VibeTalk

Voice-to-text dictation CLI — speak and it types wherever your cursor is.

Works with any app: VS Code, Cursor, terminal, browser, Slack, or any text field.

Powered by Voxtral — a local AI transcription engine. No API key. No cloud. Fully private.

Quick Start

npx vibetalk

On first run, VibeTalk will automatically:

1. Clone and build the voxtral.c engine from source (~1 MB, one-time)
2. Download the Voxtral AI model weights (~8.9 GB, one-time, saved to ~/.vibetalk/voxtral/)

After setup, every start is instant — no internet needed, everything runs on your machine.

How It Works

1. Run npx vibetalk in a terminal
2. Press Enter to start recording
3. Speak — your words are typed in real-time wherever your cursor is focused
4. Press Enter to stop recording
5. Repeat as needed. Press Ctrl+C to quit.

Installation

# Use directly with npx (no install needed)
npx vibetalk

# Or install globally
npm install -g vibetalk
vibetalk

Platform Support

| Platform | Support | |---|---| | macOS (Apple Silicon — M1/M2/M3/M4) | ✅ Full support, Metal GPU acceleration | | macOS (Intel) | ✅ Full support, CPU | | Linux (x86_64) | ✅ Full support (requires build-essential + libopenblas-dev) | | Windows | ⚠️ Use WSL2 — see below |

Windows users

VibeTalk uses voxtral.c which does not compile natively on Windows. To use VibeTalk on Windows:

1. Open PowerShell as Administrator and run: wsl --install
2. Restart your PC and open the Ubuntu app
3. Inside WSL2, install Node.js and run: npx vibetalk

VibeTalk will show step-by-step WSL2 instructions automatically if you run it on Windows.

Linux setup

On first run, VibeTalk needs git, build-essential, and libopenblas-dev:

sudo apt-get install -y git build-essential libopenblas-dev
npx vibetalk

Configuration

# Change language (default: en)
vibetalk config set language es

# View all settings
vibetalk config list

# Reset to defaults
vibetalk config reset

Options

vibetalk [start]                Start dictation (default command)
  --device <index>             Audio input device index
  --language <lang>            Language (en, es, fr, de, etc.)
  --typing-strategy <s>        auto | robotjs | clipboard
  --typing-delay <ms>          Ms between keystrokes (default: 12)
  --use-hotkey                 Use global hotkey instead of Enter
  --hotkey <combo>             Hotkey combo (default: ctrl+shift+v)
  --verbose                    Debug logging

vibetalk config set <k> <v>    Set config value
vibetalk config get <key>      Get config value
vibetalk config list           Show all settings
vibetalk config reset          Reset to defaults
vibetalk devices               List audio input devices

Requirements

• Node.js 18+
• macOS or Linux (Windows via WSL2)
• A microphone
• git and a C compiler (make, cc)
• ~10 GB disk space for the model (one-time download)

Permissions

On macOS, you may need to grant:

• Microphone access to your terminal app
• Accessibility access for keyboard simulation (System Settings > Privacy & Security)

How It Works (Technical)

1. Captures microphone audio via PvRecorder (16kHz, 16-bit PCM)
2. Pipes raw audio to voxtral.c — a pure C implementation of the Voxtral realtime model
3. Receives finalised transcript lines from voxtral's stdout
4. Types confirmed text using robotjs
5. Falls back to clipboard paste if robotjs is unavailable

License

MIT