JP6498141B2 - Acoustic signal analyzing apparatus, method, and program - Google Patents

Description

  The present invention relates to an acoustic signal analyzing apparatus, method, and program, and more particularly to an acoustic signal analyzing apparatus, method, and program for analyzing observation data of a fundamental frequency locus of an acoustic signal indicating a singing voice.

  Conventionally, two machine learning fields called non-parametric Bayes method and kernel method are known as analysis methods.

<Kernel averaging method>
When expressing the probability distribution for complex data, in principle, it can be configured by adding probabilities to all kinds of data, but it becomes more difficult as the number of possible cases of data increases. Go. Especially for infinite numbers, it is extremely difficult to express them on a computer. The kernel averaging method of Non-Patent Document 1 gives a technique to express it on a computer while maintaining the same expression ability as the original probability distribution, and can be approximated well from a small amount of finite data. Are known.

<Signal decomposition by gamma process>
If observation data is considered to be composed of potentially meaningful combinations of dictionaries, it is possible to construct an infinite dictionary in principle by constructing a dictionary model using the gamma process described in Non-Patent Document 2. It is possible to construct an analysis algorithm that behaves to learn only as few dictionaries as possible, despite its size.

A. Smola, A. Gretton, L. Song, B. Scholkopf. (2007). A Hilbert Space Embedding for Distributions. Algorithmic Learning Theory: 18th International Conference. Springer: 13-31. Matthew D. Hoffman, David M. Blei, Perry R. Cook, Bayesian Nonparametric Matrix Factorization for Recorded Music, International Conference on Machine Learning, 2011.

  Conventionally, research has been conducted to extract potential and meaningful features from very information-rich data such as singing voices, but it is possible to extract features that correspond to the individuality of each person from those features. There wasn't.

  The present invention has been made in consideration of the above circumstances, and an object thereof is to provide an acoustic signal analyzing apparatus, method, and program capable of extracting features corresponding to the individuality of each singer from a singing voice. To do.

In order to achieve the above object, the acoustic signal analysis apparatus of the present invention provides observation data of an acoustic signal indicating a singing voice when each of N singers sang at least once for each of L kinds of sheet music. An acoustic signal analyzing apparatus for analyzing, a similarity between a pair (x, s) of a fundamental frequency trajectory x representing a fundamental frequency at each time of an acoustic signal representing a singing voice and a score vector s representing a pitch at each time of the score predetermined the M kernel K _m as criteria to measure the degree to each of the M kernel K each weight a _m of _m, and the n number each L type singer n of music s _l The singer n sings the score s _l obtained from each pair (x, s) of the fundamental frequency trajectory x obtained from the observation data when sung at least once and the score vector s. Hilbert space of pairs (x, s _l) at the time The D number of the expected value mu _{n, l,} the score common for probability distributions, singer n of each pair (x, s) represented by using the dictionary q _d of the D pieces of voice singing distribution predetermined each weight b _d dictionaries q _{d, n,} and each of the weights c _l a singer common the D pieces of dictionary q _d for the score s _l, determined from _d, singer n Whereas score s _l Of the M kernels K _m so as to minimize the objective function expressed using the distance between the expected value μ _{n, l} ^* of the pair (x, _sl ) in the Hilbert space and each of the weights a _m, the weight b _d of each of the music common said for each of the singers n of the n number D number of dictionary q _d, and _n, singer common the respect to the L type of music s _l weight c _l of each of the D pieces of dictionary q _d, is configured to include a parameter estimation unit that estimates a _d.

The acoustic signal analysis method of the present invention is an acoustic signal analysis apparatus for analyzing observation data of an acoustic signal indicating a singing voice when each of N singers sings at least once on each of L types of sheet music. In the signal analysis method, the parameter estimation unit between a pair (x, s) of a fundamental frequency trajectory x representing a fundamental frequency at each time of an acoustic signal indicating a singing voice and a score vector s representing a pitch at each time of the score of the M kernel K _m that predetermined as criteria to measure the similarity, each of the weights a _m of the M kernel K _m, and each of the singers n of said n number of L type music s _l A singer n obtained from each pair (x, s) of the fundamental frequency trajectory x obtained from the observation data obtained when singing at least once for each and the score vector s is _assigned to the score _sl . pair at the time of singing (x, s _l The probability distribution of the expected value mu _{n, l} in the Hilbert space, each pair represented by using the dictionary q _d of the D pieces of voice singing distribution predetermined (x, s), score common for the singer n the D number each of the dictionary q _d weights b _{d, n,} and each of the singers common the D pieces of dictionary q _d for the score s _l weight c _l, obtained from _d, singer n musical score of pairs when singing against s _l (x, s _l) so as to minimize the objective function expressed by using the distance between the expected value mu _{n, l} on Hilbert space ^* of, the M and weight a _m each kernel K _m, the weight b _d of each of the music common for each D number of dictionary q _d of singer n of the n _number, and _n, singing for the L type of music s _l The weights c _{l, d} of each of the D dictionaries q _d common to the users are estimated.

