Word Boundary Estimation of isiZulu Continuous Speech

Typical automatic speech recognition algorithms assume that spoken words are completely isolated; this is not always the case in real-world applications. Consequently, there is a need for intra-sentence segmentation of words uttered in a continuous speech. Word boundary estimation is a method that was proposed for the separating of words. There has been extensive research on western languages, especially English. However, limited investigations have been done for indigenous languages. IsiZulu is the most spoken indigenous language in South Africa. This paper proposes the use of prosodic features and higher-order statistic (HOS) features to classify IsiZulu speech segments i.e. noise, voice and boundary. For automatic classification, supervised learning was used.