Deep learning in automatic detection of dysphonia: Comparing acoustic features and developing a generalizable...

Area of Science:

• Artificial intelligence
• Machine learning
• Speech science

Background:

• Auditory-perceptual voice assessment is subjective and limited by rater reliability.
• Deep learning (DL) offers potential for consistent and accessible voice analysis.
• The performance of DL models on various acoustic features for dysphonia detection is not well-established.

Purpose of the Study:

• To develop a generalizable DL framework for identifying dysphonia.
• To evaluate the effectiveness of different acoustic features within a DL model.
• To compare DL model performance against traditional machine learning approaches.

Main Methods:

• Collected sustained phonation recordings from 238 dysphonic and 223 healthy Chinese Mandarin speakers.
• Extracted Mel frequency cepstral coefficients (MFCCs) and mel-spectrograms from normalized audio segments.
• Utilized a convolutional neural network (CNN) for binary classification and cross-validation.

Main Results:

• The mel-spectrogram feature achieved the highest performance, with 97.2% AUC and 92% accuracy.
• The DL framework significantly outperformed two baseline machine learning models on both Chinese and German test sets.
• Optimal classification using all segments of both vowels yielded 95% accuracy on the Chinese set and 92% on the German set.

Conclusions:

• DL provides a feasible method for the automatic detection of dysphonia.
• Mel-spectrograms are a preferred acoustic feature for DL-based voice analysis.
• The developed DL framework can be applied to vocal health screening and remote voice monitoring.