JP2015064602A - Acoustic signal processing device, acoustic signal processing method, and acoustic signal processing program - Google Patents

Abstract

PROBLEM TO BE SOLVED: To realize an acoustic signal processing device that improves performance of separating a voice signal.SOLUTION: Likelihood calculation means in an acoustic signal processing device extracts an acoustic feature showing each feature of a voice signal and a non-voice signal, from an acquired acoustic signal, and uses the acoustic feature to calculate likelihood showing plausibility that the voice signal is included in the acoustic signal. Spectral extraction means analyzes a frequency of the acoustic signal, and extracts a spectral feature. First basis matrix generation means uses the spectral feature to generate a first basis matrix showing a feature of the non-voice signal. Second basis matrix generation means uses the likelihood to identify a component that is strongly associated with the voice signal in the first basis matrix, and generates a second basis matrix after excluding the component.

Description

Embodiments described herein relate generally to an acoustic signal processing device, an acoustic signal processing method, and an acoustic signal processing program.

An acoustic signal acquired by a microphone in speech recognition or the like includes not only a speech signal that is a user's speech but also non-speech signals such as background noise and music. A technique for extracting only a desired signal from an acoustic signal mixed with an audio signal and a non-audio signal is a sound source separation technique, and one of them is a system using non-negative matrix factorization. In the case of separating an audio signal from an acoustic signal, in this method, first, a base matrix of the non-speech signal is created from a spectrogram of the acoustic signal in a section having a high probability of including the non-speech signal. Next, using the basis matrix of the non-speech signal, the basis matrix and coefficient matrix of the speech signal are created from the spectrogram of the acoustic signal to be separated. Then, the spectrogram of the speech signal is estimated by the product of the basis matrix and coefficient matrix of the speech signal. Finally, the spectrogram of the estimated speech signal is converted into a time signal to separate the speech signal from the acoustic signal.

However, in this method, if the audio signal is mixed in the acoustic signal when the base matrix of the non-speech signal is obtained, the base matrix of the non-speech signal cannot be created accurately. As a result, there is a problem that the separation performance of the audio signal is deteriorated.

JP 2009-128906 A

Eric Scheirer, Malcolm Slaney, “Construction and Evaluation of a Robust Multifeature Speech / Music Discriminator,” Proc. Of ICASSP, April 1997, Vol 2, p.1331-1334.

The problem to be solved by the present invention is to realize an acoustic signal processing device that improves the separation performance of an audio signal.

The likelihood calculation means in the acoustic signal processing device of the embodiment extracts acoustic features representing the characteristics of the speech signal and the non-speech signal from the acquired acoustic signal, and the acoustic signal is included in the acoustic signal using the acoustic feature. A likelihood representing the likelihood is calculated. The spectrum extraction means extracts a spectrum feature by frequency analysis of the acoustic signal. The first basis matrix creating means creates a first basis matrix representing the feature of the non-speech signal using the spectral feature. The second basis matrix creation means uses the likelihood to identify a component highly relevant to the speech signal in the first basis matrix, and creates a second basis matrix by excluding this component.

The block diagram which shows the acoustic signal processing apparatus of 1st Embodiment. The figure which shows the hardware constitutions of the acoustic signal processing apparatus of embodiment. The figure which shows the processing content in the 2nd basis matrix preparation part of embodiment. The figure which shows the processing content in the spectrum estimation part of embodiment. The flowchart of the acoustic signal processing apparatus of embodiment. The flowchart of the 2nd basis matrix preparation part of embodiment. The figure which shows the estimation result of the spectrogram of the audio | voice signal of embodiment.

  Hereinafter, embodiments of the present invention will be described with reference to the drawings.

(First embodiment)
The acoustic signal processing apparatus according to the first embodiment separates an audio signal from an acoustic signal including an audio signal and a non-audio signal. For example, the recognition performance of voice recognition can be improved by using the acoustic signal processing apparatus of the present embodiment as preprocessing for voice recognition.

