Modal and non-modal voice quality classification using acoustic and electroglottographic features.

Area of Science:

• Speech and Audio Processing
• Biomedical Engineering
• Machine Learning for Signal Processing

Background:

• Accurate voice quality classification is crucial for diagnosing voice disorders.
• Evaluating different acoustic and physiological voice features is essential for improving classification accuracy.
• Mel-frequency cepstral coefficients (MFCCs) are widely used but their effectiveness across different voice signals needs further investigation.

Purpose of the Study:

• To compare the performance of the COVAREP feature set against MFCCs for voice quality classification.
• To evaluate feature effectiveness across acoustic, glottal inverse filtered (GIF), and electroglottographic (EGG) voice signals.
• To assess the utility of various machine learning classifiers for speaker-independent voice quality analysis.

Main Methods:

• Extracted features including COVAREP (glottal source, frequency warped cepstrum, harmonic model) and MFCCs from voice recordings.
• Utilized acoustic, GIF, and EGG waveforms as input signals.
• Employed Support Vector Machines, Random Forests, Deep Neural Networks, and Gaussian Mixture Models for classification.
• Implemented a leave-one-speaker-out cross-validation strategy for speaker independence.

Main Results:

• The full COVAREP feature set achieved the highest classification accuracy of 79.97%.
• Harmonic model features, a subset of COVAREP, performed strongly with 78.47% accuracy.
• Static+dynamic MFCCs achieved 74.52% accuracy, effectively classifying modal, breathy, and strained qualities from acoustic and GIF signals.
• The EGG waveform showed reduced classification performance compared to other signals.

Conclusions:

• The COVAREP feature set, especially harmonic model features, demonstrates superior performance for voice quality classification compared to MFCCs.
• MFCCs remain a viable option for classifying specific voice quality dimensions from acoustic and GIF signals.
• Speaker-independent classification models can effectively differentiate voice qualities, with variations based on feature sets and signal types.