Article details

Research area
Speech recognition

Location
ICASSP 2010

Date
2010

Author(s)
Anoop Deoras, Frederick Jelinek, Yi Su

Language Model Adaptation Using Random Forests

Synopsis:

In this paper we investigate random forest based language model adaptation. Large amounts of out-of-domain data are used to grow the decision trees while very small amounts of in-domain data are used to prune them back, so that the structure of the trees are suitable for the desired domain while the probabilities in the tree nodes are reliably estimated. Extensive experiments are carried out and results are reported on a particular task of adapting Broadcast News language model to the MIT computer science lecture domain. We show 0.80% and 0.60% absolute WER improvement over language model interpolation and count merging techniques, respectively.

Read/download now