The grammar-based NLP methods that have taken part in the evaluation are of two types. The first type, b(B,N), consists of two phases. In the first phase the word graph is made smaller by selecting the N-best paths from the word graph, using the acoustic scores and a language model consisting of bigrams (B=bi) or trigrams (B=tri) (with bigrams for backing-off). Only those transitions of the word graph remain which are part of at least one of those N-best paths. In the second phase the parser is applied, using acoustic scores and a language model of trigrams (again with bigrams for backing-off).
The second type of method is f(B,N). In this case, if the
word graph contains less than N transitions, then the full word graph
is input to the parser, and acoustic scores and a language model of
trigrams (bigrams for backing-off) is applied to select the best
analysis. If the word graph contains more than N transitions, then
method b(B,1) is applied.