Automatic Punctuation Annotation in Czech Broadcast News Speech

This paper reports our initial experiments with automatic punctuation annotation from speech. We have focused on Czech broadcast news speech. The task can be defined as a classification of each inter-word boundary into one of target classes. We considered comma, sentence boundary and “no punctuation” as the target classes. We employed two statistical models – prosodic model and language model. The prosodic model expresses relationships between prosodic quantities (such as pitch, speaking rate or loudness) and punctuation marks. We tested two implementations of this model – decision tree and multi-layer perceptron. Hidden-event N -gram models were employed for language modeling. Instead of using an ordinary word-based model, we replaced infrequent word forms by their morphological tags and trained a mixed model. Scores from both models can be combined. The model combining language model with the decision tree yielded superior results. Testing on true words we achieved classification accuracy 95.2% and F -measure 78.2%.