System and method for discriminative training of regression deep neural networks
View Patent ↗A method, computer program product, and computer system for transforming, by a computing device, a speech signal into a speech signal representation. A regression deep neural network may be trained with a cost function to minimize a mean squared error between actual values of the speech signal representation and estimated values of the speech signal representation, wherein the cost function may include one or more discriminative terms. Bandwidth of the speech signal may be extended by extending the speech signal representation of the speech signal using the regression deep neural network trained with the cost function that includes the one or more discriminative terms.
1. A computer-implemented method comprising:
transforming, by a computing device, a speech signal into a speech signal representation;
training, with a cost function, a regression deep neural network to minimize a mean squared error between actual values of the speech signal representation and estimated values of the speech signal representation, wherein the cost function includes one or more discriminative terms; and
extending bandwidth of the speech signal by extending the speech signal representation of the speech signal using the regression deep neural network trained with the cost function that includes the one or more discriminative terms;
wherein at least one of: (i) the one or more discriminative terms include at least one of a fricative-to-vowel power ratio and a function thereof; (ii) the one or more discriminative terms preserve relations of statistics between different phoneme classes in the actual values of the speech signal representation and the estimated values of the speech signal representation; and (iii) the method further comprising reproducing an average power ratio at an output of the regression deep neural network.
2. The computer-implemented method of claim 1 wherein the speech signal representation is obtained by decomposing the speech signal into a spectral envelope and an excitation signal, and wherein the spectral envelope is extended using the regression deep neural network trained with the cost function.
3. The computer-implemented method of claim 1 wherein the cost function preserves a power ratio between the different phoneme classes in the actual values of the speech signal representation and the estimated values of the speech signal representation.
4. The computer-implemented method of claim 1 wherein the cost function preserves the power ratio between the different phoneme classes using a weighted sum of K power ratio errors between the different phoneme classes.
5. A computer program product residing on a non-transitory computer readable storage medium having a plurality of instructions stored thereon which, when executed across one or more processors, causes at least a portion of the one or more processors to perform operations comprising:
transforming a speech signal into a speech signal representation;
training, with a cost function, a regression deep neural network to minimize a mean squared error between actual values of the speech signal representation and estimated values of the speech signal representation, wherein the cost function includes one or more discriminative terms; and
extending bandwidth of the speech signal by extending the speech signal representation of the speech signal using the regression deep neural network trained with the cost function that includes the one or more discriminative terms;
wherein at least one of: (i) the one or more discriminative terms include at least one of a fricative-to-vowel power ratio and a function thereof; (ii) the one or more discriminative terms preserve relations of statistics between different phoneme classes in the actual values of the speech signal representation and the estimated values of the speech signal representation; and (iii) the operations further comprising reproducing an average power ratio at an output of the regression deep neural network.
6. The computer program product of claim 5 wherein the speech signal representation is obtained by decomposing the speech signal into a spectral envelope and an excitation signal, and wherein the spectral envelope is extended using the regression deep neural network trained with the cost function.
7. The computer program product of claim 5 wherein the cost function preserves a power ratio between the different phoneme classes in the actual values of the speech signal representation and the estimated values of the speech signal representation.
8. The computer program product of claim 5 wherein the cost function preserves the power ratio between the different phoneme classes using a weighted sum of K power ratio errors between the different phoneme classes.
9. A computing system including one or more processors and one or more memories configured to perform operations comprising:
transforming a speech signal into a speech signal representation;
training, with a cost function, a regression deep neural network to minimize a mean squared error between actual values of the speech signal representation and estimated values of the speech signal representation, wherein the cost function includes one or more discriminative terms; and
extending bandwidth of the speech signal by extending the speech signal representation of the speech signal using the regression deep neural network trained with the cost function that includes the one or more discriminative terms;
wherein at least one of: (i) the one or more discriminative terms include at least one of a fricative-to-vowel power ratio and a function thereof; (ii) the one or more discriminative terms preserve relations of statistics between different phoneme classes in the actual values of the speech signal representation and the estimated values of the speech signal representation; and (iii) the operations further comprising reproducing an average power ratio at an output of the regression deep neural network.
10. The computing system of claim 9 wherein the speech signal representation is obtained by decomposing the speech signal into a spectral envelope and an excitation signal, and wherein the spectral envelope is extended using the regression deep neural network trained with the cost function.
11. The computing system of claim 9 wherein the cost function preserves a power ratio between the different phoneme classes in the actual values of the speech signal representation and the estimated values of the speech signal representation, and wherein the cost function preserves the power ratio between the different phoneme classes using a weighted sum of K power ratio errors between the different phoneme classes.