Speechworks International, Inc.

Etienne Barnard

Michael Stuart Phillips

Jean-Guy Dahan

Michael J Metzger

Mark Fanty

Krishna K Govindarajan

A method and apparatus are provided for automatically training or modifying one or more models of acoustic units in a speech recognition system. Acoustic models are modified based on information about a particular application with which the speech recognizer is used, including speech segment alignment data for at least one correct alignment and at least one wrong alignment. The correct alignment correctly represents a phrase that the speaker uttered. The wrong alignment represents a phrase that the speech recognition system recognized that is incorrect. The segment alignment data is compared by segment to identify competing segments and those that induced the recognition error. When an erroneous segment is identified, acoustic models of the phoneme in the correct alignment are modified by moving their mean values closer to the segment's acoustic features. Concurrently, acoustic models of the phoneme in the wrong alignment are modified by moving their mean values further from the acoustic features of the segment of the wrong alignment. As a result, the acoustic models will converge to more optimal values based on empirical utterance data representing recognition errors.

Method and apparatus for discriminative training of acoustic models of a speech recognition system

A method and apparatus are provided for automatically recognizing words of spoken speech using a computer-based speech recognition system according to a dynamic semantic model. In an embodiment, the speech recognition system recognizes speech and generates one or more word strings, each of which is a hypothesis of the speech, and creates and stores a probability value or score for each of the word strings. The word strings are ordered by probability value. The speech recognition system also creates and stores, for each of the word strings, one or more keyword-value pairs that represent semantic elements and semantic values of the semantic elements for the speech that was spoken. One or more dynamic semantic rules are defined that specify how a probability value of a word string should be modified based on information about external conditions, facts, or the environment of the application in relation to the semantic values of that word string. A speech recognition application, upon receiving the keyword-value pairs, instructs the speech recognizer to modify one or more of the probability values, according to the dynamic semantic rules. The dynamic semantic rules are applied to the word strings and the keyword-value pairs. The speech recognizer modifies one or more of the probability values, re-orders the word strings, and returns control to the application. As a result, the speech recognizer may adjust dynamically to a changing likelihood that a speaker uttered a particular utterance, as the application executes, depending on the context of the application and the external factors.

Dynamic semantic control of a speech recognition system

A telephone-based interactive speech recognition system is retrained using variable weighting and incremental retraining. Variable weighting involves changing the relative influence of particular measurement data to be reflected in a statistical model. Statistical model data is determined based upon an initial set of measurement data determined from an initial set of speech utterances. When new statistical model data is to be generated to reflect new measurement data determined from new speech utterances, a weighting factor is applied to the new measurement data to generate weighted new measurement data. The new statistical model data is then determined based upon the initial set of measurement data and the weighted new measurement data. Incremental retraining involves generating new statistical model data using prior statistical model data to reduce the amount of prior measurement data that must be maintained and processed. When prior statistical model data needs to be updated to reflect characteristics and attributes of new speech utterances, statistical model data is generated for the new speech utterances. Then the prior statistical model data and the statistical model data for the new measurement data are processed to generate the new statistical model data.

Automatically retraining a speech recognition system

An approach for automatically determining the accuracy of a pronunciation dictionary in a speech recognition system involves comparing an expected pronunciation representation for a particular word from a pronunciation dictionary to one or more actual pronunciations of the particular word. An accuracy score for each of the phonemes that constitute the pronunciation of the particular word is determined from the comparison of the expected and actual pronunciations for the particular word. The accuracy score is evaluated against specified accuracy criteria to determine whether the expected pronunciation for the particular word satisfies the specified accuracy criteria. If the expected pronunciation does not satisfy the specified accuracy criteria for the particular word, then the expected pronunciation for the particular word in the pronunciation dictionary is identified as requiring updating. Manual or automated update mechanisms may then be employed to update the identified expected pronunciation representations to reflect the actual pronunciations.

Growing community of inventors

Pittsburgh, PA, United States of America

Etienne Barnard