Simple-TTS: End-to-End Text-to-Speech Synthesis with Latent Diffusion

TL;DR — We introduce an end-to-end text-to-speech (TTS) latent diffusion model as a simpler alternative to more complicated pipelined approaches for TTS synthesis.

Abstract

We propose an end-to-end text-to-speech (TTS) latent diffusion model as a simpler alternative to more complicated pipelined approaches for TTS synthesis. In particular, we show that one can adapt a recently proposed text-to-image diffusion architecture, U-ViT, as an excellent backbone for audio generation. We identify and explain the changes required for this adaptation and demonstrate that latent diffusion is an effective approach for end-to-end speech synthesis, without the need for phonemizers, forced aligners, or complex multi-stage pipelines. Despite its simplicity, our proposed approach, Simple-TTS, outperforms more complex models that rely on explicit alignment components and significantly outperforms the best open-source multi-speaker TTS system. We will open-source Simple-TTS upon acceptance, making it the strongest system publicly available to the community. Due to its straight-forward design, we expect that Simple-TTS can easily be adapted to many diverse TTS settings --- opening the stage to repeat the success of Stable Diffusion in computer vision, in audio generation.