An information-theoretic analysis of auditory features in noisy environments
Current engineering efforts such as automatic speech recognition (ASR) systems and cochlear implants tend to perform much worse than the human auditory system in noisy environments. We aim to address this problem by developing auditory-inspired features that are robust to noise. Here, we review several general principles that appear to be important for the noise robustness of the auditory system, such as precisely timed inhibition. These principles are combined with a basic model of the auditory nerves (a half-rectified Gammatone filterbank) to create a new set of auditory features. To assess the quality of these features, we use information theoretic measures. Altogether this new class of features may allow us to both improve current understanding of general principles in the auditory system, as well as finding new features that could be directly applied to ASR systems.