🎙️ Maya — A Fully Offline Voice Assistant

Talk to a local LLM. No cloud, no API keys, no data leaving your machine.

Maya is a 100% local, real-time voice assistant. Speech recognition, language reasoning, and speech synthesis all run on your own hardware — making it a practical playground for LLMs on constrained / edge devices.

---

How It Works

🎤 Mic  →  Silero VAD  →  faster-whisper (STT)  →  Ollama LLM  →  Kokoro (TTS)  →  🔊 Speaker
          (detects        (transcribes              (generates       (speaks the
           speech)         to text)                  a reply)         response)

The pipeline listens continuously, detects when you start and stop speaking using voice-activity detection, transcribes your speech, sends it to a local LLM with conversational memory, and speaks the reply back — all without an internet connection.

---

Features

• Real-time voice-activity detection — Silero VAD records only when you're actually speaking and stops on silence
• Accurate offline transcription — faster-whisper (small, int8) tuned for low latency on CPU
• Local LLM reasoning — Ollama (mistral:7b) with a witty assistant persona and 10-turn conversational memory
• Natural offline speech — Kokoro TTS (af_bella voice) for the response
• Text mode fallback — flip a flag to type instead of speak

---

Project Structure

maya_v1/ — the working voice assistant

File	Role
listener.py	Silero VAD + faster-whisper — captures mic audio and transcribes speech
ollama_connection.py	Main loop: STT → LLM → TTS with chat history (run this)
ausdio_tts.py	Kokoro TTS — synthesizes and plays the reply
load.py	Startup animation helper

maya_v2/ — experimental agentic features

File	Role
langgraph_test.py	LangGraph + Ollama agent workflow with tool calling
apps_test.py	Desktop app launching via AppOpener
master.py	LangChain prompt-pipeline demo

---

Setup

# 1. Install Ollama and pull the model
ollama pull mistral:7b

# 2. Install Python dependencies
pip install -r requirements.txt

# 3. Run the assistant
python maya_v1/ollama_connection.py

Windows note: pyaudio and sounddevice may need platform-specific wheels and working audio drivers. faster-whisper downloads its model on first run.

---

Configuration

Key knobs live at the top of maya_v1/listener.py:

Setting	Default	Notes
WHISPER_MODEL	small	tiny/base/small/medium/large-v3
WHISPER_DEVICE	cpu	switch to cuda for GPU
VAD_THRESHOLD	0.5	higher = stricter speech detection
SILENCE_DURATION	1.2s	silence before recording stops

Swap the LLM by changing model = 'mistral:7b' in ollama_connection.py.

---

Roadmap

• Merge v2's LangGraph agent into the main voice loop (tool use by voice)
• Streaming TTS for lower response latency
• Wake-word detection