Learning pronunciation dictionary from speech data

An algorithm and first results from investigations in automatically learning pronunciation variations from speech data are presented. Pronunciation dictionaries establish an important feature in state-of-the-art speech recognition systems. In most systems only simple dictionaries containing the canonical pronunciation forms are implemented. However, for a good recognition performance more sophisticated dictionaries including pronunciation variations are essential. The generation of such dictionaries by hand is an extremely time consuming task, and the introduction of errors and inconsistencies is probable. The authors show an approach for automatically generating suitable pronunciation dictionaries from the speech database itself, as they are desirable not only for speech recognition tasks but also for speech technology and phonologic research in general. The only knowledge sources besides the database are the (unlabeled) signals and their transliterations on word level. First experiments yielding promising results have been performed with the software system DataLab, which integrates the recognition system of the TU Dresden.