Segment-based classes for language modeling within the field of CSR

Abstract

In this work, we propose and formulate two different approaches for the language model integrated in a Continuous Speech Recognition System. Both of them make use of class-based language models where classes are made up of segments or sequences of words. On the other hand, an interpolated model of a class-based language model and a word-based language model is explored as well. The experiments carried out over a spontaneous dialogue corpus in Spanish, demonstrate that introducing segments of words in a class-based language model a better performance of a Continuous Speech Recognition system can be achieved. © Springer-Verlag Berlin Heidelberg 2007.