F5-TTS

Non-autoregressive flow-matching TTS from Kai Yu's X-LANCE lab (ACL 2025): a Diffusion Transformer that "fakes fluent and faithful speech" without duration models, text encoders, or phoneme alignment. One of the most-adopted open TTS models anywhere — ~798K monthly HuggingFace downloads, 15K★, and a large finetune ecosystem spanning dozens of languages (e.g. the Arabic Habibi-TTS, Jan 2026).