Authors: Samiya Alkhairy¹
¹MIT, KACST
Background: Gammatone-family filters are widely used in computational auditory models, speech technology, and hearing device design. The filter parameters used to mimic human hearing have remained largely unchanged since 1990, based on simultaneous masking data influenced by suppression. Whether recent forward masking and physiological data warrant updated values has not been systematically investigated.
Method: We developed a characteristics-based framework using a sharp-filter approximation valid across three filter classes of the Gammatone family. We derived closed-form expressions relating measurable characteristics—such as quality factors, peak group delay, phase accumulation, and spectral convexity—to the filter parameters. We then applied this framework to recent physiological and psychoacoustic data to identify the most informative characteristic ratios for estimation.
Results: The ratio $g=Q_{erb}/N_{beta}$ was found to be highly sensitive and yielded a point estimate of $B_u=7.2$ using chinchilla auditory nerve fiber data, which is substantially higher than the historical value of 4. Combined with human forward masking quality factors, this yields peak group delays significantly larger than historical estimates at 4 kHz. Relative errors remained below 2% across all filter classes, requiring only 2-3 scalar constants for the full filterbank.
Conclusion: The updated parameters have significant implications for predicted temporal envelope resolution, binaural ITD extraction windows, and cochlear implant channel bandwidth-temporal resolution trade-offs. The framework enables systematic parameter updates as new data become available and can serve as the base for level-dependent models such as the gammachirp.
