Effective lexical tree search for large vocabulary continuous speech recognition

In this paper, we present an e cient calculation of the factored LM probabilities for speeding up the large vocabulary continuous speech recognition. We introduced a novel technique based on the independent calculation of the factored LM probability. The basic idea of the proposed method is that each factored LM probability is calculated on-demand for a new combination of a previous word hypothesis and a LM look-ahead tree node, instead of calculating all the factored LM probabilities over the tree at a time. The speaker-independent continuous speech recognition experiment was performed for 20 speakers on a 60k word newspaper dictation task. As a result, the proposed method achieved 25% improvement in speed.