Phonetic Classification Using Hierarchical, Feed-forward, Spectro-temporal Patch-based Architectures

A preliminary set of experiments are described in which a biologically-inspired computer vision system (Serre, Wolf et al. 2005; Serre 2006; Serre, Oliva et al. 2006; Serre, Wolf et al. 2006) designed for visual object recognition was applied to the task of phonetic classification. During learning, the system processed 2-D wideband magnitude spectrograms directly as images, producing a set of 2-D spectrotemporal patch dictionaries at different spectro-temporal positions, orientations, scales, and of varying complexity. During testing, features were computed by comparing the stored patches with patches from novel spectrograms. Classification was performed using a regularized least squares classifier (Rifkin, Yeo et al. 2003; Rifkin, Schutte et al. 2007) trained on the features computed by the system. On a 20-class TIMIT vowel classification task, the model features achieved a best result of 58.74% error, compared to 48.57% error using state-of-the-art MFCC-based features trained using the same classifier. This suggests that hierarchical, feed-forward, spectro-temporal patch-based architectures may be useful for phonetic analysis. 1 This memo is identical to previous CSAIL Technical Report MIT-CSAIL-TR-2007-007 dated January 2007, except for the addition of some new references.