NUVA: A Naming Utterance Verifier for Aphasia Treatment.

Area of Science:

• Neurology
• Speech-Language Pathology
• Artificial Intelligence

Background:

• Anomia, or word-finding difficulties, is a primary symptom of aphasia, a language disorder often resulting from stroke.
• Current assessment of anomia relies on manual evaluation by speech and language therapists (SLTs) during picture naming tasks.
• Automated speech recognition (ASR) and AI advancements have seen limited application in developing automated systems for this crucial assessment.

Purpose of the Study:

• To develop and evaluate NUVA, an automated utterance verification system utilizing deep learning.
• To classify naming attempts as 'correct' or 'incorrect' for individuals with aphasia (PWA).
• To provide an objective and efficient tool for assessing anomia in PWA.

Main Methods:

• Development of NUVA, a deep learning-based system for classifying speech attempts.
• Testing NUVA on eight native British-English speaking PWA using picture naming tasks.
• Performance evaluation using 10-fold cross-validation and comparison against a baseline ASR and human SLT ratings.

Main Results:

• NUVA achieved a performance accuracy ranging from 83.6% to 93.6% on PWA.
• The system demonstrated a 10-fold cross-validation mean accuracy of 89.5%.
• NUVA's performance surpassed a baseline Google speech-to-text service and was comparable to independent SLT assessments.

Conclusions:

• NUVA represents a significant advancement in the automated assessment of anomia in aphasia.
• The system offers a promising, accurate, and potentially more efficient alternative to manual SLT assessments.
• This AI-driven approach can aid in the diagnosis and monitoring of treatment effectiveness for PWA.