Enhance, Remix, Compress: Comodulation Effects of Dynamic Range Compression in Speech Enhancement pipelines

Background: Dynamic range compression (DRC) is an essential feature of hearing aids for improving audibility, but in situations with non-stationary noise, it can distort speech cues due to comodulation effects. Recent studies indicated that separate compression of sources may improve intelligibility, but this requires effective, real-time separation of speech and noise. Here we assess whether remixing sources before or after compression can mitigate these detrimental effects.

Methods: A word recognition task was conducted to evaluate speech intelligibility with 14 native Greek speakers with normal hearing. The stimuli consisted of GrHarvard sentences, mixed with restaurant noise at –3 dB SNR. A dual-path recurrent neural network was used to separate speech and noise signals. A wideband DRC was employed with a ratio of 2.78:1, a knee point at 50 dB SPL, an attack time of 5 ms, and a release time of 100 ms. Stimuli were presented diotically at 65 dB SPL via headphones. Experimental conditions assessed differences in speech intelligibility when DRC was applied before or after mixing speech and noise signals, and when applied to estimated sources separated by the neural network, either before or after remixing at a higher SNR of +3 dB.

Results: Objective analysis showed that intelligibility, as quantified by eSTOI, improved by 6% with ideal sources and 4% with enhanced sources when speech and noise were compressed separately rather than jointly. Listening tests are currently ongoing for the remaining participants, and the final results will include statistical comparisons across processing conditions (RM-ANOVA and t-tests).

Conclusions: This study highlights the importance of DRC placement in speech enhancement pipelines, particularly in realistic non-stationary noise conditions. Our findings may inform future hearing aid designs, encouraging source-specific DRC for SNR enhancement.