Study: DNN Denoising in Hearing Aids May Preserve Mandarin Lexical Tone Perception in Noise

New research published in the American Journal of Audiology suggests deep neural network denoising can help hearing aid users maintain categorical perception of Mandarin tones in challenging listening environments, with the clearest benefits observed at the individual level.

A collaborative study from researchers at the Phonak Audiology Research Center (PARC) Shanghai, Shanghai Jiao Tong University, and Sonova researchers in Switzerland has examined whether deep neural network (DNN) denoising in hearing aids can preserve the perception of Mandarin lexical tones in noisy listening conditions — a question with meaningful clinical implications for hearing care professionals working with Mandarin-speaking patients.

The findings, published in the American Journal of Audiology, suggest DNN denoising can support the maintenance of categorical tone identification for hearing aid users in noise, while exhibiting what the authors describe as a “do no harm” characteristic with respect to the acoustic cues critical for tone distinction.

Why Lexical Tones Matter

In Mandarin Chinese, pitch is not merely a prosodic feature—it carries lexical meaning. Mandarin has four lexical tones, distinguished primarily by differences in fundamental frequency (F0) contours, or the pattern of pitch change over time. For example, the syllable yican mean “one” with a flat pitch (Tone 1) or “aunt” with a rising pitch (Tone 2). Accurate perception of these distinctions is essential for understanding spoken Mandarin.

Previous research has shown that Mandarin speakers with mild-to-moderate sensorineural hearing loss can maintain categorical perception of lexical tones in quiet, although the precision of boundaries between tones may be reduced. What remained less understood was how performance holds up in noise—particularly for hearing aid users—and whether DNN-based noise reduction could alter the very F0 contours listeners rely on to distinguish tones.

Study Design and Methodology

The study, led by researchers at PARC Shanghai in collaboration with Professor Wang from Shanghai Jiao Tong University, focused on categorical perception of Mandarin Tones 1 and 2, which were selected in part because their F0 contours can overlap, making them especially difficult to discriminate in noise.

Hearing aid users were assessed under two conditions—DNN denoising enabled versus disabled—with the two hearing aid programs otherwise matched, allowing researchers to isolate the contribution of DNN-based processing. Participants were evaluated at two signal-to-noise ratios (SNRs): 0 dB SNR and −5 dB SNR in cafeteria noise.

Researchers measured the slope of the tone identification function, where a steeper slope reflects a sharper, more decisive boundary between Tone 1 and Tone 2.

Group-Level vs Individual-Level Findings

At the group level, results were mixed. In quiet, hearing aid users with DNN off demonstrated categorical tone perception comparable to participants with normal hearing. As the listening environment became more demanding, tone identification deteriorated across groups. In the overall analysis of the noise conditions, when DNN processing was disabled, tone identification slopes were significantly shallower than those observed in participants with normal hearing. With DNN processing enabled, slopes did not differ significantly from the normal-hearing group. However, the direct comparison between DNN-on and DNN-off conditions was not statistically significant at the group level.

This prompted the researchers to look more closely at individual-level performance, where clearer patterns emerged. At 0 dB SNR, significantly more listeners showed categorical perception with DNN on than with it off. At the more challenging −5 dB SNR, however, the difference between conditions was not statistically significant, indicating that DNN benefits may depend on the degree of noise difficulty.

Preserving Acoustic Cues for Tone Identification

A core question of the study was whether DNN noise reduction could reduce background noise without distorting the F0 contours needed for tone identification. The authors interpret the findings as suggesting that DNN denoising facilitated perceptual separation of speech from cafeteria noise while preserving the F0 contours needed for tone identification—helping prevent the degradation of categorical perception observed with conventional processing in noise, without introducing harmful distortions to the cues carrying tone information.

Clinical Relevance and Scope

The authors note that the findings are specific to categorical perception of Mandarin Tones 1 and 2 under the conditions investigated and should not be extrapolated to all Mandarin speech or other tonal languages. Within those parameters, the study addresses a gap in the evidence base: categorical perception in hearing aid users, particularly in noise, had remained largely unexplored prior to this work.

For hearing care professionals working with Mandarin-speaking hearing aid users, the research adds new evidence to inform clinical decision-making around advanced hearing aid signal processing and its potential role in preserving linguistically meaningful acoustic features.

The full study—Wang Y, Tian X, Du X, Guan J, Kuehnel V, Launer S & Liu C (2026). Effects of deep neural network-enhanced hearing aids on categorical perception of Mandarin tones. American Journal of Audiology. doi:10.1044/2026_AJA-25-00294—is available via PubMed.