Accuracy Analysis of Generalized Pronunciation Variant Selection in ASR Systems

Automated speech recognition systems work typically with pronunciation dictionary for generating expected phonetic content of particular words in recognized utterance. But the pronunciation can vary in many situations. Besides the cases with more possible pronunciation variants specified manually in the dictionary there are typically many other possible changes in the pronunciation depending on word context or speaking style, very typical for our case of Czech language. In this paper we have studied the accuracy of proper selection of automatically predicted pronunciation variants in Czech HMM ASR based systems. We have analyzed correctness of pronunciation variant selection in forced alignment of known utterances used as an ASR training data. Using the proper pronunciation variant, more exact transcriptions of utterances were created for further purposes, mainly for the more accurate training of acoustic HMM models. Finally, as the target and the most important application are LVCSR systems, the accuracy of LVCSR results using different levels of automated pronunciation generation were tested.