Back to results
US 6,275,802 B1Granted
Search algorithm for large vocabulary speech recognition
AssigneeLernout & Hauspie Speech Products N.V.
InventorFilip Van Aelten
Life of this patent
Estimating the term…
Abstract
Automatic speech recognition word sequence hypotheses are generated using an interleaved forward-backward search. A forward search pass uses relatively simple models for a given block period of time. A backward search pass then goes back over the previous block period using more complex models and the recognition hypotheses generated by the forward search pass. The backward search pass employs a word dependent n-best search having a flat model state organization.
Claims34
A method of a speech recognition system for processing a sequence of parameters that represent an input speech signal, the method comprising:
- a. at selected locations in the sequence, searching over a current block time period having a duration sufficient that at least one word in the input speech signal is likely included, with a forward search pass that:
- i. compares the current block time period parameters to selected word models from a first set of word models having sequences of model states, and
- ii. for each state in a set of selected model states, generates a most likely forward pass recognition hypothesis ending in the state and corresponding to the input speech signal;
- b. performing a backward search pass, back through the sequence of parameters within the current block time period, that:
- i. compares the current block time period parameters to selected word models from a second set of word models having sequences of model states, and to the most likely forward pass recognition hypotheses,
- ii. generates a current word graph including
- (1) a set of word graph recognition hypotheses, of at least one word, that end in selected model states, each recognition hypothesis having an associated occurrence probability score representing the likelihood of that recognition hypothesis corresponding to the input speech signal, and
- (2) nodes which connect adjacent words in the current word graph;
- and
- iii. prunes any generated word graph recognition hypothesis that has an occurrence probability score less than a selected threshold;
- c. updating any preceding word graph by linking recognition hypotheses of the preceding word graph to the current word graph;
- and
- d. repeating steps (a)-(c) for the next block time period until the end of the utterance.
10 more elements
A speech recognition system comprising:
- a. an input segmenter that processes an input speech signal into a sequence of representative parameters;
- b. a forward search comparator in communication with the input segmenter that, at selected locations in the sequence, searches a current block time period having a duration sufficient that at least one word in the input speech signal is likely included, with a forward search pass that:
- i. compares the current block time period parameters to selected word models from a first set of word models having sequences of model states,
- ii. for each state in a set of selected model states, generates a most likely forward pass recognition hypothesis ending in the state and corresponding to the input speech signal;
- c. a backward search comparator in communication with the forward search comparator that performs a backward search pass back though the sequence of parameters within the current block time period that:
- i. compares the current block time period parameters to selected word models from a second set of word models having sequences of model states, and to the forward pass recognition hypotheses,
- ii. generates a current word graph including
- (1) a set of word graph recognition hypotheses of at least one word that end in selected model states, each recognition hypothesis having an associated occurrence probability score representing the likelihood of that recognition hypothesis corresponding to the input speech signal, and
- (2) nodes which connect adjacent words in the current word graph, and
- iii. prunes any generated word graph recognition hypothesis that has an occurrence probability score less than a selected threshold;
- iv. updates any preceding word graph by linking recognition hypotheses of the preceding word graph to the current word graph;
- and
- d. a block period controller in communication with the forward search comparator and the backward search comparator that controls processing of the sequence of representative parameters in successive block periods.
10 more elements
Dependencies are inferred from the claim text.