Convolutional Neural Network-based Speech Enhancement for Cochlear Implant Recipients.

Area of Science:

• Biomedical Engineering
• Signal Processing
• Auditory Neuroscience

Background:

• Existing speech enhancement algorithms show limited success in improving speech intelligibility for cochlear implant (CI) users.
• Current methods often fail to adequately process complex acoustic environments relevant to CI users.

Purpose of the Study:

• To develop advanced speech enhancement algorithms tailored for cochlear implant users.
• To improve speech intelligibility in noisy environments by utilizing a CI-specific feature space.

Main Methods:

• Speech enhancement was performed in a cochlear filter-bank feature space using convolutional neural networks (CNNs).
• Three CNN architectures were investigated: vanilla CNN, spectral-subtraction-style CNN (SS-CNN), and Wiener-style CNN (Wiener-CNN).
• Causal variations of the CNNs were developed to minimize processing delays for real-time CI application.

Main Results:

• All proposed CNN architectures, including causal and non-causal forms, demonstrated significant improvements over existing baseline systems.
• The causal Wiener-CNN achieved superior performance, yielding the best Envelope Coefficient Measure (ECM).
• The developed algorithms show potential for real-time implementation on the CCi-MOBILE research platform.

Conclusions:

• CNN-based speech enhancement in a CI-specific feature space is a promising approach for improving CI user hearing.
• Causal Wiener-CNN offers a robust solution for real-time speech enhancement, enhancing intelligibility in naturalistic settings.
• These algorithms represent a viable advancement for CI users navigating complex acoustic environments.