Towards Automatic Speech-Language Assessment for Aphasia Rehabilitation

Aphasia is a common neurological disorder that can severely impact a person’s communication abilities. Speech-based technology has the potential to reinforce traditional aphasia therapy through the development of automatic speech-language assessment systems. Such systems can provide clinicians with supplementary information to assist with progress monitoring and treatment planning, and can provide support for on-demand auxiliary treatment. However, current technology cannot support this type of application due to two major limitations. First, the majority of speech-language assessment techniques assume the availability of manually labeled transcripts, which are time consuming to obtain and typically not available in real-world clinical applications. Second, automatic speech recognition (ASR) traditionally has poor performance on aphasic speech, resulting in inaccurate transcripts that prevent the automation of these techniques. The focus of this dissertation is on the development of computational methods that can accurately assess aphasic speech across a range of clinically-relevant dimensions without the need for manual transcripts. The dissertation is organized into three parts: • Part I: The first part focuses on novel techniques for assessing qualitative aspects of intelligibility in constrained aphasic speech. In this problem setup, speech production occurs in controlled environments, lexical content is restricted, and the target prompt for each utterance is known. While the speech-language impairments associated with aphasia often prevent exact verbalization of the prompts, this constraint greatly simplifies ASR and allows for more accurate transcript generation. We show that transcripts for constrained aphasic speech can be generated automatically with