15th International Congress of Phonetic Sciences (ICPhS-15)
In our HMM's based ASR system we have used the models of Polish triphones. To deal with triphones for which there are no examples in the training data we introduced decision tree-based clustering. In the paper we tried to analyze the clusters with the identical triphones and the questions set forming the decision trees obtained in our experiments with the ASR system. The results of our investigations of the equivalent contexts leading to the acoustically similar triphones (they share common clusters for all states) should be useful from the point of view of speech recognition, text to speech systems and applied phonetics.
Bibliographic reference. Grocholewski, Stefan (2003): "Some investigations of the decision trees from the ASR, TTS and applied phonetics point of view", In ICPhS-15, 1341-1344.