Deep neural network algorithms for noise reduction and their application to cochlear implants
Cochlear implants (CIs) fail to provide the same level of benefit in noisy settings as in quiet. Current noise reduction solutions in hearing aids and CIs only remove predictable, stationary noise, and are ineffective against realistic, non-stationary noise such as multi-talker interference. Recent developments in deep neural network (DNN) models have achieved noteworthy performance in speech enhancement and separation. However, little work has investigated the potential of DNN models in removing multi-talker interference. The research in this regard is even more scarce for CIs. Here, we implemented two DNN models that are well suited for applications in speech audio processing, including (1) recurrent neural network (RNN) and (2) SepFormer. The models were trained with a dataset developed in house ([~] 30 hours), and then tested with thirteen CI listeners. Both RNN and SepFormer models significantly improved CI listeners speech intelligibility in noise without compromising the perceived quality of speech. These models not only increased the intelligibility in stationary non-speech noise, but also introduced substantially more improvements in non-stationary speech noise, where conventional signal processing strategies fall short with little benefits. These results show the promise of using DNN models as a solution for listening challenges in multi-talker noise interference.