Objective Intelligibility Assessment by Automated Segmental and Suprasegmental Listening Error Analysis

Area of Science:

• Speech-Language Pathology
• Computational Linguistics
• Biomedical Engineering

Background:

• Subjective intelligibility assessments are preferred but labor-intensive.
• Objective transcript scoring methods require significant manual effort.
• Existing objective metrics may not fully capture intelligibility degradation.

Purpose of the Study:

• To develop an automated transcript scoring approach for objective speech intelligibility assessment.
• To create a holistic metric reflecting segmental and suprasegmental contributions to intelligibility.
• To ensure the automated metrics correlate with human perceptual ratings.

Main Methods:

• Orthographic transcription of 63,840 phrases from 73 speakers with dysarthria.
• Development of algorithms for automated analysis of phoneme and lexical segmentation errors.
• Validation against manual labels and linear regression for predicting perceptual ratings.

Main Results:

• Automated metrics achieved 0.90 correlation with manual phoneme error labels.
• 100% accuracy in identifying and coding lexical segmentation errors.
• Automated metrics significantly predicted variance in perceptual severity and word accuracy.

Conclusions:

• The developed automated approach offers a promising objective measure for speech intelligibility.
• It effectively identifies intelligibility degradation across multiple analytical levels.
• This method has the potential to reduce manual labor in speech analysis.