Comparative study on noise-augmented training and its effect on adversarial robustness in ASR systems

2024-09-03

Computer Speech and Language

AI Summary

This study investigates whether noise-augmented training can simultaneously enhance both adversarial robustness and speech intelligibility of automatic speech recognition (ASR) systems. We conduct systematic experiments comparing multiple noise sources (white Gaussian noise, babble, music), injection strategies (time-domain vs. frequency-domain), and noise intensities, under adversarial attacks guided by projected gradient descent (PGD) and CTC loss, evaluated on Conformer and fine-tuned Whisper models. We propose the first quantitative model linking noise-augmentation strategies to adversarial robustness and introduce a “robustness–intelligibility” trade-off evaluation framework, moving beyond conventional accuracy-only metrics. On LibriSpeech and VoxCeleb, our approach achieves an average 12.7% improvement in adversarial accuracy with <1.5% word error rate (WER) degradation. Results demonstrate that frequency-domain noise injection significantly outperforms time-domain injection, revealing key mechanisms and practical boundaries for robustness enhancement.

Technology Category

Machine Learning: Adversarial Learning & RobustnessComputer Vision: Adversarial Attacks & RobustnessNatural Language Processing: Safety and Robustness

Application Category

Search and Retrieval-Augmented AI: Vertical and domain-specific searchSemantics and Knowledge: Methods to enhance, augment, integrate or synergize semantic models such as knowledge graphs and LLMsSecurity and Privacy: Large-scale security measurements

Problem

• Investigating noise-augmented training effects on ASR adversarial robustness
• Comparing robustness across four ASR architectures under different augmentations
• Evaluating model resilience against white-box and black-box adversarial attacks

Innovation

• Noise-augmented training improves ASR adversarial robustness
• Comparative analysis of four ASR architectures under augmentation
• Background noise and speed variations enhance attack resistance