The acoustic signal processing apparatus according to the present embodiment is a sound source separation technique using non-negative matrix factorization (Patent Document 1). From the spectrogram (spectral feature) of an acoustic signal in a section with a high probability that a non-speech signal is included, A first basis matrix representing the characteristics of the audio signal is created. Then, using the speech likelihood representing the likelihood that the speech signal is included, the second basis matrix is created by excluding the component highly relevant to the speech signal from the first basis matrix. Using this second basis matrix, a third basis matrix and a first coefficient matrix representing the characteristics of the speech signal are calculated, and a spectrogram of the speech signal is obtained by the product of the third basis matrix and the first coefficient matrix. presume. Finally, the speech signal is separated from the acoustic signal by converting the spectrogram of the estimated speech signal into a time signal.

As described above, the acoustic signal processing device according to the present embodiment creates a second basis matrix in which a component that is highly relevant to the speech signal is excluded from the first basis matrix of the created non-speech signal. As a result, a base matrix that accurately represents the characteristics of the non-speech signal can be created, and as a result, the separation performance of the speech signal can be improved.

(Overall block configuration)
FIG. 1 is a block diagram illustrating an acoustic signal processing apparatus according to the first embodiment. The acoustic signal processing device of the present embodiment extracts an acoustic feature representing the characteristics of an audio signal and a non-audio signal from the acoustic acquisition unit 101 that acquires an audio signal including an audio signal and a non-audio signal. A likelihood calculation unit 102 that calculates a speech likelihood that represents a likelihood that a sound signal is included in an acoustic signal and a non-voice likelihood that represents a likelihood that a non-speech signal is included using an acoustic feature, and a speech likelihood And a speech / non-speech discrimination unit 103 that discriminates speech / non-speech for each predetermined section of the acoustic signal using a non-speech likelihood, and a spectral feature extraction unit 104 that extracts a spectrogram by Fourier transforming the acoustic signal, A first basis matrix creation unit 105 that creates a first basis matrix that represents a feature of a non-speech signal using a spectrogram extracted from an acoustic signal in a section determined to be a non-speech signal; and speech likelihood A second basis matrix creation unit 106 that creates a second basis matrix by removing a component of a column vector that is highly relevant to the audio signal from the first basis matrix; and a spectrogram is used as the second basis matrix. A non-negative matrix factorization is performed using the matrix to calculate a third basis matrix and a first coefficient matrix representing the characteristics of the speech signal, and the product of the third basis matrix and the first coefficient matrix is used to calculate an acoustic signal. A spectral feature estimation unit 107 that estimates a spectrogram of a speech signal included in the signal, and an inverse conversion unit 108 that converts the spectrogram of the estimated speech signal into a time signal by inverse Fourier transform.

(Hardware configuration)
The acoustic signal processing apparatus of the present embodiment is configured by hardware using a normal computer as shown in FIG. 2, and a CPU (Central Processing Unit) that controls the entire apparatus.
And a control unit 201 such as a ROM (Read Only Memory) for storing various data and various programs.
) And RAM (Random Access Memory) storage unit 202, HDD (Hard Disk Drive) storing various data and various programs, CD (Compact Disk) drive storage unit 203 and the like, and user instruction input An operation unit 204 such as a keyboard and a mouse to be received, a communication unit 205 that controls communication with an external device, a microphone 206 that acquires a user's utterance, and a bus 208 that connects them are provided.

In such a hardware configuration, the following functions are realized when the control unit 201 executes various programs stored in the storage unit 202 such as the ROM or the external storage unit 203.

(Function of each block)
The acoustic acquisition unit 101 acquires an acoustic signal input to the microphone 206. This acoustic signal includes not only a voice signal that is a user's voice but also a non-voice signal such as background noise and music. Note that the acoustic signal of the present embodiment is a digital signal sampled at 16 kHz. Note that the acoustic acquisition unit 101 may acquire an acoustic signal from an external device connected by the external storage unit 203 or the communication unit 205 instead of the microphone 206.

The likelihood calculating unit 102 extracts acoustic features representing the features of the audio signal and the non-audio signal from the acoustic signal acquired by the acoustic acquisition unit 101. Then, using this acoustic feature, the likelihood that the sound signal includes the sound signal (speech likelihood) and the likelihood that the sound signal includes the non-speech signal (
Non-voice likelihood) is calculated. Specifically, the likelihood calculator 102 converts the acoustic signal into a length of 25 ms (
400 samples) and an interval of 8 ms (128 samples). A Hamming window is used for frame division. Next, an acoustic feature is extracted from each frame. As the acoustic feature, not only a spectrum obtained by Fourier transforming the acoustic signal of each frame but also a cepstrum-based feature amount such as an LPC cepstrum or MFCC can be used. Then, the extracted acoustic features are collated with a speech model and a non-speech model learned in advance, and speech likelihood and non-speech likelihood of each frame are calculated.

