Deep Learning-Based Noise Reduction Approach to Improve Speech Intelligibility for Cochlear Implant Recipients

Area of Science:

• Audiology
• Artificial Intelligence
• Signal Processing

Background:

• Cochlear implant (CI) recipients often struggle with speech perception in noisy environments due to degraded signal quality.
• Traditional noise reduction (NR) techniques may not sufficiently restore speech intelligibility under challenging acoustic conditions.
• Developing advanced NR methods is crucial for enhancing the quality of life for CI users.

Purpose of the Study:

• To evaluate the clinical effectiveness of a novel deep learning-based noise reduction (NR) approach, termed (NC + DDAE), for Mandarin-speaking CI recipients.
• To compare the performance of the (NC + DDAE) NR approach against classical NR techniques under various noisy conditions and low signal-to-noise ratio (SNR) levels.
• To assess the impact of the proposed NR method on speech recognition and intelligibility.

Main Methods:

• The study utilized a deep learning NR approach comprising a noise classifier (NC) and a deep denoising autoencoder (DDAE).
• Mandarin sentences were corrupted with two types of maskers, two-talker babble, and construction jackhammer noise at 0 and 5 dB SNR levels.
• Performance was evaluated using objective measures (normalized covariance measure) and subjective speech recognition tests with nine Mandarin-speaking CI recipients using their clinical speech processors.

Main Results:

• The (NC + DDAE) NR approach demonstrated higher intelligibility scores compared to two classical NR techniques under challenging listening conditions.
• Objective evaluations and listening tests confirmed the superior performance of the deep learning approach, even under matched and mismatched training-testing conditions.
• The proposed method showed better noise suppression and less distortion of speech envelope information.

Conclusions:

• The novel deep learning-based (NC + DDAE) NR approach significantly enhances speech recognition for Mandarin CI recipients in noisy environments.
• This advanced NR technique offers superior noise suppression and preserves crucial speech envelope information more effectively than conventional methods.
• The findings suggest that the (NC + DDAE) approach holds potential for integration into CI signal processors to improve speech perception.