E3 TTS

Abstract. We propose Easy End-to-End Diffusion-based Text to Speech, a simple and efficient end-to-end text-to-speech model based on diffusion. E3 TTS directly takes plain text as input and generates an audio waveform through an iterative refinement process. Unlike many prior work, E3 TTS does not rely on any intermediate representations like spectrogram features or alignment information. Instead, E3 TTS models the temporal structure of the waveform through the diffusion process. Without relying on additional conditioning information, E3 TTS could support flexible latent structure within the given audio. This enables E3 TTS to be easily adapted for zero-shot tasks such as editing without any additional training. Experiments show that E3 TTS can generate high-fidelity audio, approaching the performance of a state-of-the-art neural TTS system.

Model Overview

To take advantage of recent advances in large language model development, we built our system on top of a pre-trained BERT model. The BERT model takes subword input and does not rely on any other representations of speech, such as phonemes or graphemes. Then it is followed by a 1D U-Net structure, which consists of a series of downsampling and upsampling blocks connected by residual connections. The entire model is non-autoregressive and directly outputs a waveform. The speaker and alignment are dynamically determined during the diffusion process. Our model achieves comparable results to the two-stage framework on a proprietary dataset.