Show HN: Talk to your Mac offline – sub-second Voice AI (Apple Silicon and MLX) (github.com)

A new open-source demo showcases an offline-first, full‑duplex voice assistant tuned for Apple Silicon that completes a full round trip (speech → LLM → TTS) in under one second while supporting barge‑in. It runs on-device (VAD, Whisper STT, an MLX LLM, and Kokoro TTS) sharing a single MLX runtime for efficient inference, streams sentence-level LLM outputs with immediate, cancellable TTS playback, and preserves context across interruptions with a rolling audio buffer. The project emphasizes privacy and real‑time responsiveness, making it a practical reference for building snappy, offline conversational agents. Technically, the demo combines a client-side AudioWorklet correlator for robust echo suppression and barge‑in detection with Pipecat SmartTurn end‑of‑utterance detection, Silero VAD for segmentation, and a FastAPI WebSocket server to orchestrate the pipeline. Code modules include audio_buffer.py, vad_detector.py, transcriber.py, llm_handler.py, and tts_handler, plus a browser UI (index.html, correlator.js). Requirements are pinned for MLX + Kokoro; it can be run locally with a Python venv and server.py. Licensed MIT, the project is a useful blueprint for low‑latency, privacy‑preserving voice agents and demonstrates how unified runtimes (MLX) on Apple Silicon can enable practical on‑device multimodal inference.