Use of S-shaped input-output functions for noise suppression in cochlear implants

The aim of this study is to assess the influence of the shape of the acoustic-to-electric mapping function on speech recognition in noise by cochlear implant listeners.

A new acoustic-to-electric mapping function is proposed for cochlear implant users in noisy environments. The proposed s-shaped mapping function was expansive for low input levels up to a knee point level and compressive thereafter. The knee point of the mapping functions changed dynamically and was set proportional to the estimated noise floor level. The performance of the mapping function was evaluated on a sentence recognition task using IEEE sentences embedded in +5 to 10 dB SNR multitalker babble and in +5 dB SNR speech-shaped noise. Nine postlingually deafened cochlear implant users participated in the study.

Results indicated that the same s-shaped mapping function did not yield significant improvements for all cochlear implant users. Significant benefits in speech intelligibility were observed, however, when the s-shaped mapping function was optimized to individual cochlear implant users. Significantly higher performance was achieved with the s-shaped mapping functions than the conventional log mapping function used by cochlear implant users in their daily strategy, in both multitalker (+5 and +10 dB SNR) and continuous speech-shaped (+5 dB SNR) conditions.

These results clearly indicate that the shape of the nonlinear acoustic-to-electric mapping can have a significant effect on speech intelligibility in noise when it is optimized to individual cochlear implant users. The log functions currently used in most implant processors for mapping acoustic to electric amplitudes are not the best mapping functions to use in noisy environments. This is largely because compressive functions tend to amplify low-level segments of speech along with noise, thereby decreasing the spectral contrast and effective dynamic range. In contrast, the s-shaped mapping functions, which are partly compressive and partly expansive depending on the signal level, are more suitable for noisy environments and can produce significantly higher performance than the log-mapping functions.