Almost every open speech model ships assuming an NVIDIA GPU, so running TTS or transcription on a Mac usually means a slow CPU fallback or a cloud bill. This project rebuilds the whole speech stack — synthesis, transcription, and voice conversion — directly on Apple's MLX runtime, so a MacBook's unified memory does the work locally with no CUDA and no API keys.
What Sets It Apart
- Breadth over a single model: 20+ TTS engines (Kokoro, Qwen3-TTS, OmniVoice) and 15+ STT engines (Whisper, Qwen3-ASR, Parakeet) share one MLX-native interface, so swapping models is a config change rather than a rewrite.
- Quantization from 3-bit to 8-bit lets multi-billion-parameter voices fit in the RAM of a consumer Mac — the difference between "runs" and "won't load."
- An OpenAI-compatible REST endpoint means existing speech clients point at localhost unchanged, and a Swift package drops the same inference into iOS/macOS apps.
- Voice cloning, adjustable speed, and a 3D-visualizer web UI come bundled rather than as separate tools.
Who It's For
Great fit if you live on Apple Silicon and want local, private speech — offline transcription, on-device TTS for an app, or trying the latest voice models without renting a GPU. Look elsewhere if you deploy on Linux/NVIDIA servers, need Windows support, or depend on a model that hasn't been ported to MLX yet; the project tracks the MLX ecosystem and a single maintainer's release pace.
