JP2009063773A - Speech feature learning device and speech recognition device, and method, program and recording medium thereof - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To improve recognition accuracy, even when the number of distributions of mixed Gaussian distribution increases. <P>SOLUTION: A static feature amount and a dynamic feature amount are extracted from input speech (2, S202). A trajectory x<SB>t</SB>is synthesized (104, S204) by using a linear dynamic system and a dictionary data stored in a dictionary data base. The trajectory x<SB>t</SB>is assigned to a maximized function stored in a maximized function data base 112, and a likelihood function P(Y<SB>1:T</SB>) is generated (108, S206). The likelihood function P(Y<SB>1:T</SB>) is approximated and an approximation likelihood function P'(Y<SB>1:T</SB>) is generated(1164, S207). An approximation error e' is calculated by approximating an error e between the approximation likelihood function P'(Y<SB>1:T</SB>) and the likelihood function P(Y<SB>1:T</SB>). The approximation likelihood function P'(Y<SB>1:T</SB>) and the approximation error e are added, and a corrected likelihood function P''(Y<SB>1:T</SB>) is generated, and a likelihood is calculated from the corrected likelihood function P''(Y<SB>1:T</SB>). <P>COPYRIGHT: (C)2009,JPO&INPIT

Description

  The present invention relates to, for example, a speech feature learning apparatus, a speech recognition apparatus, a method thereof, a program thereof, and a recording medium using a hidden Markov model.

  There are many speech recognition techniques using acoustic models. The acoustic model is, for example, an HMM (Hidden Markov Model). In the conventional speech recognition using the HMM, the trajectory (time series of the average value of the HMM) used as a reference when calculating the likelihood is not smooth at the state transition portion of the HMM, causing deterioration in recognition accuracy. In the methods of Non-Patent Document 1 (Prior Art 1), Non-Patent Document 2 (Prior Art 2), and Non-Patent Document 3 (Prior Art 3), the relationship between the static feature quantity and the dynamic feature quantity of the input speech is used. Then, smooth trajectories were synthesized to improve speech recognition accuracy. However, since these methods assumed that each state of the HMM has only one Gaussian distribution, it was difficult to expand to a mixed Gaussian distribution type HMM having a plurality of Gaussian distributions. Therefore, there is Non-Patent Document 4 (Prior Art 4) dealing with a mixed Gaussian distribution type HMM. The method of Non-Patent Document 4 obtains a state series and a distribution series that maximize the likelihood of HMM for input speech by a Viterbi algorithm using HMM, synthesizes a trajectory for the state series and the distribution series, and inputs This is a technique for calculating the likelihood between speech and trajectory. In this method, since only one distribution series trajectory is obtained, there is a problem that the recognition accuracy deteriorates when the number of distributions of the mixed Gaussian distribution increases. Therefore, Patent Document 1 (Prior Art 5) describes a technique using a switching linear dynamic system to realize trajectory synthesis for a mixed Gaussian distribution type HMM. The speech feature learning device and speech recognition device described in Patent Document 1 will be briefly described below.

  FIG. 1 shows a functional configuration example of the speech feature learning device 5 of the conventional technology 5, FIG. 2 shows a functional configuration example of the speech recognition device 11 of the conventional technology 5, and FIG. 3 shows the speech feature learning device 5 and the speech recognition device 11 The integrated structure is shown. First, the learning mode by the speech feature learning device 5 will be described with reference to FIGS. 1 and 3. In addition, in the arrows between the components in FIG. 3, the broken arrows indicate the movement of information in the learning mode, and the practical arrows indicate the movement of information in the recognition mode.

