We examine the practical constraints imposed on the task of sentence boundary detection in speech recognizer output, by the requirements of a system that supports large-scale commercial off-line transcription of dictations. We develop and evaluate a method that observes these constraints, reformulating the best technique previously reported in order to allow the use a smoothing technique directly tailored to boundary prediction. We then show how this method can be generalized and improved upon, demonstrating significantly better performance in three different domains.