asiff00/On-Device-Speech-to-Speech-Conversational-AI

Implements a modular multi-threaded pipeline combining Pyannote VAD, Whisper speech recognition, Ollama/LM Studio language models, and Kokoro TTS—all locally on CPU. Uses priority-based text chunking that streams TTS output as language model tokens arrive, and strategically injects filler words into LLM prompts to reduce perceived latency by 50-70%. Achieves ~1.5s end-to-end response time on mid-range hardware through queue-based inter-component coordination and user interrupt handling.

242 stars.