Improving a tone labeling algorithm for Sesotho

Abstract

We report on a study that aimed to improve an existing tone label prediction algorithm for Sesotho, an official language of South Africa. Tone is an important prosodic feature of Sesotho, since speakers use tone to distinguish meaning. In order to implement tone in a Text-to-Speech system for Sesotho, a tone modeling algorithm must receive as input the tone labels of the syllables of each word. Then it can predict the appropriate intonation of the word. Since Sesotho does not mark tone labels in orthography, the labels have to be predicted according to the tonal rules of the language. The existing tone label prediction algorithm has two drawbacks, namely it implements three tonal rules only and is restricted to the clitic phrase domain. In our study, we developed an algorithm that implements four additional tonal rules and addresses all parts of speech. The results show that the latter algorithm significantly improves the existing one by increasing the number of matched tone labels.

Rapid Development of TTS Corpora for Four South African Languages

“…Underlying tone can be marked on the lexical level, however, in sentence context the realised surface tone is a function of underlying tone and other linguistic structures (morphological, syntactic, etc.) [15]. A tonal feature was therefore not included, and requires further study.…”