Comparison of skewness-based salient event detector algorithms in speech

Abstract

In this work, we compare two skewness-based salient event detector algorithms, which can detect transients in human speech signals. Speech transients are characterized by rapid changes in signal energy. The purpose of this study was to compare the identification of transients by two different methods based on skewness calculation in order to develop a method to be used in studying the processing of speech transients in the human brain. The first method, the skewness in variable time (SKV) finds transients using a cochlear model. The skewness of the energy distribution for a variable time window is implemented on artificial neural networks. The second method, the automatic segmentation method for transient detection (RoT) is more speech segmentation-based and developed for detecting transient-speech segment ratio in spoken records. In the current study, the test corpus included Hungarian and English speech recorded from different speakers (2 male and 2 female for both languages) Results were compared by the F-measure, the Jaccard similarity index, and the Hamming distance. The results of the two algorithms were also tested against a hand-labeled corpus annotated by linguistic experts for an absolute assessment of the performance of the two methods. Transient detection was tested once for onset events alone and, separately, for onset and offset events together. The results show that in most cases, the RoT method works better on the expert labeled databases. Using F measure with +-25ms window length the following results were obtained when all type of transient events were evaluated: 0,664 on English and 0,834 on Hungarian. Otherwise, the two methods identify the same stimulus features as the transients also coinciding with those hand-labeled by experts.