Diagnosis of pathological speech with streamlined features for long short-term memory learning

Area of Science:

• Artificial Intelligence in Medicine
• Speech Pathology Diagnostics
• Biomedical Signal Processing

Background:

• Accurate pathological speech diagnosis is vital for effective treatment and improving patient quality of life.
• Rising global incidence of speech disorders necessitates efficient and reliable diagnostic tools.
• Advanced research in speech pathology is crucial for developing better intervention strategies.

Purpose of the Study:

• To introduce novel features for deep learning in analyzing short voice signals for pathological speech detection.
• To enhance the precision and reliability of diagnostic procedures for voice conditions.
• To enable more targeted treatment approaches for speech disorders.

Main Methods:

• Incorporation of time-space and time-frequency features for deep learning analysis.
• Utilized long short-term memory (LSTM) networks for voice signal analysis.
• Employed a data balancing strategy with a publicly available voice database.

Main Results:

• Achieved 90% accuracy, 93% sensitivity, 87% specificity, 88% precision, and an F1 score of 0.90.
• Demonstrated a high area under the ROC curve of 0.96.
• Outperformed existing methods using wavelet-time scattering coefficients and other feature types.

Conclusions:

• Time-frequency and time-space features show significant promise for AI-driven speech pathology diagnosis.
• The proposed approach can enhance accuracy and enable real-time pathological speech assessment.
• Facilitates more targeted and effective therapeutic interventions for individuals with speech disorders.