F5-TTS

F5-TTS is a voice-cloning text-to-speech model that can mimic a short reference recording and generate fluent speech. It is aimed at developers and researchers building voice applications.

F5-TTS is an open-source, zero-shot voice-cloning text-to-speech system from a research group (X-LANCE at SJTU and collaborators). Given a short reference audio clip and its transcript, it can synthesize new speech in that voice reading any text you provide, without per-voice training. It is built on a diffusion-transformer architecture with flow matching and a technique called Sway Sampling that improves inference quality, and the repo also includes an E2-TTS reproduction. It ships with a Gradio web app for basic TTS, multi-style and multi-speaker generation, and even a voice-chat mode, plus a CLI and support for training and fine-tuning. It is primarily a developer and research tool that needs PyTorch and ideally a GPU. Importantly, the code is MIT-licensed but the pre-trained models are released under a non-commercial CC-BY-NC license because of their training data, which limits commercial use out of the box.