JP2014013324A - Sound model adaptation device, sound model adaptation method, and sound model adaptation program - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide a sound model adaptation device, a sound model adaptation method, and a sound model adaptation program capable of performing noise adaptation of a sound model in lower computations amount without degrading adaptation accuracy.SOLUTION: A first noise adaptation part 20-1 increases sound model, which is adapted for noise, and performs noise adaptation. A second noise adaptation part 20-2 performs noise adaptation using linear approximation. An adaptation system selection part 10 selects the first noise adaptation part 20-1 or the second noise adaptation part 20-2 based on statistics value of noise, which is adapted to a sound model and the sound model.

Description

  The present invention relates to an acoustic model adaptation apparatus, an acoustic model adaptation method, and an acoustic model adaptation program for adapting an acoustic model to noise.

  Since the performance of the speech recognition apparatus is significantly deteriorated due to the influence of noise in actual operation, a noise resistance method is required. The cause of performance degradation is caused by the difference between the audio signal used during acoustic model learning (hereinafter referred to as learning data) and the audio signal that is to be recognized in actual operation (hereinafter referred to as test data). There is a discrepancy between the acoustic model and the test data. There is a model adaptation method as a noise proofing method for speech recognition aimed at suppressing such inconsistency.

  In the model adaptation method, the noise statistic (hereinafter referred to as noise statistic) included in the test data is reflected in the acoustic model, thereby bringing the mixed Gaussian distribution constituting the acoustic model closer to the distribution created by the test data. The noise statistic is, for example, the average or variance of the noise feature. As a model adaptation method, for example, there is a VTS (Vector Taylor Series) adaptation method (see, for example, Non-Patent Document 1). The VTS adaptation method performs first-order Taylor approximation of a nonlinear function that defines the relationship between speech, noise, and noise-added speech in an acoustic feature space such as the MFCC (Mel-Frequency Cepstial Coefficient), and provides a clean acoustic model (learning data). The acoustic model learned using clean speech is adapted to noise. As a result, the VTS adaptation method eliminates the complexity resulting from the non-linear function and performs low-computation noise adaptation using only linear computation.

  However, in the VTS adaptation method, the error of Taylor approximation becomes large and the adaptation accuracy is deteriorated in noise adaptation of a Gaussian distribution having an average in a region where the variance is large or the influence of nonlinearity is large. Therefore, learning is performed in advance by increasing the number of Gaussian distributions of the acoustic model to be adapted to the number necessary for recognition. Thereby, the variance of each Gaussian distribution becomes small, and the error by linear approximation can be made small. However, the disadvantage is that the amount of calculation increases by preparing in advance an acoustic model with an increased number of Gaussian distributions.

  As a method for suppressing degradation of adaptation accuracy when an acoustic model with an increased number of Gaussian distributions cannot be prepared, there is a UT (Unsented Transform) adaptation method (see, for example, Patent Document 1). In the UT adaptation method, a set of samples called “sigma points” is generated for each Gaussian distribution, noise is applied to each sample point, and a noise adaptive Gaussian distribution is generated. The generation of sample points follows the process of increasing the number of Gaussian distributions. As a result, the UT adaptation method can apply noise to the acoustic model with higher accuracy than the VTS adaptation method when an acoustic model with an increased number of Gaussian distributions cannot be prepared.

JP 2010-077865 A

A. Acero, L.M. Deng, T.A. Kristjansson, and J.M. Zhang, “HMM Adaptation using Vector Taylor Series for Noise Speech Recognition”, in Proc. ICSLP, Vol. 3, pp. 869-872, 2000.

  However, some of the Gaussian distributions constituting the clean acoustic model have a small adaptation error even if noise adaptation using linear approximation is performed. For this reason, for all Gaussian distributions of the clean acoustic model, a method that increases the Gaussian distribution and implements the VTS adaptation method or a UT adaptation method that generates sigma points requires a high accuracy but requires a large amount of computation. Applying is wasteful of computational complexity.

  Accordingly, an object of the present invention is to provide an acoustic model adaptation device, an acoustic model adaptation method, and an acoustic model adaptation program that can noise-adapt an acoustic model with a lower amount of computation without deteriorating adaptation accuracy. .

  An acoustic model adaptation apparatus according to the present invention is an acoustic model adaptation apparatus that generates a noise acoustic model by adapting an acoustic model to noise, and includes a first noise adaptation unit that performs noise adaptation by increasing the number of acoustic models adapted to noise. Select the first noise adaptor or the second noise adaptor based on the second noise adaptor that adapts to noise using linear approximation and the acoustic model and the noise statistic that adapts the acoustic model And an adaptive method selection unit.

  An acoustic model adaptation method according to the present invention is an acoustic model adaptation method for generating a noise acoustic model by adapting an acoustic model to noise, and based on the acoustic model and a noise statistic for adapting the acoustic model. It is characterized in that it is selected whether to perform noise adaptation by increasing the number of acoustic models adapted to, or to perform noise adaptation using linear approximation, and to perform noise adaptation based on the selection.

