Automatic intelligibility classification of sentence-level pathological speech.

Area of Science:

• Computational linguistics
• Speech processing
• Biomedical engineering

Background:

• Pathological speech, characterized by distortions from diseases or insults, presents challenges for automatic intelligibility assessment.
• Variability in pathological speech makes computational analysis difficult, hindering diagnosis and treatment design.

Purpose of the Study:

• To develop novel sentence-level features for capturing abnormal variations in pathological speech.
• To propose a posterior smoothing scheme to refine classification results.
• To evaluate the effectiveness of feature and decision fusion for intelligibility classification.

Main Methods:

• Novel sentence-level features were developed for prosody, voice quality, and pronunciation in pathological speech.
• A post-classification posterior smoothing scheme was introduced to refine classification.
• Feature-level and subsystem decision fusions were performed for final intelligibility decisions.
• Performance was evaluated on NKI CCRT (head and neck cancer) and TORGO (cerebral palsy, ALS) databases.

Main Results:

• Individual feature sets (voice quality, prosody, pronunciation) showed significant discriminating power for binary intelligibility classification.
• The proposed posterior smoothing reduced classification errors.
• Smoothed posterior score fusion yielded the best classification performance, with average recalls of 73.5% (unweighted) and 72.8% (weighted).

Conclusions:

• The proposed features and smoothing scheme effectively improve automatic pathological speech intelligibility assessment.
• Fusion strategies enhance classification accuracy, offering a promising tool for clinical applications.
• This approach aids experts in diagnosing speech disorders and designing personalized treatment plans.