Improved variable preselection list length estimation using NNs in a large vocabulary telephone speech recognition system

In very large vocabulary hypothesis-verification systems, the fine acoustic matcher is usually the most time consuming, so that the main concern is reducing the preselection list length as much as possible. Traditionally, these systems use a too high fixed preselection list length, increasing computational demands over the really needed. The idea we are proposing is estimating a different preselection list length for every utterance, so that we can lower the average computational effort needed for the recognition process. As we will show, it’s even possible that the resulting system outperforms the fixed length one in error rate, even when reducing computational cost. This paper presents a detailed study on a NN based approach to variable preselection list length estimation. The main achievement has been a relative decrease in error rate of up to 40%, while getting a relative decrease in average preselection list length of up to 31%.