Identification of Primary and Collateral Tracks in Stuttered Speech

Proceedings of the 12th Conference on Language Resources and Evaluation (LREC 2020), pages 1681–1688 Marseille, 11–16 May 2020 c⃝ European Language Resources Association (ELRA), licensed under CC-BY-NC 1681 Identification of primary and collateral tracks in stuttered speech Rachid Riad1,2∗, Anne-Catherine Bachoud-L´evi2, Frank Rudzicz3, Emmanuel Dupoux1 1 CoML/ENS/CNRS/EHESS/INRIA/PSL Research University, Paris, France 2 NPI/ENS/INSERM/UPEC/PSL Research University, Cr´eteil, France 3 University of Toronto/Vector Institute/St Michaels Hospital/Surgical Safety Technologies Inc, Toronto, Canada rachid.riad@ens.fr, bachoud@gmail.com, frank@spoclab.com, emmanuel.dupoux@gmail.com Abstract Disfluent speech has been previously addressed from two main perspectives: the clinical perspective focusing on diagnostic, and the Natural Language Processing (NLP) perspective aiming at modeling these events and detect them for downstream tasks. In addition, previous works often used different metrics depending on whether the input features are text or speech, making it difficult to compare the different contributions. Here, we introduce a new evaluation framework for disfluency detection inspired by the clinical and NLP perspective together with the theory of performance from (Clark, 1996) which distinguishes between primary and collateral tracks. We introduce a novel forced-aligned disfluency dataset from a corpus of semi-directed interviews, and present baseline results directly comparing the performance of text-based features (word and span information) and speech-based (acoustic-prosodic information). Finally, we introduce new audio features inspired by the word-based span features. We show experimentally that using these features outperformed the baselines for speech-based predictions on the present dataset. Keywords: evaluation metrics, disfluency, stuttering, speech processing, audio features 1. Introduction Around 6% percent of spoken words in non-pathological speech are categorised as disfluent (Tree, 1995) and this increases with the cognitive load of the speaker (Bortfeld et al., 2001; Lindstr¨om et al., 2008). Speaking in real time is a demanding activity, subject to cognitive constraints and pragmatic settings. Under time pressure, a word may not be retrieved, part of a sentence may be revised, unfilled and filled pauses may be inserted, words or part of words may be repeated. Some of these deviations can be viewed as the symptoms of sentence planning problems (McRoberts and Clark, 1996) or as the results of some strategies (Clark and Wasow, 1998) unfolded under speaker’s control to signal something. Stuttering is a severe case of speech pathology, that inter- rupts at a higher rate the flow of speech than in typical Speech Production. Indeed, in addition to the ’classic’ dis- fluencies, stutterers can produce other forms of disfluency that are both quantitatively and qualitatively distinguishable from typical forms (Lickley, 2017). Speech pathologists need to quantify all the disfluency events for clinical screening but also to assess potential treatments (Yaruss, 1997). A large number of factors and speaking settings influence stuttering behaviours and oc- currences of disfluencies: interlocutor’s characteristics (ex: age, relationship with the speaker), conversational settings (at home, at the hospital, at work), speaking tasks (ex: read- ing, dialogues, descriptions of scenes). The clinical assess- ment still rely heavily on subjective and one on one eval- uation (Yaruss and Quesal, 2006). An automatic, reliable procedure would provide Speech Pathologists an objective comparison between clinical facilities …

Content truncated to 500 words