Research on Pathological Voice Recognition Based on XGBoost.

Area of Science:

• Speech analysis
• Biomedical engineering
• Machine learning

Background:

• Increasing prevalence of voice disorders necessitates advanced diagnostic tools.
• Acoustic detection offers an objective, non-invasive approach to voice disorder analysis.
• Speech signal analysis is a key area for pathological voice recognition research.

Purpose of the Study:

• To develop and evaluate a machine learning-based classifier for pathological voice recognition.
• To explore the efficacy of nonlinear dynamic parameters extracted via wavelet packet analysis for voice disorder classification.
• To compare the performance of the XGBoost algorithm against Support Vector Machines (SVM) for this task.

Main Methods:

• Utilized 101 continuous /a/ vowels from the German SVD database.
• Applied wavelet packet technology for time-frequency analysis.
• Extracted four nonlinear dynamic parameters: approximate entropy, sample entropy, fuzzy entropy, and permutation entropy.
• Employed the XGBoost machine learning algorithm for classification, validated with five-fold cross-validation and ROC curve analysis.

Main Results:

• The XGBoost classifier achieved an accuracy of 0.857, an F1 score of 0.875, and an AUC value of 0.944.
• These performance metrics surpassed those obtained using a Support Vector Machine (SVM) classifier.
• Nonlinear dynamic parameters proved effective features for pathological voice recognition.

Conclusions:

• XGBoost demonstrates superior performance in pathological voice recognition compared to SVM.
• Wavelet packet-derived nonlinear dynamic features are valuable for classifying voice disorders.
• The proposed method offers a promising approach for objective and non-invasive diagnosis of pathological voices.