Signal processing algorithm-integrated deep neural network-based speech recognition apparatus and learning method thereof

1. A speech recognition method performed by a deep neural network (DNN)-based speech recognition apparatus implementable by a computer, the speech recognition method comprising:

performing a model parameter learning method comprising:

converting a signal processing algorithm into a signal processing deep neural network (DNN), the signal processing algorithm being for extracting a feature parameter from a speech input signal of a time domain, converting the signal processing algorithm including:

converting pre-emphasis, windowing, and discrete Fourier transform operations of the signal processing algorithm into a first single linear operation (W 1 ),

converting a mel-frequency filter bank and a logarithm operation of the signal processing algorithm into a single linear and log operation (W 2 ), and

converting a discrete cosine transform operation of the signal processing algorithm into a second single linear operation (W 3 );

initializing the signal processing DNN with a signal processing algorithm coefficient;

converting a speech input signal (X) to a first value (X 1 ) using the first single linear operation (W 1 );

converting the first value (X 1 ) to a second value (X 2 ) using the single linear and log operation (W 2 );

converting the second value (X 2 ) to a third value (X 3 ) using the second single linear operation (W 3 );

outputting the third value (X 3 );

using the third value (X 3 ) as an input with respect to a classification DNN;

fusing the signal processing DNN and the classification DNN to create a deep learning model;

receiving the speech input signal; and

learning, using the speech input signal of the time domain, a model parameter in the deep learning model; and

performing speech recognition using the deep learning model having the learned model parameter to output a word recognized by the deep learning model.

2. The speech recognition method of claim 1 , wherein the converting a signal processing algorithm comprises: converting a plurality of linear operations forming the signal processing algorithm into a single linear operation by using a matrix inner product or outer product.

3. The speech recognition method of claim 1 , wherein the fusing the signal processing DNN and a classification DNN comprises: inputting a feature parameter output from the signal processing DNN to the classification DNN.

4. The speech recognition method of claim 1 , wherein the learning a model parameter comprises: adjusting a model parameter to generate a target output value with respect to the speech input signal of the time domain.

5. The speech recognition method of claim 4 , wherein the learning a model parameter comprises: determining a model parameter with which an error of an output value with respect to the speech input signal of the time domain is minimized by using a back-propagation learning algorithm.

6. A signal processing algorithm-integrated deep neural network-based speech recognition apparatus including at least one processor and a non-volatile memory storing a code executed by the at least one processor, wherein the processor:

converts a signal processing algorithm into a signal processing deep neural network (DNN), the signal processing algorithm being for extracting a feature parameter from a speech input signal of a time domain, converting the signal processing algorithm including:

converting pre-emphasis, windowing, and discrete Fourier transform operations of the signal processing algorithm into a first single linear operation (W 1 ),

converting a mel-frequency filter bank and a logarithm operation of the signal processing algorithm into a single linear and log operation (W 2 ), and

converting a discrete cosine transform operation of the signal processing algorithm into a second single linear operation (W 3 ),

initializes the signal processing DNN with a signal processing algorithm coefficient;

converts a speech input signal (X) to a first value (X 1 ) using the first single linear operation (W 1 );

converts the first value (X 1 ) to a second value (X 2 ) using the single linear and log operation (W 2 );

converts the second value (X 2 ) to a third value (X 3 ) using the second single linear operation (W 3 );

outputs the third value (X 3 ); and

uses the third value (X 3 ) as an input with respect to a classification DNN

fuses the signal processing DNN and the classification DNN to form a deep learning model,

receives a speech input signal of the time domain,

learns, using the speech input signal of the time domain, a model parameter in the deep learning model, and

performs speech recognition using the deep learning model having the learned model parameter to output a word recognized by the deep learning model.

7. The signal processing algorithm-integrated deep neural network-based speech recognition apparatus of claim 6 , wherein the processor converts a plurality of linear operations included in the signal processing algorithm into a single linear operation by using a matrix inner product or outer product.

8. The signal processing algorithm-integrated deep neural network-based speech recognition apparatus of claim 6 , wherein the processor inputs the feature parameter output from the signal processing DNN to the classification DNN.

9. The signal processing algorithm-integrated deep neural network-based speech recognition apparatus of claim 6 , wherein the processor adjusts a model parameter to generate a target output value with respect to the speech input signal of the time domain.

10. The signal processing algorithm-integrated deep neural network-based speech recognition apparatus of claim 9 , wherein the processor determines a model parameter with which an error of an output value with respect to the speech input signal of the time domain is minimized by using a back-propagation learning algorithm.