Language model adaptation with MAP estimation and the perceptron algorithm

Abstract

In this paper, we contrast two language model adaptation approaches: MAP estimation and the perceptron algorithm. Used in isolation, we show that MAP estimation outperforms the latter approach, for reasons which argue for combining the two approaches. When combined, the resulting system provides a 0.7 percent absolute reduction in word error rate over MAP estimation alone. In addition, we demonstrate that, in a multi-pass recognition scenario, it is better to use the perceptron algorithm on early pass word lattices, since the improved error rate improves acoustic model adaptation.