🎙️ A high-performance, privacy-focused voice assistant that runs entirely on your Mac, optimized for Apple Silicon.

✨ Real-time Voice AI • 🔒 100% Local • ⚡ Apple Silicon Optimized

MacEcho is a sophisticated voice assistant built from the ground up for macOS, leveraging the power of Apple Silicon's Neural Engine and unified memory architecture. It provides real-time voice interaction with complete privacy - all processing happens locally on your device.

• macOS 12.0 or later (Apple Silicon recommended)
• Python 3.9+
• Homebrew (for audio dependencies)

# Install audio dependencies
brew install portaudio

# Clone the repository
git clone https://github.com/realtime-ai/mac-echo.git
cd mac-echo

# Create virtual environment
python3 -m venv venv
source venv/bin/activate

# Install requirements
pip install -r requirements.txt

# Basic voice assistant (runs continuously)
python examples/agent_usage_example.py --mode=run

# Demo mode (30-second demo)
python examples/agent_usage_example.py --mode=demo

# Custom configuration
MACECHO_LLM__MODEL_NAME="mlx-community/Qwen2.5-7B-Instruct-4bit" python examples/agent_usage_example.py

• 100% Local Processing - Complete privacy with no cloud dependencies
• Apple Silicon Optimized - Leverages MLX framework for maximum performance on M-series chips
• Real-time Streaming - Sub-second response times with streaming audio pipeline
• Multilingual Support - Automatic language detection (English, Chinese, Japanese, Korean)
• Context-Aware Conversations - Maintains conversation history across interactions

• Voice Activity Detection (VAD) - Silero VAD for accurate speech detection with configurable thresholds
• Speech Recognition (ASR) - SenseVoice model with excellent accuracy and language auto-detection
• Neural Language Models - Qwen model family via MLX with 4-bit quantization support
• Text-to-Speech (TTS) - CosyVoice for natural-sounding speech synthesis with multiple voices

• Event-Driven Messaging - Asynchronous message passing with priority queues
• Modular Pipeline Design - Easily swap or extend components
• Streaming Sentencizer - Real-time sentence boundary detection for immediate TTS
• Interrupt Handling - Graceful interruption during speech generation
• Frame-Based Processing - 32ms audio frames for ultra-low latency
• Comprehensive Configuration - Flexible Pydantic-based settings management

┌─────────────┐     ┌─────────────┐     ┌─────────────┐     ┌─────────────┐
│   Audio In  │ ──> │     VAD     │ ──> │     ASR     │ ──> │     LLM     │
└─────────────┘     └─────────────┘     └─────────────┘     └─────────────┘
                                                                     │
                                                                     v
┌─────────────┐     ┌─────────────┐     ┌─────────────┐     ┌─────────────┐
│  Audio Out  │ <── │     TTS     │ <── │ Sentencizer │ <── │   Response  │
└─────────────┘     └─────────────┘     └─────────────┘     └─────────────┘

                          ▲                       │
                          └───── Message Bus ─────┘
                               (Event-Driven)

MacEcho uses a hierarchical configuration system with multiple sources:

1. Command-line arguments
2. Environment variables (MACECHO_ prefix)
3. .env file
4. Configuration files (JSON/YAML)
5. Default values

Create a .env file in the project root:

# Core Settings
MACECHO_DEBUG=false
MACECHO_APP_NAME="My Assistant"

# Audio Configuration
MACECHO_AUDIO_RECORDING__SAMPLE_RATE=16000
MACECHO_AUDIO_RECORDING__CHANNELS=1

# Model Selection
MACECHO_LLM__MODEL_NAME="mlx-community/Qwen2.5-7B-Instruct-4bit"
MACECHO_LLM__MAX_TOKENS=1000
MACECHO_LLM__TEMPERATURE=0.7

# Voice Settings
MACECHO_VAD__THRESHOLD=0.7
MACECHO_TTS__VOICE_ID="中文女"

See Configuration Guide for detailed options.

from macecho.agent import Agent
from macecho.config import get_config

# Load configuration
config = get_config()

# Create and run agent
agent = Agent(config)
agent.run()

from macecho.llm import MLXQwenChat

# Create chat model with context
chat = MLXQwenChat(
    model_name="mlx-community/Qwen2.5-14B-Instruct-4bit",
    context_enabled=True,
    max_context_rounds=10,
    system_prompt="You are a helpful coding assistant."
)

# Stream response
response = chat.chat_with_context(
    user_message="How do I sort a list in Python?",
    stream=True
)

from macecho.message import MessageQueue, MessageType, MessagePriority

# Create message queue
queue = MessageQueue()

# Subscribe to ASR messages
@queue.subscribe(MessageType.ASR)
async def handle_transcription(message):
    print(f"Transcribed: {message.data['text']}")

# Send high-priority message
await queue.send_message(
    MessageType.LLM,
    {"text": "Process this urgently"},
    priority=MessagePriority.HIGH
)

On Apple Silicon (M1/M2/M3):

• First response: < 1 second
• VAD latency: < 50ms per frame
• ASR processing: ~200ms for 3-second audio
• LLM token generation: 20-50 tokens/second (model dependent)
• TTS synthesis: Real-time factor < 0.3

Audio Input Not Working

# List audio devices
python -c "import pyaudio; p=pyaudio.PyAudio(); print([p.get_device_info_by_index(i)['name'] for i in range(p.get_device_count())])"

# Set specific device
export MACECHO_AUDIO_RECORDING__DEVICE_INDEX=1

Model Download Issues

# Models are cached in ~/.cache/modelscope/
# Clear cache if corrupted
rm -rf ~/.cache/modelscope/hub/iic/SenseVoiceSmall

# Set custom model directory
export MACECHO_STORAGE__MODELS_DIR=/path/to/models

Memory Issues

• Use smaller quantized models (4-bit recommended)
• Reduce context window size
• Disable model warmup for testing

We welcome contributions! Areas of interest:

• 🎯 Additional language model support
• 🌍 More languages for ASR/TTS
• 🔊 Alternative TTS engines
• 🧪 Test coverage improvements
• 📚 Documentation enhancements

MIT License - see LICENSE file for details.

MacEcho builds upon excellent open-source projects:

• MLX - Apple's machine learning framework
• SenseVoice - Speech recognition
• CosyVoice - Text-to-speech
• Silero VAD - Voice activity detection

Made with ❤️ for the Mac community