Tunable Q-factor Wavelet Transform-Based Features in the Classification of Phonation Types in the Singing and...

Area of Science:

• Acoustic phonetics
• Signal processing
• Machine learning

Background:

• Phonation, the production of audible sounds using the laryngeal and respiratory systems, allows for various voice types like breathy, neutral, and pressed.
• These phonation types are crucial in both singing and speaking, influencing vocal quality and communication.

Purpose of the Study:

• To propose and evaluate a novel feature extraction method for classifying phonation types.
• To assess the effectiveness of tunable Q-factor wavelet transform (TQWT) features for discriminating between breathy, neutral, and pressed phonation.

Main Methods:

• Voice signals were decomposed into sub-bands using the tunable Q-factor wavelet transform (TQWT).
• Shannon wavelet entropy was calculated for each sub-band.
• A feed forward neural network was trained using these entropy values to classify phonation types.

Main Results:

• The proposed TQWT-based features demonstrated superior performance compared to six existing state-of-the-art features.
• The method achieved high classification accuracies: 91% for singing voices and 82% for speaking voices.
• TQWT features effectively discriminated between breathy, neutral, and pressed phonation types.

Conclusions:

• The tunable Q-factor wavelet transform (TQWT) provides effective features for phonation type classification.
• This approach offers a significant advancement in accurately identifying different voice qualities in both singing and speaking applications.