Auditory Filterbank Based Frequency-Warping Invariant Features for Automatic Speech Recognition

Auditory filterbanks have a long history in the preprocessing stage of automatic speech recognition systems, with the most prominent examples being the mel frequency cepstral coefficients (MFCCs). In this paper, we study the usefulness of auditory-filterbank analyses as a preprocessor for the generation of frequency-warping invariant features. The results indicate, that gammatone-filterbank analyses following the equivalent rectangular bandwidth (ERB) scale yield the most robust feature sets. The performance improvements are most significant when the vocal tract lengths in the training and test sets differ, which is important when, for example, children speech is to be recognized with a system that was mainly trained on adult data.