Deep learning applications in telerehabilitation speech therapy scenarios

Area of Science:

• Speech and Language Pathology
• Biomedical Engineering
• Artificial Intelligence

Background:

• Automatic Speech Recognition (ASR) is valuable in speech therapy for dysarthria, but atypical speech presents significant challenges.
• Standard ASR systems fail with dysarthric speech due to high variability and lack of specialized datasets, especially for non-English languages.
• Existing ASR technology requires adaptation for effective use in speech disorder treatment.

Purpose of the Study:

• To develop a speaker-dependent ASR system for recognizing isolated keywords in the speech of native Italian speakers with dysarthria.
• To leverage a mobile application for collecting audio data from individuals undergoing speech therapy for articulation exercises.
• To address the limitations of standard ASR in handling the unique characteristics of dysarthric speech.

Main Methods:

• Collected audio data from Italian speakers with dysarthria using a mobile app during articulation exercises.
• Trained a convolutional neural network (CNN) model using the collected atypical speech data.
• Employed a speaker-dependent approach for keyword spotting within the dysarthric speech samples.

Main Results:

• The trained CNN model demonstrated effectiveness in recognizing a small set of keywords within atypical speech.
• The speaker-dependent method achieved reliable keyword spotting for individuals with dysarthria.
• The developed ASR system shows promise for practical application in speech therapy.

Conclusions:

• The study successfully developed an ASR system tailored for Italian dysarthric speech, overcoming common recognition barriers.
• The system facilitates personalized telerehabilitation by enabling remote monitoring and treatment of dysarthria patients.
• This approach supports speech language pathologists in providing effective, technology-assisted interventions for articulation disorders.