Hybrid speech recognition systems using both bigram and grammar models yield improved performance compared with the use of either model alone, but performance is sub-optimal because the grammar is abandoned for sentences that fail to parse overall. By merging bigrams (in general n~grams) and grammars into a single framework we aim to combine the advantages of both, in particular structural capacity and trainability, in a robust recognition system. A substring parser allows whatever grammar structure is present to participate in scoring the candidate sentences. Extended bigrams using an information criterion capture remote dependencies and reduce perplexity. The first version of a consolidated model using these methods is described.
Keywords: Language modelling, robust parsing.