Abstract
A pathological voice is one that exhibits abnormal quality due to a dysfunction or disease of the vocal mechanism. Pathological voice classification systems enable automated detection of voice disorders and play an increasingly important role in Human-Machine Interaction (HMI) technologies. By classifying speaker attributes, including voice pathologies, these systems enhance the accessibility and adaptability of assistive devices, telemedicine platforms, and voice-controlled interfaces. However, environmental noise can significantly impact the performance of such systems, particularly in real-world settings where recording conditions are less controlled. This study investigates how varying levels of white noise affect pathological voice classification, using a dataset of real clinical recordings collected in collaboration with a hospital, as the benchmark. We analyze changes in classification accuracy, feature importance rankings, and inter-feature correlations as noise levels increase. Our findings reveal that although noise influences feature rankings, especially at noise levels above 50 dB, most of the top 10 features remain stable across different noise levels, underscoring their robustness. We also observe that training on clean data and testing on noisy data up to moderate noise levels yields similar performance to training and testing on data with the same noise levels, a finding that is crucial for real-world applicability. Furthermore, we find that adding noise increases the correlation among features, which may contribute to decreased classification performance by potentially confounding the model. These insights, derived from a dataset created with the support of a hospital and authentic pathological voice recordings, highlight the importance of considering environmental noise in developing robust HMI systems and offer guidance for feature selection and system optimization in noisy conditions.