Training AI as an Assistant: Automating Dysarthria Assessment with Improved Accuracy and Efficiency

Background:

Speech therapists often work with patients suffering from dysarthria due to neurological conditions such as stroke and traumatic brain injury. To obtain an objective measure of speech intelligibility, therapists typically rely on transcribing verbal output. However, this process is time-consuming and impractical for routine clinical practice. As a result, intelligibility is often estimated based on clinical judgment, which is subjective and varies across therapists (Hirsch et al, 2022). Hence there is a need for a more consistent and objective method to assess speech intelligibility.

Aim:

To improve assessment accuracy and efficiency, transcription tools with Artificial Intelligence (AI) integration such as Otter.ai and descript.com were tested. However, these tools frequently autocorrect speech, disregarding the speaker’s intended words and failing to account for local accents. Hence, the aim of this project is to develop a more refined solution for evaluating dysarthric speech, accounting for accent differences and ensuring a more accurate intelligibility assessment.

Methods:

An AI-based framework was built to automate dysarthric speech evaluation, focusing on intelligibility rather than autocorrection. The system identifies mispronunciations, localizes errors, and classifies them into specific categories. The AI was trained and tested on a dataset of dysarthric speech samples from stroke survivors. The system was designed to be sensitive to local accent variations, ensuring that intelligibility ratings remain reliable across diverse speech patterns.

Results:

The AI framework demonstrated high accuracy in detecting substitution errors (up to 70%), while it performed moderately in detecting deletion and insertion errors. Results using Whisper-based Automatic Speech Recognition models showed a strong correlation between the AI-derived clarity scores and therapist-provided ratings, suggesting that the model can consistently evaluate dysarthric speech. However, mispronunciations related to repetition and prosody proved more challenging for the AI, highlighting areas for future improvement.

Conclusion:

Integrating AI into dysarthria assessment can significantly enhance both accuracy and efficiency. The findings suggest that AI can provide consistent and objective feedback, complementing traditional speech therapy practices.

References:

Hirsch, M. E., Thompson, A., Kim, Y., & Lansford, K. L. (2022). The Reliability and Validity of Speech-Language Pathologists’ Estimations of Intelligibility in Dysarthria. Brain Sciences, 12(8), 1011. https://doi.org/10.3390/brainsci12081011