Hybrid deep learning based assessment and resynthesis of stuttered speech

Abstract

Stuttering is a speech fluency disorder with precedents of sound prolongations silent blocks repetition of syllables words and phrases that disrupt the fluent rendition of speech Disfluency is common even among healthy individuals which could sometimes be mistaken for stuttering So it is important to categorize disfluencies into Stutter Like Disfluencies SLD and Typical Disfluencies TD The SLDs are crucial stutter markers which include word medial repetitions prolongations and silent blocks while the TDs include natural disfluencies like the filled pauses whole word repetitions phrase repetitions and revisions Presently the assessment of stuttering severity levels as mild or severe is being carried out by the Speech Language Pathologists SLPs by a manual count of SLDs This procedure is cumbersome and may be prone to error Alternatively machine learning models may be used for disfluency classification However the machine learning models yield reduced accuracy of classification than the deep learning models which require huge training data for improving accuracy So this research work has proposed hybrid deep learning models that yield high accuracy of classification even when trained over a sparse speech dataset at an optimal computational time complexity Further a hybrid deep learning generative model has been proposed to resynthesize speech for fluency enhancement newline