Restoring auditory discrimination in noise: mismatch negativity evidence for a deep neural network-based denoising...

Understanding speech in noise is a primary challenge for individuals with sensorineural hearing loss (SNHL). While deep neural network (DNN)-based noise reduction in hearing aids shows behavioral promise, objective neurophysiological evidence for its efficacy is lacking. This study investigated whether a DNN denoising algorithm enhances the automatic neural discrimination of speech in older adults with SNHL.

We recorded the mismatch negativity (MMN), an index of automatic auditory change detection, from seventeen older adults with SNHL using a within-subject design. The MMN was elicited by a phonemic contrast (/ba/ vs. /la/) in quiet and in noise, with the DNN system activated (DNNON) and deactivated (DNNOFF).

In noise, DNN activation significantly shortened MMN peak latency and elicited a robust MMN, which was virtually absent in the DNNOFF condition. Furthermore, DNN activation normalized the neural response, eliminating the compensatory correlation between hearing loss and MMN amplitude observed in the DNNOFF condition.

These findings provide compelling neurophysiological evidence that DNN-based noise reduction enhances the neural representation of speech in challenging listening environments. By accelerating processing speed, rescuing neural discrimination responses, and normalizing cortical activities, the DNN denoising algorithm establishes a clear neural basis for its perceptual benefits, highlighting the MMN as a sensitive validation tool.