A corpus of spontaneous telephone transactions between call centre operators of a pizza company and its customers is examined for disfluencies (fillers and speech repairs) with the aim of improving automatic speech recognition. From this, a subset of the customer orders is selected as a test set. An architecture is presented which allows filled pauses and repairs to be detected and corrected. A language repair module removes fillers and reparanda and transforms utterances containing them into fluent utterances. An experiment on filled pauses using this module and architecture is then described. A speech recognition grammar for recognising fluent speech is used to provide a baseline. This grammar is then enriched with filled pauses, based on their placement in relation to syntactic boundaries. Evaluation is done at the level of understanding, using a metric on feature structures. Initial results indicate that incorporating filled pauses at syntactic boundaries improves the recognition results for spontaneous continuous speech containing disfluencies.