agentvox 0.2.0

Edge-based voice assistant using Gemma LLM with STT and TTS capabilities

🎙️ AgentVox

Edge-based voice assistant using Gemma LLM with real-time Speech-to-Text and Text-to-Speech capabilities

🆕 What's New (v0.2.0)

• Real-time Streaming STT/TTS: Replaced traditional STT/TTS with RealtimeSTT and RealtimeTTS for lower latency
• Live Transcription: See what you're saying in real-time as you speak
• Streaming TTS: Faster response times with streaming audio synthesis
• Improved Performance: Better voice activity detection and faster processing
• Configurable TTS Speed: Adjust speech speed with --tts-speed parameter

Key Features

• Real-time Speech Recognition (STT): Live transcription using RealtimeSTT with Whisper
• Conversational AI (LLM): Local LLM based on Llama.cpp (Gemma 3 12B)
• Streaming Speech Synthesis (TTS): Real-time voice synthesis with RealtimeTTS using Coqui XTTS v2
• Complete Offline Operation: All processing is done locally, ensuring privacy
• Voice Cloning: Clone any voice with a short audio sample

Installation

1. Install via pip

pip install agentvox

Or install from source:

git clone https://github.com/yourusername/agentvox.git
cd agentvox
pip install -e .

For NVIDIA CUDA Users

If you have an NVIDIA GPU and want to use CUDA acceleration, you need to rebuild llama-cpp-python with CUDA support:

# Rebuild llama-cpp-python with CUDA support
CMAKE_ARGS="-DGGML_CUDA=on" pip install llama-cpp-python --upgrade --force-reinstall --no-cache-dir

This will significantly improve LLM inference performance on NVIDIA GPUs.

2. Download Model

# Automatically download Gemma model (~7GB)
agentvox --download-model

The model will be saved in ~/.agentvox/models/ directory.

Usage

Basic Usage

# Start voice conversation
agentvox

Speak into your microphone and the AI will respond with voice.

Voice Configuration

# Voice cloning with Coqui-TTS (default engine)
agentvox --speaker-wav speaker_sample.wav

# Record your own voice sample for cloning
agentvox --record-speaker
# Then use the recorded sample
agentvox --speaker-wav speaker_ko.wav

# Adjust TTS speed (1.0 is normal, higher is faster)
agentvox --tts-speed 1.5  # 50% faster
agentvox --tts-speed 1.3  # 30% faster (default)
agentvox --tts-speed 1.0  # Normal speed
agentvox --tts-speed 0.8  # 20% slower

Advanced Configuration

STT (Speech Recognition) Parameters

# Recognize speech in different languages
agentvox --stt-language en

# Increase beam size for more accurate recognition (default: 5)
agentvox --stt-beam-size 10

# Adjust VAD sensitivity (default: 0.5)
agentvox --stt-vad-threshold 0.3

# Adjust minimum speech duration in ms (default: 250)
agentvox --stt-vad-min-speech-duration 200

# Adjust minimum silence duration in ms (default: 1000)
agentvox --stt-vad-min-silence-duration 800

# Change Whisper model size (tiny, base, small, medium, large)
agentvox --stt-model small

LLM (Language Model) Parameters

# Generate longer responses (default: 512)
agentvox --llm-max-tokens 1024

# More creative responses (higher temperature, default: 0.7)
agentvox --llm-temperature 0.9

# More conservative responses (lower temperature)
agentvox --llm-temperature 0.3

# Adjust context size (default: 4096)
agentvox --llm-context-size 8192

# Adjust top-p sampling (default: 0.95)
agentvox --llm-top-p 0.9

Device Configuration

# Auto-detect best available device (default)
agentvox

# Explicitly use CPU
agentvox --device cpu

# Explicitly use CUDA GPU
agentvox --device cuda

# Explicitly use Apple Silicon MPS
agentvox --device mps

The system automatically detects the best available device:

• NVIDIA GPU with CUDA → cuda
• Apple Silicon → mps
• Otherwise → cpu

Combined Examples

# English recognition + fast speech + longer responses
agentvox --stt-language en --tts-speed 1.5 --llm-max-tokens 1024

# High accuracy STT + creative responses + voice cloning
agentvox --stt-beam-size 10 --llm-temperature 0.9 --speaker-wav voice_sample.wav

# Use custom model path with fast TTS
agentvox --model /path/to/your/model.gguf --tts-speed 1.4

Python API Usage

from agentvox import VoiceAssistant, ModelConfig, AudioConfig

# Configuration
model_config = ModelConfig(
    stt_model="base",
    llm_temperature=0.7,
    tts_speed=1.0,  # Adjust TTS speed
    speaker_wav="voice_sample.wav"  # Optional: voice cloning
)

audio_config = AudioConfig()

# Initialize voice assistant
assistant = VoiceAssistant(model_config, audio_config)

# Start conversation
assistant.run_conversation_loop()

Using Individual Modules

from agentvox import STTModule, LLMModule, TTSModule, ModelConfig

config = ModelConfig()

