A Modified MASPER Training Procedure for ASR Systems and its Performance on Slovak MOBILDAT Database

This article discusses modifications to the MASPER training procedure for HMM models, and its application in a construction of ASR system based on the Slovak MOBILDAT database. First, a general description of MASPER and REFREC based training methods is given from a historical prospective, followed by the proposed modification regarding the handling of truncated speech, mispronounced words, and unintelligible speech by inserting general background model. Various structures of background models and their gradual enhancement steps were tested using classical test procedures defined by MASPER. For the best ones, they exhibited improved results especially for more complex models (context dependent) compared to the reference methods. In those cases more than 5% relative improvements were observed.