This paper discusses a new framework for the evaluation of the detection of disfluency in speech and natural language processing. We argue for the supervised identification of two tracks of communication, primary and collateral tracks, inspired by the theory of performance from H. Clark, Using Language.
This setting enables a direct quantitative comparison between the detection techniques from the Natural Language Processing(NLP) and Speech Technologies(ST) communities. It finally provides comparison metrics of the models that can be used by speech pathologists, HCI engineers, (psycho)-linguists, whom can have all different needs based on the model predictions.
Finally, we tackle this problem of disfluency identification in adults stuttered speech in the context of semi-directed interviews. We compare word-based prediction and frame-base predictions, using semantic, span and acoustic-prosodic information.