Detection of Pathological Voice Using Cepstrum Vectors: A Deep Learning Approach

Area of Science:

• Computational linguistics
• Medical informatics
• Signal processing

Background:

• Computerized voice disorder detection is crucial for early screening.
• Endoscopic confirmation is the current standard but invasive.
• Developing automated methods aids in pre-screening voice pathologies.

Purpose of the Study:

• To propose and evaluate a deep learning approach for pathological voice detection.
• To compare the performance of deep neural networks (DNN) against other machine learning algorithms.
• To assess the utility of DNN for voice disorder screening.

Main Methods:

• Retrospective collection of 60 normal and 402 pathological voice samples.
• Extraction of Mel frequency cepstral coefficients from sustained vowel samples.
• Evaluation of DNN, support vector machine, and Gaussian mixture model using cross-validation.

Main Results:

• DNN demonstrated superior performance over support vector machine and Gaussian mixture model.
• DNN achieved 94.26% accuracy in males and 90.52% in females for voice pathology detection.
• DNN validation on the MEEI database yielded 99.32% accuracy.

Conclusions:

• The proposed DNN algorithm effectively utilizes acoustic features to distinguish normal from pathological voices.
• DNN shows promise as an efficient tool for voice disorder screening.
• Further research into DNN applications in clinical and laboratory settings is warranted.