A Trellis Based Fast Lattice Generating Algorithm

Abstract

Lattice is widely used as a kind of the search results in Large Vocabulary Continuous Speech Recognition (LVCSR). A new lattice-generation algorithm is presented in this paper. The algorithm is based on a classical forward-backward decoding method, which is proved to be highly efficient. Moreover, some improvements have been done to satisfy the requirements in the lattice decoding. Two Chinese mandarin large-scale speech recognition tasks are used to evaluate the proposed algorithm and the experimental results show that our algorithm can both improve decoding speed and save decoding space significantly without sacrificing the recognition accuracy, compared with the widely used Lattice decoding method as.