The speech / non-speech discrimination unit 103 uses the speech likelihood and non-speech likelihood for each frame calculated by the likelihood computation unit 102 to determine whether the speech signal is speech or non-speech for each predetermined section. Determine. In the present embodiment, the length of the predetermined section is set to 400 ms (for 50 frames). The speech / non-speech discrimination unit 103 calculates speech / non-speech by calculating the average value of speech likelihood and non-speech likelihood for 50 frames and comparing the magnitudes thereof. In addition, voice / non-voice may be determined by comparing each average value with a preset threshold value. The method described in Non-Patent Document 1 can also be used.

The spectral feature extraction unit 104 generates an acoustic signal having a length of 25 ms (400 samples) and an interval of 8
The frame is divided into ms (128 samples), and a spectrum is extracted by performing Fourier transform (frequency analysis) on the acoustic signal of each frame. In this embodiment, 11 for each frame.
After zero padding of two points, 512-point discrete Fourier transform is executed.

The first basis matrix creation unit 105 creates a first basis matrix H that represents the features of the non-speech signal using the spectrum (spectrogram) of a plurality of frames extracted by the spectrum feature extraction unit 104. Here, in the present embodiment, the first basis matrix H is created using the spectrogram extracted from the acoustic signal of the section determined as non-speech by the speech / non-speech discrimination unit 103.

The first basis matrix creation unit 105 first creates a matrix Z having the number of rows F and the number of columns T in which all the elements of each F-dimensional spectrum extracted from the T frames are stored. Next, this matrix Z is converted into the number of rows F
Approximation is performed by the product of the first basis matrix H having the number of columns D and the coefficient matrix U having the number of rows D and the number of columns T. here,
D represents the number of bases and is set to 32 empirically. The first basis matrix H and the coefficient matrix U are derived by an iterative method based on the square error of the product of the first basis matrix H and the coefficient matrix U and the matrix Z. At this time, random values can be used for the initial matrixes of the first basis matrix H and the coefficient matrix U.

The second basis matrix creation unit 106 uses the speech likelihood calculated by the likelihood calculation unit 102 to remove the second column vector component that is highly relevant to the speech signal from the first basis matrix H. A basis matrix B is created.

The operation of the second basis matrix creation unit 106 will be described with reference to FIG. In FIG. 3, reference numeral 301 denotes a matrix Z composed of elements of each four-dimensional spectrum extracted from four frames, a first basis matrix H having 4 rows and 4 columns, and 4 rows and 4 columns. It is approximated by the product of the coefficient matrix U of Equation 4. 302 represents the speech likelihood corresponding to each frame (column vector) of the matrix Z. The threshold value 302 is a value for discriminating a frame including an audio signal, and a predetermined value is set in advance. By comparing the threshold and the speech likelihood, in this example, it is determined that the sound signal is included in the acoustic signal of the t-th frame. Therefore, the second basis matrix creation unit 106 creates the second basis matrix B by excluding the component due to the t-th frame from the first basis matrix H.

First, the second basis matrix creation unit 106 performs normalization processing for each row of the coefficient matrix U to obtain a normalized coefficient matrix U ′. Normalization is a process for comparing the sizes of the elements of the column vectors, and the elements of each row are executed so that the average is 0 and the variance is 1. Next, the second basis matrix creation unit 1
In 06, the element having the maximum value is selected from the column vectors of the coefficient matrix U ′ corresponding to the t-th frame. The element is specified as in the X row and the Y column of U ′. In this example, the element 303 in the second row and the third column is the maximum. Next, the second basis matrix creation unit 106 calculates the column vector of the X column in which the product of the element of the X row and the Y column of the coefficient matrix U in the product of the first basis matrix H and the coefficient matrix U is calculated. Is excluded. In this example, since the element in the second row and third column of U ′ is selected, the first
The second base matrix B is created by excluding the second column vector from the base matrix H.

As described above, the acoustic signal processing apparatus according to the present embodiment creates the second basis matrix B by excluding the column vector components that are highly relevant to the audio signal from the first basis matrix H. As a result, the second basis matrix B representing the characteristics of the non-speech signal can be accurately created.

