Analysis of a simple bipos language model-attempt at a strategy toimprove language models for speech recognition

摘要

A speech recognizer has to choose, at each point in the utterance,the words among all the words in the vocabulary, that are the mostlikely. To that end, it uses an acoustic model and a language model andthe author focuses on the language model. The bipos model is presentedand analysed. A method is introduced called probability decomposition tomeasure which part of the model is performing particularly well orpoorly. Based on this analysis, the author modifies the modeling ofunknown words and this leads to a reduction in the entropy of at least14% (up to 21%). Other conclusions obtained from the analysis are alsogiven. An attempt at a strategy to improve language models in general isgiven. To that end, the author defines a class of models called statelanguage models. This class contains most currently employed models.However, these currently used models cover only a small area in thespace of all possible state language models. A more systematic study ofthis space is proposed in order to improve current language models. Astatistical method, called classification and regression trees ispresented as a tool for this purpose