much-miller

Local AI assistant with speech interface.

Named after Much the Miller's Son, one of Robin Hood's Merry Men - a loyal, humble companion rather than a flashy presence.

Status: Working Prototype

The core voice loop is now functional:

1. Wake word detection - openWakeWord (listens for "hey jarvis", "alexa", etc.)
2. TTS response - Piper says "Hello Romilly"
3. Speech transcription - RealtimeSTT with faster-whisper
4. Voice commands - "play radio 3", "play radio 4", "play news", "stop"
5. End keyword - Say "over" to return to wake word listening

Quick Start

# Activate virtual environment
source venv/bin/activate

# Run with device name (partial match)
python -m much_miller.main Samson

# Or set device in environment
export MUCH_MILLER_DEVICE=Samson
python -m much_miller.main

Environment Variables

Variable	Description
MUCH_MILLER_DEVICE	Audio input device name or index
MUCH_MILLER_MODEL_PATH	Path to Piper TTS model (ONNX file)

List Audio Devices

python -m much_miller.main
# Shows available input devices if no device specified

Architecture

┌─────────────────────────────────────────────────────────────┐
│                     much-miller (trend)                      │
├─────────────────────────────────────────────────────────────┤
│  openWakeWord    RealtimeSTT      Piper TTS    BBC Radio    │
│  (wake word)     (faster-whisper) (speech out) (mpv streams)│
│                  CPU inference                               │
└─────────────────────────────────────────────────────────────┘

Components

Component	Library	Purpose
Wake word	openWakeWord	Lightweight, runs on CPU, multiple wake words
Transcription	RealtimeSTT + faster-whisper	Real-time STT with VAD
TTS	Piper	Fast local text-to-speech
Radio	mpv	BBC radio streaming via HLS
Audio	sounddevice + PyAudio	Device enumeration and capture

Dependencies

openwakeword      # Wake word detection
RealtimeSTT       # Real-time speech-to-text
piper-tts         # Text-to-speech
sounddevice       # Audio device enumeration
pyaudio           # Audio capture for openWakeWord

Project Structure

src/much_miller/
├── main.py                 # Main entry point - wake word + transcription + commands
├── radio/
│   ├── ports/              # Abstract interfaces
│   │   └── radio_player.py
│   └── adapters/           # Implementations
│       ├── bbc_radio_player.py
│       └── fake_radio_player.py
└── wake_word/
    ├── ports/              # Abstract interfaces
    │   ├── audio_recorder.py
    │   └── speaker.py
    └── adapters/           # Implementations
        ├── sounddevice_recorder.py
        ├── piper_speaker.py
        ├── fake_recorder.py
        └── fake_speaker.py

Development

# Install dependencies
source venv/bin/activate
uv pip install -r requirements.txt
uv pip install -r requirements-test.txt

# Run tests
pytest

# Type check
pyright src/

Roadmap

See docs/much-rearchitecture-plan.md for the full architecture plan.

Completed (Phase 1)

• Validate OpenAI Whisper on polwarth (GPU)
• Test RealtimeSTT on trend (CPU)
• Test openWakeWord
• Clean up old VAD-based code
• Working main.py with wake word → TTS → transcription → "over" loop
• BBC Radio control (play/stop Radio 3, Radio 4, World Service, News)

Next Steps

• Train custom "hey much" wake word (blocked: openWakeWord training broken for Python 3.12)
• Build transcription-service (FastAPI on polwarth)
• Add intent routing (Claude API)

Related Projects

• the-great-dictator - Browser-based dictation app
• video-transcriber - Video → markdown transcription
• arxiv-librarian - Research paper monitoring
• john-dee - Personal knowledge search

Infrastructure

Host	Hardware	Role
trend	Intel Linux (no GPU)	Development, microphone, Much prototype
polwarth	Intel Linux, NVIDIA 3060	GPU transcription (24/7 once Ollama moves to xavier)

---

Updated 2 February 2026 - Radio control added