Memory efficient decoding graph compilation with wide cross-word acoustic context

Abstract

We present an efficient technique for the compilation of static decoding graphs which can utilize full word of left cross-word context. We put an emphasis on memory efficiency, in particular to be able to deploy this technique on platforms with limited resources. The incremental application of the composition process efficiently produces a weighted finite state acceptor which is globally deterministic and minimized with the maximum memory need during the composition essentially the same as that needed for the final graph.