Look-ahead techniques for fast beam search

In this paper, we present two efficient look-ahead pruning techniques in beam search for large vocabulary continuous speech recognition. Both techniques, the language model look-ahead and the phoneme look-ahead, are incorporated into the word conditioned search algorithm using a bigram language model and a lexical prefix tree [5]. The paper present the following novel contributions: We describe a method for language model (LM) look-ahead pruning which is similar to [1, 9]. We show special techniques to reduce the memory and computational requirements. These techniques are based on a compressed LM look-ahead tree. To compute the LM look-ahead tree probabilites in an efficient way, we present a backward dynamic programming scheme. We present a phoneme look-ahead pruning technique to increase the efficiency of the acoustic pruning. In particular, we refine the acoustic pruning strategy by a 1and 2-phoneme look-ahead, respectively. We report results for both look-ahead pruning methods on the the 20,000-word North American Business (NAB’94) task. As a result, the combination of bigram look-ahead and 1phoneme look-ahead reduces the search space by a factor of 10 without loss in recognition accuracy in comparison with the baseline search using a unigram language model look-ahead as described in [2]. The computational costs can be reduced by a factor of 5 on a SGI workstation (Indy R4400).