JP2001249681A - Device and method for adapting model, recording medium, and pattern recognition device - Google Patents

Abstract

PROBLEM TO BE SOLVED: To improve recognition performance. SOLUTION: A voiceless acoustic model correction part 7 performs adaptation of a voiceless acoustic model, which is an acoustic model presenting a voiceless state, based on voice data observed in a section immediately before a voice recognition section to be voice-recognized, and the degree of freshness showing freshness of the acoustic data.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a model adaptation apparatus, a model adaptation method, a recording medium, and a pattern recognition apparatus, and more particularly, to a model adaptation apparatus and a model adaptation method suitable for performing, for example, speech recognition. , A recording medium, and a pattern recognition device.

[0002]

2. Description of the Related Art Conventionally, there has been known a method of recognizing a word or the like uttered in a noise environment. A typical method is a PMC (Parallel Model Combination).
Method, SS / NSS (Spectral Subtraction / Nonlinear Spectral
Subtraction) method, SFE (Stochastic Feature Extraction)
There is a law.

[0003] Since the PMC method directly incorporates information of environmental noise into an acoustic model, the recognition performance is good, but the calculation cost is high. That is, in the PHC method, since a high-level operation is required, the scale of the apparatus is increased, and the time required for the processing is increased. In the SS / NSS method, the environmental noise is removed at the stage of extracting the feature amount of the voice data, so the calculation cost is lower than the PMC method.
Widely used. In the SFE method, similar to the SS / NSS method,
At the stage of extracting the feature amount of the audio signal including the environmental noise, the environmental noise is removed, and a feature amount represented by a probability distribution is extracted. In the SFE method, the feature of speech is extracted as a distribution in the feature space. Thus, the SS / NSS method in which the feature of speech is extracted as a point in the feature space, and the PMC method. different.

In any of the above-described methods, after the feature amount of the voice is extracted, it is determined which of the acoustic models corresponding to a plurality of words or the like registered in advance is most suitable. Then, a word corresponding to the most suitable acoustic model is output as a recognition result.

The details of the SFE method are described in Japanese Patent Application Laid-Open No. Hei 11-133992 (Japanese Patent Application No. 9-300979) previously filed by the present applicant. PMC method, SS / NSS
For comparison of the performance of the SFE method and the SFE method, see, for example, "H.
Pao, H. Honda, K. Minamino, M. Omote, H. Ogawa and NI
wahashi, Stochastic Feature Extraction for Improvi
ng Noise Robustness in Speech Recognition, Proceed
ings of the 8th Sony Research Forum, SRF98-234, p
p. 9-14, October 1998 "," N. Iwahashi, H. Pao, H. Hond
a, K. Minamino and M. Omote, Stochastic Features for
Noise Robust in Speech Recognition, ICASSP'98 Pro
ceedings, pp.633-636, May, 1998 "," N. Iwahashi, HP
ao (presented), H.Honda, K.Minamino and M.Omote, No
ise Robust Speech Recognition Using Stochastic Rep
resentation of Features, ASJ'98-Spring Proceeding
s, pp.91-92, March, 1998 "," N.iwahashi, H.Pao H.Ho
nda, K. Minamino and M. Omote, Stochastic Represetat
ion of Feature for Noise Robust Speech Recognitio
n, Technical Report of IEICE, pp.19-24, SP97-97 (19
98-01) and the like.

[0006]

However, the above-mentioned SFE
In the law, etc., the environmental noise is not directly reflected in the speech recognition stage, that is, since the information of the environmental noise is not directly taken into the silent acoustic model, the silent When there is a section, there is a problem that the recognition performance is reduced due to the silent section.

More specifically, because information on environmental noise is not directly taken into the silent acoustic model,
If the time from the start of the speech recognition to the start of the utterance becomes long, there is a problem that the recognition performance is reduced.

The present invention has been made in view of such a situation. By updating (correcting) a silent acoustic model using information of environmental noise, utterance is started from the time when speech recognition is started. It is possible to prevent the recognition performance from deteriorating as the time until the time becomes longer.

[0009]

According to the present invention, there is provided a model adapting apparatus for adapting a predetermined model used for pattern recognition based on extracted data in a predetermined section and freshness indicating the freshness of the extracted data. It is characterized by having an adaptation means.

