JP2003241776A - Speech analyzing method and apparatus therefor, and speech analyzing program and recording medium therefor - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To accurately find an articulatory motion track in voicing from a given speech. <P>SOLUTION: Vectors generated by adding parameters of a speed and acceleration to position parameters of articulation motion and a speech spectrum, a hidden Markov model (HMM) constituted as a dynamic model for the articulation motion, and linear functions for determining speech spectrums from articulation parameters as speech generation models by states of the HMM, are prepared. The speech spectrum of an input speech is compared with a speech spectrum pattern in an articulation-to-sound code book to select a plurality of pairs of articulation code parameters xC in the increasing order of distances, and the speech generation models, those xC, and the speech spectrum are used to determine the state series (q) maximizing the appearance probability of the speech spectrum. For the (q), an articulation parameter maximizing the appearance probability of the speech spectrum of the speech generation model is generated from the given speech signal. <P>COPYRIGHT: (C)2003,JPO

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a voice analysis method for an input voice signal, which determines each position information of articulatory organs (generating organs) such as jaw, tongue and lips when the voice is generated, that is, articulatory movement. , A technique relating to the apparatus and the voice analysis program and the recording medium.

[0002]

2. Description of the Related Art As a method of inversely estimating an articulatory movement when an input voice is generated, there is a method using an articulatory-acoustic codebook. This is because a large number of data pairs obtained by simultaneously observing articulatory motion and voice are stored in a codebook, and by referring to this codebook, the articulatory data corresponding to the nearby voice data is extracted to obtain the articulatory motion. Estimate (Schroeter, J. and Sondhi, MM, “Spee
ch coding based on physiological models of speech
production, ”Advance in Speech Signal Processing,
edited by S.Furui and MMSondhi (Dekker, New Yor
k), 231-267, 1992).

On the other hand, the change state before and after the articulatory data obtained by referring to the articulatory / acoustic codebook is examined, and a dynamic constraint for eliminating the articulatory data, which does not occur in normal articulatory motion, is incorporated. Another method has been proposed (S. Suzu
ki, T. Okadome and M. Honda, “Determination of artic
ulatory positions from speech acoustics by applyin
g dynamic articulatory constraints, ”Proc. of the
5th ICSLP, 2251-2254, 1998). In this way, if the articulatory movement of the articulatory organ at the time of utterance can be analyzed, the analysis result can be used in various voice fields such as artificial voice generation and language learning.

Although it is not the inverse estimation of articulatory motion, there is an algorithm for generating a voice parameter from a hidden Markov model (hereinafter referred to as HMM), which takes into consideration the velocity and acceleration parameters that represent the temporal change of the voice parameter as a dynamic feature. Proposed (K. Tokuda, T. Masuko, T. Yamada,
T. Kobayashi and S. Imai, “An algorithm for speechge
neration from coutinuous mixture HMMs with dynamic
features, "Proc.EUROSPEECH, 757-760, 1995). This is used, for example, to give a phoneme sequence, obtain a speech spectrum from the corresponding HMM, and perform speech synthesis.

[0005]

The conventional inverse estimation of articulatory motion is to obtain articulatory data of voice data close to a given voice by using an articulatory-acoustic pair codebook created from learning data. It is unlikely that each piece of data in the data series will match any of the audio data in the codebook, but rather most input audio data will be slightly different from the audio data in the codebook. . Therefore, the obtained articulatory data series, that is, the articulatory movement was not accurate.

[0006]

According to the present invention, an acoustic / articulatory pair codebook storing a pair of a voice parameter vector and an articulatory parameter vector, and a probabilistic dynamic state transition of a time series of the articulatory parameter vector. And a model storage unit that stores a voice generation model that determines the occurrence probability of the voice parameter vector for the articulatory parameter vector expressed by Select at least one articulatory parameter vector of the voice parameter vector close to the input voice parameter vector by referring to the paired chord vector, and refer to the input voice parameter vector and the approximate articulatory parameter vector and the model storage unit. And the voice corresponding to these Using a formation model and the articulation model,
An articulation parameter vector sequence that maximizes the appearance probability of the input speech parameter vector sequence is obtained.

