Towards A Clinical Tool For Automatic Intelligibility Assessment.

Area of Science:

• Speech Processing
• Computational Linguistics
• Biomedical Engineering

Background:

• Automatic assessment of pathological speech intelligibility is an under-explored area.
• Current subjective intelligibility tests are inconsistent, costly, and unreliable.
• Existing automatic methods are primarily for telecommunications, not pathological speech.

Purpose of the Study:

• To propose and evaluate a novel algorithm for automatic intelligibility assessment of pathological speech.
• To address the limitations of current subjective and objective methods for pathological speech.

Main Methods:

• The algorithm captures multi-scale perceptual cues correlated with speech intelligibility.
• Nonlinear classifiers are trained at various time scales.
• Ensemble learning methods combine decisions from multiple classifiers for a final intelligibility assessment.

Main Results:

• Preliminary results demonstrate a marked improvement in intelligibility assessment accuracy.
• The proposed algorithm outperforms existing published baseline results for pathological speech.

Conclusions:

• The developed algorithm offers a more reliable and potentially cost-effective solution for assessing pathological speech intelligibility.
• This approach advances the field of speech processing for clinical applications.