F5-TTS

F5-TTS is a text-to-speech system for generating speech from text, with support for basic TTS, multi-style and multi-speaker generation, and voice chat. It is aimed at lower-latency inference and includes a CLI for running inference from the command line.

The model uses a Diffusion Transformer with ConvNeXt V2, plus Sway Sampling for inference-time flow step sampling. It also includes E2 TTS as a Flat-UNet Transformer reproduction, and offers custom inference with more language support. Runtime deployment is available with Triton and TensorRT-LLM.

F5-TTS ships code under the MIT License, while the pre-trained models use CC-BY-NC. It provides Docker usage and training and finetuning guidance with Hugging Face Accelerate. Published base models are available on Hugging Face, Model Scope, and Wisemodel.