Transducer composition for "on-the-fly" lexicon and language model integration

Abstract

We present the use of a specialized composition algorithm that allows the generation of a determinized search network for ASR in a single step. The algorithm is exact in the sense that the result is determinized when the lexicon and the language model are represented as determinized transducers. The composition and determinization are performed simultaneously, which is of great importance for "on-the-fly" operation. The algorithm pushes the language model weights towards the initial state of the network. Our results show that it is advantageous to use the maximum amount of information as early as possible in the decoding procedure.