Back to results
US 2011/0257976 A1Published application
Robust Speech Recognition
AssigneeMicrosoft Corporation
InventorQiang Huo
Life of this patent
Estimating the term…
Abstract
Speech recognition includes structured modeling, irrelevant variability normalization and unsupervised online adaptation of speech recognition parameters.
Claims20
A method comprising:
- receiving an unknown utterance;
- generating a first sequence of feature vectors from the unknown utterance;
- removing, in a first recognition pass and using a parameter, at least a portion of irrelevant information from the first sequence of feature vectors resulting in a second sequence of feature vectors;
- recognizing speech in the first recognition pass based at least in part on the second sequence of feature vectors to obtain a first result;
- updating the parameter used in the removing based at least in part on the first result;
- removing, in a second recognition pass, using the updated parameter, at least a portion of irrelevant information from the first sequence of feature vectors resulting in a third sequence of feature vectors;
- and
- recognizing speech in the second recognition pass based at least in part on the third sequence of feature vectors to determine a second result.
5 more elements
A computer-implemented method comprising:
- training a generic set of hidden Markov models using training data and at least one vector mapping function, the at least one vector mapping function removing at least a portion of information not relevant to recognizing speech from feature vectors derived from the training data;
- and
- training the at least one vector mapping function based at least in part on the training data while training the set of hidden Markov models.
A device comprising:
- a processor coupled to computer-readable storage media containing instructions executable by the processor;
- and
- an irrelevant information removal component implemented by the processor to receive a sequence of feature vectors and remove at least a portion of acoustic information not relevant to recognizing a unit of speech from the sequence.
Dependencies are inferred from the claim text.