huggingface/speech-to-speech

Implements a cascaded four-stage pipeline (VAD → STT → LLM → TTS) with pluggable components across all stages, supporting models from Hugging Face Hub, MLX for Apple Silicon acceleration, and external libraries like Whisper, MeloTTS, and ChatTTS. Deployable locally, via TCP sockets, or WebSockets with multi-language support and streaming capabilities. Optimizes for low-latency inference on consumer hardware including CUDA and Apple Silicon with torch.compile support.

4,541 stars. Actively maintained with 90 commits in the last 30 days.