A hybrid approach for binary and multi-class classification of voice disorders using a pre-trained model and ensemble...

Area of Science:

• Artificial Intelligence
• Speech Processing
• Bioacoustics

Background:

• Accurate classification of voice disorders is crucial for diagnosis and treatment.
• Current AI methods face challenges in achieving high accuracy, particularly for multi-class classification.
• Existing feature extraction techniques may not capture all relevant vocal characteristics.

Purpose of the Study:

• To propose and evaluate a novel two-stage hybrid AI framework for enhanced voice disorder classification.
• To achieve state-of-the-art accuracies in both binary and multi-class voice disorder classification.
• To compare the proposed method against established techniques using a standard voice database.

Main Methods:

• A two-stage hybrid approach combining deep learning features with traditional classifiers.
• Stage 1: Feature embedding extraction from voice spectrograms using a pre-trained VGGish model.
• Stage 2: Classification using Support Vector Machine (SVM), Logistic Regression (LR), Multi-Layer Perceptron (MLP), and Ensemble Classifier (EC).

Main Results:

• The hybrid VGGish-SVM model achieved the highest accuracy in multi-class classification for male (77.81%) and combined (70.53%) speakers.
• For binary classification, VGGish-SVM and VGGish-EC showed top performance for male and female speakers, respectively.
• The proposed hybrid approach consistently outperformed MFCC, MFCC-glottal, wav2vec, and HuBERT-based methods.

Conclusions:

• The novel hybrid AI framework demonstrates significant potential for improving voice disorder classification accuracy.
• The VGGish feature embeddings combined with classifiers offer a robust approach for complex voice analysis.
• This work provides a foundation for developing advanced automated tools to aid clinical voice disorder assessment.