Adapting to adverse acoustic environment in speech processing using playback training data

• US 20030191636A1
• Filed: 04/05/2002
• Published: 10/09/2003
• Est. Priority Date: 04/05/2002
• Status: Active Grant

• Current Assignee: Intel Corporation
• Original Assignee: Intel Corporation
• Inventors: Zhou, Guojun
• Granted Patent: US 7,072,834 B2
• Time in Patent Office: Days
• US Class Current: 704/226
• 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 21/0216 characterised by the method used for estimating noise

Abstract

An arrangement is provided for an automatic speech recognition mechanism to adapt to an adverse acoustic environment. Some of the original training data, collected from an original acoustic environment, is played back in an adverse acoustic environment. The playback data is recorded in the adverse acoustic environment to generate recorded playback data. An existing speech model is then adapted with respect to the adverse acoustic environment based on the recorded playback data and/or the original training data.

Claims

1. A method, comprising:

playing back at least some of original training data, collected in a first acoustic environment, in an adverse acoustic environment to generate playback data;

recording the playback data in the adverse acoustic environment to generate recorded playback data; and

generating an adapted speech model from an existing speech model with respect to the adverse acoustic environment based on the recorded playback data and the original training data.

2. The method according to claim 1, wherein said generating an adapted speech model includes at least one of:

re-training the existing speech model using the recorded playback data and the original training data to generate the adapted speech model, the existing speech model being derived based on the original training data with respect to the first acoustic environment;

or

adapting the existing speech model based on discrepancy between the original training data and the recorded playback data to generate the adapted speech model.

3. The method acording to claim 2, wherein the existing speech model includes at least one of:

an acoustic model for describing the acoustic realization of at least one speech sound;

a background model for describing an acoustic environment; and

a transformation for altering some property of a given function when the transformation is applied to the function.

4. The method according to claim 3, wherein adapting the existing speech model to generate the adapated speech model comprises:

estimating the discrepancy between the original training data and the recorded playback data; and

updating the existing speech model based on the discrepancy to generate the adapted speech model.

5. The method according to claim 1, further comprising:

receiving input speech collected in the adverse acoustic environment; and

processing the input speech using the adapted speech model that is updated with respect to the adverse acoustic environment.

6. A method for deriving an adapted speech model in an adverse acoustic environment, comprising:

playing back at least some of original training data, collected in a first acoustic environment, in an adverse acoustic environment to generate playback data;

recording the playback data in the adverse acoustic environment to generate recorded playback data; and

re-training an existing speech model using the recorded playback data to generate an adapted speech model which adapts the existing speech model, derived based on the original training data with respect to the first acoustic environment, to the adverse acoustic environment.

7. The method according to claim 6, further comprising:

determining the at least some of the origial data prior to said playing back; and

retrieving the recorded playback data prior to said re-training.

8. A method for deriving an adapted speech model in an adverse acoustic environment, comprising:

playing back at least some of original training data, collected in a first acoustic environment, in an adverse acoustic environment to generate playback data;

recording the playback data in the adverse acoustic environment to generate recorded playback data;

estimating the discrepancy between the original training data and the recorded playback data; and

adapting an existing speech model based on the discrepancy to generate an adapted speech model with respect to the adverse acoustic environment.

9. The method according to claim 8, wherein said adapting includes:

revising the existing speech model to generate the adapted model;

or

generating the adapated speech model based on the existing speech model.

10. The method according to claim 9, wherein said generating comprises:

deriving a transformation based on the existing speech model and the discrepancy; and

applying the transformation to the existing speech model to produce the adapted speech model.

11. A system, comprising:

an original training data set collected in a first acoustic environment for training at least one existing speech model;

a training data sampling and playback mechanism for selecting at least some of the original training data and for playing back the at least some of the original data in an adverse acoustic environment to generate playback data; and

an adverse acoustic environment adaptation mechanism for deriving at least one adapted speech model, adapting to the adverse acoustic environment, based on the playback data and the original training data.

12. The system according to claim 11, wherein the adverse acoustic environment adaptation mechanism comprises at least one of:

a speech model re-training mechanism for re-training an existing speech model using the playback data to derive the adapted speech model with respect to the adverse acoustic environment;

or

a discrepancy based speech model adaptation mechanism for adapting the existing speech model, derived based on the original training data with respect to the first acoustic environment, to obtain the adapted speech model based on discrepancy between the original training data and the playback data.

13. The system according to claim 12, further comprising:

an automatic speech recognition mechanism for processing input speech, collected in the adverse acoustic environment, using the adapated speech model.

