Developed and integrated Irodori-TTS v3 into the mlx-audio repository, focusing on enhancing text-to-speech capabilities with automatic output length estimation and faster inference. Leveraging Python and deep learning techniques, the work introduced a duration predictor and Sway Sampling, allowing the system to estimate and trim audio output automatically when duration is unspecified. The implementation included updates to configuration and model logic, as well as comprehensive documentation to support onboarding and maintenance. Backward compatibility with previous model checkpoints was maintained, ensuring seamless upgrades. This integration improved throughput and user experience for generated speech while supporting scalable, high-quality audio processing workflows.