Enabling Off-the-Shelf Disfluency Detection and Categorization for Pathological Speech

Area of Science:

• Computational linguistics
• Speech processing
• Artificial intelligence in healthcare

Background:

• Speech disfluencies, like pauses and repetitions, disrupt speech flow and can indicate health conditions, especially cognitive impairment.
• Current disfluency modeling research prioritizes detection over categorization, with specific challenges in classifying repetitions and revisions.

Purpose of the Study:

• To evaluate Bidirectional Encoder Representations from Transformers (BERT) against other models for speech disfluency detection and categorization.
• To propose and assess a novel fine-tuning approach using triplet loss to improve the distinction between repetitions and revisions.

Main Methods:

• Comparative analysis of BERT and other models on standard disfluency datasets.
• Implementation of a secondary fine-tuning task for BERT utilizing triplet loss to differentiate speech repairs.
• Application of fine-tuned models to analyze small corpora of pathological speech.

Main Results:

• BERT, particularly with triplet loss, demonstrated superior performance in both disfluency detection and categorization compared to previous methods.
• The proposed method showed significant improvements in classifying challenging categories like repetitions and revisions.
• The study presents the first analysis of fine-tuning BERT on available disfluency data for off-the-shelf use on pathological speech.

Conclusions:

• BERT with triplet loss offers a powerful and improved approach for speech disfluency analysis.
• This method shows promise for clinical applications, aiding in the assessment of conditions like cognitive impairment through speech analysis.
• The fine-tuning strategy enables effective application of advanced models to specialized pathological speech datasets.