JP2011133780A - Signal analyzing device, signal analyzing method and signal analyzing program - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To provide a signal analyzing device that achieves hierarchical factorization. <P>SOLUTION: The signal analyzing device includes: a means which determines a data matrix Y for storing time-frequency components of acoustic signal data by performing time-frequency analysis on read acoustic signal data; a means which sets initial values of a spectrum base parameter, a power envelope base parameter and a power envelope base activity value to be determined when the data matrix Y is approximated by the product of a base matrix H and a coefficient matrix U using a non-negative matrix factorization; a means which calculates a spectrogram model value obtained by expressing as a convolution mixture format the spectrum base activity time series corresponding to each row of the coefficient matrix U, using the parameter with initial value; and a means which continues updating the value till convergence of the parameter to be determined, and outputs the parameter value to be determined at the time of convergence. <P>COPYRIGHT: (C)2011,JPO&INPIT

Description

  The present invention relates to a signal analysis apparatus, a signal analysis method, and a signal analysis program for analyzing an acoustic signal in order to decompose the acoustic signal into signals for each note (note in a musical score).

It is not easy to separate and extract individual acoustic signals from a mixed signal in which a plurality of acoustic signals are superimposed. Such a problem is called sound source separation. In particular, signal separation for monaural signals is a typical defect setting problem and is difficult to solve without making any assumptions. Many approaches have been studied for monaural signal separation so far, but the approach that has been highlighted as an effective approach in recent years is the application of the principle of non-negative matrix factorization (NMF). (For example, refer to Patent Document 1). In this approach, a non-negative data matrix Y in which spectra (magnitudes of frequency components) at each time of an observation signal are arranged as a column vector is expressed as a product of a non-negative base matrix H and a non-negative coefficient matrix U. Approximate.

  As a result, the spectral basis functions constituting the entire spectrum to be observed are stored in each column of the basis matrix H, and the spectral basis representing how large the specific spectral basis function is activated at each time. A time series of activity values is stored in one row of the coefficient matrix U. As described above, a decomposition expression of the signal can be obtained. This method is characterized by the fact that it uses the assumption that the observed signal consists only of sounds with a limited type of spectrum for the sound source separation problem. This is an effective solution for signals that

The decomposition expression of the spectrogram by non-negative matrix factorization is the spectrogram
As
Like
It is obtained by deciding. Here, ω and t are indexes corresponding to the frequency and time, respectively.

_{y t: = (Y 1,} , ···, Y Ω, t) T, h i: = (H 1, i, ···, H Ω, i) When ^T Equation (2)
As can be seen from the above, the observation data y _t at all t is to be regarded as being composed of at most I types of “parts” h ₁ ,..., H _I. It is understood that this is a matter of determining how to place each part most appropriately. What each matrix obtained in this way represents is easier to understand when looking at FIG. In each column vector of H, a spectrum repeatedly appearing in the music is regarded as a typical component part and is expressed. Therefore, if the music spectrogram is composed only of a limited pattern spectrum determined by the type of musical instrument and the scale, each column vector of H is a spectrum that roughly corresponds to a specific musical scale of the specific musical instrument. On the other hand, each row vector of U represents how much intensity each spectrum part is activated at what time.

P. Smaragdis and J. C. Brown, “Non-negative matrix factorization for music transcription,” in Proceedings of the IEEE Workshop on Applications of Signal Processing to Audio and Acoustics (WASPAA), 2003, pp. 177.180.

  As mentioned above, non-negative matrix factorization approximates a data matrix that stores the observed spectrum by the product of the base matrix that stores the spectrum base in each column and the coefficient matrix that stores the time series of spectrum base activities in each row. Thus, a spectrum basis function is automatically acquired from a set of observed spectra, and the observed spectrum is decomposed for each spectrum basis function.

  However, in order to decompose a music signal into signals for each note (notes in a musical score), it is necessary to decompose the spectrum base activity time series itself into a spectrum base activity time series corresponding to each note event. Obviously, the conventional non-negative matrix factorization has a problem that there is no function for acquiring such a hierarchical decomposition expression.

  The present invention has been made in view of such circumstances, and an object thereof is to provide a signal analysis device, a signal analysis method, and a signal analysis program capable of obtaining the hierarchical decomposition expression as described above.

