The project addresses one of the outstanding issues in speech technology: the ability to synthesize prosodically coherent and cohesive conversational speech, appropriately incorporating long-distance contextual and situational dependencies. The objective is to deliver a novel speech synthesis platform explicitly using encoded prosodic information as a source of conversation dynamics that helps maintain context-dependent cohesion and coherence in human-machine interaction and synthesized dialogues. In a reciprocal fashion, the system trained on a large data set of conversational speech will be reinterpreted as a complex statistical model and contribute to our theoretical understanding of features and wide-range interdependencies shaping conversational prosody. The design of the system will be informed by our expertise in prosodic analysis and deep learning.

2023

2027