lucasnewman/f5-tts-mlx

Leverages MLX for efficient inference on Apple Silicon, enabling ~4-second generation on M-series MacBooks through a non-autoregressive flow-matching architecture with diffusion transformers and ConvNeXT v2-based text alignment. Supports zero-shot voice cloning from reference audio, quantized 4-bit and 8-bit model variants for memory-constrained environments, and integrates seamlessly with MLX language models via stdout piping for end-to-end speech synthesis pipelines.

611 stars and 4,144 monthly downloads. Used by 2 other packages. No commits in the last 6 months. Available on PyPI.