Word recognition system which stores two models for some words and allows selective deletion of one such model

• US 5,920,837 A
• Filed: 06/26/1997
• Issued: 07/06/1999
• Est. Priority Date: 11/13/1992
• Status: Expired due to Term

• Current Assignee: Nuance Communications, Inc. (Microsoft Corporation)
• Original Assignee: Dragon Systems, Inc. (Microsoft Corporation)
• Inventors: Gould, Joel M., McGrath, Frank J., Roberts, Jed M.
• Primary Examiner(s): Dorvil, Richemond
• Application Number:
  US08/882,814
• Time in Patent Office: 740 Days
• Field of Search: 704/251, 704/243, 704/245, 704/231, 704/244, 704/275, 704/270, 704/255, 704/272, 704/273, 704/252, 704/250, 704/246, 704/247
• US Class Current: 704/251
• CPC Class Codes: G09B 19/04 Speaking with audible presentation of the material to be studied G09B5/04
• CPC Class Codes: G10L 15/063 Training
• CPC Class Codes: G10L 15/18 using natural language modelling
• CPC Class Codes: G10L 15/1815 Semantic context, e.g. disambiguation of the recognition hypotheses based on word meaning
• CPC Class Codes: G10L 15/22 Procedures used during a speech recognition process, e.g. man-machine dialogue
• CPC Class Codes: G10L 15/285 Memory allocation or algorithm optimisation to reduce hardware requirements
• CPC Class Codes: G10L 2015/0631 Creating reference templates; Clustering
• CPC Class Codes: G10L 2015/0638 Interactive procedures
• CPC Class Codes: G10L 2015/228 of application context

Abstract

A computerized word recognition system, such as a speech recognition system, stores word models of a first and second set for each of a plurality of vocabulary words. The system has a user interface which enable a user to selectively prevent the use of a second set'"'"'s word model for a selected word. Often the first set of word models are spelled word models, such as models represented by a sequence of phonetic component models, each of is derived from similar speech sounds occurring in different words. In such systems the second set of word models are custom words models derived largely from word signals which are presumed to correspond only to the model'"'"'s associated word. In many embodiments, the user interface allows a user to select to stop using a selected word'"'"'s custom, or second set, model by selecting a menu or control window of a user interface. It is preferred that the system automatically create custom models when word signals presumed to correspond to a given word score poorly against the word'"'"'s spelled model. It is also preferred that the system respond to a command to delete a word'"'"'s custom model by increasing, in subsequent adaptive training of word components for the word'"'"'s spelled model, the weight given to information from such subsequent word signals relative to model information previously associated with such word components. This is done, in such a case, to compensate for the fact that past training data for the word has probably been corrupted.

Claims

1. A computerized method for performing word recognition comprising the steps of:

receiving user generated word signals representing words;

storing a plurality of vocabulary words, including storing, for each such word, pattern information identifying word signals which represent that word, and including storing at least a first and second set of such pattern information for each of a plurality of such vocabulary words;

performing word recognition upon received word signals by pattern matching them against pattern information, including said first or second sets of pattern information, for each of a plurality of said vocabulary words in an active vocabulary to produce scores indicating the relative likelihood that each of said plurality of words corresponds to a given word signal;

providing a user interface specifically designed to enable a user to both select a given vocabulary word and to generate a command to stop using the second set of pattern information, if any, associated with that selected vocabulary word in said recognition step; and

responding to the generation of such a model-changing command for said selected vocabulary word by causing the second set of pattern information associated with that selected vocabulary word, if any, to be substantially ignored in said recognition step and for causing said selected word'"'"'s first set of pattern information, if any, to be used by said recognition step instead.

2. A computerized method as in claim 1 wherein:

said receiving of user generated word signals includes receiving acoustic word signals representing the sounds of spoken words; and

said recognition is a speech recognition.

3. A computerized method as in claim 1 wherein

said step of providing a user interface specifically designed to enable a user to both select a given vocabulary word and to generate said model-changing command includes projecting a menuitem or control window on a computer screen and generating said model-changing command when the user selects said menu item or control window.

