Methods and apparatus for generating and using garbage models for speaker dependent speech recognition purposes

• US 5,842,165 A
• Filed: 04/30/1997
• Issued: 11/24/1998
• Est. Priority Date: 02/29/1996
• Status: Expired due to Term

• Current Assignee: Google LLC (Alphabet Inc.)
• Original Assignee: NYNEX Science & Technology, Inc. (Verizon Communications Inc.)
• Inventors: Vysotsky, George J., Raman, Vijay R.
• Primary Examiner(s): Dorvil, Richemond
• Application Number:
  US08/846,617
• Time in Patent Office: 573 Days
• Field of Search: 704/231, 704/243, 704/245, 704/200, 704/244, 704/255, 704/256, 704/252, 704/250, 379/88
• US Class Current: 704/255
• CPC Class Codes: G10L 15/065 Adaptation
• CPC Class Codes: G10L 15/20 Speech recognition techniques specially adapted for robustness in adverse environments, e.g. in noise, of stress induced speech G10L21/02 takes precedence
• CPC Class Codes: G10L 15/22 Procedures used during a speech recognition process, e.g. man-machine dialogue
• CPC Class Codes: G10L 15/26 Speech to text systems G10L15/08 takes precedence
• CPC Class Codes: G10L 15/34 Adaptation of a single recogniser for parallel processing, e.g. by use of multiple processors or cloud computing
• CPC Class Codes: G10L 2015/088 Word spotting
• CPC Class Codes: G10L 2015/223 Execution procedure of a spoken command
• CPC Class Codes: H04M 1/271 controlled by voice recognition
• CPC Class Codes: H04M 2201/40 using speech recognition speech recognition per se G10L15/00
• CPC Class Codes: H04M 3/42 Systems providing special services or facilities to subscribers specially adapted for wireless communication networks H04W4/00
• CPC Class Codes: H04M 3/42204 Arrangements at the exchange for service or number selection by voice at the terminal H04M1/27
• CPC Class Codes: H04M 3/44 Additional connecting arrangements for providing access to frequently-wanted subscribers, e.g. abbreviated dialling at the subscriber's set H04M1/27; automatic redialling H04M3/424

Abstract

Methods and apparatus for the generation of speaker dependent garbage models from the very same data used to generate speaker dependent speech recognition models, e.g., word models, are described. The technique involves processing the data included in the speaker dependent speech recognition models to create one or more speaker dependent garbage models. The speaker dependent garbage model generation technique involves what may be described as distorting or morphing of a speaker dependent speech recognition model to generate a speaker dependent garbage model therefrom. One or more speaker dependent speech recognition models may then be combined with the generated speaker dependent garbage model to produce an updated garbage model. The scoring of speaker dependent garbage models is varied in accordance with the present invention as a function of the number of speech recognition models from which the speaker dependent garbage model was created. In one embodiment, the number of speaker dependent speech recognition models which are used in generating a speaker dependent garbage model is limited to a preselected maximum number which is empirically determined.

Claims

1. A method of generating a speaker dependent garbage model for use in a speaker dependent speech recognition operation, comprising the steps of:

receiving a first signal including a first utterance from a speaker for which a speaker dependent speech recognition model is to be generated;

generating a speaker dependent speech recognition model from the first utterance, the speaker dependent speech recognition model including a plurality of feature vectors; and

processing the feature vectors included in the speaker dependent speech recognition model to generate a plurality of feature vectors representing the speaker dependent garbage model.

2. The method of claim 1, wherein the step of processing the feature vectors included in the speaker dependent speech recognition model includes the steps of:

generating a mean feature vector; and

replicating the mean feature vector a preselected number of times to generate the garbage model.

