Locating and correcting erroneously recognized portions of utterances by rescoring based on two n-best lists

A method of repairing speech recognized by a recognition engine of the type which generates an n-best list of hypotheses and scores for each hypothesis in response to the speech to be recognized, said method comprising the steps of:

• receiving from a recognition engine a first n-best list of hypotheses, and scores for each hypothesis, generated in response to a primary utterance to be recognized;
• locating an error within the hypothesis having the highest score;
• parsing each hypothesis in said first n-best list to produced a scored paraphrase for each;
• receiving from said recognition engine a second n-best list of hypotheses, and scores for each hypothesis, generated in response to a secondary utterance which is a paraphrase of said primary utterance;
• parsing each hypothesis in said second n-best list to produced a scored paraphrase for each;
• combining the scores for the paraphrases from the first n-best list with the scores for the paraphrases from the second n-best list;
• selecting the paraphrase having the highest combined score;
• and
• replacing the located error with the selected paraphrase.

A speech recognition and repair apparatus for use in a speech recognition system of the type including a recognition engine which generates an n-best list of hypotheses and scores for each hypothesis in response to the speech to be recognized, said apparatus comprising:

• means for receiving from a recognition engine a first n-best list of hypotheses, and scores for each hypothesis, generated in response to a primary utterance to be recognized;
• means for locating an error within the hypothesis having the best score;
• means for generating control signals from said first n-best list, said control signals being input to the recognition engine to constrain the selection of hypotheses;
• means for receiving from the recognition engine a second n-best list of hypotheses, and scores for each hypothesis, generated in response to an event independent of the primary utterance and in accordance with said control signals;
• means for combining the scores for the hypotheses in the first n-best list with the scores for the hypotheses in the second n-best list;
• and
• means for replacing the located error with the hypothesis having the best combined score.