Speech-to-Speech Translation with Discrete-Unit-Based Style Transfer

Abstract. Direct speech-to-speech translation (S2ST) with discrete self-supervised representations has achieved remarkable accuracy, but is unable to preserve the speaker timbre of the source speech. Meanwhile, the scarcity of high-quality speaker-parallel data poses a challenge for learning style transfer during translation. We propose an S2ST framework with style-transfer capability on the basis of discrete self-supervised speech representations and codec units. The acoustic language model we introduce for style transfer leverages self-supervised in-context learning, acquiring style transfer ability without relying on any speaker-parallel data, thereby overcoming the issue of data scarcity. By using extensive training data, our model achieves zero-shot cross-lingual style transfer on previously unseen source languages. Experiments show that our model generates translated speeches with high fidelity and style similarity. Audio samples are available at http://stylelm.github.io/ .

Model Overview

Our method comprises three consecutive stages, utilizing two distinct types of discrete units: 1) speech-to-semantic-unit translation stage S₁, which converts source audio into semantic units of the translated speech; 2) acoustic unit modeling stage S₂, generating target acoustic units conditioned on the semantic output from the preceding stage and the acoustic units of the source speech as style prompt; 3) unit-to-wave generation stage S₃, producing translated speech that maintains consistent style with the source. We provide details about these two types of units and the three stages in the following subsections.

The training procedure of S₂ adopts a self-supervised training paradigm, where the first three seconds of each audio sample is truncated as prompt, and the acoustic language model is trained to predict the acoustic units of the remaining part conditioned on its semantic units and the prompt acoustic units with cross-entropy loss. This in-context learning approach enables the model to grasp the correspondence in acoustic characteristics between the two parts and acquire style transfer ability. During inference, we use semantic tokens from the previous stage and acoustic units of source speech as the style prompt to realize cross-lingual style transfer.

Table of Contents