This research adapts powerful large language models to support letter-by-letter communication for people with speech impairments.

Key Innovations:

• Fine-tuned language models for character-level prediction rather than standard subword tokens
• Created a specialized dataset of sentences rated by usefulness for AAC users
• Developed practical techniques to make predictions both accurate and computationally efficient
• Improved communication speed and accuracy for AAC users

Medical Impact: This research directly addresses a critical need for the 2 million Americans with severe speech disabilities, potentially transforming how AAC devices assist users in expressing themselves more naturally and efficiently.

Adapting Large Language Models for Character-based Augmentative and Alternative Communication