Recovering Capitalization and Punctuation Marks on Speech Transcriptions

This work addresses two metadata annotation tasks, involved in the production of rich transcripts: automatic capitalization, and punctuation marks recovery. The main focus concerns broadcast news, using both manual and automatic speech transcripts. Different capitalization models were analysed and compared, and results support the ideia that generative approaches capture the structure of written corpora better, while the discriminative approaches are robust to ASR errors and suitable for dealing with speech transcripts. The so-called language dynamics has been addressed, and results indicate that the capitalization performance is affected by the temporal distance between the training and testing data. In what concerns the punctuation task, this study covers the three most frequent marks: full stop, comma, and question mark, combining lexical, acoustic, and prosodic information. Much of the research described here is language independent, but a special focus is given to the Portuguese language. This work provides the first evaluation results of these two tasks over European Portuguese broadcast news data.