Enhancing robustness for speech recognition through bio-inspired auditory filter-bank

One of the important properties observed in basilar membrane filtering, aimed to improve robustness of the human car is lateral inhibition-based level-dependent frequency resolution. However, this particular property has not been extensively considered for improving robustness of the speech processing systems. In this work, an auditory filter-bank which includes lateral inhibition based on input stimulus providing a good fit to human auditory masking is used for improving robustness of the speech recognition system. The gammachirp auditory filter is the real part of the analytic gammachirp function which has been shown to provide an accurate description for the asymmetric and lateral inhibition observed in the basilar membrane filtering. The gammachirp is characterised with asymmetry in the low frequency tail of auditory filter response and models level dependent properties such as decrease in gain and a shift in the centre frequency of the filter with increase in level. The speech recognition experiments using the standard HTK framework are performed on standard Aurora-5 digit task database, both simulated and real data recorded with distant microphones in a hands-free mode at a real meeting room. The gammachirp-based features show reliable and consistent improvements when compared to the conventional features used for speech recognition.