  The present invention provides a signal data storage means storing acoustic signal data, and a data matrix Y storing time frequency components of the acoustic signal data by time frequency analysis on the acoustic signal data read from the signal data storage means. Spectral basis parameters, power envelope basis parameters, and power to be obtained when the data matrix Y is approximated by a product of a basis matrix H and a coefficient matrix U using a time frequency analysis means to be obtained and a non-negative matrix factorization method. Corresponding to each row of the coefficient matrix U by using an initial value setting means for setting initial values of each envelope base activity value, and a spectrum base parameter, a power envelope base parameter, and a power envelope base activity value in which the initial values are set Spectral basis activity Model calculation means for calculating a spectrogram model value representing a time series in the form of convolution mixture, the spectrogram model value, the data matrix, the spectrum basis parameter, the power envelope basis parameter, and the power envelope basis Update means for updating the spectrum base parameter, the power envelope base parameter, and the power envelope base activity value for which the initial value is set using the values of the activity values, the spectrum base parameter, and the power envelope The updating of the value is continued by the updating means until the basis parameter and the power envelope basis activity value converge, and at the time of convergence, the spectrum basis parameter, the power envelope Bottom parameter, characterized in that an output means for outputting a value of the power envelope basal activity value and the spectrogram model.

  The present invention is a signal analysis method for causing a computer of a signal analysis apparatus having signal data storage means in which acoustic signal data is stored to perform signal analysis processing, wherein the acoustic signal data read from the signal data storage means A time-frequency analysis step for obtaining a data matrix Y storing a time-frequency component of the acoustic signal data by a time-frequency analysis on the sound signal, and using a non-negative matrix factorization technique, An initial value setting step for setting initial values of spectrum base parameters, power envelope base parameters, and power envelope base activity values to be obtained when approximated by the product of U, and the spectral base parameters and powers for which the initial values are set Envelope basis parameters and power envelope A model calculation step of calculating a spectrogram model value representing a spectrum base activity time series corresponding to each row of the coefficient matrix U in a convolution mixed format using a bottom activity value, the spectrogram model value, and the data A matrix, the spectrum basis parameter, the spectrum basis parameter in which the initial value is set using the values of the power envelope basis parameter and the power envelope basis activity value, the power envelope basis parameter, and the power envelope. An updating step for updating a base activity value, and the updating until the spectrum base parameter, the power envelope base parameter, and the power envelope base activity value converge. Continues to update the value by step, at the time of convergence, the spectral basis parameter, the power envelope basal parameters, and having an output step of outputting a value of the power envelope basal activity value and the spectrogram model.

  The present invention is a signal analysis program for causing a computer of a signal analysis apparatus having signal data storage means in which acoustic signal data is stored to perform signal analysis processing, the acoustic signal data read from the signal data storage means A time-frequency analysis step for obtaining a data matrix Y storing a time-frequency component of the acoustic signal data by a time-frequency analysis on the sound signal, and using a non-negative matrix factorization technique, An initial value setting step for setting initial values of spectrum base parameters, power envelope base parameters, and power envelope base activity values to be obtained when approximated by the product of U, and the spectral base parameters and powers for which the initial values are set Envelope basis parameters and power envelope A model calculation step of calculating a spectrogram model value expressing a spectrum base activity time series corresponding to each row of the coefficient matrix U in the form of a convolution mixture using a loop basis activity value; a value of the spectrogram model; Using the data matrix, the spectrum basis parameter, the power envelope basis parameter and the power envelope basis activity value, the spectrum basis parameter in which the initial value is set, the power envelope basis parameter, An update step of updating a power envelope basis activity value, and before the spectral basis parameter, the power envelope basis parameter and the power envelope basis activity value converge Updating the value by the updating step, and causing the computer to perform an output step of outputting the spectrum basis parameter, the power envelope basis parameter, the power envelope basis activity value, and the spectrogram model value at the time of convergence. It is characterized by.

  According to the present invention, detection of a specific sound from an acoustic signal in which a plurality of sounds are mixed, extraction of a specific sound from an acoustic signal in which a plurality of sounds are mixed, and acoustic in which a plurality of sounds are mixed An effect is obtained that a parameter of a signal analysis result can be used for processing a specific sound from a signal.

