ProsodyBERT: Self-Supervised Prosody Representation for Style-Controllable TTS

We propose ProsodyBERT, a self-supervised approach to learning prosody representations from raw audio. Different from most previous works, which use information bottlenecks to disentangle prosody features from speech content and speaker information, we perform an offline clustering of speaker-normalized prosody-related features (energy, pitch, their dynamics, etc.) and use the cluster labels as targets for HuBERT-like masked unit prediction. A span boundary loss is also introduced to capture long-range prosodic information. We demonstrate the effectiveness of ProsodyBERT on a multi-speaker style-controllable text-to-speech (TTS) system. Experiments show that the TTS system trained with ProsodyBERT features can generate natural and expressive speech samples, surpassing the model supervised by energy and pitch on subjective human evaluation. Also, the style and expressiveness of synthesized audio can be controlled by manipulating the prosody features. In addition, We achieve new state-of-the-art results on the IEMOCAP emotion recognition task by combining our prosody features with HuBERT features, showing that ProsodyBERT is complementary to popular pretrained speech self-supervised models.