Looking at alternatives within the framework of n-gram based language modeling for spontaneous speech recognition

摘要

This paper presents different methods using a weighted mixture of word and word-class language models in order to perform language model adaptation. A general language model is built from the whole training corpus, then several numbers of clusters are created according to a word co-occurrence measure and finally, word models as well as word-class models are built from each cluster. The general language model is then combined with one or several other models chosen according to a minimum perplexity criterion. Results show an absolute reduction of the word error rate of 1.40% and 0.49% on average for two different test sets of the "Corpus of Spontaneous Japanese."