Method and apparatus for optimizing a speech recognition result
View Patent ↗According to one embodiment, an apparatus for optimizing a speech recognition result comprising: a receiving unit configured to receive a speech recognition result; a calculating unit configured to calculate a pronunciation similarity between a segment of the speech recognition result and a key word in a key word list; and a replacing unit configured to replace the segment with the key word in a case that the pronunciation similarity is higher than a first threshold.
1. An apparatus for optimizing a speech recognition result, comprising:
a computer that, by using computer executable instructions,
receives a speech recognition result from a speech recognition engine,
calculates a phoneme acoustic distance between a phoneme sequence of a segment of the speech recognition result and a phoneme sequence of a key word in a keyword list,
calculates a tone acoustic distance between a tone sequence of the segment and a tone sequence of the key word,
calculates a weighted average of the phoneme acoustic distance and the tone acoustic distance,
calculates an average acoustic distance which is obtained by dividing the weighted average by the number of characters, syllables or phonemes of the key word,
calculates a language model score of the segment, based on a language model score of each word in the segment,
replaces the segment with the key word in a case that the average acoustic distance is lower than a first threshold and the language model score is lower than a second threshold, and
outputs the speech recognition result of which the segment is replaced with the keyword to the speech recognition engine.
2. The apparatus according to claim 1 , wherein
the computer calculates the average acoustic distance between a segment of the speech recognition result, in which a language model score of the segment is lower than the second threshold, and a key word in the key word list.
3. The apparatus according to claim 1 , wherein
the computer calculates the phoneme acoustic distance between the phoneme sequence of the segment and the phoneme sequence of the key word by using a phoneme confusion matrix as a weight.
4. The apparatus according to claim 1 , wherein
the computer calculates the tone acoustic distance between the tone sequence of the segment and the tone sequence of the key word by using a tone confusion matrix as a weight.
5. A method for optimizing a speech recognition result, comprising steps of:
receiving a speech recognition result from a speech recognition engine;
calculating a phoneme acoustic distance between a phoneme sequence of a segment of the speech recognition result and a phoneme sequence of a key word in a keyword list;
calculating a tone acoustic distance between a tone sequence of the segment and a tone sequence of the key word;
calculating a weighted average of the phoneme acoustic distance and the tone acoustic distance;
calculating an average acoustic distance which is obtained by dividing the weighted average by the number of characters, syllables or phonemes of the key word;
calculating a language model score of the segment, based on a language model score of each word of the segment;
replacing the segment with the key word in a case that the average acoustic distance is lower than a first threshold and the language model score is lower than a second threshold; and
outputting the speech recognition result of which the segment is replaced with the keyword to the speech recognition engine.