Enhancing the Performance of Pathological Voice Quality Assessment System Through the Attention-Mechanism Based...

Area of Science:

• Computational Linguistics
• Artificial Intelligence
• Speech Pathology

Background:

• Current voice quality assessment relies on subjective auditory-perceptual evaluations by physicians, leading to variability in diagnosis.
• Subjectivity and time intervals in diagnosis can impact the accuracy of pathological voice assessment.
• An accurate, automated system is needed to enhance the reliability and efficiency of voice quality evaluations.

Purpose of the Study:

• To develop and evaluate a novel computerized system for pathological voice quality assessment.
• To improve the accuracy and consistency of voice disorder diagnosis using deep learning.
• To create a system that mimics professional doctors' evaluation of voice parameters.

Main Methods:

• A self-attention-based bidirectional long-short term memory (SA BiLSTM) deep learning model was developed.
• The model incorporated various pitches (low, normal, high) and vowels (/a/, /i/, /u/) to enhance learning.
• The system was trained to learn the complex, high-dimensional aspects of voice quality evaluation.

Main Results:

• The SA BiLSTM system demonstrated superior performance compared to baseline deep neural network and convolution neural network systems.
• The macro average F1 scores for grade, roughness, and breathiness were significantly higher with the proposed system (0.768, 0.820, 0.815) than baseline systems.
• The experimental results confirm the enhanced accuracy of the SA BiLSTM model in classifying voice quality parameters.

Conclusions:

• The proposed SA BiLSTM system, utilizing specific pitches and vowels, provides a more accurate method for voice quality assessment.
• This advanced system can significantly aid clinical voice evaluations, leading to improved patient benefits from voice therapy.
• The findings suggest a promising future for AI-driven tools in objective pathological voice analysis.