14. A system for deriving an adapted speech model in an adverse acoustic environment, comprising:

a training data sampling and playback mechanism for selecting at least some of original training data and for generating playback data in an adverse acoustic environment based on the selected original training data; and

an adverse acoustic environment adaptation mechanism comprising a speech model re-training mechanism for re-training an existing speech model using the playback data to derive an adapted speech model with respect to the adverse acoustic environment.

15. The system according to claim 14, wherein the speech model re-training mechanism includes at least some of:

an acoustic model re-training mechanism for re-training an acoustic model to adapt to the adverse acoustic environment based on the playback data; and

a background model re-training mechanism for re-training a background model to adapt to the adverse acoustic environment based on the playback data.

16. A system for deriving an adapted speech model, comprising:

a training data sampling and playback mechanism for selecting at least some of original training data and for generating playback data in an adverse acoustic environment based on the selected original training data; and

an adverse acoustic environment adaptation mechanism comprising a discrepancy based speech model adaptation mechanism for adapting an existing speech model, derived based on the original training data with respect to the first acoustic environment, to obtain an adapted speech model based on discrepancy between the original training data and the playback data.

17. The system according to claim 16, wherein the discrepancy based speech model adaptation mechanism comprises:

a discrepancy detection mechanism for detecting discrepancy betwee the original training data, collected in the first acoustic environment, and the playback data, collected in the adverse acoustic environment; and

at least one of;

an acoustic model update mechanism for adaping an existing acoustic model established based on the original training data, to derive an adapted acoustic model, which is part of the adapted speech model;

a background model update mechanism for revising an existing background model established based on the original training data, to derive an adapted background model, which is part of the adapted speech model; and

a transformation update mechanism for generating a transformation based on the discrepancy and the existing speech model, the transformation being applied to transform the existing speech model to the adapted speech model.

18. A machine-accessable medium encoded with data, the data, when accessed, causing:

playing back at least some of original training data, collected in a first acoustic environment, in an adverse acoustic environment to generate playback data;

recording the playback data in the adverse acoustic environment to generate recorded playback data; and

generating an adapted speech model from an existing speech model with respect to the adverse acoustic environment based on the recorded playback data and the original training data.

19. The medium according to claim 18, wherein said generating an adapted speech model includes at least one of:

re-training the existing speech model using the recorded playback data and the original training data to generate the adapted speech model, the existing speech model being derived based on the original training data with respect to the first acoustic environment;

or

adapting the existing speech model based on discrepancy between the original training data and the recorded playback data to generate the adapted speech model.

20. The medium acording to claim 19, wherein the existing speech model includes at least one of:

an acoustic model for describing the acoustic realization of at least one speech sound;

a background model for describing an acoustic environment; and

a transformation for altering some property of a given function when the transformation is applied to the function.

21. The medium according to claim 20, wherein adapting the existing speech model to generate the adapated speech model comprises:

estimating the discrepancy between the original training data and the recorded playback data; and

updating the existing speech model based on the discrepancy to generate the adapted speech model.

22. The medium according to claim 18, the data, when accessed, further causing:

receiving input speech collected in the adverse acoustic environment; and

processing the input speech using the adapted speech model that is updated with respect to the adverse acoustic environment.

23. A machine-accessable medium encoded with data for deriving an adapted speech model in an adverse acoustic environment, the data, when accessed, causing:

playing back at least some of original training data, collected in a first acoustic environment, in an adverse acoustic environment to generate playback data;

recording the playback data in the adverse acoustic environment to generate recorded playback data; and

re-training an existing speech model using the recorded playback data to generate an adapted speech model which adapts the existing speech model, derived based on the original training data with respect to the first acoustic environment, to the adverse acoustic environment.

24. The medium according to claim 23, the data, when accessed, further causing:

determining the at least some of the origial data prior to said playing back; and

retrieving the recorded playback data prior to said e-training.

25. A machine-accessable medium encoded with data for deriving an adapted speech model in an adverse acoustic environment, the data, when accessed, causing:

playing back at least some of original training data, collected in a first acoustic environment, in an adverse acoustic environment to generate playback data;

recording the playback data in the adverse acoustic environment to generate recorded playback data;

estimating the discrepancy between the original training data and the recorded playback data; and

adapting an existing speech model based on the discrepancy to generate an adapted speech model with respect to the adverse acoustic environment.

26. The medium according to claim 25, wherein said adapting includes:

revising the existing speech model to generate the adapted model;

or

generating the adapated speech model based on the existing speech model.

27. The medium according to claim 26, wherein said generating comprises:

deriving a transformation based on the existing speech model and the discrepancy; and

applying the transformation to the existing speech model to produce the adapted speech model.