Evaluation of an Automatic Speech Recognition Platform for Dysarthric Speech

Abstract

Introduction: The use of commercially available automatic speech recognition (ASR) software is challenged when dysarthria accompanies a physical disability. To overcome this issue, a mobile and personal speech assistant (mPASS) platform was developed, using a speaker-dependent ASR software.

Objective: The aim of this study was to evaluate the performance of the proposed platform and to compare mPASS recognition accuracy to a commercial speaker-independent ASR software. In addition, secondary aims were to investigate the relationship between severity of dysarthria and accuracy and to explore people with dysarthria perceptions on the proposed platform.

Methods: Fifteen individuals with dysarthric speech and 20 individuals with nondysarthric speech recorded 24 words and 5 sentences in a clinical environment. Differences in recognition accuracy between the two systems were evaluated. In addition, mPASS usability was assessed with a technology acceptance model (TAM) questionnaire.

Results: In both groups, mean accuracy rates were significantly higher with mPASS compared to the commercial ASR for words and for sentences. mPASS reached good levels of usefulness and ease of use according to the TAM questionnaire.

Conclusions: Practical applicability of this technology is realistic: the mPASS platform is accurate, and it could be easily used by individuals with dysarthria.

Keywords: Automatic speech recognition; Disabilities and handicaps; Dysarthria; Intelligibility; Speech therapy.