  An acoustic model adaptation program according to the present invention is an acoustic model adaptation program in an acoustic model adaptation apparatus that generates a noise acoustic model by adapting an acoustic model to noise. Based on the statistics, select whether to perform noise adaptation by increasing the number of acoustic models that adapt to noise, or to perform noise adaptation using linear approximation, and execute the process of performing noise adaptation based on the selection It is characterized by that.

  According to the present invention, an acoustic model can be noise-adapted with the same degree of adaptation accuracy as compared with a high-calculation and high-accuracy method, that is, without lowering the adaptation accuracy.

It is a block diagram which shows the structure in 1st Embodiment of the acoustic model adaptation apparatus by this invention. It is a flowchart which shows an example of operation | movement of the acoustic model adaptation apparatus in 1st Embodiment. It is a block diagram which shows the structure in 2nd Embodiment of the acoustic model adaptation apparatus by this invention. It is a flowchart which shows an example of operation | movement of the 1st noise adaptation part in 2nd Embodiment. It is explanatory drawing which shows an example of a structure of the acoustic model of a tree structure which shows the relationship between the Gaussian distribution set used for recognition, and the Gaussian distribution set used in the 1st noise adaptation part. It is a block diagram which shows the structure in 3rd Embodiment of the acoustic model adaptation apparatus by this invention. It is a flowchart which shows an example of operation | movement of the adaptive system selection part 303 in 3rd Embodiment. It is a block diagram which shows the minimum structure of the acoustic model adaptation apparatus by this invention. It is a block diagram which shows the other minimum structure of the acoustic model adaptation apparatus by this invention.

Embodiment 1. FIG.
A first embodiment of the present invention will be described below with reference to the drawings.

  FIG. 1 is a block diagram showing the configuration of the acoustic model adaptation apparatus according to the first embodiment of the present invention. As shown in FIG. 1, the acoustic model adaptation apparatus 100 includes a noise statistic acquisition unit 101, a Gaussian distribution acquisition unit 102, an adaptation method selection unit 103, a first noise adaptation unit 104, and a second noise adaptation. Unit 105 and a Gaussian distribution storage unit 106.

  As shown in FIG. 1, the acoustic model adaptation device 100 is connected to a clean acoustic model storage device 1 and a noise statistic storage device 2 that store information input by the acoustic model adaptation device 100. The acoustic model adaptation device 100 is connected to a noise adaptive acoustic model storage device 3 that stores information output from the acoustic model adaptation device 100.

  The noise statistic acquisition unit 101, the Gaussian distribution acquisition unit 102, the adaptation method selection unit 103, the first noise adaptation unit 104, the second noise adaptation unit 105, and the Gaussian distribution storage unit 106 are included in the acoustic model adaptation apparatus 100. This is realized by a CPU or the like provided.

  FIG. 2 is a flowchart illustrating an example of the operation of the acoustic model adaptation device 100 according to the first embodiment.

  As shown in FIG. 2, the noise statistic acquisition unit 101 acquires a noise statistic from the noise statistic storage device 2 (step S101). The Gaussian distribution acquisition unit 102 acquires the Gaussian distribution parameters constituting the clean acoustic model one by one from the clean acoustic model storage device 1 (step S102). Based on the noise statistic acquired in step S101 and the Gaussian distribution parameter acquired in step S102, the adaptive scheme selection unit 103 uses the scheme of the first noise adaptation unit 104 as the adaptation scheme, or second It is selected whether to use the method of the noise adaptation unit 105 (step S103). That is, it is selected which of the first noise adaptation unit 104 and the second noise adaptation unit 105 is to perform noise adaptation.

  When the adaptation scheme selection unit 103 selects the scheme of the first noise adaptation unit 104 (Yes in step S103), the first noise adaptation unit 104 performs noise adaptation on the Gaussian distribution parameter (step S104). When the adaptation scheme selection unit 103 selects the scheme of the second noise adaptation unit 105 (No in step S103), the second noise adaptation unit 105 performs noise adaptation on the Gaussian distribution parameter (step S105).

  The Gaussian distribution storage unit 106 stores the noise-adapted Gaussian distribution parameter (hereinafter referred to as a noise adaptive acoustic model) in the noise adaptive acoustic model storage device 3 (step S106).

  Next, the detail of each component with which the acoustic model adaptation apparatus 100 in this embodiment is provided is demonstrated.

  First, the details of the clean acoustic model storage device 1 and the noise statistic storage device 2 for storing information input by the acoustic model adaptation device 100 and the noise adaptive acoustic model storage device 3 for storing information output by the acoustic model adaptation device 100 will be described. explain. Next, the noise statistic acquisition unit 101, the Gaussian distribution acquisition unit 102, the adaptation method selection unit 103, the first noise adaptation unit 104, the second noise adaptation unit 105, and the Gauss, which are components of the acoustic model adaptation device 100. Details of the distribution storage unit 105 will be described.

