Towards Automated Vocal Mode Classification in Healthy Singing Voice-An XGBoost Decision Tree-Based Machine Learning...

Area of Science:

• Acoustics and Audio Signal Processing
• Artificial Intelligence in Music Performance
• Voice Science and Technology

Background:

• Auditory-perceptual assessment of singing voice is standard but suffers from inconsistent reliability.
• Machine learning (ML) offers potential for objective, automated voice analysis.
• Previous ML models show promise in classifying pathological and healthy voices.

Purpose of the Study:

• To develop and evaluate an XGBoost machine learning classifier for automated vocal mode classification in healthy singing voices.
• To compare the performance of the ML model against human auditory-perceptual assessments.
• To identify key acoustic features for accurate singing voice classification.

Main Methods:

• An XGBoost decision tree classifier was trained using various acoustic features: mel-frequency cepstrum coefficients (MFCCs), glottal features, voice quality features, and alpha-ratios.
• The model was tested on distinguishing metallic vs. non-metallic singing and general vocal modes in male and female singers.
• Performance was benchmarked against 41 professional singers assessing 64 vocal samples.

Main Results:

• The ML classifier achieved high accuracy in distinguishing metallic (92% F1-score for males, 87% for females) and vocal modes (70% F1 for males, 69% for females).
• Key features for classification were MFCCs and alpha-ratios, with models using only these showing comparable performance.
• The automated system's performance approximated or was subpar to human expert assessment.

Conclusions:

• XGBoost models show significant potential for automated singing voice analysis, particularly using MFCCs and alpha-ratios.
• Current AI models do not yet match human perceptual discrimination accuracy but represent an improvement over prior automated methods.
• Further research may enhance AI capabilities for reliable and objective singing voice assessment.