Handling disfluencies in spontaneous language models

In automatic speech recognition, a statistical language model (LM) predicts the probability of the next word on the basis of previously recognized words. For recognition of dictated speech this works reasonably well as (1) large amounts of training data are available which allow for a reliable estimation of the LM (2) sentences are typically well-formed. For spontaneous speech however neither condition is fulfilled. First, written transcripts of spontaneous speech are scarce and definitely insufficient for training good statistical LMs. Second, spontaneous speech is riddled with disfluencies which render the prediction context less uniform. Both factors contribute to the poor performance of an automatic speech recognizer on spontaneous input. In this talk, we will present the first results in our research on language modeling for spontaneous speech. More precisely, the choice of LM prediction context for three typical disfluencies was investigated: (1) repetitions e.g. "That is what what I think" (2) hesitations e.g. "That is what uh I think" (3) sentence restarts e.g. "That is what... Yeah I think so" We will show how adjusting the LM prediction context influences the recognition performance and we will interpret these results.Abstractbook 13th meeting of computational linguistics in The Netherlands - CLIN2002, pp. 26, November 29, 2002, Groningen, The Netherlandsstatus: publishe