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

Area of Science:

• Speech synthesis
• Pathological speech
• Machine learning

Background:

• Pathological speech synthesis presents challenges in controlling speech severity.
• Existing text-to-speech (TTS) systems struggle with generating diverse pathological speech characteristics.
• Generating multi-severity datasets for single-speaker pathological speech is difficult.

Purpose of the Study:

• To develop a novel pathological text-to-speech (TTS) synthesis system capable of controlling speech severity.
• To improve the naturalness, controllability, and stability of synthesized pathological speech.
• To leverage existing speaker embeddings (x-vectors) for severity conditioning.

Main Methods:

• Utilized data augmentation to create a single-speaker multi-severity training dataset.
• Employed x-vectors as a conditioning variable for speech severity control.
• Modified the GradTTS architecture to improve duration modeling for pathological speech.

Main Results:

• The proposed GradTTS system demonstrated superior performance compared to the baseline TransformerTTS.
• Objective and subjective evaluations confirmed the system's effectiveness.
• The system successfully generated more natural, controllable, and stable pathological speech samples.

Conclusions:

• The developed GradTTS system offers effective control over pathological speech severity.
• The integration of x-vectors and architectural modifications enhances pathological TTS capabilities.
• This work advances the field of controllable and natural-sounding pathological speech synthesis.