TomChat

Fast, local speech-to-text with a global hotkey (CLI version)

TomChat is a lightweight Linux CLI application that converts speech to text and types it into any application. Press a hotkey, speak, and your words appear. All processing happens locally - no cloud services required.

Features

• Local Speech Recognition - NVIDIA Parakeet TDT 0.6B with ~6% WER (better than Whisper)
• Voice Activity Detection - Silero VAD auto-stops recording after silence
• Global Hotkey - Works in any application (Caps Lock by default)
• Text Refinement - Optional Ollama integration for fixing transcription errors
• Lightweight - Pure Rust CLI, minimal footprint

Requirements

• OS: Linux (Ubuntu 22.04+ recommended)
• RAM: 4GB minimum
• Disk: ~200MB for models (INT8)

Quick Start

1. Install Dependencies

sudo apt update
sudo apt install -y \
    build-essential \
    libasound2-dev \
    libssl-dev \
    pkg-config \
    wget

# Install Rust
curl --proto '=https' --tlsv1.2 -sSf https://sh.rustup.rs | sh
source ~/.cargo/env

2. Clone and Build

git clone https://github.com/sixteen-dev/tomchat.git
cd tomchat
cargo build --release

3. Download Models

# Run the download script
./scripts/download-parakeet.sh

This downloads:

• Parakeet TDT 0.6B v2 (INT8) - ~180MB speech recognition model
• Silero VAD - ~2MB voice activity detection model

After downloading, your models/ directory should contain:

models/
├── sherpa-onnx-nemo-parakeet-tdt-0.6b-v2-int8/
│   ├── encoder.int8.onnx
│   ├── decoder.int8.onnx
│   ├── joiner.int8.onnx
│   └── tokens.txt
└── silero_vad.onnx

4. Run

# Use the wrapper script (sets library path automatically)
./tomchat

# Or run directly with library path
LD_LIBRARY_PATH=./target/release ./target/release/tomchat

5. Use

1. Press Caps Lock to start recording
2. Speak into your microphone
3. Wait for auto-stop (1.5s silence) or press Caps Lock again
4. Text appears in the focused application

Configuration

Edit config.toml to customize:

[hotkey]
combination = "caps"  # Options: "caps", "ctrl+shift+space", "f24", etc.

[vad]
model_path = "./models/silero_vad.onnx"
sensitivity = "Normal"  # Low, Normal, High, VeryHigh
timeout_ms = 1500       # Auto-stop after this much silence
auto_stop = true        # Set false for manual stop only

[speech]
model_dir = "./models/sherpa-onnx-nemo-parakeet-tdt-0.6b-v2-int8"
language = "en"

[text_refinement]
enabled = false         # Enable for Ollama-based text cleanup
model_name = "gemma3:1b"
ollama_url = "http://localhost:11434"

Environment Variables

export TOMCHAT_MODEL_DIR="/path/to/models"
export TOMCHAT_HOTKEY="ctrl+alt+c"

Text Refinement (Optional)

TomChat can use Ollama to fix transcription errors:

# Install Ollama
curl -fsSL https://ollama.com/install.sh | sh

# Pull a small model
ollama pull gemma3:1b

# Enable in config.toml
[text_refinement]
enabled = true

This fixes issues like:

• "cooper nettys" → "Kubernetes"
• Missing punctuation and capitalization

Manual Model Download

If the script doesn't work, download manually:

mkdir -p models
cd models

# Parakeet model
wget https://github.com/k2-fsa/sherpa-onnx/releases/download/asr-models/sherpa-onnx-nemo-parakeet-tdt-0.6b-v2-int8.tar.bz2
tar xjf sherpa-onnx-nemo-parakeet-tdt-0.6b-v2-int8.tar.bz2
rm sherpa-onnx-nemo-parakeet-tdt-0.6b-v2-int8.tar.bz2

# Silero VAD
wget https://github.com/k2-fsa/sherpa-onnx/releases/download/asr-models/silero_vad.onnx

Troubleshooting

Hotkey Not Working

# Try a different hotkey in config.toml
combination = "ctrl+shift+space"

# Or use a function key
combination = "f24"

Audio Issues

# List audio devices
arecord -l

# Test recording
arecord -d 3 test.wav && aplay test.wav

Library Not Found

# Always use the wrapper script
./tomchat

# Or set LD_LIBRARY_PATH manually
LD_LIBRARY_PATH=./target/release ./target/release/tomchat

Debug Mode

RUST_LOG=debug ./tomchat

Project Structure

tomchat/
├── src/
│   ├── main.rs           # Entry point
│   ├── app.rs            # Application logic
│   ├── config.rs         # Configuration handling
│   ├── audio/
│   │   ├── mod.rs        # Audio capture
│   │   └── vad.rs        # Voice activity detection
│   ├── speech/
│   │   └── transcriber.rs # Parakeet transcription
│   └── text_refinement/  # Optional Ollama integration
├── scripts/
│   └── download-parakeet.sh
├── config.toml           # Default configuration
├── Cargo.toml
└── README.md

Performance

Metric	Value
Startup	~2-3s
Transcription	0.5-2s
Memory	~1.5GB
Model Size	~200MB (INT8)

Tech Stack

• sherpa-rs - Rust bindings for sherpa-onnx
• cpal - Cross-platform audio
• global-hotkey - System-wide hotkeys
• enigo - Text injection
• tokio - Async runtime

Related

• tomchat-app - GUI version with Tauri

License

MIT License - see LICENSE

---

Named after my dog Tommy - Built with Rust + sherpa-rs