Evaluation of advanced language modeling techniques for the Slovak LVCSR

Abstract

In this paper we compare several advanced language modeling techniques for the Slovak continuous speech recognition. Five different language modeling techniques were analyzed, considering their model size and perplexity, speech recognition performance and complexity of their usage in real conditions of speech recognition in Slovak. The preliminary experimental results show that the convenient n-gram models smoothed by the Witten-Bell back-off algorithm produce the best performance according to the model perplexity and recognition accuracy. Other modeling techniques including Maximum Entropy, Power Law Discounting, Hierarchical Pitman-Yor process, or Variable-order Kneser-Ney smoothed models achieved better results only in the model perplexity. However, the increased computational requirements and worse recognition performance limit their usage in the real speech recognition tasks in Slovak.