First, the feature quantity extraction unit 2 extracts static feature quantities and dynamic feature quantities from the learning speech. Here, the dynamic feature amount is a parameter representing a temporal change of the static feature amount, such as a first derivative component (Δ component) and a second derivative component (ΔΔ component) of the static feature amount (y), for example. (Δy, ΔΔy). The acoustic model learning unit 4 learns an acoustic model from the static feature amount and the dynamic feature amount. Specifically, the acoustic model learning unit 4 obtains, for example, an average value and variance of static feature quantities as feature patterns, and an average value Δu _SK and a variance value Δσ of differential coefficients of static feature quantities as dynamic feature quantities. _SK ² and the average value ΔΔu _SK and variance value ΔΔσ _SK ² of the second derivative of the static feature quantity are calculated. At the time of this storage, the number of states of the HMM and the number of Gaussian distributions (how many Gaussian distributions the Gaussian distribution of each state is expressed as) are determined. The HMM database 6 stores an average value M _SK = [u _SK , Δu _SK , ΔΔu _SK ] and variance [σ _SK ² , Δσ _SK ² , ΔΔσ _SK ² ] for each Gaussian distribution in each state of the HMM. Here, S represents an HMM state number, and K represents a Gaussian distribution number.

  The model conversion unit 8 converts the acoustic model into a switching linear dynamic system using the relationship between the static feature quantity and the dynamic feature quantity. The relation between the static feature quantity and the dynamic feature quantity is stored in the feature quantity relational expression database 10. Here, the linear dynamic system is the following state equation and observation equation.

M _SK = CX ^t + W _SK
X ^{t + 1} = AX ^t + N _t (1)
As described above, S represents an HMM state number, and K represents a Gaussian distribution number. X ^t is a collection of trajectory time series for five frames, and can be expressed as X ^t = [x _{t + 2} x _{t + 1} x _t x _t−1 x _t−2 ] ^Q. x _t represents a trajectory at time t, and “ ^Q ” represents transposition. The matrices C, A, M, and N will be described in detail in “Best Mode for Carrying Out the Invention”. As understood from Equation (1), in a mixed Gaussian distribution in which one state is composed of S states, one mixed Gaussian distribution is converted into a state equation and an observation equation, and S linear dynamic systems are obtained. Create The switching linear dynamic system is a model in which a plurality of state equations and observation equations are switched over time. The trajectory synthesis unit 14 obtains a trajectory by using a switching Kalman filter from the switching linear dynamic system stored in the switching linear dynamic system database 12 and the static feature quantity and dynamic feature quantity from the feature quantity extraction unit 2. In this process, first, the switching Kalman filter individually operates the Kalman filter using the state equation and the observation equation shown in Expression (1) in the switching linear dynamic system database.

This is shown in FIG. The example of FIG. 4 shows a case where the maximum value of K of the switching linear dynamic system is 2, that is, has two sets of state equations and observation equations. Let t be the time. In FIG. 4, it is assumed that the mixed Gaussian distribution of the trajectory up to time t-1 is obtained in advance. The trajectory distribution at this time is the sum of two Gaussian distributions. The average value and variance of the two Gaussian distributions are (average value, variance) = (x ¹ _{t−1 | t−1} , V ¹ _{t−1 | t−1} ), (x ² _{t−1 | t−1).} , V ² _{t−1 | t−1} ). In the forward processing, the distribution of t trajectories is calculated from the individual Gaussian distribution of t-1. This can be realized by operating the Kalman filters 1051 and 1052 using the state equation and the observation equation of K = 1 and K = 2 for the individual Gaussian distribution of t−1, respectively. When such an operation is performed, four trajectories are calculated from two trajectories as shown in FIG. However, if such a process is performed only for time T, 2 ^T trajectories are finally obtained. In order to prevent this, in the switching filter, the four trajectories are integrated by the integrators 1053 and 1054 to be reduced to two trajectories. By sequentially performing these operations, a trajectory close to the learning voice is obtained. Finally, it performs backward processing to smooth the trajectory. Specifically, the likelihood for learning speech y, Δy, ΔΔy calculated from the average values u ^ij , Δu ^ij , ΔΔu ^ij and variances Σ ^ij , ΔΣ ^ij , ΔΔΣ ^ij of each state distribution of the HMM is high. The weight is set so as to be, and this weight is applied. Thereby, a trajectory close to the learning voice can be finally obtained. With the introduction of the trajectory, a new variance calculation is required. The variance calculation unit 20 calculates the variance using the EM algorithm and stores it in the variance database 22.

