Robust Vocal Quality Feature Embeddings for Dysphonic Voice Detection

Area of Science:

• Speech processing
• Biomedical engineering
• Machine learning

Background:

• Approximately 1.2% of the global population experiences impaired voice production, necessitating reliable automated assessment tools.
• Current automated voice analysis methods often lack generalizability across different datasets and applications.
• There is a significant need for robust and accurate methods for dysphonic voice detection.

Purpose of the Study:

• To develop a deep learning framework for generating acoustic feature embeddings that are sensitive to vocal quality.
• To enhance the robustness of voice analysis models across diverse corpora and conditions.
• To improve the accuracy and generalizability of automatic dysphonic voice detection.

Main Methods:

• A deep learning model was trained using a combination of contrastive and classification loss functions.
• Data warping techniques were applied to input voice samples to increase model robustness.
• The framework was designed to generate acoustic feature embeddings sensitive to voice quality.

Main Results:

• The proposed method achieved high classification accuracy both within and across different corpora.
• The generated embeddings demonstrated sensitivity to voice quality and robustness across varied datasets.
• The model consistently outperformed three baseline methods on clean and deteriorated voice datasets.

Conclusions:

• The developed deep learning framework offers a robust and accurate solution for automatic dysphonic voice detection.
• The method's ability to generalize across corpora makes it suitable for diverse clinical and research applications.
• The generated acoustic embeddings hold potential for further advancements in voice quality assessment.