Obstruent classification using modulation spectrogram based features

In this paper, a new feature extraction technique based on modulation spectrogram is proposed. Modulation spectrogram gives a 2-dimensional (2-D) feature set for each obstruent segment. Since the size of feature vector given by modulation spectrogram is of very large dimension, Higher Order Singular Value Decomposition (HOSVD) theorem is used to reduce the size of feature vector. The reduced feature vector is then applied to a classifier, which classify the obstruent in three broad classes, viz., stop, affricate and fricative. Four-fold cross-validation experiments have been conducted on TIMIT database to find accuracy of obstruent classification at phoneme-level and recognition of manner of articulation of obstruents. Our experimental results show 92.22 % and 94.85 % accuracies for obstruent classification at phoneme-level and recognition of manner of articulation of obstruents, respectively, using 3-nearest neighbor classifier while with same experimental setup Mel Frequency Cepstral Coefficients (MFCC) shows 87.24 % and 93.68 % average classification accuracy of phoneme-level classification and manner of articulation level classification of obstruents, respectively.