An effectively causal deep learning algorithm to increase intelligibility in untrained noises for hearing-impaired...

Area of Science:

• Hearing Technology
• Signal Processing
• Artificial Intelligence

Background:

• Real-time operation is crucial for effective noise reduction in hearing aids.
• Causality, or processing information without future data, is essential for real-time systems.
• Effectively causal algorithms incorporate minimal future information within human perception limits.

Purpose of the Study:

• To apply effectively causal deep learning for speech separation and intelligibility enhancement in hearing-impaired listeners.
• To investigate the performance of a single-microphone, gated convolutional recurrent network for noise reduction.

Main Methods:

• A gated convolutional recurrent network was employed for complex spectral mapping.
• The network estimated both real and imaginary parts of noise-free speech, yielding magnitude and phase.
• The deep neural network was trained on diverse noise datasets and tested on unseen complex noises.

Main Results:

• Significant improvements in speech intelligibility were observed across all tested conditions.
• The greatest benefits were noted in individuals with severe to profound hearing loss.
• The algorithm demonstrated effectiveness even with untrained background noises.

Conclusions:

• Effectively causal deep learning significantly enhances speech intelligibility for hearing-impaired individuals.
• This approach is particularly beneficial in challenging listening environments with complex, unpredicted noises.
• The study validates the use of effectively causal deep learning for practical hearing assistance applications.