Continuous Speech for Improved Learning Pathological Voice Disorders

Area of Science:

• Speech pathology
• Computational linguistics
• Machine learning in healthcare

Background:

• Previous research successfully differentiated normal and abnormal voice samples.
• Further classification of specific voice disorders from speech signals remains a challenge.
• Limited studies explore continuous speech for multi-class voice disorder identification.

Purpose of the Study:

• To propose a novel framework for classifying four common voice disorders using continuous Mandarin speech.
• To compare the performance of continuous speech analysis against single-vowel analysis for voice disorder classification.
• To evaluate the effectiveness of a bi-directional long-short term memory network (BiLSTM) for modeling sequential speech features in voice disorders.

Main Methods:

• Acoustic signals were converted into mel-frequency cepstral coefficients (MFCCs).
• A bi-directional long-short term memory network (BiLSTM) was employed to capture temporal dependencies in speech.
• Experiments utilized a large-scale database of 1,045 continuous Mandarin speech recordings collected between 2012 and 2019.

Main Results:

• The proposed continuous speech framework achieved accuracy improvements of 78.12-89.27%.
• Unweighted average recall saw significant gains, ranging from 50.92-80.68%, compared to single-vowel methods.
• Principal component analysis visualized model capabilities, and sensitivity analysis provided insights into disorder-specific performance.

Conclusions:

• Continuous Mandarin speech analysis with BiLSTM demonstrates superior performance for voice disorder classification.
• The findings align with other machine learning models, confirming the robustness of the approach.
• Alternative experiments on balanced datasets reinforced the advantages of using continuous speech for learning voice disorders.