Improving auditory attention decoding in noisy environments for listeners with hearing impairment through contrastive...

Area of Science:

• Neuroscience
• Signal Processing
• Machine Learning

Background:

• Auditory attention decoding (AAD) is crucial for understanding speech in complex acoustic environments.
• Hearing impairments necessitate advanced assistive technologies for effective auditory processing.
• Electroencephalography (EEG) offers a non-invasive method for brain activity monitoring.

Purpose of the Study:

• To evaluate the efficacy of contrastive learning in improving AAD using EEG data.
• To compare contrastive learning models against baseline and non-contrastive models in challenging auditory scenarios.
• To assess the potential of EEG-based AAD for clinical applications in hearing technology.

Main Methods:

• Three models were developed: a baseline linear model (LM), a non-linear model without contrastive learning (NLM), and a non-linear model with contrastive learning (NLMwCL) using SigLIP loss.
• EEG data and speech envelopes from 34 listeners with hearing impairment were used for training and evaluation.
• Speech envelope reconstruction accuracy and attention classification accuracy were the primary metrics.

Main Results:

• The NLMwCL model consistently outperformed LM and NLM in both speech reconstruction and attention classification.
• In a 3-second window, NLMwCL achieved 68.0% attention classification accuracy, compared to 64.4% for NLM and 62.6% for LM.
• Attended speech reconstruction accuracy was highest with NLMwCL (0.105), followed by NLM (0.096) and LM (0.084).

Conclusions:

• Contrastive learning offers a significant improvement for AAD in noisy, multi-talker environments.
• EEG-based AAD holds considerable potential for developing advanced neuro-steered signal processing algorithms for hearing aids.
• This research paves the way for enhanced auditory assistance and clinical tools for individuals with hearing loss.