  The program of this invention is a program for functioning a computer as each part of an acoustic signal analyzer.

As described above, the acoustic signal analyzer of the present invention, a method, and according to the program, the predetermined M-number of kernel K _m, the obtained from each of the weights a _m, and the observation data of the kernel K _m each pair (x, s) of the score vector s with the fundamental frequency trajectory x obtained from the Hilbert space on the expected value of the pair (x, s _l) when the singer n sang relative score s _l μ _{n, l} and the probability distribution of each pair (x, s) expressed using a predetermined dictionary _d of singing voice singing singing q _d , each of the common score q _d for singer n weight b _{d, n,} and the weight c _l for each singing person common dictionary q _d for the score s _l, determined from _d, the pair (x when singer n sang relative score s _l, s _l ) Minimizes the objective function expressed using the distance to the expected value μ _{n, l} ^{* in} the Hilbert space As such, the weight a _m of each kernel K _m, each of the weights b _d music common dictionary q _d for each singer _n, and _n, each singing person common dictionary q _d for the score s _l By estimating the weights cl _{, d} , it is possible to extract features corresponding to the individuality of each singer from the singing voice.

It is a block diagram which shows the functional structure of the acoustic signal analyzer of this Embodiment. It is a flowchart which shows the content of the parameter estimation process routine in the acoustic signal analyzer of this Embodiment. It is a figure which shows an experimental result.

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

<Principle of Embodiment of the Present Invention>
The fundamental frequency trajectory x of the singing voice can be expressed as a vector in which the discrete fundamental frequency values (real numbers) for each time are arranged. Here, it is assumed that the lengths in the time direction of the fundamental frequency trajectory x are all unified and V. In addition, the score s can be expressed as a vector in which the pitches of each time are arranged. A function φ that maps the fundamental frequency trajectory x of the singing voice and the score s to the reproduction kernel Hilbert space is introduced, and a pair of the fundamental frequency trajectory x and the score s is considered as one point φ (x, s). This Hilbert space uses the kernel K (・, (x, s)) as a criterion for measuring the similarity between the fundamental frequency trajectory of the singing voice and the score pair (x, s), (x´, s´). Can be characterized.

Kernel K _m ((x, s), (x´, s´)) (m = 1, 2, ..., M) shown in the following equation (1) is used as a criterion for M pieces. Think.

(1)

Here, positive real numbers (for example, λ1 = λ2 = 1) set in advance are used as λ ₁ and λ ₂ . Due to physical constraints of the human body, the generation of singing voices can be strongly constrained. Therefore, the distance criterion for measuring the similarity of singing voices can be considered to have some typical templates. Kernel K uses real weights a _{m to} Gamma (.,.) (M = 1, 2, ..., M)

(2)

It will be expressed as

Consider a case where N singers are singing L types of music. It is assumed that the fundamental voice frequency when the nth singer sings the lth sheet music follows the probability distribution _{Pn, l} . This probability distribution is considered using the expected value of the previous reproduction kernel Hilbert space. Expected value μ _{n, l} in Hilbert space of a pair (x, s) of fundamental frequency trajectory and sheet music according to probability distribution P _{n, l} is obtained by using real numbers b _{n, l, d}

(3)

Can be expressed as

Let us consider expressing this singing voice singing probability distribution by an easy-to-handle model. It is assumed that _d q singing voice distribution dictionaries q ₁ , q ₂ ,..., Q _D are given in advance. The μ _{n, l} equivalent to the singing probability distribution for the l-th score of the n-th singer is the weight b _{d, n} (> 0) of the common dictionary q _d for the n-th singer and the l-th song _Can be approximated by the model μ _{n, l} ^* using the weight c _{l, d} (> 0) of the common dictionary q _d for

(4)

n-th of the singer is the l-th score s _l actually sang o _l pieces of observation data (fundamental frequency trajectory) with respect to x _1, x _2, ..., and the x _o μ _{n, l} is It can be calculated as follows:

(5)

Next, μ _{n, l} ^* can be approximated as follows.