  The clean acoustic model storage device 1 stores a clean acoustic model learned using clean speech as learning data. Hereinafter, the feature quantity used for learning and recognition is the MFCC 13 dimension including the C0 feature quantity corresponding to power. The C0 feature amount is an MFCC 13-dimensional zeroth-order element. Note that a vector of 39 dimensions in total composed of MFCC 13 dimensions, its primary dynamic components (13 dimensions), and secondary dynamic components (13 dimensions) may be used. In addition, if the feature-value corresponding to power is included, not only what was illustrated but all the feature-values can be used. In the following description, the mean and variance in the Gaussian distribution of the clean acoustic model are expressed as follows.

μ _{x, i} , Σ _{x, i} (i = 1,..., N)

  Here, the subscript x indicates a parameter of the clean acoustic model. The subscript i indicates the distribution ID number of the Gaussian distribution. N indicates the total number of Gaussian distributions included in the clean acoustic model.

  The noise statistics storage device 2 stores noise statistics used for adaptation. In the present embodiment, the noise statistic storage device 2 stores the average and variance of noise in the same feature quantity region as that used for recognition as noise statistic. The feature amount area refers to a set or space of feature amounts created by a certain process. In the following description, the mean and variance of noise are expressed as follows.

μ _n , Σ _n

  The noise adaptive acoustic model storage device 3 stores the noise adaptive acoustic model adapted by the acoustic model adaptation device 100. In the following description, the mean and variance, which are each Gaussian distribution parameter of the acoustic model, are expressed as follows.

μ _{y￣, i} , Σ _{y￣, i} (i = 1,..., N)

  Here, the subscript y￣ (￣ is on y. The same applies hereinafter) indicates that it is a parameter of the noise adaptive acoustic model. The subscript i indicates the distribution ID number of the Gaussian distribution.

The noise statistic acquisition unit 101 acquires the noise statistic μ _n and Σ _n stored in the noise statistic storage device 2, and the adaptive method selection unit 103, the first noise adaptation unit 104, and the second It passes to the noise adaptation unit 105.

The Gaussian distribution acquisition unit 102 stores N pieces of Gaussian distribution parameters μ _{x, i} , Σ _{x, i} (i = 1,..., N) of the clean acoustic model stored in the clean acoustic model storage device 1 one by one. Obtained and passed to the adaptive method selection unit 103.

The adaptive scheme selection unit 103 receives the clean acoustic model Gaussian distribution parameters {μ _{x, i} , Σ _{x, i} } and the noise statistics {μ, respectively passed from the Gaussian distribution acquisition unit 102 and the noise statistics acquisition unit 101. Compare _n , Σ _n }. The adaptation method selection unit 103 determines whether the first noise adaptation unit 104 should perform noise adaptation on the Gaussian distribution parameters {μ _{x, i} , Σ _{x, i} } according to the comparison result, or determines whether the second noise adaptation unit 104 In 105, it is selected whether noise adaptation should be performed. As shown below, the comparison introduces a scalar function Comp (μ _{x, i} , Σ _{x, i} , μ _n , Σ _n ), and checks whether this is a value greater than or equal to a threshold Th. Do that.

When Expression 1 is satisfied, the adaptation method selection unit 103 passes the Gaussian distribution parameter {μ _{x, i} , Σ _{x, i} } to the second noise adaptation unit 105.

When Expression 2 is satisfied, the adaptive method selection unit 103 passes the Gaussian distribution parameters {μ _{x, i} , Σ _{x, i} } to the first noise adaptation unit 104.

Next, a specific example of Comp (μ _{x, i} , Σ _{x, i} , μ _n , Σ _n ) will be described.

When the adaptive method is selected, for example, a difference in C0 feature value between the average μ _{x, i} of the Gaussian distribution and the average μ _{n of the} noise may be examined. At this time, if the C0 feature quantity of the average μ _{x, i} of the Gaussian distribution and the C0 feature quantity of the average μ _n of the noise are (μ _{x, i} ) ₀ and (μ _n ) ₀ respectively, the comparison function is expressed by the following equation (3). become that way.

  This utilizes the fact that the C0 feature amount is a feature amount related to power, and that the magnitude of the difference between the power of speech and noise affects the degree of nonlinearity of the nonlinear function indicating the noise-added speech.

Further, when an adaptive method is selected, a C0 feature value of Gaussian distribution variance may be used. This is because the adaptation error due to linear approximation also depends on the magnitude of the variance of the Gaussian distribution to be adapted. Here, (f (x, n)) ₀ is a non-linear function indicating the C0 feature amount of noise-added speech (x and n are speech and noise feature amounts, respectively). Note that f (x, n) is specifically expressed as in Expression 4. In Equation 4, D represents a DCT matrix, and D ⁻¹ represents a DCT inverse matrix.

_{(F (x, n))} 0 of _x = μ _{x, i,} the C0 feature quantity Taylor approximation formula in _n = μ n, expressed as _{(f¯μ x, i, μ n} (x, n)) 0 (￣ is placed on f. Μ _{x, i} and μ _n represent subscripts of f. The same shall apply hereinafter.) Here, f￣μ _{x, i} , μ _n is expressed as in Equation 5.

