System and method for speech recognition

An automated system with speech input control comprising:

• an audio input device configured to generate audio input data corresponding to speech input from a user;
• and
• a controller operatively connected to the audio input device and a memory, the controller being configured to:
• receive the audio input data from the audio input device;
• generate a first plurality of candidate speech recognition results corresponding to the audio input data using a first general-purpose speech recognition engine;
• generate a second plurality of candidate speech recognition results corresponding to the audio input data using a first domain-specific speech recognition engine;
• generate a third plurality of candidate speech recognition results, each candidate speech recognition result in the third plurality of candidate speech recognition results including a plurality of words included in one of the first plurality of candidate speech recognition results and at least one word included in another one of the second plurality of candidate speech recognition results;
• rank at least the third plurality of candidate speech recognition results using a pairwise ranker to identify a highest ranked candidate speech recognition result, the controller being further configured to:
• identify the highest ranked candidate speech recognition result based on a ranking score using a pairwise ranking process between a plurality of pairs of candidate speech recognition results selected from the third plurality of candidate speech recognition results using the pairwise ranker, the ranking of each pair of candidate speech recognition results further comprises the controller being configured to:
• generate a first feature vector including a feature corresponding to a bag-of-words with decay value corresponding to at least one word in the first candidate speech recognition result;
• generate a second feature vector, the first feature vector including a feature corresponding to a bag-of-words with decay value corresponding to at least one word in the second candidate speech recognition result;
• generate a third feature vector based on a difference between the first feature vector and the second feature vector;
• and
• estimate a first word error rate in the first candidate speech recognition result and a second word error rate in the second candidate speech recognition result based on the third feature vector using the pairwise ranker;
• increment a ranking score associated with the first candidate speech recognition result in response to the first word error rate being less than the second word error rate;
• and
• increment another ranking score associated with the second candidate speech recognition result in response to the first word error rate being greater than the second word error rate;
• and
• operate the automated system using the highest ranked candidate speech recognition result as an input from the user.