(6)

However, Q _d (d = 1, 2,..., D) may be any parametric probability distribution prepared in advance. Each of a _m , b _{d, n} , c _{l, d} can have a gamma process prior distribution, for example, most simply

(7)

(8)

(9)

Can be set. Model fitting can be performed by approximating μ _{n, l} by μ _{n, l} ^* , and the objective function W is

(10)

A _m , b _{d, n} , c _{l, d} (m = 1, 2, ..., M, d = 1, 2, ..., D, l = 1, 2, ..., L). For example, the steepest descent method or the stochastic steepest descent method can be considered.

  Most simply, the following updates can be repeated, for example, a predetermined number of times.

(11)

(12)

(13)

<System configuration>
Next, the configuration of the acoustic signal analyzing apparatus according to the embodiment of the present invention will be described with reference to FIG. The acoustic signal analyzing apparatus 10 according to the embodiment of the present invention analyzes observation data when each of N singers sings at least once for each of L kinds of musical scores. As shown in FIG. 1, the acoustic signal analysis device 10 includes an input unit 12, a calculation unit 14, and an output unit 16.

  By the input unit 12, the singing voice sound signal and the score when each of the N singers sang at least once for each of the L types of score are input to the calculation unit 14.

  The calculation unit 14 can be represented by a configuration including a fundamental frequency extraction unit 20, a data storage unit 22, and a parameter estimation unit 24.

  The fundamental frequency extraction unit 20 estimates and outputs a fundamental frequency trajectory x for each of the singing voice acoustic signals when each of the N singers sang at least once for each of the L types of sheet music. This process can be realized by a well-known technique, for example, literature: A de Cheveign´e and H. Kawahara, “YIN, a fundamental frequency estimator for speech and music,” Journal of the Acoustical Society of America, vol. 111, no. 4, the fundamental frequency estimation method YIN proposed in pp. 1917-1930, 2002 is used. This method uses the autocorrelation function to estimate the fundamental frequency, but post-processing such as difference function, normalization, and parabolic interpolation is used to reduce estimation errors due to double pitch error, half-pitch error, and other noises. It is an introduced method. Previous studies have shown that this is an effective technique for estimating the fundamental frequency of high pitch music and singing voices. In this embodiment, YIN is used to estimate the fundamental frequency from the singing voice acoustic signal every 5 ms and output a fundamental frequency locus.

  The data storage unit 22 is a vector x representing an extracted fundamental frequency trajectory for each of the singing voice signals when each of the N singers sang at least once for each of the L types of sheet music; A pair (x, s) with a score vector s representing a pitch at each time of the score is stored.

The parameter estimation unit 24 determines M kernels K determined in advance based on each pair (x, s) of the vector x representing the fundamental frequency locus and the score vector s stored in the data storage unit 22. _m, is determined from the weight a _m each kernel K _m, and the observed data of each pair (x, s), Hilbert space of pairs (x, s _l) when the singer n sang relative score sl The above expected value μ _{n, l} , the probability distribution of each pair (x, s) expressed using a predetermined dictionary _d of D singing voice singing distributions, and a common musical score dictionary q for singer n _d each weight b _{d a, n,} and the weight c _l for each singing person common dictionary q _d for the score s _l, determined from _d, the pair (x when singer n sang relative score sl the (10, represented with a distance between the s _l, the expected value mu _n on Hilbert space) _l ^* So as to minimize the objective function as shown in formula, and the weight a _m of each kernel K _m, the weight b _d of each of the music common dictionary q _d for each singer _n, and _n, each score s _l Estimate each weight c _{l, d} of the dictionary q _d common to the singers.

  Specifically, the parameter estimation unit 24 includes a parameter initialization unit 30, a kernel weight update unit 32, a singer dictionary weight update unit 34, a music dictionary weight update unit 36, and an end determination unit 38.

Parameter initializing unit 30, the weight a _m of each of the M kernel K _m, the weight b _d of each of the music common D number of dictionary q _d for each of the singers n of N _number, and _n, L each of the weights c _l a singer common D number of dictionary q _d for each type of music s _l, sets an initial value to the _d.

