Simulating speech processing with cochlear implants: How does channel interaction affect learning in neural networks?

Area of Science:

• Machine Learning
• Auditory Neuroscience
• Speech Processing

Background:

• Cochlear implants (CIs) are devices that restore hearing by bypassing damaged parts of the ear.
• Speech processing in CIs is limited by reduced spectral resolution and channel interaction.
• Understanding these limitations is crucial for improving CI efficacy.

Purpose of the Study:

• To develop and apply machine learning models to investigate speech processing with simulated cochlear implant signal degradation.
• To analyze how spectral resolution reduction and channel interaction impact speech perception models.
• To predict how these factors influence learning in cochlear implant users.

Main Methods:

• Utilized a perceptron and a deep convolutional network to classify speech spectrograms.
• Simulated CI signal degradation by reducing spectral resolution and modeling channel interaction.
• Trained models on both degraded and high-resolution speech data.

Main Results:

• Pre-training on high-resolution speech degraded performance when subsequently trained on CI-simulated speech.
• Channel interaction was identified as a key factor causing performance degradation.
• The findings suggest that spectral degradation conflicts with learned high-resolution speech expectations.

Conclusions:

• The spectral degradation inherent in cochlear implant processing can impede adaptation, especially for users with prior normal-hearing experience.
• Reducing channel interaction may accelerate learning for individuals implanted later in life.
• Machine learning provides a valuable tool for understanding and potentially mitigating the effects of CI signal processing limitations.