Linguistically motivated tied-state triphones for polish speech recognition

Abstract

The paper presents one of the possible approaches to build a triphone model for automatic speech recognition of Polish. Even though classifiers are well developed and described, such task is not a trivial one because of lack of enough training data and importance of calculation time spent for the training of the model. To overcome this problem, some states are typically tied using data-driven criteria. We investigate a linguistically motivated approach, where phonetically related contexts are tied. We compared recognition results of a system using this approach and of a system with no context tying on around 15 000 utterances. The results indicate a small improvement in the performance of the system.