OpenCode Speech-to-Text Plugin

Speech-to-text plugin for OpenCode with support for Moonshine (fast, edge-optimized) and Whisper (OpenAI's model) backends.

Features

• Voice Input Tool - Record audio from microphone and transcribe to text
• Automatic Silence Detection - Recording stops when you stop speaking
• Multiple Backends - Moonshine (recommended), Whisper, or Faster-Whisper
• Privacy-First - All processing is done locally, no cloud APIs
• Configurable - Choose model size, language, and recording duration

Quick Start

1. Install the Plugin

Add to your opencode.json:

{
  "plugin": ["opencode-speech-to-text"]
}

2. Install Python Dependencies

# Base dependencies (required)
uv pip install sounddevice soundfile numpy

# Choose ONE backend:

# Option A: Moonshine (Recommended - fastest, smallest)
uv pip install useful-moonshine-onnx@git+https://github.com/moonshine-ai/moonshine.git#subdirectory=moonshine-onnx

# Option B: Whisper (OpenAI's original)
uv pip install openai-whisper

# Option C: Faster-Whisper (Optimized Whisper)
uv pip install faster-whisper

3. Grant Microphone Permissions

Ensure your terminal application has microphone access in your system settings.

Usage

Using the Tool

OpenCode will automatically have access to the voice_input tool:

You: Record my voice input
Assistant: I'll record your voice now. Please speak...
[Recording starts, stops on silence]
Assistant: I heard: "Please help me refactor the authentication module"

Using the Skill

Type /voice to get guidance on using voice input.

Checking Setup

Use the voice_check tool to verify your setup:

You: Check my voice setup
Assistant: [Uses voice_check tool]
Available STT backends: moonshine
Current configuration:
- Backend: auto
- Model: tiny
- Language: en
- Max duration: 30s

Configuration

Configure via environment variables:

Variable	Description	Default
STT_BACKEND	Backend: moonshine, whisper, faster-whisper, auto	auto
STT_MODEL	Model size: tiny, base, small, medium, large	tiny
STT_LANGUAGE	Language code	en
STT_MAX_DURATION	Max recording seconds	30
STT_PYTHON_PATH	Path to Python	python3

Model Comparison

Backend	Model	Size	Speed	Quality
Moonshine	tiny	27M (190MB)	Fastest	Good
Moonshine	base	62M (400MB)	Very Fast	Better
Whisper	tiny	39M	Slow	Good
Whisper	base	74M	Slower	Better
Whisper	small	244M	Much Slower	Great

Recommendation: Use Moonshine tiny for best balance of speed and accuracy.

Supported Languages

Moonshine

• Arabic (ar), Chinese (zh), English (en), Japanese (ja), Korean (ko), Spanish (es), Ukrainian (uk), Vietnamese (vi)

Whisper

• 99+ languages (see OpenAI Whisper)

System Requirements

macOS

brew install portaudio ffmpeg

Ubuntu/Debian

sudo apt install -y portaudio19-dev python3-dev ffmpeg

Fedora

sudo dnf install portaudio-devel python3-devel ffmpeg

Troubleshooting

"No STT backend available"

Install one of the supported backends:

uv pip install useful-moonshine-onnx@git+https://github.com/moonshine-ai/moonshine.git#subdirectory=moonshine-onnx

"No microphone found"

1. Check system permissions for your terminal
2. Verify microphone is connected: python -c "import sounddevice; print(sounddevice.query_devices())"

Slow transcription

1. Use Moonshine instead of Whisper
2. Use tiny model instead of larger variants
3. Ensure you have a capable CPU/GPU

Development

# Install dependencies
bun install

# Build
bun run build

# Type check
bun run typecheck

License

MIT

Credits

• Moonshine - Fast ASR for edge devices
• OpenAI Whisper - Robust speech recognition
• VoiceMode MCP - Inspiration for the voice interface