J-TranPSP: A Joint Transition-based Model for Prosodic Structure Prediction, Word Segmentation and PoS Tagging

Text analysis is usually the first step in speech synthesis. It provides linguistic features as the inputs of the state-of-the-art acoustic models to generate speech. The linguistic features include word boundaries, part-of-speech (PoS) tags, prosodic phrase boundaries and phoneme sequences. Usually all acoustic models from different speakers share the same linguistic features. So the correctness of the linguistic features will profoundly affect the synthesized speech. In this paper, we first time applied transition-based model on prosodic structure prediction with two actions read and break. Moreover we proposed a simple but effective joint model on top of the ELECTRA encoder that simultaneously predicts word boundaries and PoS tags. Our unified model achieves comparable or better F1-scores than the state-of-the-art models on the three tasks.