Authors: Tom Gajecki1
1Â Hannover Medical School)
Background Cochlear implants (CIs) have been successful in restoring hearing for people with profound sensorineural hearing loss. While they work well in quiet conditions, speech understanding significantly declines when there is background noise. To solve this issue, CI systems have speech enhancement modules, but they are only effective in certain noisy environments, leaving room for improvement in challenging situations.
Method This study proposes a deep neural network (DNN) to replace the CI sound coding strategy, providing end-to-end speech denoising by taking raw audio and providing a denoised electrodogram. The proposed algorithm, called ‘Deep ACE,’ emulates a common CI sound coding strategy and is designed to remove unwanted noises while accurately coding acoustic signals without sacrificing processing latency. The model was optimized using a CI-specific loss function and evaluated using objective measures and listening tests with CI participants.
Results show that the proposed model achieved higher scores than baseline algorithms, and eight CI users showed the highest speech intelligibility results with the proposed deep learning-based sound coding strategy.
