Emo-BPO improves control of emotional speech in AI voices
Emo-BPO: Emotion Bidirectional Preference Optimization for Diffusion-based Emotional TTS
Sound
Summary
Text-to-speech (TTS) systems can make voices speak with emotions, but controlling those emotions precisely is hard. This paper proposes a new method called Emotion Bidirectional Preference Optimization (Emo-BPO) that better tells the system how to balance the intended emotion against other possible emotions. This approach helps produce clearer and more expressive emotional speech without needing extra training data or complex reward systems. The method works with existing TTS technologies to make emotional speech sound more natural and controllable.
What this means in practice
- •For voice assistant developers: Create digital assistants that express emotions more naturally and can finely control the emotional tone of their responses.$Commercial implications: Enables commercial voice assistant products with improved emotional expressiveness and dynamic emotional control, enhancing user engagement.
- •For game audio engineers: Generate varied emotionally expressive character voices that adjust dynamically to the game context without extra annotation effort.
Authors
Jiacheng Shi, Hongfei Du, Xinyuan Song, Y. Alicia Hong, Yanfu Zhang, Ye Gao
Abstract
Diffusion-based TTS has demonstrated strong capability in emotional speech generation. However, existing emotion alignment methods primarily reinforce target emotional expressions while overlooking modeling of competing emotional modes, limiting fine-grained controllability. This issue becomes more pronounced under classifier-free guidance (CFG), which relies on contrasting conditional trajectories to shape generation. We propose Emotion Bidirectional Preference Optimization (Emo-BPO), a simple effective framework that jointly learns emotion-aligned and emotion-contrastive score functions from reordered same-text emotional pairs. Our method requires neither auxiliary reward models nor additional annotations, and can be seamlessly integrated into existing diffusion-based TTS systems to improve emotional controllability and expressiveness.