It is a block diagram which shows the structure of one Embodiment of this invention. It is explanatory drawing which shows the decomposition | disassembly expression of the music spectrumgram by nonnegative matrix factorization (NMF).

  Hereinafter, a signal analyzer according to an embodiment of the present invention will be described with reference to the drawings. First, the basic principle of the signal analyzer according to the present invention will be described. When a data matrix is acquired from a music signal, the spectrum base activity time series stored in each row of the coefficient matrix obtained by non-negative matrix factorization is often expressed by a combination of limited types of local patterns. This is because, in music, there is a unit related to the sound duration of a note called note value, and notes with the same note value are often similar in how the sound rises and decays.

  In the present invention, in addition to the conventional function of non-negative matrix factorization that decomposes the observed spectrum for each spectral basis function while automatically acquiring the spectral basis function from the set of observed spectra, using the above-mentioned music properties, A function for decomposing the spectrum basis activity time series for each power envelope basis is realized while automatically acquiring basis functions of local patterns (hereinafter referred to as power envelope basis) mixed in the basis activity time series.

More specifically, the spectrum base activity time series corresponding to each row of the coefficient matrix U is expressed in the form of convolutional mixing. That is, if the elements in the i-th row of U are U _{i, 0} , U _{i, 1} ,.
It expresses. G _{j, 0} , G _{j, 1} ,... Is the jth power envelope base, and O _{i, j, 0} , O _{i, j, 1 ,. , J, t} are called power envelope base activities. By the way, convolutional mixing is widely recognized as a model of a mixing process of signal sources when signals coming from a plurality of signal sources are observed with a plurality of microphones. The spectral basis activity time series of the eye corresponds to the observed signal at the i-th microphone, G _{j, 1} , G _{j, 2} ,... Correspond to the signal of the j-th signal source, and O _{i, j, 1} , O _{i, j, 2} ... Correspond to impulse responses from the j-th signal source to the i-th microphone.

Next, the spectrogram model will be described. Looking at each row vector of U shown in FIG. 2, the time envelope of the spectral basis Activation h _i is notice that consists only of a limited type of local patterns. This is because each note is played with a limited number of note lengths, so the types of rise and decay patterns are also limited. So for U,
Consider a decomposition expression expressed in the form of convolutional mixing. G _{j, τ} represents the local pattern of the jth time envelope. On the other hand, O _{i, j, t} represents the activation and ideally has an image that peaks at the onset time of each note. Substituting equation (5) into the right side of equation (2),
Thus, a spectrogram model obtained by extending the NMF type model given by the equation (2) is established. Here, for the purpose of removing the arbitrary scale of decomposition,
Is assumed.

Next, an optimization algorithm will be described. We want to find H, G, and O from Y under the proposed model, as H and U were obtained from Y as shown in FIG. 2 by non-negative matrix factorization. In the following, an optimization algorithm for obtaining a hierarchical sparse representation of music according to the proposed model will be described. First, the square error criterion will be described. Here, the following optimization problem is set under the observation spectrogram Y: = ( _{Yω, t} ) _{Ω × T.}
To consider. S (G, O) is a regularization term that leads G and O to a sparse solution, where
It is defined as However, 0 <p _g ≦ 2 and 0 <p _o ≦ 2.

First, an update formula for H that lowers F is derived. If the updated values of H, G and O one step before are H ′, G ′ and O ′, respectively,
It becomes. However,
It is. Although detailed explanation is omitted, _H that minimizes F _H (H, G ′, O ′) is analytically determined.
It can be guaranteed that F (H, G ′, O ′) does not increase by updating in this way. Further, if both H ′ and U ′ are non-negative values, H is always a non-negative value.

Next, an update formula for G that lowers F is derived. As before,
It becomes. However,
It is. The second term of F _G (H ′, G, O ′) is
(The right side is apparent because it is a parabola in contact with | x | ^p at the contact ± x ′). From the above, the update formula of G is obtained using F _G (H ′, G, O ′).
I can lead.

Finally, an update formula for O that lowers F is derived. As before,
But using this, the update formula for O
I can lead.

Next, the I divergence criterion will be described. Optimization problem when modeling error is measured by I-divergence
Also consider.

