AangJnr/vox-africa-livekit

Local STT/TTS plugins for LiveKit Agents - no cloud APIs required # Local LiveKit Plugins **Run LiveKit voice agents with fully local STT and TTS - no cloud APIs required.** Custom plugins for LiveKit Agents that enable completely local speech processing using FasterWhisper for STT and Piper for TTS. ## Tested On - **OS:** Linux (Arch 6.17.7) - **GPU:** NVIDIA RTX 3060 (12GB VRAM) - **CUDA:** 12.x - **Python:** 3.10+ **Windows/Mac users:** Not yet tested. Community contributions welcome! Please report issues on GitHub. ## Why Local? | | Cloud | Local | |---|---|---| | **Quality** | Better | Good | | **Latency** | ~2.1s total | ~2.8s total | | **Cost** | ~$150/year* | ~$20/year | | **Privacy** | Data sent externally | Stays on your network | | **Control** | Vendor dependent | Full ownership | *Based on 100 hours/year: Deepgram Nova-2 ($0.35/hr) + Cartesia Sonic ($50/1M chars). ## Features - **FasterWhisperSTT** - GPU-accelerated speech-to-text - Multiple model sizes (tiny → large-v3) - ~230-460ms processing time on RTX 3060 - Configurable beam search and VAD - **PiperTTS** - Fast local text-to-speech - Multiple voice models available - ~9ms per character (~300ms for short responses) - Configurable speed, volume, pitch ## Quick Start ### Prerequisites **Required:** - uv (recommended) or pip - Python 3.10+ - Docker (for LiveKit server) - Ollama (for local LLM) - Must be running: `ollama serve` - ffmpeg or libavcodec (for audio processing) **For GPU Acceleration (recommended):** - NVIDIA GPU with 4GB+ VRAM (8GB+ recommended for larger Whisper models) - NVIDIA drivers with CUDA 11.8+ support - Note: PyTorch (~2GB download) includes bundled CUDA libraries ### 1. Clone and Install ```bash git clone github.com cd local-livekit-plugins # Install with uv (recommended) uv sync # Or with pip pip install -e ".[examples]" ``` ### 2. Download a Piper Voice Model ```bash mkdir -p models/piper && cd models/piper # Download Ryan (male US English, high quality) curl -LO https:/ …