MSStyleTTS: Multi-Scale Style Modeling with Hierarchical Context Information for Expressive Speech Synthesis

Expressive speech synthesis is crucial for many human-computer interaction scenarios, such as audiobooks, podcasts, and voice assistants. Previous works focus on predicting the style embeddings at one single scale from information within current sentence. Whereas, context in neighboring sentences multi-scale nature of human are neglected, making it challenging to convert multi-sentence text into natural expressive speech. In this paper, we propose MSStyleTTS, a modeling method synthesis, capture predict styles different levels wider range rather than Two sub-modules, including extractor predictor, trained together with FastSpeech 2 based acoustic model. The predictor designed explore hierarchical by considering structural relationships global-level, sentence-level subword-level. extracts embedding ground-truth explicitly guides prediction. Evaluations both in-domain out-of-domain audiobook datasets demonstrate that proposed significantly outperforms three baselines. addition, conduct analysis representations have never been discussed before.