On-device voice AI for Apple Silicon

Streaming speech recognition, neural voice synthesis with visemes, and tool-calling chat — all running locally on MLX.

Built on MLX · FastAPI · Next.js 16

Service status

STT…TTS…LLM…

Speech-to-Text

Whisper-class streaming ASR over WebSocket. Word-level timestamps, multilingual.

Text-to-Speech

Qwen3-TTS and Kokoro voices with Preston-Blair visemes for lipsync.

Local Chat

Qwen3-4B with OpenAI-compatible tool calling. ~104 tok/s on M-series.