Note that the second basis matrix creation unit 106 may create the second basis matrix B by excluding a plurality of column vectors. Further, instead of excluding the column vector, the value of the element of the column vector can be replaced with 0 or a value close to 0. In addition, the created second basis matrix B is stored in the storage unit 202 or the external storage unit 203, and the stored second basis matrix B is reused to execute separation processing on other acoustic signals. May be.

In the present embodiment, the component having high relevance to the speech signal is specified using the speech likelihood itself, but the ratio of the non-speech likelihood and the speech likelihood may be used. In addition, it is possible to specify a component that is highly relevant to the speech signal using only the non-speech likelihood. For example, a frame with a low non-voice likelihood can be regarded as a component highly related to a voice signal.

The spectrum feature estimation unit 107 performs non-negative matrix factorization on the spectrum feature extracted by the spectrum feature extraction unit 104 using the second basis matrix B, thereby providing a third basis matrix M representing the features of the speech signal and A coefficient matrix V of 1 is calculated. And this third basis matrix M
And the first coefficient matrix V are taken to estimate the spectral characteristics of the speech signal included in the acoustic signal.

First, as shown by 401 in FIG. 4, the spectral feature estimation unit 107 performs non-negative matrix factorization on the spectrogram X of the acoustic signal to be separated from the audio signal using the second basis matrix B. Similar to the derivation of the first basis matrix H, an iterative method based on the product of the basis matrix and the coefficient matrix and the square error of the spectrogram X is applied. At this time, the number of columns of the base matrix is set to a value E + D−1 which is larger than the number of columns D−1 of the second base matrix B, and the initial values after the E column of the base matrix are elements of the second base matrix B. The elements after the E column of the base matrix are not applicable to the iterative method. As a result of applying the iterative method, the matrix from the first column to the E column of the basis matrix is the third basis matrix M and the coefficient matrix 1
The matrix from the line E to the line E is the first coefficient matrix V, and the matrix after the line E is the second coefficient matrix W.

Next, the spectral feature estimation unit 107 estimates the spectrogram of the speech signal included in the acoustic signal by taking the product of the third basis matrix M and the first coefficient matrix V (402 in FIG. 4). In addition, the spectrogram of the non-speech signal included in the acoustic signal can be estimated by the product of the second base matrix B representing the characteristics of the non-speech signal and the second coefficient matrix W.

The spectral feature estimation unit 107 estimates the spectrogram of the speech signal estimated by the product of the third basis matrix M and the first coefficient matrix V, and the product of the second basis matrix B and the second coefficient matrix W. The spectrogram of the speech signal may be estimated by calculating the ratio of the component due to the speech signal for each spectrogram element from the spectrogram of the non-speech signal and multiplying the spectrogram to be separated by this ratio.

The inverse transform unit 108 converts the spectrogram of the estimated speech signal into a time signal by performing a discrete inverse Fourier transform. Thereby, an audio | voice signal can be isolate | separated from an acoustic signal. In addition, by outputting the separated voice signal to a subsequent voice recognition unit (not shown), it is possible to execute a voice recognition process in which the influence of the non-voice signal is removed.

(Flowchart: Overall processing)
The processing of the audio signal processing apparatus according to the present embodiment will be described using the flowchart of FIG. First, in step S501, the acoustic acquisition unit 101 acquires an acoustic signal to be separated.

In step S502, the spectrum feature extraction unit 104 extracts a spectrum from the acoustic signal divided into frames by discrete Fourier transform. The frame interval is 8 ms.

In step S503, the likelihood calculation unit 102 calculates an LPC cepstrum or MF from the acoustic signal.
Extracts cepstrum-based acoustic features such as CC. Then, the speech likelihood and the non-speech likelihood are calculated by collating the acoustic feature with the previously learned speech model and the non-speech model.

In step S504, the speech / non-speech discrimination unit 103 discriminates speech / non-speech for each predetermined section having a length of 400 ms (for 50 frames). For the determination, the speech likelihood and the non-speech likelihood calculated in step S503 are used.

In step S505, the process branches depending on whether or not the section to be determined is the first section. If it is the first interval, the basis matrix of the non-speech signal has not been created,
Regardless of the determination result in step 504, the process proceeds to step S507. If it is not the first section, the process proceeds to step S506.

In step S506, the process branches according to the determination result of the section to be determined. If the determination result is non-voice, the process proceeds to step S507. If the determination result is voice, the process proceeds to step S509.

