An Exact Word Lattice Generation Method in the Weighted Finite-State Transducer Framework

As a multiple-candidate output format of speech recognition system, word lattice is essential for applications such as keyword spotting, confidence measure, multi-pass decoding and so on. This paper analyzes the problems of generating word lattice using Weighted Finite-State Transducer (WFST) decoders, such as word boundary decision, word position pushing and redundancy existed in the word lattice. We present an efficient word lattice generation method which is able to retain all the accurate word alignment information. Furthermore, a new word-level determinization algorithm that keeps the alignment information is described to completely remove the redundant paths in the word lattice. Experiments show that the proposed determinization algorithm is effective for improving the quality of the word lattice-based confidence measure and accuracy of keyword spotting.