For example, a _m , b _{d, n} , c _{l, d} (m = 1, 2, ..., M, d = 1, 2, ..., D, l = 1, 2, ..., L) is initialized.

The initial value of a _m (m = 1, 2, ..., M) is generated from the gamma distribution of Gamma (1 / M, 1). The initial values of b _{d, n} (d = 1, 2,..., D, n = 1, 2,..., N) are generated from the gamma distribution of Gamma (1 / D, 1). c _{l, d} (l = 1, 2, ..., L, d = 1, 2, ... D) is generated from the gamma distribution of Gamma (1 / D, 1).

The kernel weight updating unit 32 stores each pair (x, s) of the fundamental frequency trajectory x and the score vector s stored in the data storage unit 22 and the kernel K _m that has been initialized or updated last time. weight _{a m (m = 1, 2} , ..., M), the weight b _d music common dictionary q _d for singer _{n, n (d = 1,} 2, ..., D, n = 1, 2, ..., N), and weight c _l a singer common dictionary q _d for the score _{s l, d (l = 1} , 2, ..., L, d = 1, 2, ... D ), The weight a _m (m = 1, 2,..., M) of the kernel K _m is updated according to the above equation (11) so as to minimize the objective function of the above equation (10).

The singer dictionary weight updating unit 34 initializes or previously updated the kernel K with each pair (x, s) of the fundamental frequency trajectory x and the score vector s stored in the data storage unit 22. _m weight _{m a} (m = 1, 2,..., M), common score _{d d} for song n _, b _{d, n} (d = 1, 2,..., D, n = 1, 2, ..., N) , and weight c _l a singer common dictionary q _d for the score _{s l, d (l = 1} , 2, ..., L, d = 1, 2, .. .D), the weight b _{d, n} (d = 1, d = 1, d) of the common score q _d for the singer n according to the above equation (12) so as to minimize the objective function of the above equation (10). 2, ..., D, n = 1, 2, ..., N).

The music dictionary weight updating unit 36, each pair (x, s) of the fundamental frequency trajectory x and the score vector s stored in the data storage unit 22, and the kernel K _m that has been initialized or updated last time. weight a _m of (m = 1, 2, ... , M), the weight b _d music common dictionary q _d for singer _{n, n (d = 1,} 2, ..., D, n = 1 , 2, ..., N), and weight c _l a singer common dictionary q _d for the score _{s l, d (l = 1} , 2, ..., L, d = 1, 2, ... Based on D), according to the above equation (13), the weights c _{1, d} (l = 1, 1) of the common dictionary q _d for the score s _l are used to minimize the objective function of the above equation (10). 2, ..., L, d = 1, 2, ... D).

  The end determination unit 38 determines whether a predetermined end condition is satisfied, and updates by the kernel weight update unit 32, update by the singer dictionary weight update unit 34, and music dictionary weight update until the end condition is satisfied. The updating by the unit 36 is repeated.

  For example, reaching the number of iterations specified in advance may be used as the termination condition. Further, as an end condition, it may be used that an error between the value of the objective function calculated using the parameter before update and the value of the objective function calculated using the parameter after update is equal to or less than a predetermined threshold. .

The output unit 16, when the end determination unit 38 determines that the end condition is satisfied, the kernel finally updated by the kernel weight update unit 32, the singer dictionary weight update unit 34, and the music dictionary weight update unit 36. K _m weights a _m (m = 1, 2,..., M), a common score q _d weights b _{d, n} (d = 1, 2,..., D, n) for the singer n = 1, 2, ..., N ), and weight c _l a singer common dictionary q _d for the score _{s l, d (l = 1} , 2, ..., L, d = 1, 2,. ..D) is output.

<Operation of acoustic signal analyzer>
Next, the operation of the acoustic signal analysis device 10 according to the present embodiment will be described. First, when the singing voice acoustic signal and the score when each of the N singers sang at least once for each of the L types of scores are input to the acoustic signal analysis apparatus 10 by the input unit 12, FIG. The parameter estimation processing routine shown in FIG.

  In step S100, the fundamental frequency extraction unit 20 estimates the fundamental frequency every 5 ms, estimates the fundamental frequency trajectory x for each of the input singing voice acoustic signals using the fundamental frequency estimation method YIN, and calculates the score vector. The pair (x, s) with s is stored in the data storage unit 22.