In step S507, the first basis matrix creation unit 105 creates a first basis matrix H and a coefficient matrix that represent the characteristics of the non-speech signal using the spectrogram extracted from the section to be determined.

In step S508, the second basis matrix creation unit 106 uses the speech likelihood calculated in step S503 to exclude the column vector component that is highly relevant to the speech signal from the first basis matrix H. 2 basis matrix B is created. Details of step S508 will be described later.

In step S509, the spectrum feature estimation unit 107 performs non-negative matrix factorization on the spectrogram extracted from the 400 ms (50 frames) section in which speech / non-speech is determined in step S504 using the second basis matrix B, A third basis matrix M representing the characteristics of the speech signal
And a first coefficient matrix V is calculated.

In step S510, the spectrum feature estimation unit 107 obtains an estimated value of the spectrogram of the speech signal by the product of the third basis matrix M and the first coefficient matrix V. The estimated number of frames of the speech spectrogram is 50.

In step S511, the inverse transform unit 108 transforms the spectrogram of the speech signal estimated in step S510 into a time signal by discrete inverse Fourier transform.

In step S512, the acoustic signal processing device determines whether or not the audio signal separation processing has been completed for all the sections. If not completed, the process proceeds to step S504, and voice / non-voice is determined for the next 400 ms section.

(Flowchart: second basis matrix generator)
Step S50 of FIG. 5 for creating the second basis matrix B using the flowchart of FIG.
8 will be described in detail. First, in step S601, the second basis matrix creation unit 106 obtains a coefficient matrix U ′ by performing normalization for each row of the coefficient matrix U.

In step S602, the second basis matrix creation unit 106 obtains a speech likelihood corresponding to each column vector of the normalized coefficient matrix U ′.

In step S603, the second basis matrix creation unit 106 compares the speech likelihood with a preset threshold value.

In step S604, the second basis matrix creation unit 106 selects a column vector of the coefficient matrix U ′ associated with the speech likelihood exceeding the threshold. When there are a plurality of frames whose speech likelihood exceeds the threshold, a plurality of column vectors are also selected.

In step S605, the second basis matrix creation unit 106 selects an element indicating the maximum value in each column vector selected in step S604. The elements are selected as in the Xth row and the Yth column of the coefficient matrix U ′. When there are a plurality of column vectors, a plurality of elements are selected.

In step S606, the second basis matrix creation unit 106 performs the first basis matrix H and the coefficient matrix U.
, The column vector of the X column for which the product with the element of the X row and the Y column of the coefficient matrix U is calculated is excluded from the first basis matrix H. Through the above processing, the second basis matrix creation unit 106 can create the second basis matrix B from which the influence of the audio signal is removed.

FIG. 7 shows an example of a speech spectrum estimated by the acoustic signal device of the present embodiment. 7 in the figure
01 shows a spectrogram of an acoustic signal in which an audio signal and a non-audio signal are mixed. In this spectrogram, black represents the component of the audio signal and gray represents the component of the non-audio signal. In the spectrogram of the estimated speech signal, the ideal result is that the black display remains and the gray display is converted to white.

Reference numerals 702 and 703 denote spectrograms of the speech signal estimated using the basis matrix of the non-speech signal created from the spectrogram of the section 705 determined to be non-speech. 702 estimates the spectrogram of the speech signal by non-negative matrix factorization using the first basis matrix H, and 703 estimates the spectrogram of the speech signal by non-negative matrix factorization using the second basis matrix B. Yes. On the other hand, reference numeral 704 denotes a spectrogram of the speech signal estimated using the base matrix of the non-speech signal created from the spectrogram of 706, which is a section having a low speech likelihood in each frame and shorter than 705. When the spectrograms of FIG. 7 are compared, black portions 702 and 704 representing the components of the audio signal are deleted. 702 is the effect of the acoustic signal included in the section 705, and 704 is the section 70.
It is considered that the accuracy of creating the base matrix of the non-speech signal has decreased because 6 does not have a sufficient length. On the other hand, the spectrogram 703 using the acoustic signal processing apparatus of the present embodiment removes the non-speech signal component while leaving the speech signal component, indicating a more ideal estimation result. I understand.

