Severity-Controllable Pathological Text-to-Speech Synthesis for Clinical Applications

The article presents a new pathological text-to-speech (TTS) synthesis system that has the ability to control speech severity using latent interpolations. Recognizing the difficulty of this task, our work uses a data augmentation technique to generate a single-speaker multi-severity training dataset required for training such a model. Furthermore, we show how x-vectors already contain information about the severity and leverage it as a conditioning variable for the synthesis. Finally, we propose modifications to the GradTTS architecture to enhance the duration modeling of pathological speech. We carry out objective and subjective evaluations to demonstrate that the proposed GradTTS system works well, and produces more natural, controllable, and stable pathological speech samples than the baseline TransformerTTS system.