Deep scattering spectrum in acoustic modeling for speech recognition

1. A method, comprising: extracting deep scattering spectral features from an acoustic input signal to generate a deep scattering spectral feature representation of the acoustic input signal;

inputting the deep scattering spectral feature representation to a speech recognition engine;

decoding the acoustic input signal based on at least a portion of the deep scattering spectral feature representation input to a speech recognition engine; and

outputting the decoded acoustic input signal;

wherein the speech recognition engine utilizes a hybrid architecture comprising a combination of a deep neural network and a convolutional neural network to decode the acoustic input signal, and further wherein features in the deep scattering spectral feature representation that have a local correlation in frequency are fed into the convolutional neural network part of the hybrid architecture and features that remove the local correlation are fed into the deep neural network part of the hybrid architecture, and a set of probabilities output by the hybrid architecture are used to evaluate fit between a set of acoustic models and the acoustic input signal; and wherein a result of the evaluation of fit between the set of acoustic models and the acoustic input signal is output;

wherein the extracting, inputting, decoding, and outputting steps are executed via a computer system comprising an acoustic signal processing unit and a memory.

2. The method of claim 1 , wherein the extracting step further comprises extracting first-order scatter features from the acoustic input signal.

3. The method of claim 2 , wherein extracting first-order scatter features from the acoustic input signal further comprises passing the acoustic input signal through a first wavelet transform, a modulus operator, and a low pass filter.

4. The method of claim 3 , wherein the extracting step further comprises extracting second-order scatter features from the acoustic input signal.

5. The method of claim 4 , wherein extracting the second-order scatter features from the acoustic input signal further comprises passing at least a portion of the acoustic input signal through a second wavelet transform.

6. The method of claim 5 , wherein the extracting step further comprises normalizing the second-order scatter features based on the first-order scatter features.

7. The method of claim 5 , wherein the extracting step is performed in the time domain.

8. The method of claim 5 , wherein the extracting step is performed in the frequency domain.

9. The method of claim 5 , wherein the extracting step is performed in a multi-resolution manner based on wavelets having different quality factors.

10. The method of claim 1 , wherein the extracting step further comprises reducing a dimensionality of the deep scattering spectral features.

11. The method of claim 10 , wherein the dimensionality reducing step further comprises performing a principal component analysis.

12. The method of claim 10 , wherein the dimensionality reducing step further comprises performing a linear discriminant analysis.

13. The method of claim 10 , wherein the dimensionality reducing step further comprises applying a linear bottleneck layer.

14. The method of claim 1 , wherein the extracting step further comprises applying at least one of a speaker adaptation and a sequence training.

15. An article of manufacture comprising a non-transitory computer readable storage medium having tangibly embodied thereon computer readable program code which, when executed, causes an acoustic signal processing unit to:

extract deep scattering spectral features from an acoustic input signal to generate a deep scattering spectral feature representation of the acoustic input signal;

input the deep scattering spectral feature representation to a speech recognition engine;

decode the acoustic input signal based on at least a portion of the deep scattering spectral feature representation via the speech recognition engine; and

output the decoded acoustic input signal;

wherein the speech recognition engine utilizes a hybrid architecture comprising a combination of a deep neural network and a convolutional neural network to decode the acoustic input signal, and further wherein features in the deep scattering spectral feature representation that have a local correlation in frequency are fed into the convolutional neural network part of the hybrid architecture and features that remove the local correlation are fed into the deep neural network part of the hybrid architecture, and a set of probabilities output by the hybrid architecture are used to evaluate fit between a set of acoustic models and the acoustic input signal; and wherein a result of the evaluation of fit between the set of acoustic models and the acoustic input signal is output.

16. An apparatus, comprising:

a memory; and an acoustic signal processing unit operatively coupled to the memory and configured to:

extract deep scattering spectral features from an acoustic input signal to generate a deep scattering spectral feature representation of the acoustic input signal;

decode the acoustic input signal based on at least a portion of the deep scattering spectral feature representation via a speech recognition engine; and

output the decoded acoustic input signal;

wherein the speech recognition engine utilizes a hybrid architecture comprising a combination of a deep neural network and a convolutional neural network to decode the acoustic input signal, and further wherein features in the deep scattering spectral feature representation that have a local correlation in frequency are fed into the convolutional neural network part of the hybrid architecture and features that remove the local correlation are fed into the deep neural network part of the hybrid architecture, and a set of probabilities output by the hybrid architecture are used to evaluate fit between a set of acoustic models and the acoustic input signal; and wherein a result of the evaluation of fit between the set of acoustic models and the acoustic input signal is output.