A digit and spelling speech recognition system for the Croatian language

Out of need for an efficient speech recognition system for the Croatian language, a model was built using the open source PocketSphinx toolkit. Firstly, a speech corpus of alphanumeric utterances was built for training. Secondly, the speech recognition model was built using PocketSphinx toolkit. Using a separate testing set, parameters for a good minimization of the word error rate were found. Useful observations were made that can be used in further improving the Croatian speech recognition model.