Memory-efficient LVCSR search using a one-pass stack decoder

This paper describes the details of a fast, memory-efficient one-pass stack decoder for efficient evaluation of the search space for large vocabulary continuous speech recognition. A modem efficient search engine is not based on a single idea, but is a rather complex collection of separate algorithms and practical implementation details, which only in combination make the search efficient in time and memory requirements. Being the core of a speech recognition system, the software design phase for a new decoder is often crucial for its later performance and flexibility. This paper tries to emphasize this point-after defining the requirements for a modem decoder, it describes the details of an implementation that is based on a stack decoder framework. It is shown how it is possible to handle arbitrary order N-grams, how to generate N-best lists or lattices next to the first-best hypothesis at little computational overhead, how to handle efficiently cross-word acoustic models of any context order, how to efficiently constrain the search with word graphs or word-pair grammars, and how to use a fast-match with delay to speed up the search, all in a single left-to-right search pass. The details of a disk-based representation of an N-gram language model are given, which make it possible to use language models (LMs) of arbitrary (file) size in only a few hundred kB of memory. On-demand N-gram smearing, an efficient improvement over the regular unigram smearing used as an approximation to the LM scores in a tree lexicon, is introduced. It is also shown how lattice rescoring, the generation of forced alignments and detailed phone-/state-alignments can efficiently be integrated into a single stack decoder. The decoder named "Nozomi"(1) was tested on a Japanese newspaper dictation task using a 5000 word vocabulary. Using computationally cheap models it is possible to achieve real-time performance with 89% word recognition accuracy at about 1% search error using only 4 MB of total memory on a 300 MHz Pentium II. With computationally more expensive acoustic models, which also cover for the Japanese language essential cross-word effects, more than 95% recognition accuracy(2) is reached. (C) 2000 Academic Press.