VIB-TTS: Making Emotion Text-to-Speech More Accurate via Variational Information Bottleneck and Emotion Dimension-Based Selection

Emotional speech synthesis continues to attract significant research interest due to its wide-ranging applications in human-computer interaction, entertainment, and assistive technologies. Despite recent advances, generating speech that is both natural and expressive while precisely matching specified emotions remains a major challenge. Existing methods typically fall into two categories, each facing significant limitations. Emotion label-based methods are easy to implement, but often lack sufficient emotional details. In contrast, reference audio-based methods capture richer emotional information, yet they face the challenges in selecting reference speech with accurate emotions. To address the aforementioned problems, we propose a novel approach, VIB-TTS, which combines the strengths of emotion label-based methods and reference audio-based methods, capturing emotional details and enhancing accuracy via emotional labels. We design an emotion enhancement module incorporating variational information bottleneck to improve the emotional accuracy of the synthesised speech. Furthermore, we select reference audio via calculated Valence-Arousal-Dominance (VAD) values to ensure high-quality emotional synthesis. The results show that our method has better emotion expression capability compared to previous methods.