Can Prosody Transfer Embeddings be Used for Prosody Assessment?

Abstract

In voice conversion, it is possible to transfer some characteris-tic components of a (target) speech utterance, such as the content, pitch, or speaker identity, from the corresponding component from another (source) utterance. This has recently been achieved by characterizing these components through neural-based vector embeddings which encode the specific information to be transferred. In the particular case of neural prosody embeddings, to the best of our knowledge, no work has ex-plored the informativeness of these embeddings for other pur-poses, such as prosody assessment or comparison of prosodic patterns. In this work, we use an intonation data set and a voice conversion corpus to explore how these neural prosody embeddings group for utterances of different intonation, content, and speaker identity. We compare these neural prosody embeddings to hand-crafted acoustic-prosodic features and to content embeddings. We found that neural prosody embeddings can achieve a geometrical separability index as high as 0.956 for highly contrastive intonations, and 0.706 for different sentence types.