Automatic Classification of Strain in the Singing Voice Using Machine Learning

Area of Science:

• Vocal acoustics and bioacoustics
• Machine learning applications in speech and voice analysis
• Singing voice science and pedagogy

Background:

• Classifying vocal strain is crucial for protecting professional singers and optimizing vocal training.
• Current methods for assessing vocal strain can be subjective and time-consuming.
• Distinguishing between normal-mild and moderate-severe strain is key for intervention.

Purpose of the Study:

• To investigate the efficacy of machine learning in automatically classifying singing voices based on perceived vocal strain.
• To compare the performance of different acoustic feature sets and machine learning classifiers for strain detection.
• To analyze singing voice samples from classical and contemporary commercial music (CCM) genres.

Main Methods:

• Analysis of 324 singing voice samples from 15 professional singers (classical and CCM).
• Extraction and comparison of three acoustic feature sets: mel-frequency cepstral coefficients (MFCCs), extended Geneva Minimalistic Acoustic Parameter Set (eGeMAPS), and wavelet scattering features.
• Utilized support vector machine (SVM) and multilayer perceptron (MLP) classifiers with recursive feature elimination for feature selection.

Main Results:

• The highest classification accuracy achieved was 86.1% using wavelet scattering features with the MLP classifier.
• The first MFCC coefficient, indicating spectral tilt, demonstrated the most significant separation between strain categories.
• Selected acoustic features and machine learning models proved effective in differentiating vocal strain levels.

Conclusions:

• Machine learning models can automatically classify perceptual vocal strain in singing voices with high accuracy.
• This approach has the potential to support vocal health monitoring and singing training programs.
• Further research with larger, diverse singer populations across multiple genres is warranted.