F _i in Equation 5 represents the Jacobian at x = μ _{x, i} and n = μ _{n with} respect to x of f (x, n). Two sigma points related to the C0 feature amount derived from the variance Σ _{x, i} of the Gaussian distribution are expressed as σ _1,0 = + √ (Σ _{x, i} ) ₀ , σ _2,0 = −√ (Σ _{x, i} ) _Set to ₀ . Then, the comparison function is expressed as shown in Equation 6. Note that (Σ _{x, i} ) ₀ is a vector in the 0th column.

It should be noted that Comp (μ _{x, i} , Σ _{x, i} , μ _n , Σ _n ) of Equation 3 and Equation 6 is _expressed as Comp ₁ (μ _{x, i} , Σ _{x, i} , μ _n , Σ _n ), Comp, respectively. ₂ (μ _{x, i} , Σ _{x, i} , μ _n , Σ _n ), and the respective linear sums may be used as comparison functions. Equation 7 shows the comparison function at that time. w ₁ and w ₂ represent weights.

  In addition, regarding Equation 6, a feature amount other than the C0 feature amount may be used. For example, if all the feature values are used, it can be expressed as Equation 8.

Here, J represents the number of dimensions of the feature quantity, and is σ _{1, j} = + √ (Σ _{x, i} ) _j , σ _{2, j} = −√ (Σ _{x, i} ) _j . Note that (Σ _{x, i} ) _j represents a column vector of the j-th column of the matrix Σ _{x, i} . Note that the optimum threshold Th in Equations 1 and 2 and the weight in combination in Equation 7 may be experimentally obtained.

Second noise adaptation unit 105 uses the VTS adaptive _{method, μ y¯, i, Σ y¯} , and outputs the _i. A conversion formula according to the VTS adaptive method when the feature amount is the MFCC 13-dimensional including the C0 feature amount corresponding to the power is expressed as follows.

The second noise adaptation unit 105 _passes {μ _{y ￣, i} , Σ _{y ￣, i} } to the Gaussian distribution storage unit 106. In addition, the conversion formula by the VTS adaptive method of the parameter {Δμ _{x, i} , ΔΣ _{x, i} } relating to the primary dynamic feature quantity and the parameter relating to the secondary dynamic feature quantity {ΔΔμ _{x, i} , ΔΔΣ _{x, i} } is as follows: It is expressed as

Here, ΔΣ _n and ΔΔΣ _n indicate the variance of the primary dynamic feature value and the secondary dynamic feature value of noise.

The first noise adaptation unit 104 outputs μ _{y ￣, i} , Σ _{y ￣, i} using the UT adaptation method. A conversion equation based on the UT adaptation method when the feature amount is the MFCC 13-dimensional including the C0 feature amount corresponding to the power is expressed as follows.

Here, S _k represents a sigma point and is expressed as in Expression 17.

μ _{s, i} , Σ _{s, i} are expressed as in Equation 18.

Here, D = 13. Further, (√Σ) _k represents a vector in the k-th column of the matrix Σ. And w _k = 1 / 4D. Then, {μ _{y￣, i} , Σ _{y, i} } is passed to the Gaussian distribution storage unit 106. In addition, the conversion formula by the UT adaptive method of the parameter {Δμ _{x, i} , ΔΣ _{x, i} } relating to the primary dynamic feature quantity and the parameter relating to the secondary dynamic feature quantity {ΔΔμ _{x, i} , ΔΔΣ _{x, i} } is as follows: It is expressed as

F ′ _i indicates a Jacobian of μ _{y￣, i} in Expression 15 regarding μ _{x, i} .

The Gaussian distribution storage unit 106 stores noise-adapted Gaussian distribution parameters {μ _{y ￣, i} , Σ _{y ￣, i} } in the noise adaptive acoustic model storage device 3.

  As described above, in the present embodiment, the UT adaptation method is applied when the difference between the average C0 feature value of the Gaussian distribution and the average C0 feature value of the noise is a certain value or more. Yes. Therefore, according to the present embodiment, the acoustic model can be noise-adapted with the same degree of adaptation accuracy and with a lower amount of computation than when the UT adaptation method is used for the total Gaussian distribution.

Embodiment 2. FIG.
Hereinafter, a second embodiment of the present invention will be described with reference to the drawings.

  FIG. 3 is a block diagram showing the configuration of the acoustic model adaptation apparatus according to the second embodiment of the present invention.

  The configuration of the acoustic model adaptation device 200 is the same as the configuration of the acoustic model adaptation device 100 in the first embodiment. However, as illustrated in FIG. 3, the acoustic model adaptation device 200 includes a detailed Gaussian distribution acquisition unit 207.

  The acoustic model adaptation apparatus 200 includes an adaptation method selection unit 203 and a first noise adaptation unit 204 instead of the adaptation method selection unit 103 and the first noise adaptation unit 104.

  In addition to the clean acoustic model storage device 1 and the noise statistic storage device 2, the acoustic model adaptation device 200 is connected to a detailed clean acoustic model storage device 7 that stores information input by the acoustic model adaptation device 200.

