Machine learning-based sound coding for cochlear implants (vor Ort)
* Presenting author
Abstract:
Sound coding strategies for cochlear implants (CIs) translate the incoming sound signal into parameters for the electrical pulse pattern delivered by the implant. Without specific noise reduction, the common envelope coding in CIs, like in the continuous interleaved sampling strategy (CIS), leads to problems with sound localization and listening in situations with noise and reverberation. Hence, this basic sound coding is accompanied by noise reduction and de-reverberation stages, which clean the sound signal prior to pulse processing, together achieving good performance in many listening situations. Statistical approaches using deep neural networks (DNN) achieve impressive results for segregating sound sources against a background of other sources and reverberation. These networks, like ConvTasNet, contain stages for filtering and amplitude extraction similar to the processing in CIs. We investigated if a DNN could be trained to replace the complete sound coding strategy of a CI. We compare the amplitude stimulation sequences for each electrode computed by a new stochastic approach and compare them to those of the classic CIS strategy. Results show that a DNN is capable of replacing existing CI coding strategies like CIS, while having the potential to include preprocessing stages for source segregation at no extra computational cost.
