WavTTS

WavTTS is a zero-shot text-to-speech model from ByteDance that generates speech directly in the raw waveform domain rather than in VAE latents or mel-spectrograms. It uses flow matching with a Diffusion Transformer (DiT) and a simple patchification strategy to cope with the long sequence length of audio. Multi-scale mel-spectrogram supervision provides perceptual guidance during training, and the authors study prediction targets and noise schedules for waveform diffusion.

According to the project page, it is the first raw-waveform generative TTS model to substantially narrow the gap with latent-space generative models. On open-source benchmarks it approaches current state-of-the-art latent zero-shot TTS systems and clearly outperforms earlier end-to-end speech generation models. The demo page shows English and Chinese zero-shot generation from a reference prompt.

A YouTube creator described it as built on F5-TTS, cloning a voice from a few seconds of audio, and released with training scripts and a roughly 10GB checkpoint. These details come from that secondary source and were not confirmed on the provided project page, which has no code, weights or paper links. No pricing, license or hosted service was found.