Towards improving the speech recognition of cochlear implant users by identification and deactivation of less informative channels
Cochlear implants (CIs) have improved the speech perception of patients suffering from severe to profound hearing loss significantly. However, CI users’ speech perception is still not satisfactory when it comes to challenging acoustic scenarios, e.g., in presence of noise. Poor spectral resolution due to channel interaction is hypothesized to be one of the reasons behind CI users’ degraded speech-in-noise perception. Therefore, mitigating channel interaction by deactivating channels with poor speech information could be beneficial for CI users. However, it is not straight forward to identify these channels. Using model simulations, this study investigates, whether speech perception in CI users may be improved by identifying and deactivating channels that deteriorate speech recognition.
Speech in general is a redundant signal: The speech information in different frequency channels is correlated. In addition to the natural across-channel correlation of the speech signal, in electric listening, the spread of electric field highly influences the channel cross correlation. Since the spread of electric field is individual, it is hypothesized that an individual procedure is required to identify independent channels that carry relevant speech information. To assess this hypothesis, a computer model of speech intelligibility for CI users was employed. The model simulates the spread of electric field in the cochlea and outputs an internal representation (IR), i.e., the post-processed spiking pattern in different auditory channels as a function of time. Individual spread of electric fields were simulated according to the clinical data of 14 CI users. To identify independent channels, the across-channel amplitude modulation correlation (AMCor) matrices based on IRs were obtained. AMCor matrices have been successfully applied in acoustic hearing to guide speech separation (Anemüller and Kollmeier, 2000). The results showed individual patterns across AMCor matrices that vary substantially with the individual pattern of field spread. This confirms the hypothesis that a selection of a subset of channels for maximizing speech information while minimizing the influence of field spread would lead to a highly individual channel selection, and that this selection may be guided by the AMCor patterns. Whether this channel selection strategy would lead to an improvement in speech reception in individual CI patients remains to be shown.
Anemüller, J., Kollmeier, B., “Amplitude Modulation Decorrelation for Convolutive Blind Source Separation”. Proceedings of the second international workshop on independent component analysis and blind signal separation, June 19-22, 2000, Helsinki, Finland, pp. 215-220.