Temporal Filterbanks in Cochlear Implant Hearing and Deep Learning Simulations

Abstract

The masking phenomenon has been used to investigate cochlear excitation patterns and has even motivated audio coding formats for compression and speech processing. For example, cochlear implants rely on masking estimates to filter incoming sound signals onto an array. Historically, the critical band theory has been the mainstay of psychoacoustic theory. However, masked threshold shifts in cochlear implant users show a discrepancy between the observed critical bandwidths, suggesting separate roles for place location and temporal firing patterns. In this chapter, we will compare discrimination tasks in the spectral domain (e.g., power spectrum models) and the temporal domain (e.g., temporal envelope) to introduce new concepts such as profile analysis, temporal critical bands, and transition bandwidths. These recent findings violate the fundamental assumptions of the critical band theory and could explain why the masking curves of cochlear implant users display spatial and temporal characteristics that are quite unlike that of acoustic stimulation. To provide further insight, we also describe a novel analytic tool based on deep neural networks. This deep learning system can simulate many aspects of the auditory system, and will be used to compute the efficiency of spectral filterbanks (referred to as “FBANK”) and temporal filterbanks (referred to as “TBANK”).