The latest release of swivid/f5-tts is 1.1.22. F5-TTS is the official code repository for a TTS (text-to-speech) project that generates fluent, faithful speech using Flow Matching, an approach that learns continuous data as a single flow and generates it directly.
The repository contains two model paths: F5-TTS, a Diffusion Transformer with ConvNeXt V2 applied, and E2 TTS, a Flat-UNet Transformer form reproduced close to the paper implementation. Sway Sampling is introduced in the README as a flow step sampling strategy that improves performance at inference time. Gradio and CLI are supported for speech generation, multi-style and multi-speaker generation, and voice chat, and the repository also includes training and fine-tuning guidance along with Docker usage instructions. The implementation language is Python and the license is MIT.
However, what changed in 1.1.22 is unknown because no release notes are given. The effect of Sway Sampling is also described only as improving performance, without any numerical comparison. The recorded scores include only momentum 3.9, and the novelty, technique, and quality scores are empty, so there is no remaining basis for judging which of the two models is better. All that remains as evidence is the supported features and the license scope. It is too early to evaluate using the extent of improvement in 1.1.22 as the basis.