  The adaptation method selection unit 203, the first noise adaptation unit 204, and the detailed Gaussian distribution acquisition unit 207 are realized by a CPU or the like included in the acoustic model adaptation device 200.

  In addition, since the outline | summary of operation | movement of the acoustic model adaptation apparatus 200 in this embodiment is the same as the outline | summary of operation | movement of the acoustic model adaptation apparatus 100 shown in FIG. 2, description is abbreviate | omitted.

  Next, the detail of each component with which the acoustic model adaptation apparatus 200 in this embodiment is provided is demonstrated.

  The detailed clean acoustic model storage device 7 stores the detailed clean acoustic model learned by further increasing the parameters of the clean acoustic model stored in the clean acoustic model storage device 1. In the following description, the mean and variance in the Gaussian distribution of the detailed clean acoustic model are respectively expressed as follows.

μ _{x, ij} , Σ _{x, ij} (i = 1,..., N, j = 1,..., N)

Here, μ _{x, ij} , Σ _{x, ij} are j-th Gaussian distribution parameters derived from a Gaussian distribution with a distribution ID i in the clean acoustic model.

The adaptive scheme selection unit 203 receives the clean acoustic model Gaussian distribution parameters {μ _{x, i} , Σ _{x, i} } and the noise statistics {μ, respectively passed from the Gaussian distribution acquisition unit 102 and the noise statistics acquisition unit 101. Compare _n , Σ _n }. The comparison may be the same method as the adaptive method selection unit 103 in the first embodiment. The adaptation method selection unit 203 determines whether the first noise adaptation unit 204 should perform noise adaptation on the Gaussian distribution parameters {μ _{x, i} , Σ _{x, i} } according to the comparison result, or the second noise adaptation unit. In 105, it is selected whether noise adaptation should be performed. When the first noise adaptation unit 204 is selected for noise adaptation, the adaptation method selection unit 203 passes the Gaussian distribution parameters {μ _{x, i} , Σ _{x, i} } of the clean acoustic model to the detailed Gaussian distribution acquisition unit 207.

The detailed Gaussian distribution acquisition unit 207 receives N _i Gaussian distribution parameters {μ _{x, ij} , Σ based on the ID number i of the Gaussian distribution parameters {μ _{x, i} , Σ _{x, i} } received from the adaptive method selection unit 203. _{x, ij} } (j = 1,..., N _i ) is acquired from the detailed clean acoustic model storage device 7. The detailed Gaussian distribution acquisition unit 207 then passes the N _i Gaussian distribution parameters {μ _{x, ij} , Σ _{x, ij} } (j = 1,..., N _i ) to the first noise adaptation unit 204.

The first noise adaptation unit 204 performs noise adaptation on N _i Gaussian distribution parameters {μ _{x, ij} , Σ _{x, ij} } (j = 1,..., N _i ). The first noise adaptation unit 204 outputs Gaussian distribution parameters μ _{y ￣, i} , Σ _{y ￣, i} that are noise-adapted.

  FIG. 4 is a flowchart illustrating an example of the operation of the first noise adaptation unit 204 in the second embodiment.

As shown in FIG. 4, the first noise adaptation unit 204 includes N _i Gaussian distributions based on the ID numbers i of the Gaussian distribution parameters {μ _{x, i} , Σ _{x, i} } received from the adaptation method selection unit 203. Parameters {μ _{x, ij} , Σ _{x, ij} } (j = 1,..., N _i ) are acquired (step S2041).

  FIG. 5 is an explanatory diagram showing an example of the configuration of a tree-structured acoustic model (tree-structured acoustic model) showing the relationship between the Gaussian distribution set used for recognition and the Gaussian distribution set used in the first noise adaptation unit 204.

First noise adaptation unit 204, for each Gaussian distribution formula 9, using the VTS adaptive method shown in Equation 10, the noise feature quantity _{{μ n, Σ n} N} i number of Gaussian distribution adapted to Parameters {μ _{y ￣, ij} , Σ _{y ￣, ij} } (j = 1,..., N _i ) are acquired (step S2042).

The first noise adaptation unit 204 integrates N _i noise-adapted Gaussian distribution parameters {μ _{y ￣, ij} , Σ _{y ￣, ij} } (j = 1,..., N _i ) into one Gaussian distribution parameter. Then, one noise-adapted Gaussian distribution parameter {μ _{y ￣, i} , Σ _{y ￣, i} } is acquired (step S2042).

Here, w ′ _j is a mixing weight, and Σ _{j = 1} ^Ni (w ′ _j ) = 1. Σ _{j = 1} ^Ni indicates the total sum from j = 1 to j = N _i . The mixing weight may be determined experimentally, or may be equal probability, that is, 1 / N _i . Then, {μ _{y￣, i} , Σ _{y, i} } is passed to the Gaussian distribution storage unit 106.

  As described above, in this embodiment, when the difference between the average C0 feature value of the Gaussian distribution and the average C0 feature value of the noise is equal to or greater than a certain value, the number of Gaussian distributions is increased to increase the VTS. I try to adapt. Therefore, according to the present embodiment, it is possible to prevent an increase in the amount of calculation in adaptation of the total Gaussian distribution. That is, the acoustic model can be noise-adapted with a lower amount of computation. Further, the adaptation accuracy is not deteriorated.

