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

2026-06-23

AI Summary

This study addresses the lack of interpretability in deep models for assessing speech intelligibility in patients with oral and oropharyngeal cancer. It presents the first application of Canonical Correlation Analysis (CCA) to investigate the interpretability of Wav2Vec 2.0 representations in pathological speech tasks, quantifying their associations with low-level eGeMAPS acoustic descriptors grouped by prosody, spectrum, and voice quality. The findings reveal that Wav2Vec 2.0 embeddings exhibit strong correlations with spectral (r = 0.77) and prosodic (r = 0.71) features, with the first MFCC dimension showing the highest correlation across all layers. Voice quality features also demonstrate substantial association (r = 0.65). These results provide empirical evidence and practical guidance for selecting acoustic features in pathological speech assessment.

Abstract

This work investigates the interpretability of a Wav2Vec 2.0based speech intelligibility assessment model for oral and oropharyngeal cancer patients through canonical correlation analysis. By measuring the correlation between the model embeddings and eGeMAPS low-level descriptors (LLDs) as an interpretable reference, we analyze how acoustic information is encoded across the model layers. The analysis is conducted at two levels: individual LLDs layer-wise, and group-level: prosodic, spectral, and voice quality. Results show that the learned representations are most strongly correlated with spectral and prosodic features, with the first MFCC coefficient yielding the highest correlations across all layers. At the group level, spectral and prosodic groups achieve correlations of 0.77 and 0.71 respectively, while voice quality reaches 0.65. Beyond model interpretability, this work also offers practical guidance on acoustic feature selection for pathological speech assessment.

Problem

• pathological speech
• acoustic features
• model interpretability
• oral cancer
• speech assessment

Innovation

• interpretability
• canonical correlation analysis
• Wav2Vec 2.0
• pathological speech
• acoustic features