[0007]

BEST MODE FOR CARRYING OUT THE INVENTION Embodiments of the present invention will be described below. First, a method of creating a model used in this embodiment is shown in FIG.
Will be described with reference to. Model creation An articulatory-acoustic pair codebook is created using the audio signals obtained by continuously uttering sentences and the articulatory data simultaneously observed by the magnetic sensor system. An audio signal from the audio input terminal 11 is cut out in a spectrum analysis unit 12 for each frame at a rate of 250 times per second by a Blackman window with a window length of 32 ms and subjected to spectrum analysis, for example, a 21st order excluding the 0th order term. The mel cepstrum coefficient of is obtained as a voice parameter. As needed, the speed parameter is detected as a temporal change by the differentiation (difference) of the voice parameter in the speed detection unit 13, and a vector having these voice parameter and the speed parameter as an element is synthesized as a voice parameter vector y. Will be output.

The position information signals in the horizontal and vertical directions at a plurality of positions of the articulatory organs observed at the same time, for example, the lower jaw, the upper / lower lip and the four positions on the tongue, respectively, in the horizontal direction and the vertical direction, are articulated. It is input to the terminals 15a to 15n, respectively captured by the position information section 17 at a rate of 250 times per second, and the respective position parameters are, if necessary, differentiated (difference) by the speed detection section 18 and the speed parameter as a temporal change. Further, if necessary, each speed parameter is differentiated (differenced) by the acceleration detection unit 19 to obtain an acceleration parameter as a temporal change. A vector having these 14 position parameters, velocity parameters, and acceleration parameters as elements is output from the synthesis unit 21 as an articulation parameter vector x. That is, in this example, the voice parameter vector y and the articulatory parameter vector x are vectors each having 42 elements as described below. y = [k, ........., k 21, k 1 ', ........., k 21'] x = [p a, ......, p n, p a ', ......, p n',
p _a ″, ..., P _n ″] In this way, a plurality of data, for example, 200,000 sets of voice parameter vector y and articulatory parameter vector x obtained at the same time point are held, and an articulatory / acoustic pair is stored. The code book 22 is constructed.

The articulatory parameter vector obtained in this way
Using the tone x and the voice parameter vector y.
The time series of the parameter vector x with stochastic dynamic transition states
The expressed articulation model and articulation parameter vector x
Sound that determines the occurrence probability of the voice parameter vector y
The model creating unit 23 creates a model λ including a voice generation model.
To do. In this example, HMM is used as the model λ. HM
The structure of M is a 3-state 1-mixture Gaussian distribution of 2 phonemes,
It is a left-to-right model without a ticket. Figure 2
3 states q₁, Q₂, Q₃There are various conditions
Articulatory parameter vector, voice parameter vector
Each Gaussian distribution for each occurrence probability
The process is a transition (loop) from the same state to the same state, and q
₁To q ₂Or q₂To q₃There are only 5 transitions to
It For each phoneme, a model is created for each different phoneme that follows.
Made However, in order to maintain accuracy, the state of each HMM
If there are 256 or less data in the
Is replaced by an HMM model of one phoneme set.

The HMM model λ is created by first maximizing the appearance probability P (x, q | λ) of the entire articulation parameter vector sequence x obtained by the continuous utterance of the sentence. In this case, a method similar to the method of creating an HMM model using only the voice parameters described in the [Prior Art] section may be used. Here, q represents a state sequence for the entire articulation parameter vector sequence x. State series q
When each one state and q _i which constitutes a probability of occurrence of articulatory parameter vector x of state q _i, the transition to state q _i probability P _t = P | and (q _i lambda), the condition q Probability of appearance of articulatory parameter vector x for _i P _x = P (x
| Q _i , λ). Therefore, P (x, q | λ) =
Each model may be created so that P (x | q, λ) P (q | λ) becomes maximum. Here, the appearance probability P (x | q, λ) of the articulatory parameter vector assumes a Gaussian distribution.

