Hidden trajectory modeling with differential cepstra for speech recognition

• US 7,805,308 B2
• Filed: 01/19/2007
• Issued: 09/28/2010
• Est. Priority Date: 01/19/2007
• Status: Active Grant

• Current Assignee: Microsoft Technology Licensing LLC (Microsoft Corporation)
• Original Assignee: Microsoft Corporation
• Inventors: Deng, Li, Yu, Dong
• Primary Examiner(s): Chawan; Vijay B
• Application Number:
  US11/655,704
• Publication Number:
  US 20080177546A1
• Time in Patent Office: 1,348 Days
• Field of Search: 704/256.5, 704/256, 704/243, 704/255, 704/261, 704/256.8, 704/240, 704/222, 704/254, 704/241, 704/270, 704/232, 704/209, 704/244
• US Class Current: 704/261
• CPC Class Codes: G10L 15/02 Feature extraction for speech recognition; Selection of recognition unit
• CPC Class Codes: G10L 25/24 the extracted parameters being the cepstrum

Abstract

A novel system for speech recognition uses differential cepstra over time frames as acoustic features, together with the traditional static cepstral features, for hidden trajectory modeling, and provides greater accuracy and performance in automatic speech recognition. According to one illustrative embodiment, an automatic speech recognition method includes receiving a speech input, generating an interpretation of the speech, and providing an output based at least in part on the interpretation of the speech input. The interpretation of the speech uses hidden trajectory modeling with observation vectors that are based on cepstra and on differential cepstra derived from the cepstra. A method is developed that can automatically train the hidden trajectory model'"'"'s parameters that are corresponding to the components of the differential cepstra in the full acoustic feature vectors.

Claims

1. A method comprising:

receiving a speech input;

generating an interpretation of the speech input using hidden trajectory modeling with one or more observation vectors that are based at least in part on cepstra and on differential cepstra derived from the cepstra wherein using the hidden trajectory modeling comprises using prediction functions for the cepstra to estimate a mean for a distribution of target vocal tract resonances, and using prediction functions for the differential cepstra to estimate a vector-valued parameter related to the mean for the distribution of target vocal tract resonances, wherein the prediction functions are applied to time differences of sequential vocal tract resonance trajectories; and

providing an output based at least in part on the interpretation of the speech input.

2. The method of claim 1, wherein the differential cepstra are each evaluated as a function of a difference between a first cepstral observation trajectory and a subsequent second cepstral observation trajectory from a time-ordered series of cepstral observation trajectories.

3. The method of claim 2, wherein the second cepstral observation trajectory is identified for a reference time frame plus a time interval, and the first cepstral observation trajectory is identified for the reference time frame minus the time interval, and the differential cepstra are each evaluated as the second cepstral observation trajectory minus the first cepstral observation trajectory, divided by two times the time interval.

4. The method of claim 1, wherein estimating the vector-valued parameter related to the mean for the distribution of target vocal tract resonances comprises.

5. The method of claim 1, wherein the hidden trajectory modeling comprises probabilistic mapping from interval-weighted differences in sequential stochastic vocal tract resonance trajectories to interval-weighted differences in sequential stochastic observation trajectories.

6. The method of claim 5, wherein the hidden trajectory modeling further comprises iteratively computing distributions of differential cepstra as functions of the probabilistic mapping from the interval-weighted differences in the sequential stochastic vocal tract resonance trajectories to the interval-weighted differences in the sequential stochastic observation trajectories.

7. The method of claim 1, wherein using the hidden trajectory modeling comprises estimating a parameter for covariant matrices of differential cepstral prediction residues.

8. The method of claim 7, wherein estimating the parameter for covariant matrices of differential cepstral prediction residues comprises assuming diagonality for the parameter and estimating its diagonal elements.

9. A computer-readable medium having computer-executable instructions which, when executed by a computing device with a processor, enable the computing device to receive, at the processor, a speech signal and use time-ordered differences in observation trajectory cepstra derived from the speech signal by the processor and time-ordered differences in vocal tract resonance trajectory cepstra derived from the speech signal by the processor to evaluate, with the processor, a probability of an observed value corresponding to a hidden trajectory valuer by evaluating a parameter-free, non-linear prediction function acting on the time-ordered differences in vocal tract resonance trajectory cepstra, and using the prediction function to evaluate a probability of the observed value corresponding to the hidden trajectory value and to output an interpretation of the speech signal based on the probability of the observed value.

10. The computer-readable medium of claim 9, further comprising using the vocal tract resonance trajectory cepstra and the time-ordered differences in the vocal tract resonance trajectory cepstra to estimate a plurality of target vocal tract resonance vector parameters.

11. The computer-readable medium of claim 9, further comprising using the vocal tract resonance trajectory cepstra and the time-ordered differences in the vocal tract resonance trajectory cepstra to estimate a plurality of target vocal tract resonance covariance matrices.

12. The computer-readable medium of claim 9, wherein the prediction function is estimated using a Taylor series approximation.

13. The computer-readable medium of claim 9, further comprising iteratively revising hidden trajectory model parameters and iteratively reevaluating the probability of the observed value corresponding to the hidden trajectory value.

14. The computer-readable medium of claim 9, further comprising iteratively using the probability of the observed value corresponding to the hidden trajectory value for a plurality of observed values and hidden trajectory values to train a speech recognition model, and using the speech recognition model to interpret a speech input and provide an output based at least in part on the interpretation of the speech input.

15. A computer that runs an executable application for a speech recognition application, the speech recognition application comprising a hidden trajectory model with parameters trained using cepstra and differential cepstra, the computer running the speech recognition application to receive a speech signal, use the hidden trajectory model to recognize speech in the speech signal and output the recognized speech.

16. The device of claim 15, wherein the speech recognition application comprises a voice user interface.