First, for the update formula of H,
Derived from the inequality of
Find H that minimizes
It is obtained analytically as follows. Similarly, for the G update formula,
For the inequality of, the update formula of O,
Derived from the inequality of
Find G and O to minimize
It is obtained analytically as follows. The second term of the equation (21) and the third term of the equation (22) are expressed as ∀ _{x> 0} , | x | ^p ≦ p | x ′ | ^p−1 (xx ′) + | x ′ | ^p (0 <P ≦ 2) (25) (the right side is obvious because it is a tangent line of | x | ^{p at} the contact point x ′).

  Next, a signal analyzer using the basic principle described above will be described. FIG. 1 is a block diagram showing the configuration of the signal analysis apparatus according to the first embodiment. The signal analyzer is configured by a computer device. In this figure, reference numeral 1 denotes a signal data storage unit that receives and stores acoustic signal data obtained by sampling and quantizing an acoustic signal. Reference numeral 2 denotes a time-frequency analysis unit that performs time-frequency analysis. Reference numeral 3 denotes an initial setting unit that performs initial setting of values. Reference numeral 4 denotes a spectrogram model calculation unit that calculates a spectrogram model. Reference numeral 5 denotes a spectrum base update unit that updates the spectrum base. Reference numeral 6 denotes a power envelope base update unit that updates the power envelope base. Reference numeral 7 denotes a power envelope base activity update unit that updates the power envelope base activity. Reference numeral 8 denotes a parameter normalization unit that normalizes parameters. Reference numeral 9 denotes a convergence determination unit that determines whether the process has converged. Reference numeral 10 denotes a parameter output unit that outputs parameters. Reference numeral 11 denotes a parameter storage unit that stores the output parameters.

Next, the operation of the signal analyzer shown in FIG. 1 will be described with reference to FIG. First, the time-frequency analysis unit 2 reads the signal data to be analyzed stored in the signal data storage unit 1 and uses a short-time Fourier transform (STFT), wavelet transform, or the like to obtain the time frequency. The time frequency component {Yω, t} given by the non-negative value is analyzed and _{0 ≦ Ω ≦ Ω−1 and 0 ≦ t ≦ T−1} are calculated. However, ω = 0,..., Ω-1, t = 0,..., T-1 are indices corresponding to the frequency and time, respectively. The time frequency analysis unit 2 outputs a matrix Y = (Y _{ω, t} ) _{Ω × T in} which the time frequency components Y _{ω, t} are stored.

Next, the initial setting unit 3 determines the spectrum basis number I, the power envelope basis number J, and the regularization parameters λ _g , λ _o , p _g , p _o . Then, the initial setting unit 3 performs non-negative matrix factorization (NMF) on Y,
Become
Is output. Using this, the initial values of the spectrum basis parameter H _{ω, i} , power envelope basis parameter G _{j, t} , and power envelope basis activity value O _{i, j, t} are respectively determined.
Output as. Here, i = 1,..., I is a spectrum basis index, and j = 1,..., J is a power envelope basis index. [•] _{a, b} represent the components of a row and b column of the matrix.

Next, the spectrogram model calculation unit 4 calculates and outputs the spectrogram model X _{ω, t} by the following procedure using H _{ω, i} , G _{j, t} , O _{i, j, t} obtained in the previous stage. . First, the spectrogram model calculation unit 4 convolves the spectrum base activity value U _{i, t} using G _{j, t} and O _{i, j, t.}
Calculated by This convolution mixing calculation is performed at high speed using a fast Fourier transform.

Next, the spectrogram model calculation unit 4 uses X _{ω, t} to calculate the sum of products using H _{ω, i} and U _{i, t} previously obtained.
Calculated by

Next, the spectrum base update unit 5 uses H _{ω, t} and X _{ω, t} and U _i obtained in the previous stage _{, t} and H _{ω, i} to calculate H _{ω, i} .
To update and output.

Then, the power envelope basal updating unit 6, Y _omega, X _omega obtained by _t and _{front, t} and H _{omega, i} and _{O i, j,} using a _{t, G j,} by the following procedure _t calculate. First, the power envelope base update unit 6 includes the i-th spectrum base H _{0, i,...} , H _{Ω-1, i} and the observed spectrum Y _{0, t ,. with t}
The
Calculated by Similarly, the power envelope base update unit 6 includes H _{0, i,...} , H _{Ω-1, i} and X _{0, t ,.}
The
Calculated by

Next, the power envelope base update unit 6
The
Calculated by This cross-correlation calculation is performed at high speed by using a fast Fourier transform. Similarly, the power envelope base update unit 6
The
Calculated by This cross-correlation operation is also calculated at high speed by using fast Fourier transform.