In this way, the state transition probability P _t and the appearance probability P _x are obtained for each state q _i as each model of the HMM, and the model storage unit 24 stores each model λ _1- . The state transition probability P _t (which is made up of a total of 5 probabilities of each loop and adjacent to each other as described above) is stored in the storage unit 26 for each of the storage units 25-1 to 25-J of λJ. Further, in the present invention, a linear function y = Ax + b is used as a conversion function y = f (x) for obtaining a voice parameter vector y using the articulation parameter vector x as a variable for each state.
For each state of the HMM model λ obtained above using a plurality of data of the articulatory parameter vector x and the voice parameter vector y belonging thereto, the articulatory-acoustic conversion function y = Ax + b Matrix A and vector b
Is determined by the voice generation coefficient generation unit 20. That is, as shown in FIG. 4, in this conversion function, the vector y on the left side has 42 elements, the vector x on the middle side also has 42 elements, and the matrix A is a 42 × 42 matrix as shown in FIG. Therefore, the constant b is also a vector having 42 elements. Therefore, for one state q _i , at least 42 pairs of the voice parameter vector y and the articulatory parameter vector x belonging thereto are used, and the voice parameter vector y ′ = Ax + b calculated using the articulatory parameter vector x and the articulatory parameter A and b that minimize the squared error between the vector x and the voice parameter vector y forming a pair are obtained.

One of the speech generation models thus obtained
Each of A and b constituting the unit is a corresponding model of the model storage unit 24.
Is stored in the coefficient storage unit 28 corresponding to the corresponding state. Also
Output of the articulatory parameter vector x for each state of each model
Current probability P_x, The appearance probability P of the voice parameter vector y_yTo
As can be obtained, the noise generator 31 (FIG. 1)
Then, y ′ = Ax + b and y for each state of each model
Error w (in this example, a vector with 42 elements) is calculated
And the average w of each element of the error w_mNoise average calculation unit 3
2 and the covariance σ of the error (remaining sound) w_xThe noise
The calculation is performed by the dispersion calculator 33. Also, for each state of each model
Mean x of the articulation parameter vector x belonging to _mTo
The articulation average calculation unit 34 calculates the
Covariance σ of vector x_xIs calculated by the covariance calculation unit 35
It Average w of each of these calculated states_m, X_mas well as
Covariance σ_w, Σ_xThe average of the corresponding states of the corresponding model
It is stored in the covariance storage unit 29. Appearance probability P_xAnd P_yHaso
X each_m, Σ_xAnd A, b, w_m, Σ_wTo ask for
it can.

Articulatory Motion Estimation Next, an embodiment of the present invention will be described in which the articulatory motion estimation for an input voice signal is performed using the above model. As shown in FIG. 5, the voice signal input from the input terminal 41 is normally temporarily stored in the signal storage unit 42, and then the spectrum analysis unit 43 performs spectrum analysis for each frame to detect a voice parameter. The speech parameters are the same as the learning speech parameters used when creating the model, and in the above example, the 21st-order mel-cepstral coefficient excluding the 0th-order term is obtained. It is also preferable to cut out the audio signal in the same manner as when creating the model. In order to match with the voice parameter vector y at the time of model creation, the velocity detecting unit 44 obtains a velocity parameter as a temporal change of the obtained voice parameter, and the synthesizing unit 45 combines these parameters into a voice parameter vector y.

This voice parameter vector y is compared with the voice parameter vector y _c in the articulatory-acoustic pair codebook 22 created as described above by the approximate articulatory parameter vector selecting unit 46, and y _c is calculated in ascending order of distance. A plurality of approximate articulatory parameter vectors x _c are selected. Regarding the HMM model λ of the model storage unit 24 described above,
For each of these approximate articulation parameter vectors x _{c1 to} x _{cR, using} the input speech parameter vector y, the appearance probabilities P _y and P _x of each state of the model, and the transition probability P _t , the appearance probability of the sequence of the input speech parameter vector y P
(Y, q | λ) = ∫P (y | x _c , q, λ) P (x _c |
The state sequence determination unit 47 determines the HMM state sequence q that maximizes q, λ) P (q | λ) dx _c by using the Viterbi algorithm.