[0010] The model adaptation method of the present invention is characterized by comprising a model adapting step of adapting a predetermined model based on extracted data in a predetermined section and freshness indicating the freshness of the extracted data.

[0011] The recording medium of the present invention stores a program having a model adaptation step of adapting a predetermined model based on extracted data in a predetermined section and freshness indicating the freshness of the extracted data. It is characterized by.

[0012] The pattern recognition apparatus of the present invention is characterized by comprising model adaptation means for adapting a predetermined model based on extracted data in a predetermined section and freshness indicating the freshness of the extracted data.

[0013] In the model adaptation apparatus and model adaptation method, recording medium, and pattern recognition apparatus of the present invention,
A predetermined model is adapted based on extracted data in a predetermined section and freshness indicating the freshness of the extracted data.

[0014]

FIG. 1 shows a configuration example of an embodiment of a speech recognition apparatus to which the present invention is applied. In this voice recognition device, the microphone 1 collects the uttered voice to be recognized together with the environmental noise and outputs it to the framing unit 2. The framing unit 2 extracts audio data input from the microphone 1 at a predetermined time interval (for example, 10 ms), and
Output as frame data. The audio data in units of one frame output from the framing unit 2 is supplied to the noise observation section extraction unit 3 and the feature extraction unit 5 as an observation vector a having each of the time-series audio data constituting the frame as a component. You.

Here, the observation vector, which is the audio data of the t-th frame, is represented as a (t).

The noise observation section extraction unit 3 buffers the audio data in frame units input from the framing unit 2 for a predetermined time (2M frames or more).
As shown in, the from the timing t _b of the speech switch 4 is turned on until the M frames only previous timing t _a as a noise observation interval Tn, extracts the observation vector a of 2M frames in the noise observation interval Tn Then, it outputs to the feature extraction unit 5 and the silent acoustic model correction unit 7. In the present embodiment, the noise observation section includes a noise observation section Tm for extracting a feature distribution described later,
2 of the noise observation section Tn for adapting the acoustic model
Each of the noise observation sections Tm and Tn is an M frame. However, the number of frames in the noise observation sections Tm and Tn does not need to be the same.

The utterance switch 4 is turned on by the user when the user starts uttering, and is turned off when ending the utterance. Therefore, the speech switch 4 is in the voice data on the timing t _b before (noise observation interval Tn), speech is not included, only ambient noise is present. Further, from the timing t _b of the speech switch 4 is turned on until the timing t _d the speech switch 4 is turned off, is a voice recognition section, the audio data in the speech recognition section is subjected to speech recognition .

The feature extraction unit 5 includes a noise observation section extraction unit 3
Based on the audio data only ambient noise in the first half of the noise observation interval Tm of the noise observation interval Tm and Tn to be input exists from input from the framing section 2, after the timing t _b of the speech recognition section The environmental noise component is removed from the observation vector a, and the feature amount is extracted. That is, the feature extraction unit 5 performs, for example, Fourier transform on the audio data as the observation vector a, obtains a power spectrum, and calculates a feature vector y having each frequency component of the power spectrum as a component. The method for calculating the power spectrum is not limited to the method based on the Fourier transform. That is, the power spectrum can also be obtained by, for example, the so-called filter bank method.

Further, the feature extraction unit 5 calculates the observation vector a
A parameter representing the distribution in the feature vector space obtained when the speech included in the speech data as is mapped to the space of the feature amount (feature vector space) (hereinafter, referred to as
Z is described as a feature vector y
And the environmental noise of the noise observation section Tm,
It is supplied to the voice recognition unit 6.

FIG. 3 shows a detailed configuration example of the feature extraction unit 5 of FIG. The observation vector a input from the framing unit 2 is supplied to the power spectrum analysis unit 11 in the feature extraction unit 5. In the power spectrum analysis unit 11, the observation vector a is Fourier-transformed by, for example, FFT (Fast Fourier Transform), whereby the power spectrum of the voice is extracted as a feature vector. Here, it is assumed that the observation vector a as the audio data of one frame is converted into a feature vector (D-dimensional feature vector) including D components.

