How speech statistics limits the number of effective channels in cochlear implants: Implications for sound-coding strategies
Typically, cochlear implant (CI) processors first analyse sound through a bank of evenly spaced filters on a log-frequency scale, before each extracted temporal envelope is used to modulate the excitatory current passed through each implanted electrode. However, factor analysis (FA) of speech shows that the information carried by the temporal modulations of speech is distributed in bands whose width grows non-monotonically with frequency. Given that the spread of excitation (SOE) along the spiral ganglion limits the number of independent channels in CI users to eight or less, optimising transmission of information should be considered in the design of sound-coding strategies. A scree plot derived from factor analysis suggests that six channels should suffice to transmit most of the speech information. This outcome may explain why speech-reception thresholds (SRTs) using the SPIRAL CI vocoder improve markedly more slowly as a function of number of channels beyond 7 simulated channels. A first study tests whether FA-inspired channel bandwidths might improve speech intelligibility. Simulations with normal-hearing listeners employed a pulsatile vocoder (Oldenburg) and adaptively measured SRTs for the following factors: two levels of SOE (200 dB/oct and 8 dB/oct) x two numbers of channels (8 or 11) x three channel definition/allocation strategies. The first strategy (CIS) used evenly-spaced channels and simulated electrodes aligned with the channel centre frequencies. The second and third strategies used FA-selected channels, with simulated electrodes selected from a fixed 22-electrode array; in the second (FAw), selected electrodes were evenly spaced, thereby limiting channel interaction, but leading to spectral warping; in the third (FAu), the selected electrodes were those with a place frequency closest to each of the channel centre frequencies, leading to an almost unwarped electrode map. Simulating SOE significantly elevated thresholds and the number of channels interacted with SOE. Although no effect of strategy was found, spectral warping would be expected to elevate SRTs. However, with 8 channels and SOE typical of CI users, the CIS and FAw strategies led to the two lowest thresholds. A second study that varied the number of activated channels (4, 5, 6, 8, 11, 15 and 22) and simulated typical or no SOE with the FAu strategy, showed that 5 channels sufficed to attain the lowest threshold. Perceptual learning experiments may help elucidate which FA-inspired strategy would be most effective in CI users.