Modeling Pathological Speech Perception From Data With Similarity Labels.

Area of Science:

• Speech Science
• Computational Linguistics
• Biomedical Engineering

Background:

• Current pathological speech intelligibility assessment relies on subjective, often unreliable, evaluations by speech-language pathologists (SLPs).
• Existing methods are costly and lack consistent intra- and inter-judge reliability.
• There is a critical need for objective, reliable, and perceptually-relevant measures of pathological speech.

Purpose of the Study:

• To develop a data-driven approach for objective evaluation of pathological speech intelligibility.
• To identify acoustic features that correlate with perceptual dimensions of intelligibility as rated by SLPs.
• To create objective measures that reliably reflect speech signal degradation.

Main Methods:

• Collected perceptual ratings from certified SLPs on pathological speech, including comparative judgments (e.g., rhythm similarity).
• Developed novel cost functions to analyze ordinal data from SLP experiments, differing from traditional machine learning data structures.
• Identified and utilized relevant acoustic features that correlate with the collected ordinal data.

Main Results:

• The proposed method successfully identified acoustic features linked to perceptual intelligibility ratings.
• Objective measures derived from these features demonstrated strong correlation with SLP assessments.
• The approach provides a more consistent and reliable evaluation of speech degradation.

Conclusions:

• A data-driven methodology can effectively create objective measures for pathological speech intelligibility.
• The identified acoustic features and developed measures offer a promising alternative to subjective SLP assessments.
• This approach has the potential to improve the consistency and reliability of pathological speech evaluation.