Here, the observation vector a of the t-th frame
The feature vector obtained from (t) is represented as y (t). In the feature vector y (t), the spectral component of the true voice is x (t), and the spectral component of the environmental noise is u.
(T). In this case, the spectral component x of the true voice
(T) is represented by the following equation (1).

(Equation 1) However, here, the environmental noise has irregular characteristics,
Also, it is assumed that the audio data as the observation vector a (t) is obtained by adding environmental noise to a true audio component.

On the other hand, environmental noise in the noise observation section Tm as speech data input from the noise observation section extraction section 3 is input to the noise characteristic calculation section 13 in the feature detection section 5. The noise characteristic calculation unit 13 obtains environmental noise characteristics in the noise observation section Tm.

That is, here, the distribution of the power spectrum u (t) of the environmental noise in the speech recognition section is:
Assuming that the noise distribution is the same as the environmental noise in the noise observation section Tm immediately before the speech recognition section, and that the distribution is a normal distribution, the noise characteristic calculating section 13 defines the normal noise of the environmental noise. An average value (average vector) and a variance (variance matrix) are obtained.

The mean vector μ ′ and the variance matrix Σ ′ can be obtained according to the following equation (2).

(Equation 2) Here, μ ′ (i) represents the i-th component of the average vector μ ′ (i = 1, 2,..., D). Further, y (t) (i) represents the i-th component of the feature vector of the t-th frame. Furthermore, Σ ′ (i,
j) represents the component of the ith row and the jth column of the variance matrix Σ ′ (j = 1, 2,..., D).

Here, in order to reduce the amount of calculation, regarding environmental noise, each component of the feature vector y
Assume that they are uncorrelated with each other. In this case, as shown in the following equation, the variance matrix Σ ′ is 0 except for the diagonal components.

(Equation 3)

In the noise characteristic calculation section 13, the average vector μ ′ and the average value Σ ′ that define the normal distribution as the environmental noise characteristics are obtained as described above, and are supplied to the characteristic distribution parameter calculation section 12. You.

On the other hand, the output of the power spectrum analyzer 11, that is, the feature vector y of the uttered voice including environmental noise is supplied to the feature distribution parameter calculator 12.
The characteristic distribution parameter calculation unit 12 calculates the distribution of the power spectrum of the true voice (the distribution of the estimated values) based on the characteristic vector y from the power spectrum analysis unit 11 and the environmental noise characteristics from the noise characteristic calculation unit 13. A characteristic distribution parameter to be represented is calculated.

That is, the feature distribution parameter calculator 12
Then, assuming that the distribution of the power spectrum of a true voice is a normal distribution, its average vector ξ and variance matrix Ψ
Are the following as the feature distribution parameters:
Is calculated according to

(Equation 4)

(Equation 5)

(Equation 6)

(Equation 7)

Here, ξ (t) (i) represents the ith component of the average vector ξ (t) in the t-th frame. E [] means an average value in []. x (t) (i) represents the ith component of the true speech power spectrum x (t) in the t-th frame. Further, u (t) (i) represents the ith component of the environmental noise power spectrum at the t-th frame, and P (u (t) (i)) represents the environmental noise power spectrum at the t-th frame. Represents the probability that the i-th component is u (t) (i).
Here, since a normal distribution is assumed as the distribution of the environmental noise, P (u (t) (i)) is expressed as shown in Expression (7).

Ψ (t) (i, j) represents the component of the ith row and jth column of the variance Ψ (t) in the tth frame. Further, V [] represents the variance in [].

In the feature distribution parameter calculation unit 12, as described above, the average vector ξ and the variance matrix 、 are distributed for each frame in the feature vector space of the true speech (here, the true speech Distribution assuming normal distribution in feature vector space)
Is obtained as a feature distribution parameter representing

Thereafter, the feature distribution parameters obtained in each frame of the speech recognition section are output to the speech recognition unit 6. That is, it is now assumed that the speech recognition section is a T frame, and the characteristic distribution parameters obtained in each of the T frames are represented by z (t) = {(t), Ψ
(T)｝ (t = 1, 2,..., T), the feature distribution parameter calculation unit 12 calculates the feature distribution parameter (series) Z = Ｚz (1), z (2),.・, Z (T)｝
Is supplied to the voice recognition unit 6.

