💬 Talk to Me

A macOS menu bar utility for local speech-to-text, powered by Whisper and NVIDIA Parakeet. No cloud, no API key — everything runs on your machine.

Built with Tauri v2 (Rust + vanilla HTML/CSS/JS).

Features

• 🎙 Dictation anywhere — Press Alt+Space to record, press again to transcribe and inject text into any app
• 🤖 Multiple engines — Whisper (via whisper.cpp) and NVIDIA Parakeet (via ONNX Runtime), including multilingual TDT models
• 📦 Model management — Browse, download, and switch models from HuggingFace directly in the app
• 🔒 100% local — No data leaves your machine, no account required
• ⚡ Apple Silicon optimized — CoreML/Metal acceleration for fast inference

Quick start

# Prerequisites: Rust toolchain, Node.js
cargo tauri dev

On first launch, open the settings window to download a model. Recommended starting points:

Model	Engine	Size	Languages
Whisper Small	whisper.cpp	~244 MB	Multilingual
Whisper Large v3 Turbo	whisper.cpp	~1.5 GB	Multilingual
Parakeet CTC 0.6B	ONNX	~700 MB	English
Parakeet TDT 0.6B v3	ONNX	~2.5 GB	25 languages (EN, FR, DE, ES…)

Models are stored in ~/Library/Application Support/TalkToMe/models/.

How it works

Alt+Space → start recording (mic capture via cpal)
Alt+Space → stop recording
   → resample to 16kHz
   → compute mel spectrogram
   → run inference (Whisper or Parakeet)
   → inject text into active app (CGEvent or clipboard)

The overlay window shows recording state and transcription progress.

Architecture

src-tauri/src/          Rust backend
├── engine/             SttEngine trait → whisper_stt.rs, onnx_stt.rs
├── audio/              Mic capture, resampling, mel spectrogram (pure Rust)
├── commands/           Tauri IPC: STT, models, settings
├── hub/                HuggingFace API, downloads, model registry
├── hotkey/             Global shortcut dispatch
└── platform/           OS abstraction (TextInjector, TextSelector traits)

src/                    Vanilla JS frontend
├── index.html          Settings window (model management, preferences)
└── overlay.html        Floating recording/transcription overlay

Designed for future TTS support (Phase 6) and cross-platform portability (Windows/Linux via platform/ trait abstraction).

Requirements

• macOS 13+ (Ventura)
• Rust toolchain
• Node.js
• Microphone access permission
• Accessibility permission (for keystroke injection, optional — falls back to clipboard)

Build

cargo tauri build       # Production .dmg

⚠️ Without an Apple Developer certificate, users will need to right-click → Open on first launch.

License

MIT