Multi-pass sentence-end detection of lecture speech

Making speech recognition output readable is an important task. The first step here is automatic sentence end detection (SED). We introduce novel F0 derivative-based features and sentence end distance features for SED that yield significant improvements in slot error rate (SER) in a multi-pass framework. Three different SED approaches are compared on a spoken lecture task: hidden event language models, boosting, and conditional random fields (CRFs). Experiments on reference transcripts show that CRF-based models give best results. Inclusion of pause duration features yields an improvement of 11.1% in SER. The addition of the F0-derivative features gives a further reduction of 3.0% absolute, and an additional 0.5% is gained by use of backward distance features. In the absence of audio, the use of backward features alone yields 2.2% absolute reduction in SER.