Improving Prosodic Features Extraction for Tone Detection in Yemba Language

In tonal languages, tone is of vital importance in differentiating between lexical words and grammatical forms. Tone recognition can improve the performance of speech recognition tasks for tonal languages by re-evaluating word hypotheses using tonal information or by including prosodic features in the acoustic model used. Little effort has been made to evaluate the effectiveness of machine learning approaches for tone detection in African low resourced languages. In this paper we propose a selection of prosodic acoustic features to deal with the linguistic specificities of the Yemba language (spoken in Cameroon) for tone detection. The following features, extracted for each frame of a given syllable, are used: pitch, energy, duration, slope of consecutive F0. Experiments have been conducted using multi-speaker models trained with Naive Bayes, LDA, QDA, SVM and decision trees. The decision trees using the cost complexity pruning method gave the best results: an accuracy of 61.82% and a F1 measure of 58.90%.