An overview of decoding techniques for large vocabulary continuous speech recognition

A number of decoding strategies for large vocabulary continuous speech recognition (LVCSR) are examined from the viewpoint of their search space representation. Different design solutions are compared with respect to the integration of linguistic and acoustic constraints, as implied by m-gram language models (LM) and cross-word (CW) phonetic contexts. This study is structured along two main axes: the network expansion and the search algorithm itself. The network can be expanded statically or dynamically while the search can proceed either time-synchronously or asynchronously which leads to distinct architectures. Three broad classes of decoding methods are briefly reviewed: the use of weighted finite state transducers (WFST) for static network expansion, the time-synchronous dynamic-expansion search and the asynchronous stack decoding. Heuristic methods for further reducing the search space are also considered. The main approaches are compared and some prospective views are formulated regarding possible future avenues. (C) 2002 Academic Press.