Moonshine Node

On-device speech-to-text CLI for Node.js using Moonshine models.

Features

• Real-time transcription - Speech-to-text with automatic voice activity detection
• File transcription - Transcribe WAV files directly from the CLI
• On-device processing - No data leaves your machine
• Multiple models - Choose between tiny (faster) or base (more accurate) models
• Streaming mode - Get partial transcriptions as you speak
• Linux audio support - Uses ALSA via @mastra/node-audio

Installation

npm install

Usage

# Basic usage (tiny model)
npx moonshine-node

# Transcribe a WAV file
npx moonshine-node --file audio.wav

# Use base model for better accuracy
npx moonshine-node --model base

# Transcribe one sentence and exit
npx moonshine-node --once

# Enable streaming/partial updates
npx moonshine-node --streaming

# Specify audio device (Linux/ALSA)
npx moonshine-node --device "sysdefault:CARD=Mini"

# List available audio devices
npx moonshine-node --list-devices

# Verbose output
npx moonshine-node --verbose

Options

| Option | Description | Default | | ---------------- | ---------------------------------- | --------- | | --help | Show help message | - | | --list-devices | List available audio input devices | - | | --file | WAV file to transcribe | - | | --device | ALSA device name | default | | --model | Model to use (tiny or base) | tiny | | --streaming | Enable streaming/partial updates | false | | --once | Transcribe one sentence and exit | false | | --verbose | Show detailed logs | false |

Controls

• Press q or Ctrl+C to quit

Models

• tiny (~14MB) - Faster, lower accuracy. Good for simple commands.
• base (~45MB) - Slower, better accuracy. Recommended for general use.

File Transcription

Transcribe WAV audio files directly:

npx moonshine-node --file audio.wav
npx moonshine-node --file audio.wav --model base

Supported format: 16-bit PCM WAV files (16kHz recommended, other sample rates are auto-resampled).

Requirements

• Node.js 18+
• Linux (for microphone input via ALSA)
• arecord (ALSA utility) installed

Architecture

• VAD: TEN VAD for voice activity detection
• STT: Moonshine ONNX models for speech-to-text
• Audio: @mastra/node-audio for microphone input
• Runtime: onnxruntime-node for inference

License

MIT