dML - Machine Learning Models

ML and NN training and inference optimized for Apple Silicon via MLX. Currently focused on voice.

All d* projects are entirely AI generated.

Structure

Directory	Description	Status
model_mlx_migration	MLX models for Apple Silicon	Usable
voice	Streaming voice I/O, 14 languages	Preview
metal_mps_parallel	Metal/MPS GPU threading	Preview

Note: model_mlx_migration will be split into separate TTS, STT, and tooling repos.

Models

Speech-to-Text (STT)

Model	Description
ZipFormer	Streaming ASR encoder (k2/icefall), 2.85% WER
Whisper large-v3-turbo	Non-streaming fallback, 1.8% WER
Silero VAD	Voice activity detection

Text-to-Speech (TTS)

Model	Description
CosyVoice3	Alibaba TTS with voice cloning, DiT flow-matching
Kokoro	Lightweight TTS, 38x realtime

Rich Audio

9 classification heads: emotion (92% acc), paralinguistics (97% acc), language ID (99% acc), pitch, singing detection, punctuation.

License

Apache 2.0