MCP server process exits.

Supported Audio Formats

Any format supported by ffmpeg: wav, mp3, flac, ogg, m4a, wma, etc.

Links

• GitHub
• PyPI

License

MIT