Automatic sentence boundary detection in conversational speech: A cross-lingual evaluation on English and Czech

Automatic sentence segmentation of speech is important for enriching speech recognition output and aiding downstream language processing. This paper focuses on automatic sentence segmentation of speech in two different languages - English and Czech. For this task, we compare and combine three statistical models - HMM, maximum entropy, and a boosting-based model BoosTexter. All these approaches rely on both textual and prosodic information. We evaluate these methods on a corpus of multiparty meetings in English, and on a corpus of broadcast conversations in Czech, using both manual and speech recognition transcripts. The experiments show that superior results are achieved when all the three models are combined via posterior probability interpolation. We observe differences in terms of model performance between English and Czech, as well as the feature usage difference in prosodic models between the two languages. Overall, the analysis is important for porting sentence segmentation approaches from one language to another.