Dementia Detection from Speech Using Machine Learning and Deep Learning Architectures.

Area of Science:

• Neurology
• Speech-Language Pathology
• Computational Linguistics

Background:

• Dementia significantly impacts memory and language, with speech deterioration being a key indicator.
• Previous dementia recognition studies often combined speech features with linguistic transcriptions.
• Identifying standalone speech features is crucial for efficient and accessible dementia screening.

Purpose of the Study:

• To identify a compact set of standalone speech features for dementia recognition.
• To evaluate the effectiveness of machine learning (ML) and deep learning (DL) models using these features.
• To establish a computationally efficient method for dementia detection.

Main Methods:

• Utilized speech samples from the Pitt corpus in the Dementia Bank.
• Extracted and analyzed prosodic, voice quality, and cepstral speech features.
• Applied ML and DL models to classify dementia based on the selected speech features.

Main Results:

• Machine learning models achieved 87.6% accuracy, outperforming deep learning models at 85%.
• The proposed compact set of speech features demonstrated high efficacy in dementia recognition.
• The ML approach showed lower time and memory consumption compared to DL.

Conclusions:

• Standalone speech features, particularly prosodic, voice quality, and cepstral, are effective for dementia recognition.
• Machine learning models offer a superior and more efficient approach than deep learning for this task.
• This research provides a promising foundation for developing accessible, speech-based dementia screening tools.