MMM-Assistant is a modular, offline-first voice assistant framework for MagicMirror².
It acts as the central orchestrator that connects hotword detection, speech recognition, local LLMs, and text-to-speech into a single assistant pipeline.

The project is designed with privacy, extensibility, and Raspberry Pi compatibility in mind.

• Run fully locally (no cloud dependency)
• Keep components loosely coupled
• Allow easy swapping of ASR, LLM, and TTS engines
• Integrate cleanly with MagicMirror’s module system
• Serve as an alternative base to cloud assistants

MMM-Assistant follows a pipeline-based architecture:

Hotword Detector
↓
Speech-to-Text (ASR)
↓
MMM-Assistant (Core Orchestrator)
↓
LLM (Local)
↓
Text-to-Speech (TTS)
↓
Audio Output

MMM-Assistant itself does not perform detection, transcription, or synthesis.
It coordinates these stages using MagicMirror notifications.

• Receive transcribed user input
• Maintain short-term conversational context
• Forward prompts to a local LLM
• Receive and store assistant responses
• Trigger TTS playback
• Act as the single “brain” of the assistant

MMM-Assistant relies on MagicMirror notification events.

Typical flow:

1. Hotword module signals listening state
2. ASR module sends final transcript
3. MMM-Assistant sends prompt to LLM backend
4. LLM response is returned
5. MMM-Assistant forwards text to TTS module

This approach keeps modules independent and replaceable.

• Local LLM via HTTP (e.g. Ollama)
• Notification-based TTS triggering
• Context-aware responses

• Streaming ASR + LLM responses
• Intent detection layer
• MQTT / Home Assistant integration
• Audio cues (listening, thinking, speaking)
• Wake-word state management

MMM-Assistant uses standard MagicMirror configuration.

Example:

{
  module: "MMM-Assistant",
  position: "top_left",
  config: {
    // Experimental – subject to change
  }
}

Configuration options will expand as the module stabilizes.

• Version: 0.0.1
• Status: Experimental
• Breaking changes expected
• Not production-ready

This release focuses on validating the architecture.

• Keep modules loosely coupled
• Prefer notifications over direct imports
• Avoid cloud-only dependencies
• Ensure Raspberry Pi compatibility

Pull requests and architectural discussions are welcome.

MIT