Automatic temporal analysis of speech: A quick and objective pipeline for the assessment of overt stuttering

Fluency disorders, such as developmental stuttering, have been characterized by behavior such as blocks, repetitions, and prolongations in speech. Accurate measurement of overt stuttering behavior can aid in diagnostic evaluation and the determination of optimal treatment for this disorder. This study proposes a method - Automatic Temporal Analysis of Speech (ATAS) - for the assessment of speech fluency based on the detection and quantification of discrete pauses and vocal events. Our ATAS metrics include speech rate, total pause time, pause count, mean pause duration, mean vocal duration, pause duration variability, and vocal duration variability. We used oral reading audio samples from a total of 35 English-speaking participants: 17 from adults who stutter (AWS) and 18 from adults who do not stutter (AWNS). AWS, in general, exhibited more pausing or hesitancy in speech compared to AWNS, as evidenced by slower speech rate, greater total pause time, higher pause count, and longer mean duration of pause events. Numerous pause and vocal metrics acquired from ATAS were correlated with a canonical measure of stuttering frequency percent syllables stuttered, suggesting that automatically detected temporal metrics of pause and vocal events within continuous speech are highly associated with overt stuttering behavior. ATAS metrics generally predicted the status of each participant as either an AWS or AWNS grouping with accuracies considerably higher than random guessing using random forest and LSTM classifiers. This pipeline may provide an alternative and complementary method that speech-language pathologists and other health professionals can use in the assessment of fluency disorders.