Slovenian large vocabulary speech recognition with data-driven models of inflectional morphology

The paper describes experiments in large vocabulary speech recognition of the highly inflective Slovenian language. The main problem of an inflective language is its high OOV (out-of-vocabulary) rate. To achieve a usable OOV rate, smaller modeling units (namely stems and endings) are used instead of words. Word decompositions are based on data-driven methods. Experiments with different-sized vocabularies were performed to show the effects of data sparsity and acoustic confusability. The most remarkable improvement is obtained with a vocabulary of 20,000 units. We compare subword-based models with word-based models. All results are computed on word level. The best results are obtained with subword trigram language models. They improve recognition for 7.5%. By using larger vocabularies, the results are not improved. The problems of acoustic confusability of subword units becomes evident. Also, the statistics of some modelling units are poorly estimated due to their low frequency of occurrence.