For example, as shown in FIG. 6, for each approximate articulatory parameter vector x _{c1 to} x _cR , a silent model, preferably each silent model in a two-phoneme model is used as an initial value, and an approximate articulatory parameter vector in each state q _i thereof. _{x cr (r = 1, ...} , R) occurrence probability P of the input speech parameter vector y for the _{(y | x cr, q i} , λ) and its approximate articulate parameter vector x _cr of the occurrence probability P (x _cr |
The product of q _i , λ) and the transition probability P (q _i | λ) is taken as the branch metric M _B, and the branch metric M _B is obtained for each transitionable state, thus obtained x _c1
A surviving path largest ones in all M _B for your ~x _cR, the adding the branch metric M _B to the path metric M _P up to that performed for the input speech parameter vector y for each frame.

In this way, the state series q in which one state of one model is associated with each y of the series of the input speech parameter vector y is obtained. According to the obtained HMM state series q, the appearance probability P (y | q, λ) = ∫P (y | x, q, λ) P (x The articulatory parameter calculation unit 48 uses the voice generation model of the model storage unit 24 to generate an articulatory motion that maximizes | q, λ) dx (1).

As described above, it is assumed that the appearance probabilities P _y and P _x are Gaussian distributions. At this time, each appearance probability can be expressed by the following equation. _Py = P (y | x, q, [lambda]) = [1 / (( ² [pi]) ^{N / 2} | [sigma] _w | ^1/2 )] * exp [-(1/2) (y-Ax-b-w) _m ) ^T σ _w ^-1 (y-Ax-b-w _m )] (2) P _x = P (x | q, λ) = [1 / ((2π) ^{M / 2} | σ _x | ^1/2 )] × exp [− (1/2) (x−x _m ) ^T σ _x ⁻¹ (x−x _m )] (3) N is the order of the vector y, M is the order of the vector x,
In the above example, both are 42, and () ^T represents the transpose of the matrix.

Substituting equations (2) and (3) into equation (1)
And finding x that maximizes equation (1)
It is only necessary to find x that minimizes the denominator of the above equation. That is J = (y-Ax-b-w_m)^Tσ_w ^-1(y-Ax-b-w_m)
+ (X-x_m)^Tσ_x ^-1(xx_m) Find x that minimizes Differentiating this equation and setting it to 0   (Σ_x ^-1+ A^Tσ_w ^-1A) x = σ_x ^-1x_m+ A^Tσ_w ^-1(ybw_m) (4) Becomes Therefore, the speech parameter vector for each frame
in the determined state sequence q to which it belongs for y
State q_iA, b, w_m, X_m, Σ_w, Σ _xThe
These values and their audio are read from the Dell memory unit 24.
Substitute the parameter vector y into equation (4) to calculate the value of x.
When calculated, the articulatory parameter vector for y of that frame is calculated.
You can get the cutle x.

The above articulation parameter vector sequence
The procedure for estimating the articulatory motion is briefly described with reference to FIG.
It An audio signal is input, where articulatory movement should be estimated
And stored in the storage unit 42 (S1)
The parameter vector y is detected for each frame (S
2). In the above example, the spectrum is analyzed for each frame and
The luke pstral coefficient is obtained as a voice parameter (S2
-1), and the temporal change of the voice parameter, that is,
The speed parameter is detected (S2-2), and both parameters are detected.
Data into a voice parameter vector. Articulation / acoustic
Refer to the codebook 22 for each voice parameter vector
Voice parameter vectors y close to _cAgainst each
Approximate articulation parameter vector x_cSelect (S3),
These voice parameter vector y and multiple approximate articulation parameters
Data vector x_cAnd is paired and stored in the signal storage unit 42.
(S4).

When the detection of each voice parameter vector y from the beginning to the end of the input voice signal and the selection of the corresponding and approximate articulation parameter vector x _c are completed, the model in the model storage unit 24 is referred to for the entire voice. The state sequence q that maximizes the appearance probability of the parameter vector is determined by the Viterbi algorithm (S5). With reference to the determined state sequence q, for each vector of the sequence of the input voice parameter vector y, the voice generation coefficient A, b of the state q _i to which it belongs, the average w _m , x _m , the covariance σ _w , σ _x is acquired from the model storage unit 24 and equation (4) is calculated to obtain the articulatory parameter vector x (S6). Further, if necessary, this series of articulatory parameter vectors may be displayed or recorded as a change state for each element of the vector.

