F5-TTS

F5-TTS is a non-autoregressive text-to-speech system that pairs flow matching with a Diffusion Transformer, using ConvNeXt V2 blocks for faster training and inference. Zero-shot voice cloning works from a short reference clip, with optional ASR transcription when the reference audio is unlabeled, and the repository bundles E2 TTS, a Flat-UNet Transformer variant, for comparison. Sway Sampling, an inference-time strategy introduced by the project, improves output quality without retraining. Trained on datasets including Emilia, WenetSpeech4TTS, LibriTTS, and LJSpeech, the models reach a real-time factor of about 0.04 on an L20 GPU. Inference options span a Gradio web UI with multi-style and multi-speaker support plus a voice chat mode driven by Qwen2.5-3B-Instruct, a command-line interface, Docker, and a TensorRT-LLM runtime. The code is MIT licensed while the pretrained checkpoints carry a CC-BY-NC license, a distinction that matters for commercial use.