# STT (Speech to Text)
stt = STTModule(config)
text = stt.transcribe("audio.wav")

# LLM (Generate text response)
llm = LLMModule(config)
response = llm.generate_response(text)

# TTS (Text to Speech)
tts = TTSModule(config)
tts.speak(response)

Available Commands During Conversation

• "exit" or "종료": Exit the program
• "reset" or "초기화": Reset conversation history
• "history" or "대화 내역": View conversation history

System Requirements

• Python 3.8 or higher
• macOS (with MPS support), Linux, Windows
• Minimum 8GB RAM (16GB recommended)
• Approximately 7GB disk space (for model storage)

Required Packages

• torch >= 2.0.0
• realtimestt (Real-time speech-to-text)
• realtimetts[coqui] (Real-time text-to-speech with Coqui engine)
• llama-cpp-python
• numpy
• pygame
• sounddevice
• soundfile
• pyaudio
• hangul-romanize (for Korean language support)

Project Structure

agentvox/
├── agentvox/              # Package directory
│   ├── __init__.py               # Package initialization
│   ├── voice_assistant.py        # Main module
│   ├── cli.py                    # CLI interface
│   └── record_speaker_wav.py     # Voice recording module
├── setup.py                      # Package setup
├── pyproject.toml               # Build configuration
├── requirements.txt             # Dependencies
├── README.md                    # Documentation
└── .gitignore                   # Git ignore file

Troubleshooting

First Run Model Download

On first run, the Coqui TTS model (XTTS v2, ~1.86GB) will be automatically downloaded. This only happens once.

Multiprocessing Errors on macOS/Windows

If you encounter multiprocessing errors, ensure your script uses:

if __name__ == "__main__":
    # Your code here

ctranslate2 Version Issues

If you get ctranslate2 compatibility errors:

pip install ctranslate2==4.4.0

PyAudio Installation Error

macOS:

brew install portaudio
pip install pyaudio

Linux:

sudo apt-get install portaudio19-dev python3-pyaudio
pip install pyaudio

Windows:

# Visual Studio Build Tools required
pip install pipwin
pipwin install pyaudio

Out of Memory

For large LLM models:

• Use smaller quantized models
• Reduce context size: --llm-context-size 2048
• Use CPU mode: --device cpu

Microphone Recognition Issues

• Check microphone permissions in system settings
• Close other audio applications
• Adjust VAD threshold: --stt-vad-threshold 0.3
• Reduce silence duration for faster response: --stt-vad-min-silence-duration 500

Speaker Echo/Feedback Issues

If the TTS output is being picked up by the microphone:

• Use headphones instead of speakers
• Reduce speaker volume
• Increase the distance between microphone and speakers
• The system automatically pauses STT during TTS playback to minimize echo

Model File Not Found

# Download model
agentvox --download-model

# Or download directly
wget https://huggingface.co/tgisaturday/Docsray/resolve/main/gemma-3-12b-it-GGUF/gemma-3-12b-it-Q4_K_M.gguf \
  -O ~/.agentvox/models/gemma-3-12b-it-Q4_K_M.gguf

Performance Optimization

Improve Response Speed

1. Use smaller STT model: --stt-model tiny or base
2. Limit LLM response length: --llm-max-tokens 256
3. Reduce beam size: --stt-beam-size 3

GPU Acceleration

• macOS: Automatic MPS support (--device mps)
• NVIDIA GPU: CUDA support (--device cuda)
• AMD GPU: Requires PyTorch with ROCm support

Developer Information

Developed by MimicLab at Sogang University

License

This project is licensed under the MIT License - see the LICENSE file for details.

Third-Party Licenses

This project uses several third-party libraries:

• RealtimeSTT: MIT License (for real-time speech-to-text)
• RealtimeTTS: MIT License (for real-time text-to-speech)
• coqui-tts: Mozilla Public License 2.0 (used by RealtimeTTS for voice synthesis)
• faster-whisper: MIT License (used by RealtimeSTT for speech recognition)
• llama-cpp-python: MIT License (for LLM inference)
• Gemma Model: Check the model provider's license terms

For complete third-party license information, see THIRD_PARTY_LICENSES.md.

Note on copyleft licenses:

• The coqui-tts library (MPL-2.0) is used as a dependency through RealtimeTTS. The MPL-2.0 license only requires that modifications to coqui-tts itself be shared, not your application code.
• The MPL-2.0 license of coqui-tts does not affect the MIT licensing of this project's source code.

Contributing

Issues and Pull Requests are always welcome!

Development Setup

# Clone repository
git clone https://github.com/yourusername/agentvox.git
cd agentvox

# Install in development mode
pip install -e .

# Run tests
python -m pytest tests/

Multilingual Support

AgentVox supports multiple languages through both STT and TTS:

Speech Recognition (STT)

Set the language with --stt-language:

• Korean: ko (default)
• English: en
• Japanese: ja
• Chinese: zh
• Spanish: es
• French: fr
• German: de
• And many more...

Text-to-Speech (TTS)

The Coqui XTTS v2 model supports multiple languages automatically. For best results:

• Use voice cloning with a native speaker's voice sample
• The model will automatically detect and use the appropriate language

Example:

# English conversation with cloned voice
agentvox --stt-language en --speaker-wav english_voice.wav

# Japanese conversation
agentvox --stt-language ja --speaker-wav japanese_voice.wav