Embodiment 3. FIG.
Hereinafter, a third embodiment of the present invention will be described with reference to the drawings.

  FIG. 6 is a block diagram showing the configuration of the acoustic model adaptation apparatus according to the third embodiment of the present invention.

  The configuration of the acoustic model adaptation device 300 is the same as the configuration of the acoustic model adaptation device 100 in the first embodiment.

  However, as illustrated in FIG. 6, the acoustic model adaptation device 300 includes an adaptation method selection unit 303 instead of the adaptation method selection unit 103. In addition, the acoustic model adaptation apparatus 300 includes a first noise adaptation unit 3041, a second noise adaptation unit 3042, and a third noise instead of the first noise adaptation unit 104 and the second noise adaptation unit 105. An adaptation unit 3043 (not shown) and a fourth noise adaptation unit 3044 are provided.

  The outline of the operation of the acoustic model adaptation apparatus 300 in the present embodiment is the same as the outline of the operation of the acoustic model adaptation apparatus 100 shown in FIG. However, branches corresponding to steps S104 and S105 shown in FIG. 2 increase according to the number of noise adaptation units.

  Note that the adaptation method selection unit 303, the first noise adaptation unit 3041, the second noise adaptation unit 3042, the third noise adaptation unit 3043, and the fourth noise adaptation unit 3044 are a CPU or the like provided in the acoustic model adaptation device 300. It is realized by.

  Next, the detail of each component with which the acoustic model adaptation apparatus 300 in this embodiment is provided is demonstrated.

The adaptive scheme selection unit 303 compares the Gaussian distribution parameter {μ _{x, i} , Σ _{x, i} } of the clean acoustic model with the noise statistic {μ _n , Σ _n }. The adaptation scheme selection unit 303 sets the Gaussian distribution parameter {μ _{x, i} , Σx _{, i} } to which noise adaptation unit from the first noise adaptation unit to the fourth noise adaptation unit according to the comparison result. Determine whether noise should be adapted. At the time of determination, the adaptive method selection unit 303 uses the scalar function Comp (μ _{x, i} , Σ _{x, i} , μ _n , Σ _n ) described in the first embodiment. In addition, the adaptive scheme selection unit 303 uses two threshold values Th ₁ and Th ₂ (where Th ₁ <Th ₂ ). Th ₁ and Th ₂ may be experimentally obtained.

  FIG. 7 is a flowchart illustrating an example of the operation of the adaptive method selection unit 303 according to the third embodiment.

As shown in FIG. 7, the adaptive scheme selection unit 303 acquires the Gaussian distribution parameters {μ _{x, i} , Σ _{x, i} } of the clean acoustic model from the Gaussian distribution acquisition unit 102 (step S3031), and acquires noise statistics. {Μ _n , Σ _n } is acquired from the unit 101 (step S3032).

First, the adaptive method selection unit 303 compares the _first threshold Th ₁ with Comp (μ _{x, i} , Σ _{x, i} , μ _n , Σ _n ) (step S3033).

When Expression 25 is satisfied (Yes in Step S3033), the adaptive method selection unit 303 passes the Gaussian distribution parameters {μ _{x, i} , Σ _{x, i} } to the first noise adaptation unit 3041 (Step S3036). Then, the first noise adaptation unit 3041 performs noise adaptation of the Gaussian distribution parameters {μ _{x, i} , Σ _{x, i} }.

Otherwise (No in step S3033), the adaptive method selection unit 303 compares the second threshold Th ₂ with Comp (μ _{x, i} , Σ _{x, i} , μ _n , Σ _n ) ( Step S3034).

When Expression 26 is satisfied (Yes in Step S3034), the adaptive method selection unit 303 passes the Gaussian distribution parameters {μ _{x, i} , Σ _{x, i} } to the second noise adaptation unit 3042 (Step S3037). Then, the second noise adaptation unit 3042 performs noise adaptation of the Gaussian distribution parameters {μ _{x, i} , Σ _{x, i} }.

If not (No in step S3034), it is considered that the power difference between the clean acoustic model parameter and the noise statistic is large. Specifically, there are two cases where the noise is large and the parameters of the clean acoustic model are large. In order to distinguish these two ways, (μ _{x, i} ) ₀ and (μ _n ) ₀ are compared (step S3035).

When Expression 27 is satisfied (Yes in Step S3035), the adaptive method selection unit 303 passes the Gaussian distribution parameters {μ _{x, i} , Σ _{x, i} } to the third noise adaptation unit 3043 (Step S3038). Then, the third noise adaptation unit 3043 performs noise adaptation of the Gaussian distribution parameters {μ _{x, i} , Σ _{x, i} }.

Otherwise (No in step S3035), the adaptation method selection unit 303 passes the Gaussian distribution parameters {μ _{x, i} , Σ _{x, i} } to the fourth noise adaptation unit 3044 (step S3039). Then, the fourth noise adaptation unit 3044 performs noise adaptation of the Gaussian distribution parameters {μ _{x, i} , Σ _{x, i} }.

