Language Model Combination and Adaptation Using Weighted Finite State Transducers

TitleLanguage Model Combination and Adaptation Using Weighted Finite State Transducers
Publication TypeConference Paper
Year of Publication2010
AuthorsLiu, X., Gales M.. J. F., Hieronymus J., & Woodland P.
Other Numbers3311

In speech recognition systems language model (LMs) are often constructedby training and combining multiple n-gram models. Theycan be either used to represent different genres or tasks found indiverse text sources, or capture stochastic properties of different linguisticsymbol sequences, for example, syllables and words. UnsupervisedLM adaptation may also be used to further improve robustnessto varying styles or tasks. When using these techniques,extensive software changes are often required. In this paper an alternativeand more general approach based on weighted finite statetransducers (WFSTs) is investigated for LM combination and adaptation.As it is entirely based on well-defined WFST operations,minimum change to decoding tools is needed. A wide range of LMcombination configurations can be flexibly supported. An efficienton-the-fly WFST decoding algorithm is also proposed. Significant

Bibliographic Notes

Proceedings of the IEEE International Conference on Acoustics, Speech, and Signal Processing (ICASSP), Dallas, Texas

Abbreviated Authors

X. Liu, M. J. F. Gales, J. L. Hieronymus, and P. C. Woodland

ICSI Research Group


ICSI Publication Type

Article in conference proceedings