  Next, the recognition mode of the prior art 5 will be described with reference to FIGS. The feature amount extraction unit 100 extracts a static feature amount (y) and a dynamic feature amount (Δy, ΔΔy) of the input speech. The trajectory synthesis unit 14 refers to the switching linear dynamic system stored in the switching linear dynamic system database 105 to generate a trajectory from phoneme, word, or sentence candidates stored in the dictionary. A plurality of trajectories are synthesized by forward processing and backward processing of the switching Kalman filter. As necessary, the model conversion unit 8 converts the acoustic model of the HMM database 6 into a switching linear dynamic system. If a linear dynamic system has already been obtained, the HMM database 6 and the model conversion unit 8 may be omitted.

The likelihood calculating unit 18 refers to the variance value stored in the variance database 22 and obtains the likelihood between the trajectory and the input speech every time a trajectory is created. The phoneme sequence having the greatest likelihood is taken as the result of speech recognition. As described above, the conventional technique 5 considers the extension to the mixed Gaussian distribution instead of the formulation of the trajectory synthesis method by the likelihood maximization used in the conventional method using the trajectory, and is based on the switching linear dynamic system. Formulation is performed.
Y.Minami, E, McDermott, A Nakamura, and S, Katagiri, ”A recognition method using synthesis-scoring that incorporates direct relations between static and dynamic feature vector time series“ Workshop for Consistent & Reliable Acoustic Cues for Sound Analysis 2001. Y. Minami, E, McDermott, A Nakamura, and S, Katagiri, “A recognition method with parametric trajectory synthesized using direct relations between static and dynamic feature vector time eries,” Proc.ICASSP, pp.957-960, 2002. Y. Minami, E, McDermott, A Nakamura, and S, Katagiri, “A Theoretical Analysis of Speech Recognition based on Feature Trajectory Models,” in Proc ICSLP, vol.I, 2004. Y. Minami, E, McDermott, A Nakamura, and S, Katagiri, “A recognition method with parametric generated from mixture distoribution HMMs,” Proc.ICASSP, pp.124-127 2003. JP 2006-171020 A

  In the process of the trajectory synthesizer 14 shown in the prior art 5, as shown in FIG. 4, the trajectories are integrated by the integrators 1053 and 1054, and the trajectories are averaged, so that there is a problem that the recognition accuracy deteriorates. . Further, as described above, since the conventional technique 4 obtains only one distribution series trajectory, there is a problem that the recognition accuracy deteriorates when the number of distributions of the mixed Gaussian distribution increases. Therefore, the present invention is for solving the problems of the prior art 4. That is, the present invention provides a speech feature learning device, a speech recognition device, a method thereof, a program thereof, and a recording medium thereof in which the recognition accuracy does not deteriorate even when the number of mixed Gaussian distributions increases. Objective.

  The speech feature learning apparatus according to the present invention includes a feature amount extraction unit, an acoustic model learning unit, an HMM database, a model conversion unit, a linear dynamic system database, a maximum likelihood calculation unit, a trajectory synthesis unit, a likelihood A degree function maximization unit and a maximization function database. The feature amount extraction unit extracts a static feature amount and a dynamic feature amount from the learning speech. The acoustic model learning unit learns an acoustic model from the static feature amount and the dynamic feature amount. The HMM database stores acoustic models. The model conversion unit converts the acoustic model into a linear dynamic system using the relationship between the static feature quantity and the dynamic feature quantity. The linear dynamic system database stores the linear dynamic system. The maximum likelihood calculation unit uses the acoustic model stored in the HMM database to determine the state series and distribution series of the acoustic model so that the likelihood of the learning speech for the acoustic model is maximized. The trajectory synthesis unit synthesizes the trajectory using the linear dynamic system, the state series, and the distribution series. The likelihood function maximization unit obtains a maximization function that maximizes a likelihood function including a static feature value, a dynamic feature value, a trajectory, and a variance based on a state sequence and a distribution sequence. The maximization function database stores the maximization function.