The first noise adaptation unit 3041 applies the UT adaptation method shown in Equations 15 and 16, and outputs {μ _{y ￣, i} , Σ _{y ￣, i} }.

The second noise adaptation unit 3042 applies the VTS adaptation method shown in Equations 9 and 10, and outputs {μ _{y ￣, i} , Σ _{y￣, i} }.

The third noise adaptation unit 3043 outputs {μ _{x, i} , Σ _{x, i} } as {μ _{y ￣, i} , Σ _{y ￣, i} }.

The third noise adaptation unit 3044 outputs {μ _n , Σ _n } as {μ _{y ￣, i} , Σ _{y ￣, i} }.

{Μ _{y￣, i} , Σ _{yΣ, i} } output from each noise adaptation unit is stored in the Gaussian distribution storage unit 106.

  As described above, according to the present embodiment, the acoustic model according to the first embodiment is provided by including the third noise adaptation unit and the fourth noise adaptation unit that are lower in computation than the VTS adaptation method. Compared to the adaptive device 100, the amount of calculation can be reduced while maintaining accuracy.

  In the present embodiment, the model adaptation apparatus including four noise adaptation units is taken as an example, but the number of noise adaptation units is not limited to four. That is, the model adaptation apparatus 300 may include any number of noise adaptation units having different calculation amounts and adaptation accuracy. For example, a number of noise adaptation units corresponding to the approximate granularity of adaptation may be provided.

  FIG. 8 is a block diagram showing the minimum configuration of the acoustic model adaptation apparatus according to the present invention. FIG. 9 is a block diagram showing another minimum configuration of the acoustic model adaptation apparatus according to the present invention.

  As shown in FIG. 8, the acoustic model adaptation device (corresponding to the acoustic model adaptation device 100 shown in FIG. 1) is an acoustic model adaptation device that generates a noise acoustic model by adapting an acoustic model to noise. The first noise adapting unit 20-1 (corresponding to the first noise adapting unit 104 in the acoustic model adapting apparatus 100 shown in FIG. 1) that increases the number of acoustic models adapted to noise and adapts to noise, and noise using linear approximation The adaptive second noise adaptation unit 20-2 (corresponding to the second noise adaptation unit 105 in the acoustic model adaptation apparatus 100 shown in FIG. 1), the acoustic model, and the noise statistic for adapting the acoustic model. Based on this, the adaptive scheme selection unit 10 that selects the first noise adaptation unit 20-1 or the second noise adaptation unit 20-2 (corresponding to the adaptation scheme selection unit 103 in the acoustic model adaptation apparatus 100 shown in FIG. 1). When Provided.

  In the above-described embodiment, the following acoustic model adaptation device is also disclosed.

(1) The adaptation method selection unit 10 determines the power difference between the speech and noise used when learning the acoustic model based on the acoustic model and the statistical amount of noise to which the acoustic model is adapted, The acoustic model adaptation apparatus that selects the first noise adaptation unit 20-1 when the difference is larger than a predetermined threshold value, and selects the second noise adaptation unit 20-2 when the difference is less than the threshold value.

  According to such a configuration, a Gaussian distribution having an average in a region where the influence of nonlinearity is large can be accurately recognized. This is because the magnitude of the difference between the power of speech and noise affects the degree of nonlinearity of the nonlinear function indicating the noise-added speech.

(2) The acoustic model includes a Gaussian distribution, and the first noise adaptation unit 20-1 generates a plurality of sigma points for each Gaussian distribution and performs noise adaptation for each of the plurality of sigma points.

  According to such a configuration, the UT adaptation method can be applied according to the Gaussian distribution to be adapted. For example, the UT adaptation method can be applied when noise-adapting a Gaussian distribution having an average in a region where the influence of nonlinearity is large, and the VTS adaptation method with a small calculation amount can be applied in other cases. Therefore, it is possible to prevent an increase in the amount of calculation in the adaptation of the total Gaussian distribution.

(3) The acoustic model includes a Gaussian distribution, and the first noise adaptation unit 20-1 (corresponding to the first noise adaptation unit 204 in the acoustic model adaptation apparatus 200 shown in FIG. 2) is derived from the Gaussian distribution. An acoustic model adaptation device that adapts multiple Gaussian distributions to noise.

  According to such a configuration, the Gaussian distribution parameter can be increased according to the adaptive Gaussian distribution. Therefore, it is possible to prevent an increase in the amount of calculation in the adaptation of the total Gaussian distribution.

