Abstract: An automatic speech recognition system includes a recognition engine, a compound lexicon, and a compounder. The recognition engine generates a recognition result having a sequence of recognized words representative of an input utterance. The words in the recognition result include compound word components which may be combined to form compound words. The engine uses a recognition vocabulary of words and a language model which, for a given position in the sequence of recognized words and for selected words in the recognition vocabulary, associates a likelihood of such word occurring at such position. The compound lexicon contains a plurality of compound word components and component connecting links structured so that links are present between components that are more likely to occur together than to occur as separate words.

Abstract: Automatic speech recognition word sequence hypotheses are generated using an interleaved forward-backward search. A forward search pass uses relatively simple models for a given block period of time. A backward search pass then goes back over the previous block period using more complex models and the recognition hypotheses generated by the forward search pass. The backward search pass employs a word dependent n-best search having a flat model state organization.