Progress made in the efficacy and viability of deep-learning-based noise reduction

Area of Science:

• Audiology
• Artificial Intelligence
• Signal Processing

Background:

• Deep learning has advanced noise reduction, improving speech intelligibility for hearing-impaired (HI) listeners.
• Previous deep learning methods for HI listeners had limitations, such as matched training/testing conditions and non-causal operation.
• Real-world application requires algorithms that generalize across diverse conditions and operate in real-time (causally).

Purpose of the Study:

• To assess intelligibility improvements from a current deep-learning-based noise reduction algorithm for HI listeners.
• To compare the performance of the current algorithm against an initial deep learning demonstration from ten years prior.
• To evaluate the algorithm's effectiveness under conditions simulating real-world use, including generalization and causal operation.

Main Methods:

• Utilized an attentive recurrent network for noise reduction.
• Employed diverse noise types, talkers, and speech corpora for training and testing to ensure generalization.
• Implemented a fully causal network architecture for real-time processing.
• Compared results with a similar study from 2013 using comparable stimuli and procedures.

Main Results:

• Significant intelligibility benefits were observed across all tested conditions for HI listeners.
• The average intelligibility improvement was 51 percentage points.
• Performance was comparable to the 2013 study, despite the current algorithm facing more demanding, real-world relevant conditions.

Conclusions:

• Current deep-learning-based noise reduction algorithms provide substantial intelligibility benefits for HI listeners.
• Advances in deep learning enable effective noise reduction even with relaxed constraints for generalization and real-time operation.
• The findings highlight significant progress in making hearing assistance technologies more practical and effective for real-world use.