(4) As shown in FIG. 9, the third noise adaptation unit 20-3 (third noise adaptation unit 3043 (not shown) in the acoustic model adaptation apparatus 300 shown in FIG. 6) outputs the acoustic model as a noise acoustic model. ), And a fourth noise adaptation unit 20-4 (corresponding to the fourth noise adaptation unit 3044 in the acoustic model adaptation apparatus 300 shown in FIG. 6) that outputs a noise statistic as a noise acoustic model. The adaptive method selection unit 10 (corresponding to the adaptive method selection unit 303 in the acoustic model adaptation apparatus 300 shown in FIG. 6) has a predetermined power difference between speech and noise used when learning the acoustic model. If the voice power is greater than the second threshold (corresponding to the threshold Th ₂ ), the third noise adaptation unit 20-3 is selected, and if it is smaller, the fourth noise adaptation unit 20 is selected. -4 to select the acoustic mode Dell adaptive device.

  According to such a configuration, it is possible to prevent an increase in the amount of calculation in adaptation of the entire Gaussian distribution. This is because, when the power difference between the clean acoustic model parameter and the noise statistic is large, noise adaptation can be performed with a smaller amount of computation.

(5) An acoustic model adaptation device that generates a noise acoustic model by adapting an acoustic model to noise, and has a plurality of noise adaptation units (for example, the model adaptation device 300 shown in FIG. Equivalent to the first noise adaptation unit 3041, the second noise adaptation unit 3042, the third noise adaptation unit 3043, and the fourth noise adaptation unit 3044), and the noise statistic adapted to the acoustic model. An acoustic model adaptation apparatus comprising: an adaptation scheme selection unit that selects any one of a plurality of noise adaptation units based on the above.

  According to such a configuration, the noise model can be noise-adapted with a smaller amount of computation without degrading the adaptation accuracy as compared with a method with high computation and high accuracy.

DESCRIPTION OF SYMBOLS 1 Clean acoustic model memory | storage device 2 Noise statistic memory | storage device 3 Noise adaptive acoustic model memory | storage device 7 Detailed clean acoustic model memory | storage device 10, 103, 203, 303 Adaptive system selection part 20-1, 104, 204, 3041 1st noise Adaptation unit 20-2, 105, 3042 Second noise adaptation unit 20-3 Third noise adaptation unit 20-4, 3044 Fourth noise adaptation unit 100, 200, 300 Model adaptation device 101 Noise statistics acquisition unit 102 Gaussian distribution acquisition unit 106 Gaussian distribution storage unit 207 Detailed Gaussian distribution acquisition unit

Claims (8)

An acoustic model adaptation device for generating a noise acoustic model by adapting an acoustic model to noise,
A first noise adaptation unit that adapts noise by increasing the number of acoustic models adapted to noise;
A second noise adaptation unit for noise adaptation using linear approximation;
An acoustic system comprising: an acoustic model and an adaptive method selection unit that selects the first noise adaptation unit or the second noise adaptation unit based on a statistical amount of noise to which the acoustic model is adapted. Model adaptation device. The adaptation method selection unit determines a power difference between the speech and the noise used when learning the acoustic model based on the acoustic model and a noise statistic for adapting the acoustic model, and determines the power difference. The acoustic model adaptation apparatus according to claim 1, wherein when the value is larger than a predetermined threshold, the first noise adaptation unit is selected, and when the value is equal to or less than the threshold, the second noise adaptation unit is selected. The acoustic model includes a Gaussian distribution,
The acoustic model adaptation device according to claim 1, wherein the first noise adaptation unit generates a plurality of sigma points for each of the Gaussian distributions and performs noise adaptation for each of the plurality of sigma points. The acoustic model includes a Gaussian distribution,
The acoustic model adaptation device according to claim 1, wherein the first noise adaptation unit adapts a plurality of Gaussian distributions derived from the Gaussian distribution to noise. A third noise adaptation unit that outputs the acoustic model as a noise acoustic model;
A fourth noise adaptation unit that outputs noise statistics as a noise acoustic model,
The adaptive method selection unit determines whether the power of the voice is larger when the difference in power between the voice used when learning the acoustic model and the noise is equal to or greater than a predetermined second threshold. The acoustic model adaptation device according to any one of claims 1 to 4, wherein a third noise adaptation unit is selected, and if it is small, a fourth noise adaptation unit is selected. An acoustic model adaptation device for generating a noise acoustic model by adapting an acoustic model to noise,
A plurality of noise adaptation units each having different computational complexity and adaptation accuracy;
An acoustic model adaptation apparatus comprising: an acoustic model and an adaptation method selection unit that selects any one of the plurality of noise adaptation units based on a noise statistic adapted to the acoustic model. An acoustic model adaptation method for generating a noise acoustic model by adapting an acoustic model to noise,
Based on the acoustic model and the noise statistic to which the acoustic model is adapted, it is selected whether noise adaptation is performed by increasing the number of acoustic models adapted to noise, or noise adaptation is performed using linear approximation. An acoustic model adaptation method characterized by performing noise adaptation based on selection. An acoustic model adaptation program in an acoustic model adaptation apparatus for generating a noise acoustic model by adapting an acoustic model to noise,
On the computer,
Based on the acoustic model and the noise statistic to which the acoustic model is adapted, it is selected whether noise adaptation is performed by increasing the number of acoustic models adapted to noise, or noise adaptation is performed using linear approximation. An acoustic model adaptation program for executing a noise adaptation process based on the selection.

Images