  The speech recognition apparatus according to the present invention includes a feature amount extraction unit, an HMM database, a linear dynamic system database, a trajectory synthesis unit, a likelihood function generation unit, and a likelihood calculation unit. In addition, the likelihood function calculation unit includes correction means for correcting the likelihood. The feature quantity extraction unit extracts a static feature quantity and a dynamic feature quantity from the input speech. The HMM database records acoustic models. The linear dynamic system database stores the linear dynamic system. The trajectory synthesis unit synthesizes a trajectory using a linear dynamic system stored in the linear dynamic system database and dictionary data stored in the dictionary database. The likelihood function generation unit assigns the trajectory to the maximization function stored in the maximization function database, and the likelihood function that is the probability density of the input speech with respect to the input of the trajectory or the likelihood with respect to the input of the trajectory Generate a likelihood function that takes the maximum value of degrees. The likelihood calculating unit calculates the likelihood by substituting the static feature amount and the dynamic feature amount into the likelihood function.

  With the above configuration, even when the number of mixed Gaussian distributions increases, recognition accuracy can be improved while ensuring smooth trajectory generation even in the state transition portion of the HMM.

  The best mode for carrying out the invention will be described below. In addition, the same number is attached | subjected to the process which performs the structure part which has the same function, and the same process, and duplication description is abbreviate | omitted. FIG. 5 shows a functional configuration example of the speech feature learning device 200, FIG. 6 shows a functional configuration example of the speech recognition device 300, and FIG. 7 shows an integrated configuration of the speech feature learning device 200 and the speech recognition device 300. FIG. 8 shows the main processing flow of the speech feature learning apparatus 200, and FIG. 9 shows the main processing flow of the speech recognition apparatus 300. First, the learning mode by the speech feature learning apparatus 200 will be described with reference to FIGS. 5, 7, and 8.

[Learning mode]
First, when a learning voice is input, the feature amount extraction unit 2 extracts a static feature amount and a dynamic feature amount (step S102). Here, the dynamic feature amount is a parameter representing a temporal change of the static feature amount, such as a first derivative component (Δ component) and a second derivative component (ΔΔ component) of the static feature amount (y), for example. (Δy, ΔΔy).

The acoustic model learning unit 4 learns an acoustic model from the static feature amount and the dynamic feature amount (step S104). Specifically, the acoustic model learning unit 4 obtains, for example, an average value and variance of static feature quantities as feature patterns, and an average value Δu _SK and a variance value Δσ of differential coefficients of static feature quantities as dynamic feature quantities. _SK ² and the average value ΔΔu _SK and variance value ΔΔσ _SK ² of the second derivative of the static feature quantity are calculated. The HMM database 6 stores an average value M _SK = [u _SK , Δu _SK , ΔΔu _SK ] and variance [σ _SK ² , Δσ _SK ² , ΔΔσ _SK ² ] for each Gaussian distribution in each state of the HMM ( Step S106). At the time of this storage, the number of states of the HMM and the number of Gaussian distributions (how many Gaussian distributions the Gaussian distribution of each state is expressed as) are determined. Here, S represents an HMM state number, and K represents a Gaussian distribution number.

  The model conversion unit 8 converts the acoustic model into a linear dynamic system using the relationship between the static feature quantity and the dynamic feature quantity (step S108). The relation between the static feature quantity and the dynamic feature quantity is stored in the feature quantity relational expression database 10 in advance. An example of a linear dynamic system will be described.

  It is preferable to convert each state and distribution of the acoustic model into the following state equation and observation equation. This is because the trajectory synthesis calculation amount by the trajectory synthesis unit 104 described below is reduced. The state equation and the observation equation are the same as the above equation (1).

ここで、θは正の大きな値である。 M _SK = CX ^t + W _SK
X ^{t + 1} = AX ^t + N _t (1)
However, t represents time, ^{X t} is a summary of the trajectory time series 5 frames ^can be expressed as _{X t = [x t + 2} x t + 1 x t x t-1 x t-2] Q . x _t represents a trajectory at time t, and “ ^Q ” represents transposition. For the matrices A and C, the following matrix is used here.

Here, θ is a large positive value.

