The long-term goal of this research project is to improve the speech processing strategies in cochlear implant under adverse environments. The proposed research work aims to improve CI users'speech understanding in noisy listening conditions where noise is additive. The outcomes of the proposed project can benefit cochlear implant users'quality of daily life. The project has the following two specific aims: (1) Evaluation of signal-dependent compression functions. Current CI speech processing strategies use logarithmic compression functions to transform the acoustic signal into electric output. Earlier research results showed that although more compressive compression functions yield slightly better phoneme recognition in quiet, less compressive compression functions perform better for phoneme recognition in noise. Those studies showed that optimal compression functions are signal dependent. Motivated by these findings, we propose to use signal-dependent compression functions in CI speech processing strategies. (2) Evaluation of an environment-optimized noise reduction method. The simplest noise reduction method is to apply a weighting function to the noisy speech temporal envelopes. To determine the optimal weighting function, however, is not an easy task. In this project we propose to obtain the optimal weighting function values using a data-driven training approach. Specifically, during the training, the noise recordings are artificially added to a large speech database. With access to both the clean speech and noisy speech temporal envelopes, optimal weighting function values for each CI spectral channel based on various preset optimization criteria can be obtained. After the training is finished, the corresponding weighting function values for each spectral channel are stored in look-up tables indexed by the selected independent variable values. When the noise reduction module is turned on in real time use, the independent variable in each spectral channel are constantly monitored and estimated, and the quantized values are used to obtain the weighting function values from the look-up table. The main advantage of this method is that the gain function can be optimized: (a) for different background environments (e.g., car noise, babble noise, etc);(b) for individual CI subjects;and (c) for each spectral channel.
