Automatic russian speech recognition using factored language models

This paper is dedicated to investigation of factored language models for an automatic speech recognition system. The factored language models with fixed and parallel generalized backoff paths were created using 5 linguistic factors (word-form, word lemma, stem, part-of-speech, and morphological tag). The created language models were used for rescoring N-best lists of hypotheses generated by the speech recognition system. The results and analysis of experiments on very large vocabulary Russian speech recognition using the factored language models are presented in the paper.