Enhanced automatic speech recognition using mapping between unsupervised and supervised speech model parameters trained on same acoustic training data

• US 8,306,819 B2
• Filed: 03/09/2009
• Issued: 11/06/2012
• Est. Priority Date: 03/09/2009
• Status: Active Grant

• Current Assignee: Microsoft Technology Licensing LLC (Microsoft Corporation)
• Original Assignee: Microsoft Corporation
• Inventors: Liu, Chaojun, Gong, Yifan
• Primary Examiner(s): Smits, Talivaldis Ivars
• Application Number:
  US12/400,528
• Publication Number:
  US 20100228548A1
• Time in Patent Office: 1,338 Days
• Field of Search: None
• US Class Current: 704/244
• CPC Class Codes: G10L 15/065 Adaptation

Abstract

Techniques for enhanced automatic speech recognition are described. An enhanced ASR system may be operative to generate an error correction function. The error correction function may represent a mapping between a supervised set of parameters and an unsupervised training set of parameters generated using a same set of acoustic training data, and apply the error correction function to an unsupervised testing set of parameters to form a corrected set of parameters used to perform speaker adaptation. Other embodiments are described and claimed.

Claims

1. A computer-implemented method, comprising:

generating an error correction function for an automatic speech recognition system, the error correction function representing a mapping between a supervised set of parameters and an unsupervised training set of parameters generated using a same set of acoustic training data; and

applying the error correction function to an unsupervised testing set of parameters to form a corrected set of parameters used to perform speaker adaptation.

2. The computer-implemented method of claim 1, wherein each set of parameters comprise one or more transformation matrices.

3. The computer-implemented method of claim 1, wherein the speaker adaptation comprises feature space adaptation or model space adaptation.

4. The computer-implemented method of claim 1, comprising:

processing acoustic training data;

matching the acoustic training data to a base acoustic model to produce an unsupervised training text transcription;

generating the unsupervised training set of parameters using the acoustic training data, the base acoustic model and the unsupervised training text transcription; and

mapping the unsupervised training set of parameters to a supervised set of parameters to form the error correction function.

5. The computer-implemented method of claim 4, comprising generating the supervised set of parameters using the acoustic training data, the base acoustic model and a supervised text transcription.

6. The computer-implemented method of claim 5, wherein the unsupervised training text transcription comprises a text transcription with errors, and the supervised text transcription comprises a text transcription without errors.

7. The computer-implemented method of claim 4, comprising mapping the unsupervised training set of parameters to the supervised set of parameters using a linear transform to form the error correction function.

8. The computer-implemented method of claim 1, comprising:

processing acoustic testing data from a test speaker;

matching the acoustic testing data to the base acoustic model to produce an unsupervised testing text transcription;

generating an unsupervised testing set of parameters using the acoustic testing data, the base acoustic model and an unsupervised testing text transcription; and

applying the error correction function to the unsupervised testing set of parameters to form the corrected set of parameters.

9. The computer-implemented method of claim 8, comprising:

processing acoustic speech data from the test speaker;

transforming processed acoustic speech data using the corrected set of parameters; and

matching acoustic speech data to the base acoustic model to produce the text transcription.

10. The computer-implemented method of claim 8, comprising:

processing acoustic speech data from the test speaker;

transforming the base acoustic model using the corrected set of parameters to form a run-time acoustic model; and

matching processed acoustic speech data to the run-time acoustic model to produce the text transcription.

11. A computer-readable storage medium storing computer-executable program instructions that when executed cause a computing system to:

generate an error correction function for an automatic speech recognition system, the error correction function representing a mapping between a supervised set of parameters and an unsupervised training set of parameters generated using a same set of acoustic training data;

adapt a base acoustic model or acoustic speech data from a test speaker using the error correction function; and

transcribe the acoustic speech data from the test speaker to produce speech recognition results.

12. The computer-readable storage medium of claim 11, further comprising stored computer-executable program instructions that when executed cause the computing system to adapt the base acoustic model using model space adaptation or adapt the acoustic speech data using feature space adaptation.

13. The computer-readable storage medium of claim 11, further comprising stored computer-executable program instructions that when executed cause the computing system to:

process acoustic testing data from the test speaker;

match the acoustic testing data to the base acoustic model to produce an unsupervised testing text transcription;

generate an unsupervised testing set of parameters using the acoustic testing data, the acoustic model and the unsupervised testing text transcription; and

apply the error correction function to the unsupervised testing set of parameters to form a corrected set of parameters.

14. The computer-readable storage medium of claim 13, further comprising stored computer-executable program instructions that when executed cause the computing system to transform the processed acoustic speech data using the corrected set of parameters, and match the transformed acoustic speech data to the acoustic model to produce the speech recognition results.

15. The computer-readable storage medium of claim 13, further comprising stored computer-executable program instructions that when executed cause the computing system to transform the base acoustic model using the corrected set of parameters, and match the processed acoustic speech data to the transformed acoustic model to produce the speech recognition results.

16. A system, comprising:

an enhanced automatic speech recognition system operative to generate an error correction function, the error correction function representing a mapping between a supervised set of parameters and an unsupervised training set of parameters generated using a same set of acoustic training data, and apply the error correction function to an unsupervised testing set of parameters to form a corrected set of parameters used to perform speaker adaptation.

17. The system of claim 16, comprising:

an audio processing component operative to process acoustic training data;

a matching component operative to match the acoustic training data to a base acoustic model to produce an unsupervised training text transcription; and

an adaptation component operative to generate the unsupervised training set of parameters using the acoustic training data, the base acoustic model and the unsupervised training text transcription, and map the unsupervised training set of parameters to a supervised set of parameters to form the error correction function.

18. The system of claim 16, comprising:

the audio processing component operative to process acoustic testing data from a test speaker;

the matching component operative to match the acoustic testing data to the base acoustic model to produce an unsupervised testing text transcription; and

the adaptation component operative to generate an unsupervised testing set of parameters using the acoustic testing data, the base acoustic model and an unsupervised testing text transcription, and apply the error correction function to the unsupervised testing set of parameters to form a corrected set of parameters.

19. The system of claim 16, comprising:

the audio processing component operative to process acoustic speech data from a test speaker;

a transforming component operative to transform the processed acoustic speech data using a corrected set of parameters; and

the matching component operative to match the transformed acoustic speech data to the base acoustic model to produce the text transcription.

20. The system of claim 16, comprising:

the audio processing component operative to process acoustic speech data from the test speaker;

a transforming component operative to transform the base acoustic model using a corrected set of parameters to form a run-time acoustic model; and

the matching component operative to match the processed acoustic speech data to the run-time acoustic model to produce the text transcription.