End-to-end deep learning classification of vocal pathology using stacked vowels

Area of Science:

• Computational linguistics
• Artificial intelligence in healthcare
• Speech pathology

Background:

• Artificial intelligence (AI) technology offers potential for classifying voice disorders using voice recordings as a screening tool.
• Previous models often rely on single vowel recordings, limiting prediction accuracy for vocal pathology.

Purpose of the Study:

• To develop and evaluate AI models that analyze multiple sustained vowel recordings simultaneously to enhance the prediction of vocal pathology.
• To compare the performance of a stacked vowel model against baseline and stacked pitch models for classifying voice disorders.

Main Methods:

• Voice samples from 687 healthy participants and 334 dysphonic patients (hyperfunctional dysphonia, laryngitis) were analyzed.
• Three 1-dimensional convolutional neural network models were trained: a baseline (single vowel), a stacked vowel model (three vowels simultaneously), and a stacked pitch model (one vowel across three pitches).

Main Results:

• The stacked vowel model achieved a higher F1 score (0.81) for multiclass classification compared to the baseline (0.77) and stacked pitch (0.78) models.
• The stacked vowel model demonstrated superior performance in classifying hyperfunctional dysphonia voice samples (F1 score 0.56) compared to other models.

Conclusions:

• Analyzing multiple sustained vowel recordings simultaneously significantly improves AI-driven screening and classification of vocal pathology.
• The stacked vowel model architecture presents a promising approach for enhancing the accuracy of AI-based voice disorder detection.