Disfluency Assessment Using Deep Super Learners

Abstract

Speech Disfluency detection and classification are critical in speech therapy because they aid in tracking disfluency progression and are a major tool in technology-assisted speech therapy. Existing methods for detecting disfluency in speech are heavily reliant on annotated data, which can be costly. Machine learning algorithms are rapidly gaining recognition for assessing speech fluency, reducing human error and minimizing therapy delays, making them a preferred method over manual diagnosis. Furthermore, these algorithms do not account for variable-length disfluent speech, limiting the scalability of detection systems. For the automated detection of speech impairment, a Pipelined Deep Learner-Dual Classifier (PDL-DC) is proposed. Individuals' speech fluency was evaluated in two stages: the classification of speech disfluencies and disorders into several groups such as revisions, prolongations, whole-word repetitions, wordmedial repetitions, and filled pauses. The second aspect of classification pertains to the assessment of fluency levels, wherein speakers are classified into three categories: healthy individuals, individuals with stuttering, and individuals with Specific Language Impairment (SLI). The proposed model's implementation of a pipelined design enables the dual validation of a subject's fluency. The proposed model demonstrates an average classification accuracy, precision, and recall of 97%.