A study of large vocabulary speech recognition decoding using finite-state graphs

The use of weighted finite-state transducers (WFSTs) has become an attractive technique for building large vocabulary continuous speech recognition decoders. Conventionally, the compiled search network is represented as a standard WFST, which is then directly fed into a Viterbi decoder. In this work, we use the standard WFST representations and operations during compiling the search network. The compiled WFST is then equivalently converted to a new graphical representation, which we call finite-state graph (FSG). The resulting FSG is more tailored to Viterbi decoding for speech recognition and more compact in memory. This paper presents our effort to build a state-of-the-art WFST-based speech recognition system, which we call GrpDecoder. Benchmarking of GrpDecoder is carried out separately on two languages - English and Mandarin. The test results show that GrpDecoder which uses the new FSG representation in searching is superior to HTK´s HDecode and IDIAP´s Juicer for both languages, achieving lower error rates for a given recognition speed.