Preservation Of Interaural Level Difference Cue In A Deep Learning-Based Speech Separation System For Bilateral And Bimodal Cochlear Implants Users

Due to the success of deep neural networks (DNNs) in speech separation, the DNN-based speech separation method has become a potentially feasible front end of cochlear implants (CIs) to reduce noises. However, most DNN-based methods neglect the demand of accurately preserved spatial cues, which are necessary for bilateral and bimodal CI users to localize sounds and benefit from spatial release from masking. In our previous study, a speech separation framework with spatial cues preservation has been designed, by restoring the distorted relative transfer function of the pre-separated speech. In this work, the framework was extended for evaluation in simulated CI hearing scenarios with bilateral and bimodal CI setups. Experiment results showed that the framework significantly reduced the interaural level difference (ILD) errors of the speech separated by the existing DNN-based method, indicating that the proposed DNN-based speech separation framework can effectively preserve the ILD cue in the CI application scenarios.