Perceptual effects of reducing algorithmic latency on deep-learning based noise reductiona).

Area of Science:

• Auditory Neuroscience
• Signal Processing
• Machine Learning

Background:

• Low latency is critical for real-time noise reduction in hearing devices.
• Deep neural networks (DNNs) offer advanced noise reduction but can introduce latency.
• Balancing DNN performance with latency is essential for practical applications.

Purpose of the Study:

• To investigate the trade-off between algorithmic latency and speech intelligibility in a DNN-based noise reduction system.
• To determine the impact of reduced latency on the performance of an attentive recurrent network (ARN).
• To establish latency requirements for human listeners in noise reduction scenarios.

Main Methods:

• Modified the algorithmic latency of a fully causal, talker-independent ARN by altering the analysis time frame.
• Evaluated the ARN's noise reduction performance on sentences in babble across various signal-to-noise ratios (SNRs).
• Assessed speech intelligibility through listening tests with participants with normal hearing and hearing loss.

Main Results:

• Significant improvements in speech intelligibility were achieved through noise reduction, particularly for listeners with hearing loss and at lower SNRs.
• Objective network performance measures showed slight improvements with increased latency.
• Crucially, human speech intelligibility remained largely unaffected as algorithmic latency decreased from 20 to 10 or 5 ms.

Conclusions:

• Reducing algorithmic latency in DNN-based noise reduction systems is feasible without compromising speech intelligibility for users.
• The study provides valuable insights into latency requirements for effective hearing aid and cochlear implant design.
• Findings support the development of low-latency, high-performance noise reduction solutions for auditory devices.