W _SK has an average of [0 0 0] ^Q ,
It is a random variable that follows a Gaussian distribution with variance Σ _SK = diag [σ _SK ² Δσ _SK ² ΔΔσ _SK ² ]. Here, diag is a function for creating a diagonal matrix having [] in diagonal elements. The average value of the HMM as follows: is substituted for M _SK.
M _SK = [u _SK Δu _SK ΔΔu _SK ] ^Q

Further, the matrix A indicates the _{x t + 2 = x t +} 2 x t + 1 = x t + 1 x t = x t x t-1 = x t-1, in this operation, indicating that changes in non-affected by the noise is not Yes. The matrix A is not limited to this. As the matrix C, any matrix can be used as long as it approximately realizes the first and second derivatives. The matrix C, the number of columns is the number of frames defined by X ^t, the number of the number of rows (in this example, 2 + 1) differential orders +1 dynamic features as long as the same number. The first line of the matrix C is a calculation for obtaining a static feature value, the second line is a calculation for obtaining a Δ feature value, and the third line is a calculation for obtaining a ΔΔ feature value.

  The maximum likelihood calculation unit 106 uses the acoustic model stored in the HMM database 6 to determine the state series and distribution series of the acoustic model so that the likelihood of the learning speech for the acoustic model is maximized (step S110). For example, a Viterbi algorithm may be used for this maximization process.

The trajectory synthesis unit 104 synthesizes the trajectory x _t using the linear dynamic system (state equation and observation equation) from the linear dynamic system database 102 and the state sequence and distribution sequence from the maximum likelihood calculation unit 106. (Step S112). Specific to the state sequence with the distribution sequence than the maximum likelihood calculation unit 106 substitutes the state equation observation equation, by solving for x _t assignment state equations and observation equations, seek x _t.

Then, the likelihood function maximization unit 110 obtains a maximization function that maximizes the likelihood function including the static feature quantity, the dynamic feature quantity, the trajectory, and the variance of the state series and the distribution series ( Step S114). Hereinafter, in detail, regarding the trajectory x _t , the likelihood function P (y _t ) at the time t can be expressed by the following equation (2).

が、定数になっている。 This likelihood function P (y _t ) is used when the trajectory x _t is generated from a linear dynamic model whose acoustic model state and distribution value are S _t K _t . Depending on which linear dynamic model the trajectory is generated from, the likelihood function is switched, and the maximization function that maximizes the product of the above likelihood function at time t is learned. Here, since it is difficult to calculate the EM algorithm for learning this function, approximate calculation is used. For this, a state series and a distribution series from the maximum likelihood calculation unit 106 are used. A trajectory of all learning data is calculated from the state series and the distribution series, and a maximization function that maximizes the likelihood function P (y _t ) is obtained. Then, the maximization function is stored in the maximization function database 112 (step S116). In the maximization function, y _t , x _t , Δy _t , Δx _t , ΔΔy _t , ΔΔx _t are variables,

Is a constant.

[Recognition mode]
Next, the recognition mode by the speech recognition apparatus 300 will be described using FIG. 6, FIG. 7, and FIG. Feature extraction unit 2, a static characteristic amount _{y t} and dynamic features of the input speech (Δy _t, ΔΔy _t) to extract (step S202). The trajectory synthesis unit 104 synthesizes all possible trajectories from phoneme, word, or sentence candidates stored in the dictionary database 16 by referring to the linear dynamic system from the linear dynamic system database 102 ( Step S204). If necessary, the model conversion unit 8 may convert the acoustic model stored in the HMM database 6 into a linear dynamic system and store it in the linear dynamic system database. If the linear dynamic system database 102 has already been obtained, the model conversion unit 8 may be omitted.

The likelihood function generation unit 108 substitutes the trajectory from the trajectory synthesis unit 104 for the maximization function stored in the maximization function database 112, and the likelihood function P ₁ that is the probability density of the input speech with respect to the input of the trajectory. (Y _{1: T} ) or a likelihood function P ₂ (Y _{1: T} ) that takes the maximum likelihood value for the trajectory input is generated (step S206). P ₁ (Y _{1: T} ) and P ₂ (Y _{1: T} ) can be defined as follows, for example.

