A disfluency study for cleaning spontaneous speech automatic transcripts and improving speech language models

The aim of this study is to elaborate a disfluent speech model by comparing different types of audio transcripts. The study makes use of 10 hours of French radio interview archives, involving journalists and personalities from political or civil society. A first type of transcripts is press-oriented where most disfluencies are discarded. For 10% of the corpus, we produced exact audio transcripts: all audible phenomena and overlapping speech segments are transcribed manually. In these transcripts about 14% of the words correspond to disfluencies and discourse markers. The audio corpus has then been transcribed using the LIMSI speech recognizer . With 8% of the corpus the disfluency words explain 12% of the overall error rate. This shows that disfluencies have no major effect on neighboring speech segments. Restarts are the most error prone, with a 36.9% within class error rate.