The block-synchronous search algorithm

Multiple-pass searches have many advantages, including the combination of the speed of small models with the accuracy of large models. In their simplest implementations, the later passes cannot begin until the first pass completes at the end of the utterance. This work describes a block-synchronous approach to multiple-pass searches, where the utterance is partitioned into fixed-length blocks. All search passes run on the current block before the next block is processed. The algorithm uses bounded memory and bounded delay after the utterance completes, since the blocks are of fixed length. There is no requirement to estimate phrase or word boundaries to limit the processing.