Returning to FIG. 1, the speech recognition unit 6 classifies the feature distribution parameter Z input from the feature extraction unit 5 into one of a predetermined number K of acoustic models and one silent acoustic model. The classification result is output as a recognition result of the input speech. That is, the speech recognition unit 6 includes, for example, an identification function corresponding to a silent section (a function for identifying whether the feature parameter Z is classified as a silent acoustic model) and an identification function corresponding to each of the predetermined number K of words. (A function for identifying which acoustic model the feature parameter Z is classified into), and the value of the identification function of each acoustic model is used as an argument with the feature distribution parameter Z from the feature extraction unit 5 as an argument. Is calculated as Then, an acoustic model (word or silence (noise)) having the largest function value (so-called score) is output as a recognition result.

FIG. 4 shows a detailed configuration example of the voice recognition section 6 of FIG. The feature distribution parameter Z input from the feature distribution parameter calculation unit 12 of the feature extraction unit 5
Are supplied to the identification function operation units 21-1 to 21-k and the identification function operation unit 21-s. Discriminant function operation unit 21
−k (k = 1, 2,..., K) stores an identification function G _k (Z) for identifying the word corresponding to the k-th of the K acoustic models. The discrimination function G _k (Z) is calculated using the feature distribution parameter Z from the extraction unit 5 as an argument. The identification function calculation unit 21-s identifies the identification function G for identifying a silent section corresponding to the silent acoustic model.
_s (Z) is stored, and the discrimination function G _s (Z) is calculated using the feature distribution parameter Z from the feature extraction unit 5 as an argument.

In the speech recognition unit 6, for example, HMM (Hi
Using the dden Markov Model) method, a word or silence as a class is identified (recognized).

The HMM method will be described with reference to FIG. In the figure, the HMM has H states q _{1 to} q _H , and as for the state transition, only the transition to itself and the transition to the state on the right are permitted. The initial state is the leftmost state q _1, the final state is the rightmost state q _H, the state transition from the final state q _H is prohibited. As described above, a model having no transition to a state located to the left of itself is called a left-to-right model, and in speech recognition, a left-to-right model is generally used.

Now, assuming that a model for identifying the k class of the HMM is a k class model, the k class model is, for example, a probability of being initially in the state q _h (initial state probability) π _k (q _h ). , At a certain time (frame) t, the state q _i , and at the next time t + 1, the probability of transition to the state q _j (transition probability) a _k (q _i , q _j ) and the state transition from the state q _i Occurs, the state q _i is the probability (output probability) b _k (q _i ) of outputting the feature vector O
(H) (h = 1, 2,...,
H).

Then, certain feature vector sequences O ₁ , O ₂ ,
.. Are given, for example, the class of the model having the highest probability of observing such a feature vector sequence (observation probability) is taken as the recognition result of the feature vector sequence.

Here, the observation probability is determined by the discriminant function G _k
(Z). That is, the discriminant function G
_k (Z) is a feature distribution parameter (series) Z = ｛z ₁ , z
₂ ,..., Z _T }, such a feature distribution parameter (sequence) Z = {z ₁ , z ₂ ,. _T ｝ is given by the following equation (8) to determine the probability of being observed.

(Equation 8)

Here, b _k ′ (q _i ) (z _j ) represents the output probability when the output is a distribution represented by z _j . For the output probability b _k (s) (O _t ), which is the probability of outputting each feature vector at the time of state transition, for example, a normal distribution function is used assuming that components on the feature vector space have no correlation. ing. In this case, when the input has a distribution represented by z _t , the output probability b _k ′ (s) (z _t )
Is the mean vector μ _k (s) and the variance matrix Σ _k (s)
The probability density is defined by a function ^{_{P k m (s) (x}} ),
And a probability density function P ^f (t) representing the distribution of the feature vector (here, the power spectrum) x of the t-th frame
Using (x), it can be obtained by the following equation (9).

(Equation 9) However, the integration interval of the integration in Expression (9) is the entirety of the D-dimensional feature vector space (here, the power spectrum space).

