Test-Time Adaptation for Automatic Pathological Speech Detection in Noisy Environments
Mahdi Amiri, Ina Kodrasi · 2024
Deep learning-based pathological speech detection approaches are gaining popularity as a diagnostic tool to support time-consuming and subjective clinical assessments. While these approaches perform well in controlled environments with clean recordings, their performance significantly degrades in realistic scenarios with background noise. In this paper, we propose a test-time adaptation framework to increase the robustness of such approaches to background noise during inference. To this end, we use a voice activity detector to extract noise-only segments from the test signal. These segments are used to augment a portion of the training/validation data, which is then exploited to fine-tune the classification models. Extensive experimental results demonstrate the effectiveness of the proposed framework in increasing robustness to noise for state-of-the-art automatic patholoaical speech detection approaches.