Suppressing Noise Disparity in Training Data for Automatic Pathological Speech Detection

TL;DR

The paper addresses the problem that noise disparity between healthy and pathological speech recordings biases automatic pathological speech detectors toward noise cues. It introduces a cross-group noise augmentation method that uses VAD-derived noise estimates and scaling factors $\alpha_h = \sqrt{\frac{P_{s_h}}{P_{s_p}}}$ and $\alpha_p = \sqrt{\frac{P_{s_p}}{P_{s_h}}}$ to synthesize augmented utterances $\hat{y}_h$ and $\hat{y}_p$ with aligned noise characteristics, aiming to equalize SNR relationships across groups. The approach is validated with CNN- and wav2vec2-based detectors on synthetic PC-GITA-derived data across three noise types and SNR settings, using both oracle and practical (VAD-based) estimations; oracle results show near-equal performance on noisy and clean tests, while practical results improve clean-test performance but reveal a gap due to estimation limitations. Overall, suppressing noise disparity helps models focus on pathology-discriminant cues, enhancing robustness of pathological speech detection in noisy real-world data.

Abstract

Although automatic pathological speech detection approaches show promising results when clean recordings are available, they are vulnerable to additive noise. Recently it has been shown that databases commonly used to develop and evaluate such approaches are noisy, with the noise characteristics between healthy and pathological recordings being different. Consequently, automatic approaches trained on these databases often learn to discriminate noise rather than speech pathology. This paper introduces a method to mitigate this noise disparity in training data. Using noise estimates from recordings from one group of speakers to augment recordings from the other group, the noise characteristics become consistent across all recordings. Experimental results demonstrate the efficacy of this approach in mitigating noise disparity in training data, thereby enabling automatic pathological speech detection to focus on pathology-discriminant cues rather than noise-discriminant ones.