RP208 - Does DNN Denoising Improve the Neural Discrimination of Speech?

Abstract: This investigation assessed whether deep-neural-network (DNN) denoising improves the neural discrimination of speech, using mismatch negativity (MMN) as an objective biomarker. Speech syllables /ba/ (standard stimulus, 85% probability) and /la/ (deviant stimulus, 15% probability) were presented in quiet and in noise via an oddball paradigm. Noise significantly delayed MMN latencies and suppressed its amplitude. Critically, activating the DNN reversed this neural degradation, significantly shortening response times while also initiating a restorative trend in MMN amplitude. Results of this study provide objective evidence that DNN denoising restores the brain's automatic speech encoding in noise, explaining the clinical efficacy of the DNN.

Summary: [Note: Intermediate instructional level assumes a fundamental understanding of evoked potentials (e.g., auditory brainstem responses).]
Deep-neural-network (DNN) denoising functions by estimating the real and complex components of the environmental noise, enabling the hearing instrument to effectively remove noise while preserving speech. It was hypothesized that by enhancing the signal-to-noise ratio (SNR) to provide cleaner speech, the DNN enhances the brain's bottom-up encoding of speech sounds.
The primary goal of this study was to determine whether the benefit of DNN-based noise reduction (previously demonstrated in the domains of speech understanding and listening effort) could be objectively observed in neural outcomes using mismatch negativity (MMN), an event-related potential (ERP) that reflects the brain’s automatic, pre-attentive processing of auditory changes.
Seventeen experienced hearing aid users (13m, 4f) aged 46 to 81 (m=69.1±11.0) with moderate to severe sensorineural hearing loss and normal cognitive function took part in this study.
MMN is evoked using an “oddball” paradigm, where a frequent “standard” stimulus is occasionally replaced by a physically different “deviant” stimulus (e.g., differing in frequency, level, or duration). MMN is obtained by subtracting the averaged ERP to standard stimuli from the averaged ERP to deviant stimuli. The evoked response to the deviant stimuli typically shows a larger negative peak (amplitude) and/or an earlier response time (latency) compared to the standard stimuli. Therefore, MMN is considered an objective measure of auditory discrimination. In the present study, the syllable /ba/ served as the standard (85% probability) and /la/ as the deviant (15% probability).
This investigation was conducted using a 2x2 within-subjects design. MMN responses were recorded across two acoustic conditions (quiet, 70 dB of collocated cafeteria noise) and two hearing aid settings (DNN on, DNN off). The SNR for the noise condition was individualized based on each participant’s SRT-75 as measured by a separate speech discrimination task. Statistical analysis revealed significant effects on both MMN peak latency and amplitude.
MMN response latencies were significantly prolonged by noise (p < .001). Critically, activating the DNN significantly shortened these latencies in both quiet (p < .05) and noise (p < .001) conditions. The effect was most pronounced in noise, where the DNN feature substantially mitigated the neural processing delay caused by background noise.
Noise significantly reduced MMN amplitude, making the response weaker (less negative). However, activating the DNN in the noise condition led to a recovery of MMN amplitudes towards the levels observed in quiet. Although this change did not meet statistical significance (p = .0852), the data do show a restorative trend in the neural responses in noise.
Taken together, these results suggest that DNN-based denoising effectively restores the brain's automatic encoding of speech input in noise. The primary and robust neural benefit observed was an increase in the processing speed of phoneme discrimination, evidenced by the significant shortening of MMN latencies. By preserving speech input while reducing noise, the DNN feature allows the brain to process auditory changes more efficiently at a pre-attentive level.

Brief Summary of Clinical Takeaways: The clinical takeaway for this session is an understanding of how event-related potentials (ERPs) can be used as a tool to objectively validate the neural benefits of advanced denoising and understand why it improves behavioral and/or self-reported speech perception in noise.

Assumptions: Intermediate instructional level assumes a fundamental understanding of evoked potentials (e.g., auditory brainstem responses).