Back to catalog
Yap MCP
MITYap MCP is an open-source on-device speech-to-text MCP server and CLI using NVIDIA Parakeet TDT v3 to transcribe local audio in Claude and Cursor.
Open-source alternative to — compare it against:
Overview
Yap MCP runs NVIDIA Parakeet TDT v3 speech weights through sherpa-onnx on the host CPU, transcribing recorded audio at roughly ten times real-time speed without a network round-trip. Packaged as a Model Context Protocol stdio server and CLI utility, it lets coding assistants like Claude Code, Cursor, Windsurf, and VS Code convert local media files into text and query desktop dictations. Aside from a one-time verified 670 MB download from Hugging Face, no audio buffers or API keys leave the machine.
Why Yap MCP instead of Superwhisper
Because Parakeet TDT v3 pairs a FastConformer encoder with a Token-and-Duration Transducer head instead of an autoregressive decoder, sherpa-onnx transcribes a one-hour recording in about six minutes on a laptop processor. Superwhisper also supports local transcription, but its fastest local models target Apple Silicon hardware, and unlimited local file transcription or custom prompt modes require a paid Pro plan ($8.49 monthly or $249.99 lifetime). AudioPen uploads voice memos to remote servers before returning a summary. Yap MCP decodes audio buffers in memory across macOS, Windows x64, and Linux x64 or arm64 without a discrete GPU, an account, or per-seat billing.
The tradeoff comes down to a standalone GUI versus direct agent integration. Superwhisper and AudioPen provide visual recorders and built-in rewrite rules. Yap MCP skips the recording window and exposes three callable tools—transcribe_audio, search_dictations, and prepare_speech_model—directly to your AI coding client. Instead of transcribing a file in another app and copying text into an editor, developers ask their assistant to transcribe a local path or search recent voice notes from the companion Yap desktop app inside the active workspace.
What it does well
- Fast CPU inference: Runs
sherpa-onnxgraphs on x64 and ARM64 processors at roughly 10x real-time speed without CUDA setup or cloud latency. - Broad media decoding: Reads WAV, MP3, M4A, AAC, MP4, FLAC, OGG, Opus, WebM, AIFF, and CAF files in memory, with optional timestamps and filler cleanup.
- Automatic European language detection: Identifies and transcribes 25 European languages—including English, German, French, Spanish, Polish, Russian, and Ukrainian—without manual flags.
