Novel LVCSR Decoder Based on Perfect Hash Automata and Tuple Structures – SPREAD –

Abstract

Abstract-The paper presents the novel design of a one-pass large vocabulary continuous-speech recognition decoder engine, named SPREAD. The decoder is based on a time-synchronous beam-search approach, including statically expanded cross-word triphone contexts. An approach using efficient tuple structures is proposed for the construction of the complete search-network. The foremost benefits are the important space savings and higher processing speed, and the compact and reduced size of the tuple structure, especially when exploiting the structure of the key. In this way, the time needed to load the ASR search-network into the memory is also significantly reduced. Further, the paper proposes and presents the complete methodology for compiling general ASR knowledge sources into a tuple structures. Additionally, the beam search is enhanced with the novel implementation of a bigram language model Look-Ahead technique, by using tuple structures and a caching scheme. The SPREAD LVCSR decoder is based on a token-passing algorithm, capable of restricting its search-space by several types of token pruning. By using the presented language model Look-Ahead technique, it is possible to increase the number of tokens that can be pruned without decoding precision loss.