In the above description, only the voice spectrum, that is, the velocity parameter may not be used as the voice parameter vector, and the acceleration parameter may be added. Similarly, as the articulation parameter vector, only the position parameter may be used, or the position parameter and its velocity parameter may be used and the acceleration parameter may be omitted.
However, it is better to add the velocity parameter depending on the case of only the position parameter, and the accuracy is further improved by adding the acceleration parameter. The order of the voice parameter vector and the articulatory parameter vector need not be the same. The approximate articulation parameter vector x _c may be selected only once. The voice analysis device shown in FIG. 5 may be caused to function by causing a computer to execute a voice analysis program. In this case, a voice analysis program for causing the computer to execute the procedure shown in FIG.
Install from M or flexible magnetic disk,
It may be downloaded via a communication line and executed by a computer.

[0022]

As described above, according to the present invention, a model including an articulation model and a voice generation model is used, and a voice generation model is used with an approximate articulation parameter vector close to the input voice parameter vector. Since the articulatory parameter vector is generated and the articulatory parameter sequence that maximizes the appearance probability of the input voice parameter sequence is obtained, the articulatory motion during voice utterance of the input voice parameter sequence can be obtained more accurately. .

Experiments are shown to evaluate the method of the present invention. In the experiment, the articulatory motion and the voice are those that are simultaneously observed from continuous vocalization using a magnetic sensor system (EMA). The articulatory data was observed using EMA at a sampling rate of 250 Hz. The articulatory parameters were the horizontal and vertical positions of a total of 7 positions of 4 points on the lower jaw, upper and lower lips, and tongue. The voice was recorded at 8 kHz sampling, cut out with a Blackman window having an analysis cycle of 4 ms and a window length of 32 ms, and 21st-order mel-cepstral analysis was performed. The articulatory parameter vector x is a (position / velocity / acceleration) vector, and the voice parameter vector y is (position / velocity).
And the vector.

The simultaneously observed articulatory and acoustic data are 358 sentences (data number 231,
949). For learning the HMM, it was created from 342 sentences randomly selected from 358 sentences. The remaining one
Six sentences are test data for evaluation of the method of the present invention. The structure of the HMM is a left-to-right two-phoneme model with three-state one-mixed Gaussian distribution and no skip. However, the states with a small amount of data were replaced with the one-phoneme model. The conversion function f (x) is approximated by a linear conversion y = Ax + b, and phoneme information at the time of utterance is assumed to be known. This phoneme information may not be necessary.

In FIG. 8, the articulatory motion trajectory (thick line) estimated by the method of the present invention and the articulatory motion observed by the male speaker from the speech signal uttering the sentence "Look through the accumulated documents in the afternoon". An example of a vertical signal of an orbit (thin line) is shown.
It can be seen from this figure that the articulatory motion trajectory (thin line) in which the articulatory motion activation (thick line) estimated by the method of the present invention is observed is well reproduced. FIG. 9 shows the results of the mean square error of the estimated trajectory of the articulatory motion and the observed trajectory with respect to the number of selections of the approximate articulatory code vector x _c in the determination of the state series q of the HMM. From this figure, the reduction in error saturates when the number of selections was 10 or more, and the minimum mean square error was 1.71 mm.
Also, from this figure, it is understood that the number of selected approximate articulation parameter vectors is preferably 5 or more. Further, the best value is 2.09 mm in the conventional method, but it is understood that the present invention can obtain higher accuracy than the conventional method even when only one approximate articulation parameter vector is selected.

[Brief description of drawings]

FIG. 1 is a block diagram showing a functional configuration of an embodiment of a model generation device used in the present invention.

FIG. 2 is a diagram showing an example of a state transition diagram of an HMM.

FIG. 3 is a diagram showing a storage example of a model storage unit 24.

FIG. 4 is a diagram showing a matrix expression of a conversion formula y = Ax + b.