(effect)
As described above, the acoustic signal processing device according to the present embodiment creates the second basis matrix by excluding the component highly relevant to the speech signal from the created first basis matrix of the non-speech signal. As a result, the basis matrix of the non-speech signal can be created accurately, and as a result, the separation performance of the speech signal can be improved.

(Modification)
The acoustic signal processing apparatus of the present embodiment is intended for separation of audio signals and non-audio signals, but it can also be applied to separation of other signals. For example, separation of a music signal and a non-music signal can be realized by a similar method.

When the likelihood calculation unit 102 calculates the likelihood using the spectrum feature, the spectrum feature extracted by the spectrum feature extraction unit 104 may be used. In this case, the likelihood calculation unit 1
02, the spectrum feature extracted by the spectrum feature extraction unit 104 can be used as an acoustic feature.

In steps S507 and S508 of FIG. 7, a base matrix of a non-speech signal is created every time it is determined as non-speech in step S506, but the basis of the non-speech signal is reached by the previous section of the target section. If the matrix has already been created, steps S507 and S5
It is also possible to estimate the spectrogram of the speech signal by using the basis matrix of the non-speech signal already created by skipping 08.

In addition, the acoustic signal processing apparatus according to the present embodiment creates a base matrix of a non-speech signal from the spectrogram of a section determined as non-speech by the speech / non-speech discrimination unit 103.
A base matrix of a non-speech signal can also be created from a spectrogram of a 00 ms interval.
Even if the speech signal is included in this section, the second basis matrix creation unit 106 excludes the speech signal component, so that the accuracy of the created base matrix of the non-speech signal can be maintained.

Note that some or all of the functions in the present embodiment described above can be realized by software processing.

Although several embodiments of the present invention have been described, these embodiments are presented by way of example and are not intended to limit the scope of the invention. These novel embodiments can be implemented in various other forms, and various omissions, replacements, and changes can be made without departing from the scope of the invention. These embodiments and modifications thereof are included in the scope and gist of the invention, and are included in the invention described in the claims and the equivalents thereof.

DESCRIPTION OF SYMBOLS 101 Symphony acquisition part 102 Likelihood calculation part 103 Speech / non-speech discrimination part 104 Spectrum feature extraction part 105 First basis matrix creation part 106 Second basis matrix creation part 107 Spectrum feature estimation part 108 Inverse conversion part 201 Control part 202 Storage part 203 External storage unit 204 Operation unit 205 Communication unit 206 Microphone 208 Bus 301 Non-negative matrix factorization 302 Speech likelihood 303 Element 401 indicating maximum value Non-negative matrix factorization 402 Speech signal spectrogram 701 Acoustic signal spectrogram 702 to 704 Estimated Spectrogram 705 of speech signal Section 706 determined as non-speech signal 706 Section with low speech likelihood

Claims (14)