In equation (9), P (s) (i)
(Ξ (t) (i), Ψ (t) (i, i)) is given by the following equation (1)
0).

(Equation 10) Where μ _k (s) (i) is the ith component of the mean vector μ _k (s), and Σ _k (s) (i, i) is
The components in the ith row and the ith column of the variance matrix Σ _k (s) are respectively represented. The output probability of the k-class model is defined by these.

As described above, the HMM is defined by the initial state probability π _k (q _h ), transition probability a _k (q _i , q _j ), and output probability b _k (q _i ) (O). However, these are calculated in advance by calculating a feature vector from the speech data for learning, and using the feature vector.

[0043] Here, as HMM, when used as shown in FIG. 5 is always so leftmost transition from the state q ₁ of starts, only the initial state probability corresponding to the state q ₁ is 1, The initial state probabilities corresponding to other states are all set to 0. Further, as is apparent from equations (9) and (10), when Ψ (t) (i, i) is set to 0, the output probability matches the output probability in the continuous HMM when the variance of the feature vector is not considered. I do.

As an HMM learning method, for example, B
Aum-Welch re-estimation method and the like are known.

Returning to FIG. Discriminant function operation unit 21-k (k
= 1, 2,..., K)
Initial state probability π _k previously obtained by learning
(Q _h ), transition probability a _k (q _i , q _j ), and output probability b
_k (q _i) expression defined by (O) stores a discriminant function G _k (Z) of (8), the feature distribution parameter Z from the feature extraction section 2 as an argument, the identification function G _k (Z ), And the function value (the above-described observation probability) G _k (Z) is calculated as
Output to the determination unit 22. The discrimination function calculation unit 21-s calculates the initial state probability π _s supplied from the silent acoustic model correction unit 7.
(Q _h ), transition probability a _s (q _i , q _j ), and output probability b
_s (q _i ) (O), which stores a discriminant function G _s (Z) similar to the discriminant function G _k (Z) in equation (8), and stores a feature distribution parameter from the feature extracting unit 2. The identification function G _s (Z) is calculated using Z as an argument, and the function value (the above-described observation probability) G _s (Z) is output to the determination unit 22.

In the decision unit 22, the discriminant function operation unit 21-1
With respect to the function values G _k (Z) (here, the function values G _s (Z) are included here) from the individual function units 21 to k and the discriminant function operation unit 21-s, for example, the following expression (11) is used. Using the decision rule shown, the feature distribution parameter Z, that is, the class (acoustic model) to which the input speech belongs is identified.

[Equation 11] Here, C (Z) represents a function for performing an identification operation (processing) for identifying a class to which the feature distribution parameter Z belongs. Further, max on the right side of the second equation of the equation (11) is a function value G _i (Z) (where i = s,
1, 2,..., K).

The deciding unit 22 calculates according to the equation (11)
When the class is determined, it is output as a recognition result of the input speech.

Returning to FIG. 1, the silent acoustic model correction unit 7
Based on the environmental noise as speech data in the latter half of the noise observation section Tn of the noise observation section Tm and Tn input from the noise observation section extraction section 3, corresponding to the silent acoustic model stored in the speech recognition section 6. to generate a discriminant function G _s (Z), this identification function G _s (Z),
The silent acoustic model stored in the speech recognition unit 6 is adapted.

More specifically, the silent sound model correction unit 7 calculates a feature vector y for each of M frames of speech data (environmental noise) in the noise observation section Tn input from the noise observation section extraction unit 3. Are observed, and a sequence of feature distribution represented by the following equation is generated in the same manner as in the case of the feature extraction unit 5.

(Equation 12) Note that the feature distribution ｛F _i (y), i = 1, 2,.
M｝ is the probability density function (Probabilistic Density Functio
n), and is also hereinafter referred to as a silent feature distribution PDF. Also, the suffix i in the silent feature distribution F _i (y)
Represents the number of frames from the first frame of the noise observation section Tn.

Next, the silence feature distribution PDF is calculated by the following equation (13).
According to the probability distribution F _s (y) corresponding to the silent acoustic model
Map to

(Equation 13) Here, V is a silent feature distribution PDF ｛F _i (y), i = 1,
, M} is a correction function (mapping function) for mapping the silent acoustic model F _s (y).

