A fully local, privacy-first voice assistant: speech-to-text, LLM reasoning, text-to-speech, and real-time voice cloning. Zero data leaves your machine.
Mic → faster-whisper (STT) → Ollama/Llama3 (LLM) → Edge TTS (TTS) → Speaker
↕
OpenVoice (voice clone)
Most voice assistants send your audio to remote servers. This project runs the entire pipeline on your hardware:
faster-whisperfor transcription (local)Ollamawith any open model (local)Edge TTSfor synthesis (minimal network, no data retention)OpenVoicefor voice cloning (local checkpoint)
- Wake word: activates on "Hey Agent"
- Real-time voice cloning: adapts to your voice from a reference file or live samples
- Smart audio quality filter: only high-quality samples (SNR + clarity score) used for cloning
- Personality modes: assistant / ami / expert / humour (switchable by voice command)
- Gradio web interface: visual feedback, conversation history, mic level indicator
- Conversation export: save transcripts to file
- Auto-reconnect: recovers from Ollama crashes
- Python 3.10+
- Ollama installed and running
- FFmpeg
- GPU recommended (faster-whisper + OpenVoice benefit from CUDA)
# 1. Install Ollama and pull a model
ollama pull llama3.2:3b # or llama3.2:1b for CPU
# 2. Clone and install
git clone https://github.com/Wassbdr/voicecall
cd voicecall
python -m venv .venv && source .venv/bin/activate
pip install -r requirements.txt
# 3. Place OpenVoice checkpoints
# openvoice_model/checkpoints/ ← download from OpenVoice repo
# 4. Run
ollama serve &
python agent.py
# → open http://localhost:7860Edit agent.py to tune:
| Variable | Default | Description |
|---|---|---|
OLLAMA_MODEL |
llama3.2:1b |
Ollama model to use |
WHISPER_MODEL |
tiny |
faster-whisper model size |
WAKE_WORD |
"hey agent" |
Wake word phrase |
SILENCE_THRESHOLD |
500 |
Mic amplitude threshold |
MIN_SAMPLES |
2 |
Min voice samples for cloning |
| Component | Library |
|---|---|
| STT | faster-whisper |
| LLM | ollama (Llama 3, Mistral, …) |
| TTS | edge-tts |
| Voice cloning | OpenVoice |
| Web UI | gradio |
MIT