Sound processing apparatus and sound processing method

1. A sound processing apparatus comprising:

one or more hardware processors; and

a memory having stored thereon instructions which, when executed by the one or more hardware processors, cause the sound processing apparatus to:

generate an audio matrix formed from absolute amplitude values of coefficients obtained by frequency-transforming an audio signal that is a signal of an environment sound including a target sound;

perform non-negative matrix factorization for the audio matrix, thereby factorizing the audio matrix into a basis spectrum matrix and an activity matrix;

classify bases included in the basis spectrum matrix into bases concerning the target sound and bases concerning noise, and classify the activity matrix into activity rows corresponding to bases concerning the target sound and activity rows corresponding to bases concerning the noise;

perform a first calculation to obtain new bases concerning the target sound by separating specific frequency band components from the bases concerning the noise classified from the basis spectrum matrix;

perform a second calculation to obtain a matrix including frequency amplitude values of the target sound as elements using the bases concerning the target sound classified from the basis spectrum matrix, the activity rows corresponding to the bases concerning the target sound and the activity rows corresponding to the bases concerning the noise, and the bases concerning the target sound obtained by the first calculation; and

generate the audio signal of the target sound using the matrix obtained by the second calculation,

wherein the second calculation obtains, as the matrix including the frequency amplitude values of the target sound as the elements, a sum of (1) a matrix product of a matrix formed from the bases concerning the target sound classified from the basis spectrum matrix and a matrix formed from the activity rows corresponding to the bases concerning the target sound classified from the activity matrix and (2) a matrix product of a matrix formed from the activity rows corresponding to the bases concerning the noise classified from the activity matrix and a matrix formed from the bases concerning the target sound obtained by the first calculation.

2. The apparatus according to claim 1 , wherein the instructions, when executed by the one or more hardware processors, further cause the sound processing apparatus to:

generate a histogram of a spectrum component of each row of the audio matrix;

obtain a boundary portion between a frequency band of the target sound and a frequency band of the noise as a threshold using the histogram; and

obtain the bases concerning the target sound by applying a high-pass filter having the threshold as a cutoff frequency to the bases concerning the noise classified from the basis spectrum matrix.

3. The apparatus according to claim 1 , wherein the first calculation specifies, from among columns of a matrix formed from the bases concerning the noise classified from the basis spectrum matrix, a column including components of the target sound, and obtains the bases concerning the target sound by applying, to the column, a high-pass filter having a cutoff frequency according to a spectrum component of the specified column.

4. A sound processing apparatus comprising:

one or more hardware processors; and

a memory having stored thereon instructions which, when executed by the one or more hardware processors, cause the sound processing apparatus to:

generate an audio matrix formed from absolute amplitude values of coefficients obtained by frequency-transforming an audio signal that is a signal of an environment sound including a target sound;

perform non-negative matrix factorization for the audio matrix, thereby factorizing the audio matrix into a basis spectrum matrix and an activity matrix;

classify bases included in the basis spectrum matrix into bases concerning the target sound and bases concerning noise, and classify activity rows included in the activity matrix into activity rows corresponding to bases concerning the target sound and bases concerning the noise;

perform a first calculation to obtain bases for which components of a high frequency band of the bases are suppressed from the bases concerning the noise classified from the basis spectrum matrix;

perform a second calculation to obtain a matrix including frequency amplitude values of the noise as elements using the activity rows corresponding to the bases concerning the noise classified from the activity matrix and the bases obtained by the first calculation;

perform a third calculation to obtain a matrix including the frequency amplitude values of the target sound as elements using the audio matrix and the matrix obtained by the second calculation; and

generate the audio signal of the target sound using the matrix obtained by the third calculation.

5. The apparatus according to claim 4 , wherein the first calculation:

generates a histogram of a spectrum component of each row of the audio matrix;

obtains a boundary portion between a frequency band of the target sound and a frequency band of the noise as a threshold using the histogram; and

applies a low-pass filter having the threshold as a cutoff frequency to the bases concerning the noise classified from the basis spectrum matrix.

6. The apparatus according to claim 4 , wherein the second calculation obtains, as the matrix including the frequency amplitude values of the noise as the elements, a matrix product of a matrix formed from the activity rows concerning the noise classified from the activity matrix and a matrix formed from the bases obtained by the first calculation.

7. The apparatus according to claim 4 , wherein the third calculation obtains the matrix including the frequency amplitude values of the target sound as the elements by subtracting the matrix obtained by the second calculation.

