gms | German Medical Science

This work builds upon the previously proposed deepACE framework, introducing a lightweight, end-to-end approach to speech denoising for cochlear implants (CIs). We refine the architecture by relocating the deep envelope detector (DED) from the skip connection to the output of the masking operation, enabling high-resolution time-frequency masking and improving noise reduction. Additionally, we incorporate adversarial training to enhance the generation of electrodograms – the electrical pulse patterns used to stimulate the auditory nerve. By leveraging the inherent simplicity of electrodogram representations, we demonstrate that these signals are computationally less complex than raw audio, making them more suitable for neural network processing.

Objective evaluations, including signal-to-noise ratio improvements and linear cross-correlation analysis, show that our adversarial deepACE model outperforms baseline approaches in generating high-quality electrodograms while maintaining a reduced parameter count. These findings suggest that the proposed method offers a promising direction for integrating advanced deep learning techniques into CI sound coding pipelines, with potential benefits for real-time applications. While further validation through listening tests is needed, this study provides initial evidence of the effectiveness of adversarial learning and simplified signal representations in advancing CI speech processing.