Image classification-driven speech disorder detection using deep learning technique.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Speech Pathology

Background:

• Speech disorders significantly impact communication, social interaction, education, and quality of life.
• Early and precise diagnosis is crucial for successful intervention, but current clinical examinations are time-consuming and subjective.
• Automated Speech Disorder Detection (SDD) models are needed to overcome the limitations of manual clinical assessments.

Purpose of the Study:

• To propose an automated speech disorder detection (SDD) model based on Mel-spectrogram image classification.
• To identify multiple speech disorders accurately and efficiently using advanced machine learning techniques.
• To enhance the accessibility and efficiency of diagnostic tools for speech impairments.

Main Methods:

• Mel-spectrograms were generated from voice samples using a Wavelet Transform (WT) hybridization technique.
• A LEVIT transformer was employed for enhanced feature extraction from the Mel-spectrograms.
• An ensemble learning (EL) approach, combining CatBoost, XGBoost, and Extremely Randomized Tree, was used for classification. Quantization-aware training (QAT) was utilized to reduce computational resources.

Main Results:

• The proposed model achieved an exceptional accuracy of 99.1% on the VOICED and LANNA datasets.
• The model demonstrated efficiency with a limited number of parameters (8.2 million).
• Shapley Additive Explanations (SHAP) values were used to ensure model interpretability.

Conclusions:

• The developed automated SDD model significantly enhances speech disorder classification accuracy and efficiency.
• The approach offers promising prospects for developing accessible and reliable diagnostic tools.
• Future research can integrate multimodal data for broader application across languages and dialects, enabling real-time clinical and telehealth deployment.