Here, maxA indicates the maximum value of A, P (A | B) is a conditional probability that is A when B is 1, and 1: T is 1,. . . , T, S _{1: T} represents an HMM state series, K _{1: T} represents an HMM distribution series, and X _{1: T} (S _{1: T} , K _{1: T} ) represents this state series and distribution. A trajectory sequence synthesized from the sequence is shown. P (S _{1: T} , K _{1: T} ) is a state transition probability and a probability of selecting a distribution obtained for the state and distribution sequence, and is used in the HMM. Use the same value as the current value. A symbol using “^” is a value for the trajectory, and a symbol not using “^” is a value for the HMM. In addition, in the expression described in the image, “^” is attached immediately above the symbol, and in the expression described in the text, “^” is attached immediately above the symbol, but these values are Equivalent. Y _t and Y _{1: T} are defined by the following equations.
Y _t = [y _t , Δy _t , ΔΔy _t ] ^Q
Y _{1: T} = Y ₁ , Y ₂ ,. . . , Y _T

In the following description, P ₁ (Y _{1: T} ) and P ₂ (Y _{1: T} ) are collectively referred to as P (Y _{1: T} ). When the number of states of the HMM is small (for example, the number of states is 1), the likelihood calculation unit 114 uses the likelihood function shown in Expression (3) or (3 ′) as a static function from the feature amount extraction unit 2. The likelihood is calculated by substituting the feature quantity and the dynamic feature quantity. A model that maximizes the likelihood is set as a final recognition result.

[Modification 1]
Next, the speech recognition apparatus 300-1 which is the modification 1 of the speech recognition apparatus 300 is demonstrated. When the number of states of the HMM is large, all the possible trajectories must be generated and compared with the input speech in Equation (3), which cannot be realized.

Therefore, the speech recognition apparatus 300-1 holds an approximation unit 1164 for approximating the likelihood function in the likelihood calculation unit 114. Likelihood function P (Y _{1: T} ) is input to likelihood approximating means 1164, and likelihood function P (Y _{1: T} ) is approximated to obtain approximate likelihood function P ′ (Y _{1: T} ) ( Step S207). A method of approximating the likelihood function P (Y _{1: T} ) will be described. First, a state sequence S _{1: T} and a distribution sequence K _{1: T} that maximize the likelihood of the acoustic model are obtained in advance by the Viterbi algorithm of the acoustic model. The method of obtaining the state series S _{1: T} and the distribution series K _{1: T} can be expressed by the following equation (4).

The trajectory x ^ _{1: T in the} equation (5) is preferably obtained by realizing a Kalman filter based on the state equation and the observation equation according to the equation (1) using a linear dynamic model. This is because the calculation amount is reduced. Using this trajectory x ^ _{1: T} , an approximate likelihood function P ′ (Y _{1: T} ) can be obtained as shown in the following equation (6).

Then, the likelihood calculation unit 114 calculates the likelihood by substituting the static feature amount and the dynamic feature amount from the feature amount extraction unit 2 into the approximate likelihood function P ′ (Y _{1: T} ). A model that maximizes the likelihood is set as a final recognition result.

This approximate likelihood function P ′ (Y _{1: T} ) is almost the same as that described in Non-Patent Document 4, but when obtaining a trajectory according to the above equation (5), a Kalman filter based on a state equation and an observation equation is realized. Is different from the prior art 4.

[Modification 2]
Next, a speech recognition device 300-2 that is a second modification of the speech recognition device 300 will be described. In the approximate likelihood function P ′ (Y _{1: T} ), the likelihood of trajectories other than the maximum state sequence and distribution sequence by the HMM shown below is ignored, and high-accuracy recognition cannot be performed. Therefore, the likelihood function calculation 114 further includes a correction unit 116. An example of the functional configuration of the correction unit 116 and the like is shown in FIG. The correction unit 116 includes an error approximation unit 1168 and an addition unit 1170. The error approximation means 1168 approximates the error e between the approximate likelihood function P ′ (Y _{1: T} ) and the likelihood function P (Y _{1: T} ) using the likelihood of the HMM, so that the approximate error e ′. Is generated. First, the error e can be expressed by the following equation (7). The error e can be expressed by e = | P (Y1 _{: T} ) −P ′ (Y1 _{: T} ) |.

