Efficient adaptations of the SphinxTrain procedure for building a robust ASR system in Slovak

In the following article we discuss the practical and theoretical aspects of the building of Slovak ASR system using the SPHINX system and its SphinxTrain adaptation procedure for CI and CD HMM models. Concerning issues are ranging from the optimal setting of the number of states per model, through the adjustment of the number of tied states for context dependent HMMspsila, number of Gaussian mixtures, and training scenarios regarding the spelled recordings and background models. All experiments and results were obtained for the MOBILDAT-SK speech database that contains 32500 recordings from 1100 speakers. Obtained CI and CD HMM models achieved WER around 5% for application words which qualifies them for a use in practical applications. Furthermore the suggested and realized modifications to the classical SphinxTrain procedure for a given database and the Slovak language brought improved overall results as well.