I'm interested in building a toy text-to-speech system, where at the very least I want to decide when to include dynamic emphasis, slight pitch shift, and the duration and boundary between words. It would be amazing if this could be done on the basis of simple rules, but it looks like most people use data-driven approaches. However, I haven't found a dataset which maps words in a text to any prosodic features (f0 frequency/duration, for example). Does either such a set of rules or such a dataset exist?



The prosodic contours of a word will vary greatly, depending on how it is used. This might be because of neighboring words in the sentence, or the information content of the word (new vs. given information). This will also vary greatly depending on dialect. People are actively researching this, however.

For example, this doctoral thesis Prosody modelling using machine learning techniques for neutral and emotional speech synthesis seems relevant. Google has also done some recent research on Expressive Speech Synthesis.

A dataset seems close to impossible, since it would need to contain every word in every single context.

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.