Spectral and textural features for automatic classification of fricatives using SVM

We report on an analysis of spectral and textural characteristics of fricatives for their classification. Fricative classification can be useful in applications such as differential manipulation of phonemes for the hearing impaired, where people have difficulties in perception of fricatives. Several acoustic time and frequency domain features were computed and examined for constructing discriminative feature vectors, which enable accurate classification of fricatives for various speakers and dialects, and for varied contexts. The best sets of features for classification were selected using a floating-search procedure. The evaluation included a data set of more than 18,000 fricatives, from more than 100 speakers. The classification stage included training a support vector machine (SVM) on small part of the data, initial classification of each signal frame (8-12 msec), and utilizing a majority vote for the feature vectors of the same phoneme. An overall accuracy of 89% and 80% was obtained for the unvoiced and voiced fricatives, respectively, and 97% for sibilants/non-sibilants discrimination.