Finally, the power envelope base update unit 6
And using G _{j, t} obtained in the previous stage, G _{j, t}
Update with The _{λ g p g | G j,} τ | pg-1 is a term related to sparse regularization term, G _j, which other as long as it has the effect of inducing the elements of _t to the sparse It may be replaced by a shape.

Next, the power envelope basal activity update unit 7, Y _omega, X _omega obtained by _t and the preceding _stage, using the _t and H _{omega, i} and _{O i, j, t, G j,} the following procedure _t Is calculated and output. First, the power envelope basal activity update unit 7, i-th spectral basis _{H 0, i, ···, H} Ω-1, observed in the _i and time t spectrum _{Y 0, t, ···, Y} Ω-1 _{, T}
The
Calculated by Similarly, the power envelope base activity update unit 7 includes H _{0, i,...} , H _{Ω-1, i} and X _{0, t ,.}
The
Calculated by

Next, the power envelope base activity update unit 7
The
Calculated by This cross-correlation calculation is performed at high speed by using a fast Fourier transform. Similarly, the power envelope base activity update unit 7
The
Calculated by This cross-correlation operation is also calculated at high speed by using fast Fourier transform.

Finally, the power envelope base activity update unit 7
The
Update with This + λ _o p _o | O ′ _{i, j, τ} | ^po−1 is a term related to the sparse regularization term, and has the effect of inducing the elements of O _{i, j, t} to be sparse. If there is any other form, it may be replaced.

Next, the parameter normalization unit 8 normalizes and outputs H _{ω, i} and G _{j, t} obtained in the previous stage. For example, when standardizing both to be 1,
To update H _{ω, i} and G _{j, t} respectively.

Next, the convergence determination unit 9 determines whether or not the previous iteration has satisfied a predetermined number of times, whether or not the parameter update change rate in the iterative calculation is less than or equal to a predetermined value, or the objective function value It is determined whether or not the rate of change has become a predetermined value or less. For example, the objective function is
Calculate according to Where S (G, O) is a regularization term that guides G, O to a sparse solution,
It is defined as If the iterative calculation has not converged, the convergence determination unit 9 outputs an instruction to recalculate the spectrogram model to the spectrogram model calculation unit 4 and receives the spectrogram model calculation unit 4 to update the spectrum base. The unit 5, the power envelope base update unit 6, the power envelope base activity update unit 7 and the parameter normalization unit 8 repeat the processing operations described above until the iterative calculation converges.

Next, the parameter output unit 9 stores parameters such as H _{ω, i} , G _{j, t} , O _{i, j, t} , X _{ω, t,} etc., in which the iterative calculation has converged, in the parameter storage unit 11. .

Next, a signal analysis apparatus according to the second embodiment will be described. First, the processing operation of the spectrum base update unit 5 in the second embodiment will be described. Spectral basis update unit 5, Y _omega, X obtained in the _t and preceding _omega, with _t and H _{omega, i} and _{U i, t,} H _ω, updates the _i by the following procedure. First, the spectrogram ratio R _{ω, t} between the observed spectrogram Y _{ω, t} and the spectrogram model X _{ω, t} is _expressed as
Calculated by

Next, the spectrum base update unit 5 uses the previously obtained R _{ω, t} and H _{ω, i} and U _{i, t} ,
Calculated by

Next, the processing operation of the power envelope base update unit 6 in the second embodiment will be described. Power envelope base updating unit _6, Y ω, X obtained in the _t and preceding _{omega, t} and H _{omega, i} and _{G j, t} and _{O i, j,} using a _{t, G j} according to the following _{procedure, t} Update. First, the power envelope base update unit 6 determines the spectrogram ratio R _{ω, t} between the observed spectrogram Y _{ω, t} and the spectrogram model X _{ω, t.}
Calculated by Next, the power envelope base update unit 6 calculates the relationship between H _{0, i,...} , H _{Ω-1, i} and R _{0, t ,.}
The
Calculate according to Next, the power envelope base update unit 6
Calculate according to This cross-correlation calculation is performed at high speed by using a fast Fourier transform.