In step S102, the weight a _m kernel _{K m (m = 1, 2} , ..., M), the weight b _d music common dictionary q _d for singer _{n, n (d = 1,} 2, .. ., D, n = 1, 2, ..., n), and weight c _l a singer common dictionary q _d for the score _{s l, d (l = 1} , 2, ..., L, d = Set initial values to 1, 2, ... D).

In step S104, each pair (x, s) of the fundamental frequency trajectory x and the score vector s stored in the data storage unit 22 and the kernel K _m initialized in the above step S102 or updated last time. weight a _m of (m = 1, 2, ... , M), the weight b _d music common dictionary q _d for singer _{n, n (d = 1,} 2, ..., D, n = 1 , 2, ..., N), and weight c _l a singer common dictionary q _d for the score _{s l, d (l = 1} , 2, ..., L, d = 1, 2, ... Based on D), the weight a _m (m = 1, 2,..., M) of the kernel K _m is updated according to the above equation (11).

In step S106, each pair (x, s) of the fundamental frequency trajectories x and the score vector s stored in the data storage unit 22 and, initialized at step S102, or was last updated, the kernel K _m weight a _m of (m = 1, 2, ... , M), the weight b _d music common dictionary q _d for singer _{n, n (d = 1,} 2, ..., D, n = 1 , 2, ..., N), and weight c _l a singer common dictionary q _d for the score _{s l, d (l = 1} , 2, ..., L, d = 1, 2, ... based on D), according to the above (12), the weight b _d music common dictionary q _d for singer _{n, n (d = 1,} 2, ..., D, n = 1, 2, .. , N).

In step S108, each pair (x, s) of the fundamental frequency trajectories x and the score vector s stored in the data storage unit 22 and, initialized at step S102, or was last updated, the kernel K _m weight a _m of (m = 1, 2, ... , M), the weight b _d music common dictionary q _d for singer _{n, n (d = 1,} 2, ..., D, n = 1 , 2, ..., N), and weight c _l a singer common dictionary q _d for the score _{s l, d (l = 1} , 2, ..., L, d = 1, 2, ... based on D), according to the above (13), the weights c _l a singer common dictionary q _d for the score _{s l, d (l = 1} , 2, ..., L, d = 1, 2,. Update ..D).

  In step S110, it is determined whether a predetermined end condition is satisfied. If the end condition is satisfied, each finally obtained parameter is output by the output unit 16 to end the process. On the other hand, if the end condition is not satisfied, the process returns to step S104.

<Example>
Next, in order to show the effect and operation of the present invention, an example using an acoustic signal analyzer according to an embodiment of the present invention will be described below.

  As an example, the result of applying the proposed algorithm of the embodiment of the present invention to song data in which five females and five males each sang 20 song phrases once is shown.

  First, as a quantitative evaluation, the mean square error to the test data was evaluated by the proposed algorithm of the embodiment of the present invention and the fitting by simple Gaussian process regression. As training data, 2 songs, 4 songs, and 6 songs were randomly selected, and the average and standard deviation when each algorithm was tested 10 times from the initial value determined by random numbers are summarized in Table 1 below.

  Next, as quantitative evaluation, characteristics corresponding to the individuality of each singer actually extracted and decomposed are shown in FIG. Here, all songs are used as training data, and the initial value of the parameter indicates one trial determined by random numbers. In FIG. 3 above, rows correspond to decomposed singing dictionaries, and columns correspond to singers. For example, it can be seen that the fifth singer is easy to use the fifth singing dictionary, and the third singer is easy to use the tenth singing dictionary. Such features that reflect the individuality of singing can have various applications such as singers' identification problems.

As described above, according to the sound signal analysis apparatus 10 according to the embodiment of the present invention, M-number of kernel K _m, and the fundamental frequency trajectory x obtained from each of the weights a _m, and the observation data of the kernel K _m each pair (x, s) of the musical score vector s is determined from an expected value mu _{n, l} on Hilbert space pairs when singer n sang relative score s _l (x, s _l), Probability distribution of each pair (x, s) expressed using a dictionary q _d of D singing voice singing distributions, weights b _{d, n of} a common score q _d for a singer _n , and a score s _l weight c _l for each singing person common dictionary q _d for _is determined from _d, the pair (x, s _l) when the singer n sang relative score s _l expected value of the Hilbert space of mu _{n ,} so as to minimize the objective function expressed by using the distance between _l ^*, and the weight a _m of each kernel K _m, songs 'S weight b _d each respective score common dictionary q _d for the _n, and _n, the weights c _l for each singing person common dictionary q _d for the score s _l, by estimating the _d, each singing from singing It is possible to extract features corresponding to the individuality of the person.

