misotts-mlx

Synthesize high-fidelity English speech and perform zero-shot voice cloning locally on macOS GPU via MLX. Use when a user requests text-to-speech, speaker voice cloning with an audio reference, custom speaker indexing, adjusting sampling parameters, or bypassing watermarks.

Evaluate, transcribe, and mathematically verify synthesized audio and speech outputs. Use when measuring Log-Mel Spectrogram MAE, Temporal Envelope Correlation, or when running Vertex AI Gemini or offline Gemma 4 multimodal quality evaluations.

Port, migrate, and validate PyTorch neural network modules to Apple Silicon MLX format. Use when translating attention layers, customizing Llama RoPE scale factors, setting up step-by-step autoregressive KV caches, or mapping PyTorch checkpoint weights to MLX safetensors.