SBIR Phase I: A tool for estimating objective outcome measures in clinical speech applications

点击翻译按钮获取中文摘要

The broader impact/commercial potential of the Small Business Innovation Research (SBIR) Phase I project is to make uniformly high quality practice by Speech-Language Pathologists (SLPs) accessible to non-ambulatory and underserved rural populations. Because the inability to engage in spoken communication is among the most debilitating of all human conditions, poor access to high quality speech-language pathology services is an important societal issue. It creates significant health disparities affecting not only the quality of peoples' lives but also the economics of their communities. Because speech-language pathology treatment is highly behavioral and intensive, any proposed solution must improve both the performance and productivity of individual SLPs. To provide the necessary scalability required for a nation-wide problem, any implementation must include networked mobile tools as well as a strong algorithmic foundation that automatically generates objective, reliable, and sensitive outcome measures. The project has the potential for significant commercial impact. There are currently 160K SLPs in the US and that number expected to exceed 240K by 2024. Following a software-as-a-service model where clinicians freely download software to their mobile device but need active a subscription to access outcome ratings, projects to a total clinical market of between $20M and $100M.The proposed project addresses the strong demand for the development of dependent objective measures of pathological speech that exists in a variety of clinical settings. Currently, perceptual (subjective) ratings of speech conducted by clinicians are the gold standard for evaluating speech improvement or decline. While it is acknowledged that subjective ratings have poor validity and reliability, objective measures have not been readily available. This project takes a novel intellectual approach to the problem by building machine-learning algorithms that objectively model experts' subjective ratings, but with levels of reliability that far exceed that of clinicians. The computational engine takes as an input a set of speech samples from a speaker and automatically evaluates the speech along clinically-standard perceptual dimensions. This yields outputs that are immediately clinical interpretable, thereby exceeding the value of norm-based objective outcomes. The project goals are, (i) to refine and extend existing technology for the evaluation of pathological speech; and (ii) to rigorously evaluate its performance and utility using beta testing. Statistical analyses will determine whether the model outperforms the human ratings with respect to reliability and sensitivity, as has been the case in pilot tests.