Adversarial Learning For End-To-End Cochlear Speech Denoising Using Lightweight Deep Learning Models

Abstract

This paper investigates an end-to-end speech signal denoising approach for cochlear implants (CIs). Building on previous work, we first explore the effect of relocating the deep envelope detector within the deep learning-based CI sound coding strategy, moving it from the skip connection to the output of the masking operation. This modification enables high-resolution time-frequency masking and optimizes noise reduction. Next, we introduce a discriminator network to further enhance the model by enforcing the generation of higher-quality electrodograms (i.e., the electric pulse patterns that stimulate the auditory nerve). This adversarial learning approach improves the generation of electrodograms and has the potential to enhance speech understanding for CI users. Objective evaluations, including signal-to-noise ratio improvement and linear cross-correlation coefficients, demonstrate that these enhancements significantly boost the performance of the end-to-end CI speech-denoising algorithm while reducing its parameter count, making it suitable for real-time applications.