An Efficient Auditory Fil Terbank Based on the Gammatone Function

This paper describes the development of an auditory filterbank to perform the initial frequency analysis in models of human hearing and speech perception. It is based on the gammatone function used by physiologists to summarise 'revcor' measurements of the impulse response of the auditory filter in small mammals. The first section shows that the amplitude characteristic of the gammatone function is very similar to that of the roex filter shape of Patterson, Nimmo-Smith, Weber and Milroy (1982), which is known to predict human masking data well (Patterson & Moore, 1986). The second section argues that the minimum-phase characteristic of the gammatone filterbank is the preferred alternative for an auditory filterbank, and it introduces a method of compensating for the strong skew on the output of the auditory filterbank. The last section presents a recursive implementation of the gammatone filter that is both accurate and efficient. The result is an auditory filterbank with a unique combination of advantages: It is based on physiological data and modelling. It predicts human masking data accurately. It is almost as fast as the simplified filterbanks currently being used as frontended processors for automatic speech recognition.