Method and system of estimating clean speech parameters from noisy speech parameters

• US 10,319,377 B2
• Filed: 02/28/2017
• Issued: 06/11/2019
• Est. Priority Date: 03/15/2016
• Status: Active Grant

• Current Assignee: TATA Consultancy Services Limited (Tata Sons Pvt Ltd.)
• Original Assignee: TATA Consultancy Services Limited (Tata Sons Pvt Ltd.)
• Inventors: Panda, Ashish, Kopparapu, Sunil Kumar
• Primary Examiner(s): Dorvil, Richemond
• Assistant Examiner(s): Villena, Mark
• Application Number:
  US15/444,759
• Publication Number:
  US 20170270952A1
• Time in Patent Office: 833 Days
• CPC Class Codes: G10L 15/02 Feature extraction for speech recognition; Selection of recognition unit
• 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 17/02 Preprocessing operations, e.g. segment selection; Pattern representation or modelling, e.g. based on linear discriminant analysis [LDA] or principal components; Feature selection or extraction
• CPC Class Codes: G10L 17/20 Pattern transformations or operations aimed at increasing system robustness, e.g. against channel noise or different working conditions
• CPC Class Codes: G10L 21/0208 Noise filtering

Abstract

A method and system is provided for estimating clean speech parameters from noisy speech parameters. The method is performed by acquiring speech signals, estimating noise from the acquired speech signals, computing speech features from the acquired speech signals, estimating model parameters from the computed speech features and estimating clean parameters from the estimated noise and the estimated model parameters.

Claims

1. A method of estimating clean speech parameters from noisy speech parameters, said method comprising processor implemented steps of:

acquiring speech signals using a speech acquisition module (202);

estimating noise from said acquired speech signals using a noise estimation module (204), wherein estimation of noise is performed through non-speech frames of the acquired speech signals;

computing noisy speech features from said estimated noise using a feature extraction module (206), wherein Mel-Frequency Cepstral Coefficients are used as said noisy speech features;

estimating noisy model parameters from said computed noisy speech features using a parameter estimation module (208); and

estimating clean parameters from said estimated noise and said estimated noisy model parameters using a clean parameter estimation module (210), wherein the estimation is performed using Reverse Psychoacoustic Compensation (RPC),

wherein the RPC is performed iteratively by;

initializing a clean model mean to a certain value of a model mean;

performing Psychoacoustic Compensation on the initialized clean model mean to obtain a compensated clean model mean;

determining if value of the compensated clean model mean is within a certain range of a model mean value;

terminating the iteration in response to determining that the value of the compensated clean model mean is within a certain range of the model mean value; and

updating clean parameters of the clean model mean in response to determining that the value of the compensated clean model mean is outside a certain range of the model mean value.

2. The method as claimed in claim 1, wherein said speech acquisition module (202) further converts said acquired speech signals from analog to digital waveforms.

3. The method as claimed in claim 1, wherein said estimation of noise using the noise estimation module is performed during training phase.

4. The method as claimed in claim 1, wherein said estimated noise and said estimated noisy model parameters are first converted into their spectral domain representations.

5. The method as claimed in claim 1, wherein the estimated clean parameters are in their spectral domain representation.

6. The method as claimed in claim 5, wherein the estimated clean parameters are converted from their spectral domain representation to feature domain representation.

7. A system of estimating clean speech parameters from noisy speech parameters, said system comprising:

a processor;

a data bus coupled to said processor; and

a computer-usable medium embodying computer code, said computer-usable medium being coupled to said data bus, said computer program code comprising instructions executable by said processor and configured for;

acquiring speech signals;

estimating noise from said acquired speech signals, wherein estimation of noise can be performed through non-speech frames of the acquired speech signals;

computing noisy speech features from said estimated noise, wherein Mel-Frequency Cepstral Coefficients are used as said noisy speech features;

estimating noisy model parameters from said computed noisy speech features;

estimating clean parameters from said estimated noise and said estimated noisy model parameters, wherein the estimation is performed using reverse psychoacoustic compensation (RPC), wherein the RPC is performed iteratively by;

initializing a clean model mean to a certain value of a model mean;

performing Psychoacoustic Compensation on the initialized clean model mean to obtain a compensated clean model mean;

determining if value of the compensated clean model mean is within a certain range of a model mean value;

terminating the iteration in response to determining that the value of the compensated clean model mean is within a certain range of the model mean value; and

updating clean parameters of the clean model mean in response to determining that the value of the compensated clean model mean is outside a certain range of the model mean value.

8. One or more non-transitory machine readable information storage mediums comprising one or more instructions which when executed by one or more hardware processors causes:

acquiring of speech signals using a speech acquisition module (202);

estimating of noise from said acquired speech signals using a noise estimation module (204), wherein estimation of noise can be performed through non-speech frames of the acquired speech signals;

computing of noisy speech features from said acquired speech signals using a feature extraction module (206), wherein Mel-Frequency Cepstral Coefficients are used as said noisy speech features;

estimating of noisy model parameters from said computed noisy speech features using a parameter estimation module (208);

estimating of clean parameters from said estimated noise and said estimated noisy model parameters using a clean parameter estimation module (210), wherein the estimation is performed using reverse psychoacoustic compensation (RPC), wherein the RPC is performed iteratively by;

initializing a clean model mean to a certain value of a model mean;

performing Psychoacoustic Compensation on the initialized clean model mean to obtain a compensated clean model mean;

determining if value of the compensated clean model mean is within a certain range of a model mean value;

terminating the iteration in response to determining that the value of the compensated clean model mean is within a certain range of the model mean value; and

updating clean parameters of the clean model mean in response to determining that the value of the compensated clean model mean is outside a certain range of the model mean value.

9. The one or more non-transitory machine readable information storage mediums of claim 8, wherein said speech acquisition module (202) further converts said acquired speech signals from analog to digital waveforms.

10. The one or more non-transitory machine readable information storage mediums of claim 8, wherein said estimation of noise using the noise estimation module is performed during training phase.

11. The one or more non-transitory machine readable information storage mediums of claim 8, wherein said estimated noise and said estimated noisy model parameters are first converted into their spectral domain representations.

12. The one or more non-transitory machine readable information storage mediums of claim 8, wherein the estimated clean parameters are in their spectral domain representation.

13. The one or more non-transitory machine readable information storage mediums of claim 12, wherein the estimated clean parameters are converted from their spectral domain representation to feature domain representation.