Talk

A global push-to-talk voice-to-text tool with optional real-time translation. Press a hotkey, speak, and the transcribed (and optionally translated) text is pasted into whatever app has focus.

Features

• Push-to-talk — Hold hotkey to record, release to transcribe
• 4 STT engines — Whisper, Canary Qwen 2.5B, Voxtral Transcribe 2, Qwen3-ASR
• 3 translation engines — NLLB-200, SeamlessM4T v2, Madlad-400 (optional, defaults off)
• Floating overlay — Shows real-time waveform near cursor with status indicators
• Auto-paste — Transcribed text is pasted into the focused app via Ctrl+V
• System tray — Model selection, translation toggle, target language picker
• Configurable — JSON config for hotkey, models, UI appearance

Quick Start

Python Prototype

cd python
pip install -e ".[dev]"
python -m src.main

Configuration

Edit config.json in the project root:

{
  "hotkey": "ctrl+shift+space",
  "default_model": "whisper",
  "translation": {
    "enabled": false,
    "target_language": "es",
    "model": "nllb-200"
  }
}

Running Tests

cd python
pytest tests/ -v

Linting

cd python
ruff check src/ tests/
ruff format src/ tests/

Architecture

Hotkey Press → Start Recording → Show Overlay (waveform)
Hotkey Release → Stop Recording → STT Engine → [Translation] → Paste

STT Models

Engine	Type	Notes
Whisper (faster-whisper)	In-process	Default, requires faster-whisper
Canary Qwen 2.5B	HTTP	Requires NeMo server
Voxtral Transcribe 2	HTTP	Requires vLLM server
Qwen3-ASR	In-process	Requires qwen-asr

Translation Models

Engine	Languages	Notes
NLLB-200	200+	Best coverage for low-resource languages
SeamlessM4T v2	100+	Highest quality for common languages
Madlad-400	400+	Maximum language count

Project Structure

Talk/
├── config.json              # Shared config
├── python/                  # Python prototype
│   ├── src/
│   │   ├── main.py          # Entry point
│   │   ├── audio.py         # Mic capture + ring buffer
│   │   ├── hotkey.py        # Global hotkey manager
│   │   ├── overlay.py       # Floating waveform UI
│   │   ├── paste.py         # Clipboard + paste
│   │   ├── tray.py          # System tray
│   │   ├── config.py        # Config loader
│   │   ├── engines/         # STT engines
│   │   └── translation/     # Translation engines
│   └── tests/
├── csharp/                  # C# port (planned)
├── rust/                    # Rust port (planned)
└── servers/                 # Model server scripts

Hardware Target

Optimized for AMD Ryzen AI Max 395 with 128GB unified memory, but works on any machine with sufficient VRAM for the selected models.

License

MIT