  In addition, as a modeling of singing voice and its potential individuality, the kernel averaging method and the nonparametric Bayes method that have been used independently as separate methods in the past (specifically, dictionary decomposition of signals by gamma process) As a technique for analyzing acoustic signals (especially singing voice data), it is possible to extract feature quantities reflecting the individuality of each singer from singing data sung by a plurality of singers on a plurality of songs. .

  The present invention is not limited to the above embodiment, and various modifications and applications can be made without departing from the gist of the present invention.

  For example, the acoustic signal analysis apparatus 10 described above has a computer system inside, but the “computer system” includes a homepage providing environment (or display environment) if a WWW system is used. Shall be.

  Further, in the present specification, the embodiment has been described in which the program is installed in advance. However, the program can be provided by being stored in a computer-readable recording medium or provided via a network. It is also possible to do. Moreover, you may comprise each part of the acoustic signal analyzer 10 of this Embodiment by hardware. Further, the database storing the parameter initial values can be realized by storage means exemplified by a hard disk device, a file server, etc., and the database may be provided inside the acoustic signal analysis device 10 or provided in an external device. May be.

DESCRIPTION OF SYMBOLS 10 Acoustic signal analyzer 12 Input part 14 Calculation part 16 Output part 20 Fundamental frequency extraction part 22 Data storage part 24 Parameter estimation part 30 Parameter initialization part 32 Kernel weight update part 34 Singer dictionary weight update part 36 Music dictionary weight update part 38 End determination part

Claims (7)

An acoustic signal analyzer for analyzing observation data of an acoustic signal indicating a singing voice when each of N singers sang at least once for each of L types of sheet music,
M defined in advance as a criterion for measuring the similarity between a pair (x, s) of a fundamental frequency locus x representing a fundamental frequency at each time of an acoustic signal representing a singing voice and a score vector s representing a pitch at each time of the score. wherein when the number of kernel K _m, the M kernel K _m each weight a _m, and each of singer n of the n people sang least once for each of the L types of music s _l The pair (x, s _l ) of the pair (x, s _l ) obtained when the singer n sings the score s _l obtained from each pair (x, s) of the fundamental frequency trajectory x obtained from the observation data and the score vector s. Expected value μ _{n, l} on Hilbert space,
Probability distribution of each pair (x, s) expressed using a predetermined dictionary _d of S singing voice singing distributions, weight b of each of the D dictionaries q _d common to the score for singer n _{d, n,} and the score s _l each of the weights c _l a singer common the D pieces of dictionary q _d for _is determined from _d, the pair (x when singer n sang relative score s _l, s _l) of so as to minimize the objective function expressed by using the distance between the Hilbert space on the expectation μ _{n, l ^*,} and the weight a _m each of the M kernel K _m, the n a weight b _{d, n} of each of the music common the D pieces of dictionary q _d for each of the singers n people, the singer common for each of said L types of music s _l said D number of dictionary q _d An acoustic signal analyzer including a parameter estimation unit for estimating each weight c _{l, d} . The parameter estimation unit includes:
The M kernel K and the weight a _m each of _m, each of the weights b _d music common said for each of the singers n of the N number D number of dictionary q _d, and _n, the L type of music the common singing person for each s _l D number of dictionary q _d each weight c _{l a,} and a parameter initializing unit for setting an initial value to each of the _d,
Wherein said fundamental frequency trajectory x obtained from the observed data score each pair of the vector s (x, s), the weight of each of said music common for each of the N number of singers n D pieces of dictionary q _d b _{d, n,} and the L type of music s _l each of the weights c _l a singer common the D pieces of dictionary q _d for _each, based on the _d, so as to minimize the objective function, wherein and kernel weight updating section for updating the weights a _m of each of the M kernel K _m,
Wherein each pair of said fundamental frequency trajectory x obtained from the observed data and the score vector s (x, s), weighting a _m of each of the M kernel K _m, and singing to said L types of music s _l The D dictionaries common to each of the N singers n so as to minimize the objective function based on the weights c _{l, d} of the D dictionaries q _d common a singer dictionary weight updating unit for updating each weight b _{d, n} of q _d ;
Each pair (x, s) of the fundamental frequency trajectory x obtained from the observation data and the score vector s, a weight a _m of each of the M kernels K _m , and each of the N singers n Based on the respective weights b _{d, n} of the D dictionaries q _d common to the scores for D, the D songs common to the singers for each of the L kinds of scores s _l are minimized. A music dictionary weight updating unit for updating each weight c _{l, d} of the dictionary q _d ;
An end determination unit that repeats the update by the kernel weight update unit, the update by the singer dictionary weight update unit, and the update by the music dictionary weight update unit until a predetermined end condition is satisfied,
The acoustic signal analysis device according to claim 1, comprising: The acoustic signal analyzing apparatus according to claim 1, wherein the objective function is represented by the following expression.

