Deep-Learning-Based Automated Classification of Chinese Speech Sound Disorders.

Area of Science:

• Computational Linguistics
• Speech Pathology
• Artificial Intelligence

Background:

• Children's speech sound disorders (SSDs) present diagnostic challenges.
• Accurate classification of SSDs is crucial for effective intervention.
• Automated analysis systems can support clinical diagnosis.

Purpose of the Study:

• To develop and evaluate a computer-based system for diagnosing and classifying four types of Chinese SSDs.
• To assess the performance of neural network models in analyzing acoustic speech data.
• To investigate the utility of Mel-frequency cepstral coefficients (MFCCs) for SSD classification.

Main Methods:

• A speech corpus of 2540 samples from 90 children (aged 3-6) with specific SSDs (stopping, backing, FCDP, affrication) was created.
• Speech samples were annotated by two speech-language pathologists (SLPs).
• Three neural network models were used for classification, with MFCCs as input features and data augmentation techniques applied.

Main Results:

• The system demonstrated accuracy in detecting analyzed pronunciation disorders across different Chinese phrase categories.
• The best multi-class classification using a single Chinese phrase achieved an accuracy of 74.4%.
• Experiments confirmed the system's ability to differentiate between normal and pathological speech patterns.

Conclusions:

• The developed system shows promise for assisting in the diagnosis and classification of children's speech sound disorders.
• Neural network analysis of acoustic features, like MFCCs, is a viable approach for SSD detection.
• Further research can refine the system for broader application in speech pathology.