| 摘要 |
The present invention relates to techniques for contextual filtering for improving an output of a speech recognizer. The techniques comprise receiving a representation of a speech utterance into a shallow parser component in the form of a first word lattice including a set of potential language matches for the utterance, where the language matches include at least one word; the parser component operative for receiving and analyzing the first word lattice to produce a second word lattice, and including a filter that assigns a probability match for a potential word match in relation to another potential word match in the utterance based on a particular filter-specific criteria; and outputting the probability matches from the shallow parser component as a portion of a second word lattice for further processing to determine most likely combinations of words present in the speech utterance.
|
