Robust Phoneme Recognition Based on Biomimetic Speech Contours

摘要

It has been previously suggested that ensembles of central auditory neurons optimize a sustained firing criterion as part of the underlying neural code for representing sound. Moreover, computational studies have shown that optimizing such a criterion yields ensembles of spectro-temporal receptive fields akin to those observed in physiological studies. In this study, we show that these emergent receptive fields contour the hjgh-energy modulations in speech, defining a boundary that distinguishes between noise-robust and easily corrupted modulations in speech-plus-noise mixtures. A simple 2D filter thus derived is shown to improve upon the performance of state-of-the-art phoneme recognition systems under both additive noise conditions and reverberation by 5.9% absolute on average.