[Psychosis speech recognition algorithm based on deep embedded sparse stacked autoencoder and manifold ensemble].

Area of Science:

• Computational linguistics
• Psychiatry
• Machine learning

Background:

• Speech feature learning is crucial for mental illness recognition.
• Deep learning excels at automatic feature extraction but struggles with small datasets.
• Traditional methods rely on expertise and lack adaptability.

Purpose of the Study:

• To propose a hybrid feature learning algorithm for psychotic speech recognition.
• To address the limitations of small sample sizes in deep learning for mental illness detection.
• To improve the accuracy and stability of psychotic speech classification.

Main Methods:

• Developed a deep embedded hybrid feature sparse stack autoencoder manifold ensemble algorithm.
• Integrated prior knowledge for initial feature extraction and embedded them into a deep network.
• Utilized L1 regularization for feature selection and a manifold projection ensemble for classification.
• Created a large-scale Chinese psychotic speech database for validation.

Main Results:

• The proposed algorithm demonstrated effective feature fusion and improved classification stability.
• Achieved a maximum accuracy improvement of 3.3% compared to other algorithms.
• The novel algorithm outperformed representative methods in psychotic speech recognition.

Conclusions:

• The embedded mixed sparse stack autoencoder and manifold ensemble method offers a significant advancement in psychotic speech recognition.
• The approach effectively enhances recognition rates for mental illness detection using speech.
• The developed large-scale database supports further research in this domain.