この式（８）の２行目の式は例えば、ＨＭＭのビタービアルゴリズムを用いて得られる。 However, the error e cannot be obtained directly. Therefore, the error approximating unit 1166 approximates the error e using the likelihood of the HMM to generate an approximate error e ′. The approximation error e ′ is expressed by the following equation (8).

The expression in the second row of the expression (8) is obtained using, for example, the HMM Viterbi algorithm.

The adding unit 1170 adds the approximate likelihood function P ′ (Y _{1: T} ) from the likelihood approximating unit 1164 and the approximate error e ′, thereby correcting the likelihood function P ″ (Y _{1: T} ). Is obtained (step S208). That is, the following formula (9) is calculated.

A correction likelihood function P ″ (Y _{1: T} ) is output from the adding means 1170. Then, the likelihood calculation unit 114 uses the static feature amount and the dynamic feature amount from the feature amount extraction unit 2 to use the function value of the corrected likelihood function P ″ (Y _{1: T} ) in Expression (9). Is obtained (step S210), and a model that maximizes the likelihood is set as a final recognition result.

In this way, the likelihood function P _{(Y 1: T)} is corrected by the correction means 116, correction likelihood function _{P '' (Y 1: T} ) calculated. As a result, the likelihood of trajectories other than the maximum state sequence and distribution sequence by HMM is also taken into consideration. Therefore, the speech recognition apparatus 300 of this embodiment can improve the recognition accuracy over the prior arts 4 and 5 even when the number of mixed Gaussian distributions is increased.

[Experimental result]
Next, experimental results comparing the speech recognition apparatus according to the present invention, the speech recognition apparatus according to the prior art 4 and the speech recognition apparatus using the HMM will be described. In the experiment, speaker-independent and task-independent recognition were performed. An environment-dependent HMM was learned in which the sampling rate was 16 kHz, the frame shift was 10 ms, and the number of Gaussian distributions in each state was 2 or 3. All male speakers of JNAS data were used as learning data. The data used is 2,0078 sentences. As evaluation data, 100 city utterances by 75 men analyzed under the same conditions as the learning data were used. The number of vocalizations is 7198. In the recognition experiment, under one condition, there is a possibility that the performance is improved by chance due to an error. In order to eliminate this point, a plurality of experiments were performed by changing the number of states to be created. FIG. 11 shows the result of the word error recognition rate when the number of mixed Gaussian distributions is 2, and FIG. 12 shows the result of the word error recognition rate when the number of mixed Gaussian distributions is 3. When the number of mixed Gaussian distributions is 2, the number of states of the created HMM is 389, 2589, 1992, and 1611. When the number of mixed Gaussian distributions is 3, the number of states of the created HMM is 1 type of 1992. Using. 11 and 12, regardless of the number of states of the created HMM, the speech recognition device of the present invention has improved recognition accuracy over the speech recognition device of the prior art 4 and the speech recognition device using the HMM. I understand that

The figure which shows the function structural example of the conventional audio | voice characteristic learning apparatus. The figure which shows the function structural example of the conventional speech recognition apparatus. The figure which shows what put together the conventional speech feature learning apparatus and speech recognition apparatus. The figure which showed the flow which calculates | requires the conventional trajectory. The figure which shows the function structural example of the audio | voice feature learning apparatus of an Example. The figure which shows the function structural example of the speech recognition apparatus of an Example. The figure which shows what put together the audio | voice feature learning apparatus and audio | voice recognition apparatus of an Example. The figure which shows the flow of the main processes of the audio | voice feature learning apparatus of an Example. The figure which shows the flow of the main processes of the speech recognition apparatus of an Example. The figure which shows the function structural example of the correction | amendment means. The figure which shows the experimental result in case mixing Gaussian distribution is 2. FIG. The figure which shows the experimental result in case mixing Gaussian distribution is 3. FIG.

