Development of a Sound Coding Strategy based on a Deep Recurrent Neural Network for Monaural Source Separation in Cochlear Implants

The aim of this study is to investigate whether a source separation algorithm based on a deep recurrent neural network (DRNN) can provide a speech perception benefit for cochlear implant users when speech signals are mixed with another competing voice. The DRNN is based on an existing architecture that is used in combination with an extra masking layer for optimization. The approach has been evaluated using the HSM sentence test (male voice) mixed with a competing voice (female voice) for a monaural speech separation task. Two DRNNs with two levels of complexity have been used. The algorithms have been evaluated in 8 normal hearing listeners using a Vocoder and in 3 CI users. Both DRNNs show a large and significant improvement in speech intelligibility using Vocoded speech. Preliminary results in 3 CI users seem to confirm the improvement observed using Vocoded simulations.