An acoustic feature representing the characteristics of the speech signal and the non-speech signal is extracted from an acoustic signal including the speech signal and the non-speech signal, and the likelihood that the speech signal is included in the acoustic signal is extracted using the acoustic feature. Likelihood calculating means for calculating the likelihood to represent;
Spectral feature extraction means for extracting spectral features by frequency analysis of the acoustic signal;
First basis matrix creating means for creating a first basis matrix representing the feature of the non-speech signal using the spectral feature;
A second basis matrix creating means for identifying a component highly relevant to the speech signal in the first basis matrix using the likelihood, and creating a second basis matrix by excluding this component; ,
An acoustic signal processing apparatus. Spectral feature estimation means for estimating a spectral feature of a speech signal or a spectral feature of a non-speech signal using the second basis matrix from a spectral feature of the acoustic signal or another acoustic signal;
The acoustic signal processing device according to claim 1, further comprising: The spectral feature estimation means performs non-negative matrix factorization using the second basis matrix,
3. A third basis matrix and a first coefficient matrix representing the characteristics of the speech signal are created, and a spectral feature of the speech signal is estimated by a product of the third basis matrix and the first coefficient matrix.
The acoustic signal processing device described. 4. The acoustic according to claim 1, wherein the second basis matrix creating unit creates the second basis matrix by excluding a column vector having high relevance to the speech signal from the first basis matrix. 5. Signal processing device. 4. The second basis matrix creation unit creates the second basis matrix by replacing a value of a column vector having high relevance with the speech signal from the first basis matrix with 0. The acoustic signal processing device described. 6. The method according to claim 1, wherein the second basis matrix creation means identifies a component having high relevance to the speech signal in the first basis matrix by comparing the likelihood with a predetermined threshold value. The acoustic signal processing device according to claim 1. An acoustic feature representing the characteristics of the voice signal and the non-speech signal is extracted from the acoustic signal, and the acoustic feature is used to determine whether the acoustic signal is a speech signal or a non-speech signal. A voice discrimination means;
The first basis matrix that represents the feature of the non-speech signal by using the spectral feature of the acoustic signal that has been determined by the speech / non-speech discrimination unit to be a non-speech signal. The acoustic signal processing device according to any one of claims 1 to 6, wherein the sound signal processing device is created. The spectral feature estimation means performs non-negative matrix factorization using the second basis matrix,
A second coefficient matrix representing a feature of a non-speech signal is created, and a spectral feature of the non-speech signal included in the acoustic signal is estimated by a product of the second base matrix and the second coefficient matrix. The acoustic signal processing device according to any one of claims 2 to 7. The acoustic signal processing device according to any one of claims 2 to 8, further comprising an inverse conversion unit that converts the spectral feature estimated by the spectral feature estimation unit into a time signal. An acoustic feature representing the characteristics of the first acoustic signal and the second acoustic signal is extracted from a first acoustic signal and a third acoustic signal including a second acoustic signal different from the first acoustic signal. And
Likelihood calculation means for calculating likelihood representing the likelihood that the first acoustic signal is included in the third acoustic signal using this acoustic feature;
Spectral feature extraction means for extracting a spectral feature by frequency analysis of the third acoustic signal;
First basis matrix creating means for creating a first basis matrix representing the characteristics of the second acoustic signal using the spectral feature;
Using the likelihood, a component having a high relevance to the first acoustic signal in the first basis matrix is specified, and a second basis matrix is generated by excluding this component and generating a second basis matrix Creating means;
An acoustic signal processing apparatus. An acoustic feature representing the characteristics of the speech signal and the non-speech signal is extracted from an acoustic signal including the speech signal and the non-speech signal, and the likelihood that the speech signal is included in the acoustic signal is extracted using the acoustic feature. Calculating a likelihood to represent;
Analyzing the frequency of the acoustic signal to extract spectral features;
Using the spectral features to create a first basis matrix representing the features of the non-speech signal;
Identifying a component highly relevant to the speech signal in the first basis matrix using the likelihood, and excluding this component to create a second basis matrix;
An acoustic signal processing method comprising: An acoustic feature representing the characteristics of the first acoustic signal and the second acoustic signal is extracted from a first acoustic signal and a third acoustic signal including a second acoustic signal different from the first acoustic signal. And
Using the acoustic feature to calculate a likelihood representing the likelihood that the third acoustic signal includes the first acoustic signal;
Analyzing the frequency of the third acoustic signal to extract spectral features;
Using the spectral features to create a first basis matrix representing the features of the second acoustic signal;
Identifying a component highly relevant to the first acoustic signal in the first basis matrix using the likelihood, and excluding this component to create a second basis matrix;
An acoustic signal method comprising: In the acoustic signal processing device,
An acoustic feature representing the characteristics of the speech signal and the non-speech signal is extracted from an acoustic signal including the speech signal and the non-speech signal, and the likelihood that the speech signal is included in the acoustic signal is extracted using the acoustic feature. Calculating a likelihood to represent;
Analyzing the frequency of the acoustic signal to extract spectral features;
Using the spectral features to create a first basis matrix representing the features of the non-speech signal;
Identifying a component highly relevant to the speech signal in the first basis matrix using the likelihood, and excluding this component to create a second basis matrix;
An acoustic signal processing program for realizing In the acoustic signal processing device,
An acoustic feature representing the characteristics of the first acoustic signal and the second acoustic signal is extracted from a first acoustic signal and a third acoustic signal including a second acoustic signal different from the first acoustic signal. And
Using the acoustic feature to calculate a likelihood representing the likelihood that the third acoustic signal includes the first acoustic signal;
Analyzing the frequency of the third acoustic signal to extract spectral features;
Using the spectral features to create a first basis matrix representing the features of the second acoustic signal;
Identifying a component highly relevant to the first acoustic signal in the first basis matrix using the likelihood, and excluding this component to create a second basis matrix;
An acoustic signal processing program for realizing