Date of Award

2022

Document Type

Open Access Dissertation

Degree Name

Doctor of Philosophy in Computer Science (PhD)

Administrative Home Department

Department of Computer Science

Advisor 1

Keith Vertanen

Committee Member 1

Scott Kuhl

Committee Member 2

Laura Brown

Committee Member 3

Elizabeth Veinott

Abstract

People with some form of speech- or motor-impairments usually use a high-tech augmentative and alternative communication (AAC) device to communicate with other people in writing or in face-to-face conversations. Their text entry rate on these devices is slow due to their motor abilities. Making good letter or word predictions can help accelerate the communication of such users. In this dissertation, we investigated several approaches to accelerate input for AAC users. First, considering that an AAC user is participating in a face-to-face conversation, we investigated whether performing speech recognition on the speaking-side can improve next word predictions. We compared the accuracy of three plausible microphone deployment options and the accuracy of two commercial speech recognition engines. We found that despite recognition word error rates of 7-16%, our ensemble of n-gram and recurrent neural network language models made predictions nearly as good as when they used the reference transcripts. In a user study with 160 participants, we also found that increasing number of prediction slots in a keyboard interface does not necessarily correlate to improved performance. Second, typing every character in a text message may require an AAC user more time or effort than strictly necessary. Skipping spaces or other characters may be able to speed input and reduce an AAC user's physical input effort. We designed a recognizer optimized for expanding noisy abbreviated input where users often omitted spaces and mid-word vowels. We showed using neural language models for selecting conversational-style training text and for rescoring the recognizer's n-best sentences improved accuracy. We found accurate abbreviated input was possible even if a third of characters was omitted. In a study where users had to dwell for a second on each key, we found sentence abbreviated input was competitive with a conventional keyboard with word predictions. Finally, AAC keyboards rely on language modeling to auto-correct noisy typing and to offer word predictions. While today language models can be trained on huge amounts of text, pre-trained models may fail to capture the unique writing style and vocabulary of individual users. We demonstrated improved performance compared to a unigram cache by adapting to a user's text via language models based on prediction by partial match (PPM) and recurrent neural networks. Our best model ensemble increased keystroke savings by 9.6%.

Download

Share

COinS