4. A computerized method as in claim 1:

wherein said first set of pattern information which may be stored for a given vocabulary word is a spelled word model, which has a spelling formed of a sequence of word component, that is, a signal pattern which occurs as part of the word signals of more than one word, each of which has corresponding pattern information representing its signal in multiple words; and

wherein said second set of pattern information which may be stored for a given word is a custom word model, which is defined by training data derived almost exclusively from word signals associated with its vocabulary word.

5. A computerized method as in claim 4 further including the steps of using one or more given word signals which have been associated with a given vocabulary word, to train a word model associated with the given vocabulary words by performing the steps of:

scoring the match between each of the given word signals against the word component pattern information defined by the spelled word model for the given vocabulary word; and

responding to such a scoring between said word signals said spelled word models which is worse than a certain measure by using information in said given word signals to create and train a custom word model for said given vocabulary word and for causing that custom model to be used in subsequent word recognition when said given vocabulary word is in said active vocabulary.

6. A computerized method as in claim 4 further including the steps of:

adaptively training the pattern information in the word components of a given vocabulary word by combining information from parts of a word signal of the given vocabulary word with pattern information for the corresponding word components in the spelled word model for the given vocabulary word, with the weight being given to information from the word signal relative to prior word component pattern information in each such combination decreasing as a function of increases in the amount of prior training the word component pattern information has received; and

responding to the generation of said model-changing command for a selected vocabulary word by increasing, in subsequent adaptive training for word components in the spelled word model of the selected vocabulary word, the weight given to information from word signals relative to pattern information previously associated with such word components.

7. A computerized apparatus for performing word recognition comprising:

means for receiving user generated word signals representing words;

means for storing a plurality of vocabulary words, including means for storing, for each such word, pattern information identifying word signals which represent that word, and including means for storing at least a first and second set of such pattern information for each of a plurality of such vocabulary words;

means for performing word recognition upon received word signals by pattern matching them against pattern information, including said first or second sets of pattern information, for each of a plurality of said vocabulary words in an active vocabulary to produce scores indicating the relative likelihood that each of said plurality of words corresponds to a given word signal;

means for providing a user interface specifically designed to enable a user to both select a given vocabulary word and to generate a command to stop using the second set of pattern information, if any, associated with that selected vocabulary word in said word recognition; and

means for responding to the generation of such a model-changing command for said selected vocabulary word by causing the second set of pattern information associated with that selected vocabulary word, if any, to be substantially ignored by said means for performing word recognition and for causing said selected word'"'"'s first set of pattern information, if any, to be used by said recognition means instead.

8. A computerized apparatus as in claim 7 wherein:

said means for receiving user generated word signals includes means for receiving acoustic word signals representing the sounds of spoken words; and

said means for performing recognition performs speech recognition.

9. A computerized apparatus as in claim 7 wherein said means for providing a user interface specifically designed to enable a user to both select a given vocabulary word and to generate said model-changing command includes means for projecting a menu item or control window on a computer screen and means for generating said model-changing command when the user selects said menu item or control window.

10. A computerized apparatus as in claim 7:

wherein said first set of pattern information which may be stored for a given vocabulary word is a spelled word model, which has a spelling formed of a sequence of word component, that is, a signal pattern which occurs as part of the word signals of more than one word, each of which has corresponding pattern information representing its signal in multiple words; and

wherein said second set of pattern information which may be stored for a given word is a custom word model, which is defined by training data derived almost exclusively from word signals associated with its vocabulary word.

11. A computerized apparatus as in claim 10 further including means for using one or more given word signals which have been associated with a given vocabulary word, to train a word model associated with the given vocabulary words, said means including:

means for scoring the match between each of the given word signals against the word component pattern information defined by the spelled word model for the given vocabulary word; and

means for responding to such a scoring between said word signals said spelled word models which is worse than a certain measure by using information in said given word signals to create and train a custom word model for said given vocabulary word and for causing that custom model to be used in subsequent word recognition when said given vocabulary word is in said active vocabulary.

