Affiliations: [a] Conversational Technologies, Plymouth Meeting, PA, USA | [b] Psycholinguistic Technologies, Inc., Jenkintown, PA, USA | [c] University of Maryland School of Medicine, Baltimore, MD, USA
Correspondence:
[*]
Address for correspondence: Deborah A. Dahl, Conversational Technologies, 1820 Gravers Road, Plymouth Meeting, PA 19462, USA. Tel.: +1 610 888 4532; E-mail: [email protected].
Abstract: Individuals with aphasia have difficulty with many language tasks. In addition, hemiplegia of the dominant side of the body, which frequently accompanies aphasia, can limit use of a writing implement, keyboard, or mouse. These factors make it difficult for individuals with aphasia to create text. Automatic speech recognition (ASR) could address this need. However, ASR accuracy for aphasic speech is limited. Work on decoupling speech production from real-time constraints has shown that some individuals with aphasia can improve the quality of their speech by using a “processing prosthesis,” software that allows users to record speech fragments and build them into larger structures by manipulating visual icons. This paper describes a study assessing the efficacy of combining the prosthesis with ASR. ASR accuracy was compared for speech samples created with and without the prosthesis by four individuals with aphasia; for three participants, ASR was markedly more accurate for utterances created with the prosthesis. While the support of the prosthesis alone does not increase ASR accuracy enough to support general dictation applications for users with aphasia, it suggests that offline support along with measures such as training and vocabulary restriction may eventually enable individuals with aphasia to use ASR for everyday tasks.