8. A sound processing method performed by a sound processing apparatus, comprising:

generating an audio matrix formed from absolute amplitude values of coefficients obtained by frequency-transforming an audio signal that is a signal of an environment sound including a target sound;

performing non-negative matrix factorization for the audio matrix, thereby factorizing the audio matrix into a basis spectrum matrix and an activity matrix;

classifying bases included in the basis spectrum matrix into bases concerning the target sound and bases concerning noise, and classifying the activity matrix into activity rows corresponding to bases concerning the target sound and activity rows corresponding to bases concerning the noise;

performing a first calculation to obtain new bases concerning the target sound by separating specific frequency band components from the bases concerning the noise classified from the basis spectrum matrix;

performing a second calculation to obtain a matrix including frequency amplitude values of the target sound as elements using the bases concerning the target sound classified from the basis spectrum matrix, the activity rows corresponding to the bases concerning the target sound, the activity rows corresponding to the bases concerning the noise classified from the activity matrix, and the obtained new bases concerning the target sound; and

generating the audio signal of the target sound using the obtained matrix including frequency amplitude values of the target sound as elements,

wherein the second calculation obtains, as the matrix including the frequency amplitude values of the target sound as the elements, a sum of (1) a matrix product of a matrix formed from the bases concerning the target sound classified from the basis spectrum matrix and a matrix formed from the activity rows corresponding to the bases concerning the target sound classified from the activity matrix and (2) a matrix product of a matrix formed from the activity rows corresponding to the bases concerning the noise classified from the activity matrix and a matrix formed from the bases concerning the target sound obtained by the first calculation.

9. A sound processing method performed by a sound processing apparatus, comprising:

generating an audio matrix formed from absolute amplitude values of coefficients obtained by frequency-transforming an audio signal that is a signal of an environment sound including a target sound;

performing non-negative matrix factorization for the audio matrix, thereby factorizing the audio matrix into a basis spectrum matrix and an activity matrix;

classifying bases included in the basis spectrum matrix into bases concerning the target sound and bases concerning noise, and classify activity rows included in the activity matrix into bases concerning the target sound and activity rows corresponding to bases concerning the noise;

obtaining bases for which components of a high frequency band of the bases are suppressed from the bases concerning the noise classified from the basis spectrum matrix;

obtaining a matrix including frequency amplitude values of the noise as elements using the activity rows corresponding to the bases concerning the noise classified from the activity matrix and the obtained bases for which components of the high frequency band of the bases are suppressed from the bases concerning the noise classified from the basis spectrum matrix;

obtaining a matrix including the frequency amplitude values of the target sound as elements using the audio matrix and the obtained matrix including frequency amplitude values of the noise as elements; and

generating the audio signal of the target sound using the obtained matrix including the frequency amplitude values of the target sound as elements.

10. A non-transitory computer-readable storage medium storing a computer program that causes a computer to:

generate an audio matrix formed from absolute amplitude values of coefficients obtained by frequency-transforming an audio signal that is a signal of an environment sound including a target sound;

perform non-negative matrix factorization for the audio matrix, thereby factorizing the audio matrix into a basis spectrum matrix and an activity matrix;

classify bases included in the basis spectrum matrix into bases concerning the target sound and bases concerning noise, and classify the activity matrix into activity rows corresponding to bases concerning the target sound and activity rows corresponding to bases concerning the noise;

perform a first calculation to obtain new bases concerning the target sound by separating specific frequency band components from the bases concerning the noise classified from the basis spectrum matrix;

perform a second calculation to obtain a matrix including frequency amplitude values of the target sound as elements using the bases concerning the target sound classified from the basis spectrum matrix, the activity rows corresponding to the bases concerning the target sound and the activity rows corresponding to the bases concerning the noise classified from the activity matrix, and the obtained new bases concerning the target sound; and

generate the audio signal of the target sound using the obtained matrix including frequency amplitude values of the target sound as elements,

wherein the second calculation obtains, as the matrix including the frequency amplitude values of the target sound as the elements, a sum of (1) a matrix product of a matrix formed from the bases concerning the target sound classified from the basis spectrum matrix and a matrix formed from the activity rows corresponding to the bases concerning the target sound classified from the activity matrix and (2) a matrix product of a matrix formed from the activity rows corresponding to the bases concerning the noise classified from the activity matrix and a matrix formed from the bases concerning the target sound obtained by the first calculation.

11. A non-transitory computer-readable storage medium storing a computer program that causes a computer to:

generate an audio matrix formed from absolute amplitude values of coefficients obtained by frequency-transforming an audio signal that is a signal of an environment sound including a target sound;

perform non-negative matrix factorization for the audio matrix, thereby factorizing the audio matrix into a basis spectrum matrix and an activity matrix;

classify bases included in the basis spectrum matrix into bases concerning the target sound and bases concerning noise, and classify activity rows included in the activity matrix into activity rows corresponding to bases concerning the target sound and activity rows corresponding to bases concerning the noise;

obtain bases for which components of a high frequency band of the bases are suppressed from the bases concerning the noise classified from the basis spectrum matrix;

obtain a matrix including frequency amplitude values of the noise as elements using the activity rows corresponding to bases concerning the noise classified from the activity matrix and the obtained bases for which components of the high frequency band of the bases are suppressed from the bases concerning the noise classified from the basis spectrum matrix;

obtain a matrix including the frequency amplitude values of the target sound as elements using the audio matrix and the obtained matrix including frequency amplitude values of the noise as elements; and

generate the audio signal of the target sound using the obtained matrix including the frequency amplitude values of the target sound as elements.