Joint compensation of multi-talker noise and reverberation for speech enhancement with cochlear implants using one or more microphones

Following speech in noisy and reverberant situations is difficult for cochlear implant (CI) users. This study investigates single- and multi-microphone deep neural network (DNN) speech enhancement algorithms on the joint task of denoising and dereverberation. The DNN algorithms were trained and tested on simulated sound scenes from behind-the-ear hearing devices. Performance was assessed using objective measures and a listening study for reverberant mixtures of speech in multi-talker babble noise. We compare results for signal distortion, predicted intelligibility and speech reception thresholds measured in a listening experiment with 15 typically hearing participants using cochlear implant simulations. Objective metrics indicated listening benefits for both single- and multi-microphone approaches while the listening study results confirmed significant improvements in speech intelligibility for the multi-microphone approaches, holding strong promise to benefit CI listeners.