3. A method of generating a speaker dependent garbage model for use in a speaker dependent speech recognition operation, comprising the steps of:

receiving a first signal including a first utterance from a speaker for which a speaker dependent speech recognition model is to be generated;

generating a speaker dependent speech recognition model from the first utterance, the speaker dependent speech recognition model including a plurality of feature vectors; and

processing the feature vectors included in the speaker dependent speech recognition model to generate a plurality of feature vectors representing the speaker dependent garbage model, the processing including the steps of;

generating a mean feature vector; and

replicating the mean feature vector

to generate the garbage model;

receiving another signal including another utterance from said speaker for which a speaker dependent speech recognition model is to be generated;

generating another speaker dependent speech recognition model from the another utterance, the speaker dependent speech recognition model including a plurality of feature vectors; and

updating the previously generated speaker dependent garbage model by combining corresponding feature vectors from the garbage model with corresponding feature vectors from the another speaker dependent speech recognition model.

4. The method of claim 3, wherein the combining step performed as part of the updating step includes a feature vector averaging operation.

10. The method of claim 3, further comprising the step of:

generating a single state speaker dependent garbage model as a function of the feature vectors included in the updated speaker dependent garbage model.

11. The method of claim 3, further comprising the steps of:

generating a scoring weight to be used in controlling the scoring of the generated speaker dependent garbage model when performing a speech recognition operation; and

adjusting the scoring weight as a function of the number of utterances used to update the generated speaker dependent garbage model.

12. A method of generating a speaker dependent garbage model for use in speech recognition applications, comprising the steps of:

receiving a signal including a first utterance representing speech for which a speaker dependent garbage model is to be created;

processing the first utterance to generate a set of speech characteristic data representing the utterance;

performing a distorting operation on the set of speech characteristic data to generate therefrom a speaker dependent garbage model.

5. The method of claim 3, further comprising the step of:

repeating the receiving, generating, and updating steps recited in claim 12 thereby updating the speaker dependent garbage model to be a function of a plurality of utterances; and

monitoring the number of utterances used in updating the garbage model.

6. The method of claim 5, further comprising the step of:

limiting the number of utterances used to create and update the speaker dependent garbage model.

7. The method of claim 5, further comprising the steps of:

generating a scoring weight to be used in controlling the scoring of the generated speaker dependent garbage model when performing a speech recognition operation; and

adjusting the scoring weight as a function of the number of utterances used to update the generated speaker dependent garbage model.

8. The method of claim 7, further comprising the step of:

generating a single state speaker dependent garbage model as a function of the feature vectors included in the updated garbage model.

9. The method of claim 8, wherein the generated single state garbage model is generated by computing a mean feature vector from a plurality of feature vectors included in the updated garbage speaker dependent model.

13. The method of claim 12,

wherein the set of speech characteristic data includes data corresponding to a plurality of different portions of the utterance;

and

wherein the step of performing a distorting operation includes performing a step to reduce the differences between the speech characteristic data corresponding to at least some of the plurality of different portions of the utterance.

14. The method of claim 13,

wherein the speech characteristic data includes feature vectors;

and

wherein the step to reduce the differences between speech characteristic data includes a feature vector averaging operation.

15. The method of claim 14,

wherein the step of processing the utterance to generate a set of speech characteristic data representing the utterance includes the step of:

generating a speaker dependent speech recognition model from the utterance including a plurality of feature vectors corresponding to different states included in the model.

16. The method of claim 15, wherein the step of:

performing a distorting operation on the set of speech characteristic data to generate therefrom a speaker dependent garbage model includes the steps of;

generating a mean feature vector from the feature vectors included in the speaker dependent speech recognition model; and

populating the states of a speaker dependent garbage model with generated mean feature vector.

17. The method of claim 16, further comprising the step of:

updating the generated garbage model using feature vectors corresponding to another utterance.

18. The method of claim 17, wherein the updating step includes the step of:

combining garbage model feature vectors with the feature vectors corresponding to the another utterance on a state by state basis.

19. The method of claim 18, further comprising the step of:

monitoring the number of utterances contributing to the creation of the garage model;

associating a scoring weight used to control scoring during a speech recognition operation with the generated garbage model; and

adjusting the scoring weight as a function of the number of utterances contributing to the creation of the garbage model.

20. The method of claim 19,

further comprising the step of monitoring the number of utterances contributing to the creation of the garbage model;

and

limiting the number of contributing utterances to a number not exceeding a preselected maximum number.