Claims (10)

A feature quantity extraction unit that extracts static feature quantities and dynamic feature quantities from learning speech;
An acoustic model learning unit that learns an acoustic model from the static feature amount and the dynamic feature amount;
An HMM database storing the acoustic model;
A model conversion unit that converts the acoustic model into a linear dynamic system using the relationship between the static feature quantity and the dynamic feature quantity;
A linear dynamic system database storing the linear dynamic system;
A maximum likelihood calculation unit for obtaining a state sequence and a distribution sequence of the acoustic model so as to maximize the likelihood of the learning speech for the acoustic model, using the acoustic model stored in the HMM database;
A trajectory synthesis unit that synthesizes a trajectory using the linear dynamic system, the state series, and the distribution series;
A likelihood function maximization unit that obtains a maximization function that maximizes a likelihood function consisting of the static feature quantity, the dynamic feature quantity, the trajectory, and the variance of the state series and the distribution series; ,
A maximization function database for storing the maximization function;
A speech feature learning apparatus comprising: A feature quantity extraction unit that extracts static feature quantities and dynamic feature quantities from input speech;
An HMM database for storing acoustic models;
A linear dynamic system database storing the linear dynamic system;
A trajectory synthesis unit that synthesizes a trajectory using a linear dynamic system stored in the linear dynamic system database and dictionary data stored in the dictionary database;
A likelihood function which is a likelihood function which is a probability density of input speech with respect to the input of the trajectory or a maximum value of likelihood with respect to the input of the trajectory by substituting the trajectory into a maximization function stored in the maximization function database. A likelihood function generator for generating a degree function;
A likelihood calculating unit that calculates the likelihood by substituting the static feature amount and the dynamic feature amount into the likelihood function;
A speech recognition apparatus comprising: The speech recognition device according to claim 2,
The likelihood calculator is
Approximating means for generating an approximate likelihood function by approximating the likelihood function using the likelihood of the acoustic model, and substituting the static feature amount and the dynamic feature amount into the approximate likelihood function A speech recognition apparatus characterized by calculating likelihood. The speech recognition device according to claim 3,
The likelihood calculation unit further includes:
By approximating the error between the approximate likelihood function and the likelihood function using the likelihood of the acoustic model, an approximate error is generated, and by adding the approximate likelihood function and the approximate error, A correction means for generating a corrected likelihood function by correcting the likelihood function, and calculating the likelihood by substituting the static feature quantity and the dynamic feature quantity into the corrected likelihood function; A speech recognition apparatus characterized by being. The speech recognition apparatus according to any one of claims 2 to 4,
The trajectory synthesizing unit synthesizes the trajectory using a state equation and an observation equation. The speech recognition apparatus according to any one of claims 2 to 5,
The trajectory synthesizing unit synthesizes the trajectory using a Kalman filter. A feature extraction process for extracting static features and dynamic features from learning speech;
An acoustic model learning process in which an acoustic model is learned from the static feature quantity and the dynamic feature quantity and stored in an HMM database;
Using the relationship between the static feature quantity and the dynamic feature quantity, the acoustic model is converted into a linear dynamic system and stored in a linear dynamic system database; and
A maximum likelihood calculation process for obtaining a state sequence and a distribution sequence of the acoustic model so as to maximize the likelihood of the learning speech for the acoustic model using the acoustic model stored in the HMM database;
A trajectory synthesis process for synthesizing a trajectory using the linear dynamic system, the state sequence, and the distribution sequence;
A maximization function for maximizing a likelihood function consisting of the static feature quantity, the dynamic feature quantity, the trajectory, and the variance of the state series and the distribution series is obtained in a maximization function database. A likelihood function maximization process to remember,
A speech feature learning method characterized by comprising: A feature extraction process for extracting static and dynamic features from input speech;
A trajectory synthesis process for synthesizing a trajectory using a linear dynamic system stored in the linear dynamic system database and dictionary data stored in the dictionary database;
A likelihood function which is a likelihood function which is a probability density of input speech with respect to the input of the trajectory or a maximum value of likelihood with respect to the input of the trajectory by substituting the trajectory into the maximization function stored in the maximization function database. A likelihood function generation process for generating a degree function;
A speech recognition method comprising: a likelihood calculation step of calculating a likelihood by substituting the static feature amount and the dynamic feature amount into the likelihood function.   A speech feature learning program and a speech recognition program for causing a computer to execute each process of the speech feature learning device and the speech recognition device according to claim 1. A computer-readable recording medium in which the voice feature learning program and the voice recognition program according to claim 9 are recorded.

Images