Corpus Based Evaluation of Entropy Rate Speech Segmentation

The sequence of estimates of the speech signal’s entropy rate is investigated as a potential basis for speech segmentation. Rising and falling edges of that entropy rate curve and its maxima and minima are considered as candidates for segment boundaries. These prominent points are compared to the phonetic segment boundaries and to acoustic landmarks. The comparison is made using the American TIMIT database and the German ‘Kiel corpus of read speech’ which are phonetically manually labelled speech corpora.