Human Language Technologies (HLT) – Intonation modelling

A detailed understanding of the intonation (or “prosody”) of a language is important for theoretical as well as practical reasons. Theoretical, since intonation is a deep regularity of languages (and their dialects), and practical, since high-quality speech synthesis (and, possibly, recognition) is impossible without such an understanding. Much has been written about the intonation of the indigenous South African languages – and the Bantu family of languages in general – but most of the underlying research was impressionistic rather than empirical.

In early work with languages from the Nguni family (initially, isiZulu and isiXhosa), we measured prosodic variables – fundamental frequency, duration and amplitude for naturally-spoken utterances as well as isolated words. We then developed rules to correlate these measurements with predictable morphological and lexical factors. However, it became clear that a more systematic approach was required, and our more recent focus has been on a narrower set of issues related to tone realization in the verbal domain. In collaboration with Dr. Sabine Zerbian of the University of the Witwatersrand, we have designed a set of phrases that allow for the detailed analysis of how high tones in Sepedi affect the properties of verbal phrases. Recording and analysis of these phrases has produced a number of interesting insights.

We plan to integrate the models that arise from our investigations into our text-to-speech systems, and also to investigate whether they can be used to improve speech recognition accuracy.