Representation Learning Based Speech Assistive System for Persons With Dysarthria.

Area of Science:

• Speech Processing
• Machine Learning
• Assistive Technology

Background:

• Dysarthric speech recognition is challenging due to articulatory deficits.
• Robust representation learning is crucial for accurately processing sequential speech patterns.
• Existing methods like Hidden Markov Models (HMM) and Deep Neural Network-HMM (DNN-HMM) have limitations.

Purpose of the Study:

• To develop a robust representation learning framework for dysarthric speech recognition.
• To enhance the accuracy of converting dysarthric speech to normal speech or text.
• To improve assistive systems for individuals with vocal impairments.

Main Methods:

• A hybrid framework combining generative and discriminative learning approaches.
• Utilizing Example Specific Hidden Markov Models (ESHMMs) for generative data representation.
• Employing a Support Vector Machine (SVM) as a discriminative classifier on fixed-dimensional score vectors.

Main Results:

• The proposed hybrid framework significantly outperforms conventional HMM and DNN-HMM approaches.
• Achieved higher recognition accuracy on the UA-Speech database.
• Demonstrated the effectiveness of score vector representation for very low intelligibility words.

Conclusions:

• The hybrid generative-discriminative approach offers a robust solution for dysarthric speech recognition.
• Example Specific Hidden Markov Models provide effective feature representation for challenging speech patterns.
• This method shows promise for improving assistive communication technologies for individuals with dysarthria.