Effects of Deep Neural Network-Enhanced Hearing Aids on Categorical Perception of Mandarin Tones.

PURPOSE: This study empirically evaluated the impact of deep neural network (DNN)-based noise reduction in hearing aids on the categorical perception of Mandarin Tones 1 and 2 for hard of hearing (HH) listeners in cafeteria noise at 0 and -5 dB signal-to-noise ratios (SNRs), compared to normal-hearing (NH) listeners and traditional hearing aid settings. METHOD: Twenty NH and 20 HH listeners participated in this study. HH listeners were fitted with commercial hearing aids in two modes: DNN noise reduction enabled (DNNon) and disabled (DNNoff; retaining directional microphone noise reduction technology). Stimuli consisted of a nine-step flat-to-rising tone continuum synthesized from the vowel /a/, presented in quiet and noise conditions. Listeners identified each stimulus as Tone 1 or 2. Linear mixed-effects models were conducted to analyze the effects of listener group, HA setting, and SNR on logistic functions of tone identification such as slopes and boundaries. RESULTS: In quiet, HH listeners with hearing aids showed categorical perception comparable to NH listeners. In noise, tone identification slopes were steeper at 0 dB than -5 dB SNR across groups. HH listeners using HA without DNN processing (DNNoff) tended to yield shallower slopes compared to NH listeners, while with DNN processing enabled (DNNon), identification slopes remained similar to those of NH listeners. Among HH listeners, categorical identification declined progressively across conditions: quiet > DNNon_0dB > DNNon_-5dB > DNNoff_0dB > DNNoff_-5dB. Slopes of tone identification functions correlated negatively with age and pure-tone average, but only in DNNon conditions. CONCLUSIONS: DNN-based noise reduction effectively maintains the categorical identification of Mandarin tones for HH listeners. By facilitating perceptual segregation and demonstrating a "do no harm" characteristic, that is, suppressing cafeteria noise while preserving essential fundamental frequency contours, the DNN algorithm prevented the severe degradation of categoricality. These findings highlight DNN processing as a valuable tool that addresses the specific perceptual needs of tonal language users in real-world noisy settings.