Clustering Aphasic Speech: A Comparative Study of Feature Extraction Techniques for Fluent and Non-Fluent Categories

Area of Science:

• Computational linguistics
• Speech-language pathology
• Machine learning

Background:

• Aphasia assessment is vital for rehabilitation, aiding in classification, subtype identification, and severity evaluation.
• Unsupervised learning offers potential for analyzing complex speech data in aphasia.

Purpose of the Study:

• To explore and compare unsupervised clustering methods for aphasic speech data.
• To evaluate different feature extraction techniques for distinguishing aphasia subtypes.

Main Methods:

• Compared five feature extraction methods (MFCCs, OpenL3, CAE, etc.) and four clustering algorithms (HDBSCAN, K-means, etc.).
• Focused on differentiating fluent/non-fluent and Broca's, Wernicke's, anomic aphasia subtypes.
• Utilized intrinsic evaluation metrics to assess clustering performance.

Main Results:

• The deep convolutional autoencoder (CAE) model proved most effective for feature extraction.
• CAE significantly improved the performance of clustering algorithms in distinguishing aphasia categories.
• Specific feature extraction methods showed varying effectiveness in capturing aphasic speech characteristics.

Conclusions:

• Unsupervised learning, particularly with advanced feature extraction like CAE, shows great promise for aphasia research.
• This approach can contribute to automated diagnosis and the development of better speech therapy tools.
• Enhanced aphasia assessment and rehabilitation can be achieved through these computational methods.