skip to content
.katana

Mutter

Local-first, multi-language voice dictation for macOS — captures mic or system audio, transcribes on-device with whisper.cpp, and inserts cleaned-up text at the cursor with zero network calls.

Mutter showcase
Role
Full-stack Development
Date
September 4, 2026
Deliverables
macOS desktop app, Local multilingual speech-to-text, System-audio (meeting) capture mode, Local history & metrics dashboard
Tools
Tauri, Rust, TypeScript, whisper.cpp, SQLite

Mutter turns speech into text entirely on-device. Press a hotkey to dictate into any app — including AI coding agents — or capture system audio Granola-style to transcribe meetings. Inference, history, and metrics all run locally, with no cloud calls, no subscriptions, and no paywall.

A Tauri + Rust app runs whisper.cpp locally for multilingual transcription (English, Yoruba, Spanish, Italian, French, Arabic), behind a swappable TranscriptionEngine/TextProcessor adapter layer so engines like Apple's on-device Speech framework can be added without a rewrite. Text is inserted at the cursor via the Accessibility API, with a clipboard-paste fallback for terminals. Every transcript is persisted to a local SQLite history store, and a glassmorphic "pill" HUD plus a settings dashboard — both Tauri webviews — surface live recording status and time-saved/WPM metrics, computed entirely locally with no telemetry.

view repo

Overview

Mutter turns speech into text entirely on-device. Press a hotkey to dictate into any app — including AI coding agents — or capture system audio Granola-style to transcribe meetings. Inference, history, and metrics all run locally, with no cloud calls, no subscriptions, and no paywall.

Mutter

Approach

A Tauri + Rust app runs whisper.cpp locally for multilingual transcription (English, Yoruba, Spanish, Italian, French, Arabic), behind a swappable TranscriptionEngine/TextProcessor adapter layer so engines like Apple's on-device Speech framework can be added without a rewrite. Text is inserted at the cursor via the Accessibility API, with a clipboard-paste fallback for terminals. Every transcript is persisted to a local SQLite history store, and a glassmorphic "pill" HUD plus a settings dashboard — both Tauri webviews — surface live recording status and time-saved/WPM metrics, computed entirely locally with no telemetry.

Features

  • Toggle-hotkey mic dictation and system-audio (meeting) capture, sharing one pipeline
  • 100% on-device whisper.cpp inference, multilingual including Yoruba
  • Escape-to-cancel state machine with a resumable countdown
  • Cursor text-insertion via the Accessibility API, clipboard-paste fallback for terminals
  • Local-only history and a metrics dashboard (time saved, WPM, totals, activity)
  • Swappable transcription-engine adapter architecture, no plugin marketplace
.see also
view all projects
say hello