Abstract
According to the auditory filter theory, it has been known that the basilar membrane of the cochlea in the inner ear is, for example, modeled as a 24-channel filter array. Because the filter shape dynamically varies in response to an input sound or vibration, there is room for verification on our proposed acoustic model, in which a waveform with an one-octave passband is reconstructed by the linkage of sinusoidal curves between successive extrema. In this study for 1-h speech and music recordings, we have evaluated the effects of spectral bandwidths, interpolation curves, and the least-squares estimation of extrema, using the normalized squared-error between an original waveform and the reconstruction. Firstly, the difference errors for an one-octave waveform remain at less than 1.2 percent, compared with the narrower filtered waveforms. Secondly, the error of the sinusoidal interpolation is less than other five methods with transcendental functions, in agreement with our previous studies in terms of the subjective score as well as the objective quantification with event-related potentials. Thirdly, the original waveform is identical to the reconstruction with the extremal estimation below 2,560 Hz, but with leveling minima below mere 160 Hz, suggesting the conservation of the sound quality despite ambiguity in shape.
