Audio signal processing apparatus and audio signal processing method

1. An audio signal processing apparatus utilized in pre-processing of speech recognition, the apparatus comprising:

a likelihood calculation unit, executed by a processor using a program stored in a storage, configured to extract audio features expressing features of a voice signal and a non-voice signal from an audio signal including the voice signal and the non-voice signal, and calculate a likelihood expressing a probability that the voice signal is included in the audio signal;

a spectral feature extraction unit, executed by the processor, configured to perform a frequency analysis to the audio signal to extract a spectral feature;

a first basis matrix producing unit, executed by the processor, configured to produce a first basis matrix expressing the feature of the non-voice signal using the spectral feature;

a second basis matrix producing unit, executed by the processor, configured to specify a component having a high association with the voice signal in the first basis matrix using the likelihood, and exclude the component from the first basis matrix to produce a second basis matrix; and

a spectral feature estimation unit, executed by the processor, configured to estimate a spectral feature of the voice signal or a spectral feature of the non-voice signal by performing nonnegative matrix factorization to the spectral feature of the audio signal using the second basis matrix,

wherein the spectral feature estimation unit produces a third basis matrix and a first coefficient matrix, which express the feature of the voice signal, by nonnegative matrix factorization in which the second basis matrix is used, and estimates the spectral feature of the voice signal included in the audio signal by a product of the third basis matrix and the first coefficient matrix, and wherein the product is used to separate the voice signal from the audio signal.

2. The apparatus according to claim 1 , wherein

the second basis matrix producing unit produces the second basis matrix by excluding a column vector having a high association with the voice signal from the first basis matrix.

3. The apparatus according to claim 1 , wherein

the second basis matrix producing unit produces the second basis matrix from the first basis matrix by replacing a value of a column vector having a high association with the voice signal with 0.

4. The apparatus according to claim 1 , wherein

the second basis matrix producing unit specifies a component having a high association with the voice signal in the first basis matrix by comparing the likelihood to a predetermined threshold.

5. The apparatus according to claim 1 , further comprising:

a voice/non-voice determination unit, executed by the processor, configured to extract the audio features expressing features of the voice signal and the non-voice signal from the audio signal, and determine whether the audio signal is the voice signal or the non-voice signal using the audio features,

wherein the first basis matrix producing unit produces the first basis matrix expressing the feature of the non-voice signal using the spectral feature of the audio signal in which the voice/non-voice determination unit determines that the audio signal is the non-voice signal.

6. The apparatus according to claim 1 , wherein

the spectral feature estimation unit produces a second coefficient matrix expressing the feature of the non-voice signal by the nonnegative matrix factorization in which the second basis matrix is used, and estimates the spectral feature of the non-voice signal included in the audio signal by a product of the second basis matrix and the second coefficient matrix.

7. The apparatus according to claim 1 , further comprising:

an inverse transformation unit, executed by the processor, configured to transform the spectral feature estimated by the spectral feature estimation unit into a temporal signal.

8. An audio signal processing apparatus utilized in pre-processing of speech recognition, the apparatus comprising:

a likelihood calculation unit, executed by a processor using a program stored in a storage, configured to extract audio features expressing features of a first audio signal and a second audio signal from a third audio signal including the first audio signal and the second audio signal, and calculate a likelihood expressing a probability that the first audio signal is included in the third audio signal;

a spectral feature extraction unit, executed by the processor, configured to perform a frequency analysis to the third audio signal to extract a spectral feature;

a first basis matrix producing unit, executed by the processor, configured to produce a first basis matrix expressing the feature of the second audio signal using the spectral feature;

a second basis matrix producing unit, executed by the processor, configured to specify a component having a high association with the first audio signal in the first basis matrix using the likelihood, and exclude the component from the first basis matrix to produce a second basis matrix; and

a spectral feature estimation unit, executed by the processor, configured to estimate a spectral feature of the first audio signal or a spectral feature of the second audio signal by performing nonnegative matrix factorization to the spectral feature of the third audio signal using the second basis matrix,

wherein the spectral feature estimation unit produces a third basis matrix and a first coefficient matrix, which express the feature of the first audio signal, by nonnegative matrix factorization in which the second basis matrix is used, and estimates the spectral feature of the first audio signal included in the third audio signal by a product of the third basis matrix and the first coefficient matrix and wherein the product is used to separate the voice signal from the audio signal.

9. The apparatus according to claim 8 , wherein

the second basis matrix producing unit produces the second basis matrix by excluding a column vector having a high association with the first audio signal from the first basis matrix.

10. The apparatus according to claim 8 , wherein

the second basis matrix producing unit produces the second basis matrix from the first basis matrix by replacing a value of a column vector having a high association with the first audio signal with 0.

11. The apparatus according to claim 8 , wherein

the second basis matrix producing unit specifies a component having a high association with the first audio signal in the first basis matrix by comparing the likelihood to a predetermined threshold.

12. The apparatus according to claim 8 , further comprising:

a first/second determination unit, executed by the processor, configured to extract the audio features expressing features of the first audio signal and the second audio signal from the third audio signal, and determine whether the third audio signal is the first audio signal or the second audio signal using the audio features,

wherein the first basis matrix producing unit produces the first basis matrix expressing the feature of the second audio signal using the spectral feature of the third audio signal in which the first/second determination unit determines that the third audio signal is the second audio signal.

13. The apparatus according to claim 8 , wherein

the spectral feature estimation unit produces a second coefficient matrix expressing the feature of the second audio signal by the nonnegative matrix factorization in which the second basis matrix is used, and estimates the spectral feature of the second audio signal included in the third audio signal by a product of the second basis matrix and the second coefficient matrix.

14. The apparatus according to claim 8 , further comprising:

an inverse transformation unit, executed by the processor, configured to transform the spectral feature estimated by the spectral feature estimation unit into a temporal signal.

15. The apparatus according to claim 8 , wherein the first audio signal is a voice signal and the second audio signal is a noise signal.

16. An audio signal processing method utilized in pre-processing of speech recognition, the method comprising:

extracting audio features expressing features of a first audio signal and a second audio signal from a third audio signal including the first audio signal and the second audio signal;

calculating a likelihood expressing a probability that the first audio signal is included in the third audio signal;

performing a frequency analysis to the third audio signal to extract a spectral feature;

producing a first basis matrix expressing the feature of the second audio signal using the spectral feature;

specifying a component having a high association with the first audio signal in the first basis matrix using the likelihood;

excluding the component from the first basis matrix to produce a second basis matrix; and

estimating a spectral feature of the first audio signal or a spectral feature of the second audio signal by performing nonnegative matrix factorization to the spectral feature of the third audio signal using the second basis matrix,

wherein

a third basis matrix and a first coefficient matrix, which express the feature of the first audio signal, are produced by nonnegative matrix factorization in which the second basis matrix is used, and

the spectral feature of the first audio signal included in the third audio signal, is estimated by a product of the third basis matrix and the first coefficient matrix and wherein the product is used to separate the voice signal from the audio signal.

17. The method according to claim 16 , further comprising:

extracting the audio features expressing features of the first audio signal and the second audio signal from the third audio signal,

determining whether the third audio signal is the first audio signal or the second audio signal using the audio features,

wherein the first basis matrix expressing the feature of the second audio signal using the spectral feature of the third audio signal in which the third audio signal is the second audio signal is determined, is produced.