12. A computerized apparatus as in claim 10 further including:

means for adaptively training the pattern information in the word components of a given vocabulary word by combining information from parts of a word signal of the given vocabulary word with pattern information for the corresponding word components in the spelled word model for the given vocabulary word, with the weight being given to information from the word signal relative to prior word component pattern information in each such combination decreasing as a function of increases in the amount of prior training the word component pattern information has received; and

means for responding to the generation of said model-changing command for a selected vocabulary word by increasing, in subsequent adaptive training for word components in the spelled word model of the selected vocabulary word, the weight given to information from word signals relative to pattern information previously associated with such word components.

13. A computer program for performing word recognition stored on a computer readable memory comprising:

instructions for receiving user generated word signals representing words;

instructions for storing a plurality of vocabulary words, including instructions for storing, for each such word, pattern information identifying word signals which represent that word, and including instructions for storing at least a first and second set of such pattern information for each of a plurality of such vocabulary words;

instructions for performing word recognition upon received word signals by pattern matching them against pattern information, including said first or second sets of pattern information, for each of a plurality of said vocabulary words in an active vocabulary to produce scores indicating the relative likelihood that each of said plurality of words corresponds to a given word signal;

instructions for providing a user interface specifically designed to enable a user to both select a given vocabulary word and to generate a command to stop using the second set of pattern information, if any, associated with that selected vocabulary word in said word recognition; and

instructions for responding to the generation of such a model-changing command for said selected vocabulary word by causing the second set of pattern information associated with that selected vocabulary word, if any, to be substantially ignored by said instructions for performing word recognition and for causing said selected word'"'"'s first set of pattern information, if any, to be used by said recognition instructions instead.

14. A computerized apparatus as in claim 13 wherein:

said instructions for receiving user generated word signals includes instructions for receiving acoustic word signals representing the sounds of spoken words; and

said instructions for performing recognition performs speech recognition.

15. A computerized apparatus as in claim 13 wherein said instructions for providing a user interface specifically designed to enable a user to both select a given vocabulary word and to generate said model-changing command includes instructions for projecting a menu item or control window on a computer screen and instructions for generating said model-changing command when the user selects said menu item or control window.

16. A computerized apparatus as in claim 13:

wherein said first set of pattern information which may be stored for a given vocabulary word is a spelled word model, which has a spelling formed of a sequence of word component, that is, a signal pattern which occurs as part of the word signals of more than one word, each of which has corresponding pattern information representing its signal in multiple words; and

wherein said second set of pattern information which may be stored for a given word is a custom word model, which is defined by training data derived almost exclusively from word signals associated with its vocabulary word.

17. A computerized apparatus as in claim 16 further including instructions for using one or more given word signals which have been associated with a given vocabulary word, to train a word model associated with the given vocabulary words, said instructions including:

instructions for scoring the match between each of the given word signals against the word component pattern information defined by the spelled word model for the given vocabulary word; and

instructions for responding to such a scoring between said word signals said spelled word models which is worse than a certain measure by using information in said given word signals to create and train a custom word model for said given vocabulary word and for causing that custom model to be used in subsequent word recognition when said given vocabulary word is in said active vocabulary.

18. A computerized apparatus as in claim 16 further including:

instructions for adaptively training the pattern information in the word components of a given vocabulary word by combining information from parts of a word signal of the given vocabulary word with pattern information for the corresponding word components in the spelled word model for the given vocabulary word, with the weight being given to information from the word signal relative to prior word component pattern information in each such combination decreasing as a function of increases in the amount of prior training the word component pattern information has received; and

instructions for responding to the generation of said model-changing command for a selected vocabulary word by increasing, in subsequent adaptive training for word components in the spelled word model of the selected vocabulary word, the weight given to information from word signals relative to pattern information previously associated with such word components.