Adversarial Training for Efficient End-to-End Speech Denoising in Cochlear Implants
Adversarial Training for Efficient End-to-End Speech Denoising in Cochlear Implants
Authors: Tom Gajecki¹, Waldo Nogueira¹
¹Hannover Medical School
Understanding speech in noisy environments, like trying to follow a conversation at a busy party, known as the cocktail party problem, remains a major challenge for people with cochlear implants (CIs). Traditional approaches often rely on front-end speech enhancement, which cleans the audio signal before it reaches the CI signal processing stages. However, these methods can fall short because they don’t consider how CIs convert sound into electrical stimulation.
Building on our previous work in end-to-end CI processing, which directly maps raw audio to the implant’s electrical stimulation patterns (electrodograms), we propose a refined deep learning approach that can be applied regardless of the specific sound coding strategy. End-to-end models offer a more integrated solution by optimizing the entire process, from sound input to neural stimulation, in one unified system.
Our main innovation is the addition of adversarial training, a technique borrowed from the field of generative modeling. We introduce a lightweight network that helps guide the main model during training to produce cleaner and more natural stimulation patterns. This extra network is used only during training, so it does not increase the size or slow down the final system.
Results across a range of noisy conditions show that our enhanced model consistently improves speech clarity compared to earlier methods, without adding extra computational cost. These findings highlight the potential of adversarial end-to-end models to make hearing in complex environments easier and more natural for CI users.
