lucasnewman/nanospeech

Implements end-to-end flow matching for joint text alignment and waveform generation without auxiliary models like forced aligners, with dual ~1,500-line implementations in PyTorch and MLX for experimental flexibility. The 82M parameter model trains efficiently on commodity hardware (H100 in days) using only public domain data, achieving 3-5x realtime inference on Apple Silicon and modern GPUs. Supports voice cloning from reference audio and integrates with WebDataset for scalable multi-GPU training via PyTorch Accelerate.

186 stars and 616 monthly downloads. No commits in the last 6 months. Available on PyPI.