ABSTRACT

In this article, we propose an aggregation of denoising diffusion probabilistic models (DDPMs) onto an end-to-end text-to-speech system to learn a distribution of reference speaking styles in an unsupervised manner. By applying a few steps of a forward noising process to an embedding extracted from a reference mel spectrogram, we make profit of its information to reduce the diffusion backward chain and reconstruct an improved style embedding with only a few reverse steps, performing style transfer. Additionally, a combined spectrogram reconstruction and denoising loss allows for conditioning of the acoustic model on the synthesized style embeddings. A subjective perceptual evaluation is conducted to evaluate naturalness and style transfer capability of the proposed approach. A MUSHRA test shows a 5 point increment on the mean of naturalness ratings and a paired comparison (ABX test) on style transfer reveals preference of the raters (43%) of our proposed approach over state-of-the-art models (29%).