FIG. 5 is a block diagram showing a functional configuration of an embodiment of a voice analysis device according to the present invention.

FIG. 6 is an explanatory diagram for obtaining a state series q by a Viterbi algorithm.

FIG. 7 is a flowchart showing an embodiment of a voice analysis method according to the present invention.

FIG. 8 is a diagram showing an articulatory trajectory according to an experimental result of the method of the present invention and an actually observed articulatory trajectory.

FIG. 9 is a diagram showing the relationship between the number of selected approximate articulatory parameter vectors in the experimental results of the method of the present invention and the mean square error between the estimated orbit and the observed orbit.

Claims (15)

[Claims] 1. A process of storing an input voice signal in a storage unit, a process of extracting the input voice signal from the storage unit and performing spectrum analysis for each frame to generate an input voice parameter vector, and a voice parameter vector and its articulation. With reference to the acoustic / articulatory pair codebook that stores pairs with parameter vectors, at least one articulatory parameter vector containing at least position information of a plurality of articulatory organs, which is a voice parameter vector close to each input voice parameter vector, is stored. The process of selecting as an approximate articulatory parameter vector, the process of storing the input voice parameter vector in the storage unit, the articulatory model expressing the time series of the articulatory parameter vector by probabilistic dynamic state transitions, and the voice parameter vector for the articulatory parameter vector. Speech Generation Model for Determining Occurrence Probability With reference to the stored model storage unit, using the input voice parameter vector and the approximate articulation parameter vector, and the voice generation model and articulation model corresponding to them, the articulation that maximizes the appearance probability of the input voice parameter vector sequence. And a process of calculating articulatory parameters for obtaining a parameter vector sequence. 2. A time-series state transition probability of an articulation parameter vector forming the articulation model, an appearance probability in that state, and a function for determining a voice parameter vector from the articulation parameter vector forming the voice generation model. The model of the model storage unit is configured as a hidden Markov model including a coefficient for each state and an appearance probability for each state of a voice parameter vector with respect to the articulatory parameter vector. The speech analysis method according to Item 1. 3. The state of the hidden Markov model in which the appearance probability of the input speech parameter vector sequence is maximized in the articulatory parameter calculation process using each of the approximated articulation parameter vectors, the input speech parameter vector, and the hidden Markov model. For each state of the determined state sequence, the articulation parameter vector that maximizes the appearance probability of the input speech parameter vector sequence is calculated for each state of the determined state sequence using the input speech parameter vector and the coefficient. The method of claim 2, further comprising: 4. The function is a linear function, and the coefficient is a first-order coefficient A and a constant b, and the average and covariance of the articulatory parameter vector belonging to each state and the coefficient A, The mean and covariance of the error when using b are stored in the model storage unit, and the process of calculating the articulatory parameter vector is performed in the above-mentioned state A to which each input voice parameter vector belongs,
4. The speech analysis method according to claim 3, wherein the calculation is performed using b, the mean and covariance of the articulation parameter vector, and the mean and covariance of the error. 5. The voice analysis method according to claim 1, wherein at least one of the articulatory parameter vector and the voice parameter vector includes an element representing a temporal change of the parameter. 6. A signal storage unit for storing an input voice signal, an input voice parameter vector, etc., an acoustic / articulatory pair codebook storing a pair of a voice parameter vector and its articulatory parameter vector, and an articulatory parameter vector A model storage unit that stores a voice generation model that determines the occurrence probability of the voice parameter vector for the articulatory model that represents the sequence by probabilistic dynamic state transitions, and spectrum analysis of the input voice signal for each frame. , A spectrum analysis means for generating an input speech parameter vector, and an articulation parameter vector including at least position information of a plurality of articulatory organs, which is a speech parameter vector close to the input speech parameter vector with reference to the acoustic / articulatory pair codebook. At least one approximate articulatory parameter The input voice parameter vector and the approximate articulatory parameter vector with reference to the model storage unit, and the voice generation model and articulatory model corresponding to these, the appearance probability of the input voice parameter vector sequence. And an articulatory parameter calculation means for obtaining an articulatory parameter vector sequence that maximizes 7. A time-series state transition probability of an articulatory parameter vector that constitutes the articulatory model, an appearance probability in that state, and a function for determining a voice parameter vector from the articulatory parameter vector that constitutes the voice generation model. The model of the model storage unit is configured as a hidden Markov model including a coefficient for each state and an appearance probability for each state of a voice parameter vector with respect to the articulatory parameter vector. Item 6. The voice analysis device according to item 6. 