Towards Artificial Speech Therapy: A Neural System for Impaired Speech Segmentation

Area of Science:

• Speech Processing
• Biomedical Engineering
• Machine Learning

Background:

• Accurate segmentation of impaired speech is crucial for identifying speech disorders.
• Existing methods may not effectively isolate speech segments critical for therapeutic analysis.

Purpose of the Study:

• To develop and compare neural system-based techniques for segmenting impaired speech.
• To identify specific speech sections containing information about potential speech impairments relevant to therapists.

Main Methods:

• Developed two segmentation models: one using four artificial neural networks, and another using a support vector machine (SVM).
• The SVM was trained using a nested algorithm combining metaheuristics (like compact differential evolution - CDE) and convex optimization.
• Evaluated performance against Gaussian mixture models and deep learning techniques.

Main Results:

• The SVM model, particularly with a radial basis function, effectively detected speech portions of therapeutic interest.
• A hybrid approach combining population-based methods and CDE for training yielded the best performance.
• Both proposed models outperformed existing Gaussian mixture model and deep learning segmentation techniques.

Conclusions:

• The proposed SVM-based segmentation technique is effective for analyzing impaired speech.
• The nested training algorithm, especially with hybrid metaheuristics, optimizes performance for specific segmentation tasks.
• This approach offers a promising tool for speech therapists in identifying and analyzing speech impairments.