Finally, the power envelope base update unit 6
To update G _{j, t} . The _{2λ g p g | G j,} τ | pg-1 is a term related to sparse regularization term, G _j, which other as long as it has the effect of inducing the elements of _t to the sparse It may be replaced by a shape.

Next, the processing operation of the power envelope base activity update unit 7 in the second embodiment will be described. Power envelope basal activity update unit 7, Y _omega, X obtained in the _t and preceding _{omega, t} and H _{omega, i} and _{G j, t} and _{O i, j,} using a _{t, O i} by the following steps _{, J, t} are updated. First, the power envelope base activity update unit 7 calculates a spectrogram ratio R _{ω, t} between the observed spectrogram Y _{ω, t} and the spectrogram model X _{ω, t.}
Calculated by Next, the power envelope base activity update unit 7 calculates the relationship between H _{0, i,...} , H _{Ω-1, i} and R _{0, t ,.}
The
Calculate according to Next, the power envelope base activity update unit 7
The
Calculate according to This cross-correlation calculation is performed at high speed by using a fast Fourier transform.

Finally, the power envelope base activity update unit 7
To update O _{i, j, t} . This 2λ _o p _o | O _{i, j, τ} | ^po−1 is a term related to the sparse regularization term, and has the effect of inducing the elements of O _{i, j, t} to be sparse. For example, it may be replaced with other forms.

Next, the processing operation of the convergence determination unit 9 in the second embodiment will be described. The convergence determination unit 9 determines whether or not the iterative calculation has satisfied a predetermined number of times, or whether or not the parameter update change rate has become a predetermined value or less in the iterative calculation, or whether the change rate of the objective function value is a predetermined value. It is determined whether or not the following has occurred. For example, the objective function is
Calculate according to Where S (G, O) is a regularization term that guides G and O to a sparse solution.
It is defined as

  As described above, in order to decompose the acoustic signal into signals for each note, information on the sound rise and attenuation patterns is further obtained for the decomposition element (matrix U) used in the conventional non-negative matrix decomposition (NMF). Detection of a specific sound from an acoustic signal in which a plurality of sounds are mixed by estimating each parameter of the model using a new model that introduces decomposition (expression (4) and expression (5)) that can be expressed, It can be used for extraction of a specific sound from an acoustic signal in which a plurality of sounds are mixed, processing of a specific sound from an acoustic signal in which a plurality of sounds are mixed, and the like.

  1, the time frequency analysis unit 2, the initial setting unit 3, the spectrogram model calculation unit 4, the spectrum base update unit 5, the power envelope base update unit 6, the power envelope base activity update unit 7, the parameter normalization unit 8, A program for realizing the functions of the convergence determination unit 9 and the parameter output unit 10 is recorded on a computer-readable recording medium, and the program recorded on the recording medium is read into a computer system and executed to perform signal analysis. Processing may be performed. Here, the “computer system” includes an OS and hardware such as peripheral devices. The “computer system” includes a WWW system having a homepage providing environment (or display environment). The “computer-readable recording medium” refers to a portable medium such as a flexible disk, a magneto-optical disk, a ROM, and a CD-ROM, and a storage device such as a hard disk built in the computer system. Further, the “computer-readable recording medium” refers to a volatile memory (RAM) in a computer system that becomes a server or a client when a program is transmitted via a network such as the Internet or a communication line such as a telephone line. In addition, those holding programs for a certain period of time are also included.

  The program may be transmitted from a computer system storing the program in a storage device or the like to another computer system via a transmission medium or by a transmission wave in the transmission medium. Here, the “transmission medium” for transmitting the program refers to a medium having a function of transmitting information, such as a network (communication network) such as the Internet or a communication line (communication line) such as a telephone line. The program may be for realizing a part of the functions described above. Furthermore, what can implement | achieve the function mentioned above in combination with the program already recorded on the computer system, what is called a difference file (difference program) may be sufficient.

  Detection of a specific sound from an acoustic signal in which multiple sounds are mixed, extraction of a specific sound from an acoustic signal in which multiple sounds are mixed, and detection of a specific sound from an acoustic signal in which multiple sounds are mixed It can be used in applications where it is essential to perform processing.

