Local Speech-to-Text MCP Server

What is Local Speech-to-Text MCP Server?

A high-performance Model Context Protocol (MCP) server that provides local speech-to-text transcription using whisper.cpp, optimized for Apple Silicon. It is designed for users who need private, offline transcription with speaker diarization and support for multiple audio formats.

How to use Local Speech-to-Text MCP Server?

Install Node.js 18+, whisper.cpp, ffmpeg, and optionally Python 3.8+ for speaker diarization. Clone the repository, run npm install, npm run build, and npm run setup:models. Set the HF_TOKEN environment variable for speaker diarization. Add the server path to your MCP client configuration. Available tools include transcribe, transcribe_long, transcribe_with_speakers, list_models, health_check, and version.

Key features of Local Speech-to-Text MCP Server

• 100% local processing with complete privacy
• 15x+ real‑time transcription speed on Apple Silicon
• Speaker diarization to identify and separate multiple speakers
• Universal audio support with automatic format conversion (MP3, M4A, FLAC, etc.)
• Multiple output formats: txt, json, vtt, srt, csv
• Low memory footprint (<2GB) and TypeScript implementation

Use cases of Local Speech-to-Text MCP Server

• Transcribe meetings or interviews with speaker identification
• Process long audio files with automatic chunking via transcribe_long
• Perform privacy‑sensitive transcription without cloud uploads
• Batch convert audio to text in various subtitle or data formats

FAQ from Local Speech-to-Text MCP Server

What are the prerequisites?

Node.js 18+, whisper.cpp (install via brew install whisper-cpp), ffmpeg for audio conversion (brew install ffmpeg), and Python 3.8+ with a HuggingFace token for speaker diarization.

What audio formats are supported?

Native whisper.cpp formats: WAV and FLAC. Many others (MP3, M4A, AAC, OGG, WMA, etc.) are automatically converted to 16kHz mono via ffmpeg.

Does Local Speech-to-Text MCP Server require an internet connection?

No, transcription runs entirely locally. Speaker diarization requires a one‑time HuggingFace token setup, but processing remains offline.

How does performance compare to WhisperX?

On Apple Silicon, it achieves 15.8x real‑time speed (vs WhisperX 5.5x) and uses under 2GB memory (vs ~4GB). GPU acceleration via Apple Neural Engine is supported.

How do I enable speaker diarization?

Set the HF_TOKEN environment variable with a free HuggingFace token and accept the license for pyannote/speaker‑diarization‑3.1 at huggingface.co.