Acoustic Modelling for Croatian Speech Recognition and Synthesis

Abstract

This paper presents the Croatian context-dependent acoustic modelling used in speech recognition and in speech synthesis. The proposed acoustic model is based on context-dependent triphone hidden Markov models and Croatian phonetic rules. For speech recognition and speech synthesis system modelling and testing the Croatian speech corpus VEPRAD was used. The experiments have shown that Croatian speech corpus, Croatian phonetic rules and hidden Markov models as the modelling formalism can be used to develop speech recognition and speech synthesis systems in parallel for a highly flective and free order language like Croatian. We propose an evaluation procedure for speech synthesis, which combines an objective and a subjective evaluation approach and we present the achieved evaluation results. The proposed procedures for Croatian acoustic modelling were developed as parts of speech interfaces in a spoken dialog system for a limited weather forecast domain.