Method of speech recognition resistant to convolutive distortion and additive distortion

• US 7,165,028 B2
• Filed: 09/20/2002
• Issued: 01/16/2007
• Est. Priority Date: 12/12/2001
• Status: Active Grant

• Current Assignee: Texas Instruments, Inc.
• Original Assignee: Texas Instruments, Inc.
• Inventors: Gong, Yifan
• Primary Examiner(s): Opsasnick; Michael N.
• Application Number:
  US10/251,734
• Publication Number:
  US 20030115055A1
• Time in Patent Office: 1,579 Days
• Field of Search: 704/233, 704/256
• US Class Current: 704/233
• 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

Abstract

A speech recognizer operating in both ambient noise (additive distortion) and microphone changes (convolutive distortion) is provided. For each utterance to be recognized the recognizer system adapts HMM mean vectors with noise estimates calculated from pre-utterance pause and a channel estimate calculated using an Estimation Maximization algorithm from previous utterances.

Claims

1. A method of speech recognition comprising the steps of:

receiving speech utterances and sensing noise during speech pauses;

providing models for recognizing speech;

estimating additive noise during pauses in speech to provide additive noise estimate;

adapting the models to additive noise and convolutional channel bias to provide adapted models with adapted model states;

comparing input speech utterances with the adapted models for recognizing the speech and providing an alignment between an input speech utterance and recognized models and

estimating the convolutional channel bias by an iterative statistical maximum-likelihood method using said alignment and said additive noise estimate that maximizes the channel bias based on the probability likelihood of each speech data input frame feature vector of said input speech utterance mapping to the existing adapted model states to generate the maximum-likelihood channel bias estimate.

2. The method of claim 1 wherein parameters of the models are based on log spectral parameters.

3. The method of claim 1 wherein parameters of the models are mel frequency cepstral parameters.

4. The method of claim 1 wherein said estimating step includes calculating the average of the log spectral domain during speech pauses.

5. The method of claim 1 wherein said adapting step includes adapting log spectral domain model parameters to a log spectral domain noise estimate and a log spectral domain channel estimate.

6. The method of claim 1 wherein said convolutional channel bias estimate includes estimating the log spectral domain channel bias by the maximum-likelihood method using said output alignment and said additive noise estimate to generate the maximum-likelihood spectral domain channel bias estimate.

7. The method of claim 1 wherein the models include first order time derivative features and the adapting step includes adaptation of first order time derivative features.

8. The method of claim 7 wherein the first order time derivative features are first order time derivatives of log spectral domain parameters.

9. The method of claim 1 wherein the models include second order time and the adapting step includes adaptation of second order time derivative features.

10. The method of claim 9 wherein said second order time derivative features are second order time derivatives of log spectral domain parameters.

12. The recognizer of claim 1 wherein parameters of the models are based on log spectral parameters.

13. The recognizer of claim 1 wherein parameters of the models are mel frequency cepstral parameters.

14. The recognizer of claim 1, wherein said convolutional channel bias estimator estimates the log spectral domain channel bias by the maximum-likelihood method using said output alignment and said additive noise estimate to generate the maximum-likelihood spectral domain channel bias estimate.

15. The recognizer of claim 1 wherein the models include first order time derivative features and the adapter includes adaptation of first order time derivative features.

16. The recognizer of claim 15 wherein the first order time derivative features are first order time derivatives of log spectral domain parameters.

17. The recognizer of claim 1 wherein the models include second order time derivative features and the adapter includes adaptation of second order time derivative features.

18. The recognizer of claim 17 wherein said second order time derivative features are second order time derivatives of log spectral domain parameters.

19. The recognizer of claim 1, wherein said noise estimate is determined by calculating the average of the log spectral domain during speech pauses.

20. The recognizer of claim 19, wherein said model adapter subsystem adapts log spectral domain model parameters to a log spectral domain noise estimate and a log spectral domain channel estimate.

11. A speech recognizer comprising:

a microphone and /or sensor for receiving speech utterances and sensing noise during pauses;

models for recognizing speech;

an additive noise estimator for estimating additive noise during pauses in speech to provide an additive noise estimate;

a model adapter subsystem for adapting the models to additive noise and convolutional channel bias to provide adapted models with adapted model states;

a recognizer subsystem which compares input speech utterances with the adapted models for recognizing the speech and outputs an alignment between an input speech utterance and recognized models and

a convolutional channel bias estimator that estimates the convolutional channel bias by a an iterative statistical maximum-likelihood method using said alignment and said additive noise estimate that maximizes the channel bias based on the probability likelihood of each speech data input frame feature vector of said input speech utterance mapping to the existing adapted model states to generate a maximum-likelihood channel bias estimate.