However, Q _d is a probability distribution of pairs (x, s) represented using the dictionary q _d , and O _l represents the number of all pairs (x, s) obtained from the observed data. An acoustic signal analysis method in an acoustic signal analyzer for analyzing observation data of an acoustic signal indicating a singing voice when each of N singers sings at least once for each of L types of sheet music,
As a criterion for the parameter estimation unit to measure the similarity between a pair (x, s) of a fundamental frequency trajectory x representing a fundamental frequency at each time of a sound signal indicating a singing voice and a score vector s representing a pitch at each time of the score. predetermined the M kernel K _m, of at least one each of the weights a _m of the M kernel K _m, and each of the singers n of the n number for each of the L types of music s _l A pair (x) obtained when a singer n sings a score s _l obtained from each pair (x, s) of the fundamental frequency trajectory x obtained from the observation data when sung and the score vector s. , Sl _l ) the expected value μ _{n, l} on the Hilbert space,
Probability distribution of each pair (x, s) expressed using a predetermined dictionary _d of S singing voice singing distributions, weight b of each of the D dictionaries q _d common to the score for singer n _{d, n,} and the score s _l each of the weights c _l a singer common the D pieces of dictionary q _d for _is determined from _d, the pair (x when singer n sang relative score s _l, s _l) of so as to minimize the objective function expressed by using the distance between the Hilbert space on the expectation μ _{n, l ^*,} and the weight a _m each of the M kernel K _m, the n a weight b _{d, n} of each of the music common the D pieces of dictionary q _d for each of the singers n people, the singer common for each of said L types of music s _l said D number of dictionary q _d An acoustic signal analysis method for estimating each weight c _{l, d} . By estimating by the parameter estimation unit,
Parameter initialization section, the M kernel K and the weight a _m each of _m, each of the weights b _d music common said for each of the singers n of the N number D number of dictionary q _d, and _n the initial value set L type music s _l each of the weights c _l a singer common the D pieces of dictionary q _d for _each, in each of the _d,
A kernel weight updating unit includes the D dictionaries common to the score for each pair (x, s) of the fundamental frequency trajectory x obtained from the observation data and the score vector s, and each of the N singers n. q each weight b _d of _{d, n,} and the L type of music s _l each of the weights c _l a singer common the D pieces of dictionary q _d for _each, based on the _d, minimizing the objective function as of, and updates the weights a _m of each of the M kernel K _m,
Singer dictionary weight updating section, each pair of said fundamental frequency trajectory x obtained from the observation data and the score vector s (x, s), weighting a _m of each of the M kernel K _m, and the For each of the N singers n so as to minimize the objective function based on the weights c _{l, d} of each of the D dictionaries q _d common to the singers for L kinds of sheet music s _l Update the weights b _{d, n} of each of the D dictionaries q _d common to the score,
Music dictionary weight updating section, each pair of said fundamental frequency trajectory x obtained from the observation data and the score vector s (x, s), each of the weights a _m of the M kernel K _m, and the N Common to singers for the L types of musical scores s _l so as to minimize the objective function based on the weights b _{d, n} of the D dictionaries q _d common to the musical scores for each of the human singers n Update the weights c _{l, d} of each of the D dictionaries q _d of
The termination determination unit includes repeating update by the kernel weight update unit, update by the singer dictionary weight update unit, and update by the music dictionary weight update unit until a predetermined termination condition is satisfied. 4. The acoustic signal analysis method according to 4. The acoustic signal analysis method according to claim 4 or 5, wherein the objective function is expressed by the following equation.

However, Q _d is a probability distribution of pairs (x, s) represented using the dictionary q _d , and O _l represents the number of all pairs (x, s) obtained from the observed data.   The program for functioning a computer as each part of the acoustic signal analyzer of any one of Claims 1-3.