8. The state of a hidden Markov model in which the appearance probability of an input speech parameter vector sequence is maximized, wherein the articulation parameter calculation means uses each of the approximate articulation parameter vectors, the input speech parameter vector, and the hidden Markov model. For each state of the determined state series, the articulation parameter vector that maximizes the appearance probability of the input speech parameter vector sequence is calculated using the input speech parameter vector and the coefficient. The voice analysis device according to claim 7, further comprising: 9. The function is a linear function, and the coefficients are a first-order coefficient A and a constant b, and the average and covariance of the articulatory parameter vector belonging to each state and the coefficient A, The mean and covariance of the error when using b are stored in the model storage unit, and the means for calculating the articulatory parameter vector is A for the above-mentioned state to which each of the input voice parameter vectors belongs,
9. The speech analysis apparatus according to claim 8, further comprising: b, means for calculating using the mean and covariance of the articulation parameter vector and the mean and covariance of the error. 10. The voice analysis device according to claim 6, wherein at least one of the articulatory parameter vector and the voice parameter vector includes an element representing a temporal change of the parameter. 11. A signal storage unit for storing an input voice signal, an input voice parameter vector, etc., an acoustic / articulatory pair codebook storing a pair of a voice parameter vector and its articulatory parameter vector, and an articulatory parameter vector A model storage unit that stores a voice generation model that determines the occurrence probability of the voice parameter vector for the articulatory model that represents the sequence by probabilistic dynamic state transitions, and spectrum analysis of the input voice signal for each frame. , A spectrum analysis means for generating an input speech parameter vector, and an articulation parameter vector including at least position information of a plurality of articulatory organs, which is a speech parameter vector close to the input speech parameter vector with reference to the acoustic / articulatory pair codebook. At least one approximate articulatory parameter Means for selecting the input voice parameter vector sequence using the input voice parameter vector and the approximate articulatory parameter vector, and the corresponding voice generation model and articulatory model with reference to the model storage unit. A voice analysis program for causing a computer to function as a voice analysis device, comprising: an articulation parameter calculation unit that obtains a maximum articulation parameter vector sequence. 12. A time-series state transition probability of an articulatory parameter vector that constitutes the articulatory model, an appearance probability in that state, and a function that determines a voice parameter vector from the articulatory parameter vector that constitutes the voice generation model. The coefficient for each state, including the appearance probability for each state of the voice parameter vector for the articulatory parameter vector, the model of the model storage unit is configured as a hidden Markov model by these, the articulatory parameter calculation means, Using each of the approximate articulation parameter vectors, the input speech parameter vector, and the hidden Markov model, means for determining the sequence of states of the hidden Markov model that maximizes the appearance probability of the input speech parameter vector sequence, and the above determined For each state in the state series, enter the above Using voice parameter vector and the coefficient, the input speech parameter vector sequence according to claim 11, wherein the speech analysis program that probability of occurrence, characterized in that it comprises means for calculating each articulation parameter vector becomes maximum. 13. The function is a linear function, and the coefficient is a first-order coefficient A and a constant b, and the average and covariance of the articulatory parameter vector belonging to each state and the coefficient A, The mean and covariance of the error when using b are stored in the model storage unit, and the means for calculating the articulatory parameter vector is A for the above-mentioned state to which each of the input voice parameter vectors belongs,
13. The speech analysis program according to claim 12, further comprising: b, means for calculating using the mean and covariance of the articulation parameter vector and the mean and covariance of the error. 14. The voice analysis program according to claim 11, wherein at least one of the articulatory parameter vector and the voice parameter vector includes an element representing a temporal change of the parameter. 15. A computer-readable recording medium in which the voice analysis program according to claim 11 is recorded.