  DESCRIPTION OF SYMBOLS 1 ... Signal data storage part, 2 ... Time frequency analysis part, 3 ... Initial setting part, 4 ... Spectrogram model calculation part, 5 ... Spectral base update part, 6 ... Power envelope Base update unit, 7 ... Power envelope base activity update unit, 8 ... Parameter normalization unit, 9 ... Convergence determination unit, 10 ... Parameter output unit, 11 ... Parameter storage unit

Claims (3)

Signal data storage means storing acoustic signal data;
Time frequency analysis means for obtaining a data matrix Y storing time frequency components of the acoustic signal data by time frequency analysis for the acoustic signal data read from the signal data storage means;
When the data matrix Y is approximated by a product of a base matrix H and a coefficient matrix U using a non-negative matrix factorization method, initial values of spectrum base parameters, power envelope base parameters, and power envelope base activity values to be obtained are obtained. An initial value setting means for setting a value;
A spectrogram model representing a spectrum basis activity time series corresponding to each row of the coefficient matrix U in the form of a convolution mixture using the spectrum basis parameter, the power envelope basis parameter and the power envelope basis activity value to which the initial values are set. Model calculation means for calculating a value;
Using the spectrogram model values, the data matrix Y, the spectrum basis parameters, the power envelope basis parameters, and the power envelope basis activity values, the spectrum basis parameters for which the initial values are set, Updating means for updating the power envelope basis parameter and the power envelope basis activity value;
The updating means continues updating the values until the spectrum basis parameters, the power envelope basis parameters, and the power envelope basis activity values converge, and at the time of convergence, the spectrum basis parameters, the power envelope basis parameters, the power envelope An output means for outputting a base activity value and a value of the spectrogram model. A signal analysis method for causing a computer of a signal analysis apparatus provided with signal data storage means in which acoustic signal data is stored to perform signal analysis processing,
A time frequency analysis step for obtaining a data matrix Y storing a time frequency component of the acoustic signal data by time frequency analysis with respect to the acoustic signal data read from the signal data storage means;
When the data matrix Y is approximated by a product of a base matrix H and a coefficient matrix U using a non-negative matrix factorization method, initial values of spectrum base parameters, power envelope base parameters, and power envelope base activity values to be obtained are obtained. An initial value setting step for setting a value;
A spectrogram model representing a spectrum basis activity time series corresponding to each row of the coefficient matrix U in the form of a convolution mixture using the spectrum basis parameter, the power envelope basis parameter and the power envelope basis activity value to which the initial values are set. A model calculation step for calculating a value;
Using the spectrogram model values, the data matrix Y, the spectrum basis parameters, the power envelope basis parameters, and the power envelope basis activity values, the spectrum basis parameters for which the initial values are set, An updating step for updating the power envelope basis parameter and the power envelope basis activity value;
The update step continues to update the spectrum basis parameter, the power envelope basis parameter, and the power envelope basis activity value until convergence, and at the time of convergence, the spectrum basis parameter, the power envelope basis parameter, the power envelope. An output step of outputting a base activity value and a value of the spectrogram model. A signal analysis program for causing a computer of a signal analyzer having signal data storage means in which acoustic signal data is stored to perform signal analysis processing,
A time frequency analysis step for obtaining a data matrix Y storing a time frequency component of the acoustic signal data by time frequency analysis with respect to the acoustic signal data read from the signal data storage means;
When the data matrix Y is approximated by a product of a base matrix H and a coefficient matrix U using a non-negative matrix factorization method, initial values of spectrum base parameters, power envelope base parameters, and power envelope base activity values to be obtained are obtained. An initial value setting step for setting a value;
A spectrogram model representing a spectrum basis activity time series corresponding to each row of the coefficient matrix U in the form of a convolution mixture using the spectrum basis parameter, the power envelope basis parameter and the power envelope basis activity value to which the initial values are set. A model calculation step for calculating a value;
Using the spectrogram model values, the data matrix Y, the spectrum basis parameters, the power envelope basis parameters, and the power envelope basis activity values, the spectrum basis parameters for which the initial values are set, An updating step for updating the power envelope basis parameter and the power envelope basis activity value;
The update step continues to update the spectrum basis parameter, the power envelope basis parameter, and the power envelope basis activity value until convergence, and at the time of convergence, the spectrum basis parameter, the power envelope basis parameter, the power envelope. An output step for outputting a base activity value and a spectrogram model value to the computer.