Profile based compression of n-gram language models

A profile based technique for compression of n-gram language models is presented. The technique is intended to be used in combination with existing techniques for size reduction of n-gram language models such as pruning, quantisation and word class modelling. The technique is here evaluated on a large vocabulary embedded dictation task. When used in combination with quantisation, the technique can reduce the memory needed for storing probabilities by a factor of 10 or more with only a small degradation in word accuracy. The structure of the language model is well suited for "best-first" type decoding styles, and is here used for guiding an isolated word recogniser by predicting likely continuations at word boundaries.