Vector-Quantized Feedback Recurrent Autoencoders for the Compression of the Stimulation Patterns of Cochlear Implants at Zero Delay

Cochlear Implants (CIs) are surgically implanted hearing devices that allow to restore a sense of hearing in people suffering from moderate to profound hearing loss. Modern CIs offer wireless streaming of audio to the signal processor of the CI to improve speech understanding in complex acoustic environments. To conserve energy in this wireless streaming, proprietary source coding of the stimulation patterns of CIs was proposed, achieving state-of-the-art results with respect to bitrate, latency and intelligibility of the coded stimulation patterns. This work investigates vector-quantized feedback recurrent autoencoders (VQ FRAE) to improve source coding of the stimulation patterns of CIs. The VQ FRAE is optimized with respect to the non-differentiable STOI using simultaneous perturbation stochastic approximation. With this approach, a state-of-the-art bitrate of 4.69 kbit/s was achieved, while maintaining zero latency and little to no degredation of intelligibility. The FRAE outperforms audio codecs like Opus with respect to bitrate, intelligibility and latency.