On the robustness of non-intrusive speech quality model by adversarial examples

TL;DR

This paper demonstrates that deep learning speech quality models are vulnerable to subtle adversarial attacks and shows that adversarial training can improve their robustness against such perturbations.

Contribution

It reveals the vulnerability of deep speech quality models to adversarial examples and evaluates adversarial training as a method to enhance their stability.

Findings

Deep models can be fooled by minimal perturbations as low as -30 dB.

Adversarial training improves the robustness of speech quality prediction models.

Vulnerability exposes potential reliability issues in deploying these models in real-world scenarios.

Abstract

It has been shown recently that deep learning based models are effective on speech quality prediction and could outperform traditional metrics in various perspectives. Although network models have potential to be a surrogate for complex human hearing perception, they may contain instabilities in predictions. This work shows that deep speech quality predictors can be vulnerable to adversarial perturbations, where the prediction can be changed drastically by unnoticeable perturbations as small as $-30$ dB compared with speech inputs. In addition to exposing the vulnerability of deep speech quality predictors, we further explore and confirm the viability of adversarial training for strengthening robustness of models.