Vladimir Sejnoha

Method for generating a speech recognition model for a non-vocabulary utterance

The method disclosed herein facilitates the generation of a recognition model for a non-standard word uttered by a user in the context of a large vocabulary speech recognition system in which standard vocabulary models are represented by sequences of probability distributions for various acoustic symbols. Along with the probability distributions, a corresponding plurality of converse probability functions are precalculated which represent the likelihood that a particular probability distribution would correspond to a given input acoustic symbol. For a non-standard word uttered, a corresponding sequence of acoustic symbols is generated and, for each such symbol in the sequence, the most likely probability distribution is selected using the converse probability functions. For successive symbols in the utterance, a corresponding sequence of custom converse probability functions are generated, each of which is a composite of weighted contributions from the corresponding precalculated converse probability function and the converse probability functions corresponding to time-adjacent symbols in the utterance. The resulting sequence of custom converse probability functions identify a corresponding sequence of probability distributions which constitute a model of the word uttered, which model incorporates contextual information from the utterance.

User adaptable speech recognition system

A speech recognition system is disclosed which comprises a core speech recognition program and a plurality of utility program modules for adjusting various recognition parameters such as gain, sensitivity and acceptance threshold and for improving the training of word models. The system further provides a decision tree and utility controlling program module which can be invoked by a user confronting problems during the running of the core program. The system utilizes user input to traverse the decision tree and to initiate appropriate ones of the utility program modules thereby to alter the on-going behavior of the core recognition program.

Speech recognition system using arbitration between continuous speech and isolated word modules

In the speech recognition system disclosed herein, an input utterance is submitted to both a large vocabulary isolated word speech recognition module and a small vocabulary continuous speech recognition module. The two recognition modules generate respective scores for identified large vocabulary models and for sequences of small vocabulary models. The score provided by the continuous speech recognizer is normalized on the basis of the length of the speech input utterance and an arbitration algorithm selects among the candidates identified by the recognition modules. Preferably, the competing scores from the two recognizers are scaled by a factor or factors empirically trained to minimize incursion by each of the vocabularies on correct results from the other vocabulary.

Word model candidate preselection for speech recognition using precomputed matrix of thresholded distance values

In the large vocabulary speech recognition system disclosed herein, a preliminary screening of vocabulary models is provided by applying high speed distance measuring functions. The distance measuring functions utilize subsampled or otherwise reduced representations of the unknown speech segment and the vocabulary models. The initial screening functions achieve very high speed by precalculating, for each utterance, a comparison table of distance values which can be used for all vocabulary models. The building of each comparison table is facilitated by a method which utilizes default values as initial entries and only adjusts entries which are meaningfully different from the default value.