What Does a Pathological Speech Assessment Model Know about Acoustic Features? A Case Study on Oral and Oropharyngeal Cancer Patients

Researchers analyzed how a Wav2Vec 2.0-based machine learning model interprets acoustic features in speech from oral and oropharyngeal cancer patients. Using canonical correlation analysis, they found the model's learned representations most strongly correlate with spectral and prosodic features, providing practical insights for improving pathological speech assessment systems.

This research advances our understanding of how deep learning models process acoustic information relevant to medical diagnosis. The study reveals that neural speech models naturally prioritize spectral characteristics and prosodic patterns when assessing intelligibility in cancer patients, with the first MFCC coefficient emerging as the strongest predictor across model layers. This finding validates the acoustic features clinicians have long understood as diagnostically important, while simultaneously demonstrating that modern neural architectures independently discover these patterns without explicit instruction.

The work addresses a critical gap in interpretable machine learning for healthcare applications. As AI systems increasingly support clinical decision-making, understanding what these models actually learn becomes essential for building clinician trust and identifying potential failure modes. By correlating model embeddings to established acoustic descriptors through canonical correlation analysis, the researchers provide a methodological template for auditing other pathological speech models.

For the speech-processing and healthcare AI communities, these findings streamline feature engineering for pathological speech tasks. Rather than exhaustively testing all possible acoustic features, developers can prioritize spectral and prosodic analysis given their demonstrated importance. The quantified correlations—0.77 for spectral, 0.71 for prosodic, and 0.65 for voice quality groups—offer benchmarks for future models. This guidance reduces computational overhead while maintaining diagnostic accuracy, accelerating deployment of speech-based screening tools in resource-constrained clinical settings. The research ultimately demonstrates how interpretability analysis strengthens both scientific understanding and practical application of AI in healthcare diagnostics.

• →Wav2Vec 2.0 models prioritize spectral and prosodic acoustic features when assessing pathological speech intelligibility
• →First MFCC coefficient shows highest correlations across all model layers, validating its clinical importance
• →Canonical correlation analysis provides a replicable method for auditing neural speech model interpretability
• →Spectral group features achieve 0.77 correlation while voice quality features achieve 0.65, establishing diagnostic hierarchies
• →Findings enable more efficient feature selection for developing pathological speech assessment systems