Various methods can be used for this mapping depending on the description of the silence feature distribution PDF. For example, the following equation can be adopted.

[Equation 14] Where β _i (F ₁ (y), F ₂ (y),..., F
_M (y), M) is a weighting function for the silent feature distribution obtained from the first frame of the noise observation section Tn, and is hereinafter referred to as β _i . The weighting function β _i is given by the following equation (1)
This satisfies the condition of 6).

(Equation 15)

Here, the probability distribution F of the silent acoustic model
Assuming that _s (y) is a normal distribution, and that the components constituting the feature vector of each frame are uncorrelated, the silent feature distribution PDF @ F _i (y), i =
The covariance matrix Σ _i of 1, 2,..., M｝ is a diagonal matrix. However, as a precondition for this assumption, the covariance matrix of the silent acoustic model must also be a diagonal matrix.

If the components constituting the feature vector y of each frame in the noise observation section Tn are uncorrelated, the silent feature distribution PDF @ F _i (y), i = 1, 2,.
·, M} becomes the normal distribution G with mean and variance corresponding to each component (E _i, Σ _i). Where E _i is F
_i (y) is the average value (expected value), and Σ _i is the covariance matrix of F _i (y). That is, the average of the silent feature distribution obtained from each frame in the noise observation section Tn is μ _i , and the variance is σ _i
² , the probability density function of the silent feature distribution is represented by a normal distribution G (μ _i , σ _i ² ) (i = 1, 2,...,
M).

Based on the above assumption, the normal distribution G (μ _s ) approximating the silent acoustic model F _s (X) by the following various methods using the average μ _i and the variance σ _i ² corresponding to each frame. , Σ _s ² ) (corresponding to G _s (Z) described above)
Can be calculated.

The normal distribution G (μ _s ,
A first method of calculating σ _s ² ) is a silent feature distribution ｛G (μ
_i , σ _i ² ), i = 1, 2,.
As shown in 7), the average of all μ _{i is taken} as the average value μ _s of the silent acoustic model, and as shown in equation (18), the average of all σ _i ^{2 is taken} as the variance σ _{i of the} silent acoustic model. ^It is a method to be ² .

(Equation 16) Here, a and b are coefficients for which an optimal value is determined by simulation.

The normal distribution G (μ _s ,
A second method for calculating σ _s ² ) is a silent feature distribution ｛G (μ
_i , σ _i ² ), i = 1, 2,..., M}, using only the expected value μ _i , and according to the following equations (19) and (20), the average value μ _s of the silent acoustic model and the variance This is a method of calculating σ _i ² .

[Equation 17] Here, a and b are coefficients for which an optimal value is determined by simulation.

The normal distribution G (μ _s ,
A third method for calculating σ _s ² ) is a silent feature distribution ｛G (μ
_i , σ _i ² ), i = 1, 2,..., M}, to calculate the average μ _s and the variance σ _s ² of the silent acoustic model.

In this method, each silent feature distribution G
Let X _i be the probability statistic of (μ _i , σ _i ² ).

(Equation 18)

Here, the normal distribution G (μ
_s, Σ_s ^Two) Is the probability statistic X_sThen the probability statistic X_s
Is a probability statistic X as shown in the following equation (22)._iAnd weight
Function β _iCan be represented by a linear combination of Note that the weights
Number β_iSatisfies the condition of equation (16).

[Equation 19]

Then, the normal distribution G (μ
_s , σ _s ² ) is expressed as shown in the following equation (23).

(Equation 20)

In equation (23), the weight function β _i
Can be generally set to, for example, 1 / M. In this case, the average value μ _s and the variance σ _s ² of the equation (23) are, for example,
As shown by the following equation, it is obtained by using a predetermined coefficient.

(Equation 21) Here, a and b are coefficients for which an optimal value is determined by simulation.

The normal distribution G (μ _s ,
In the fourth method for calculating σ _s ² ), the silent feature distribution ｛G
Assume a statistical population Ω _i = {f _{i, j} } corresponding to the probability statistic X _i of (μ _i , σ _i ² ), i = 1, 2,. here,

(Equation 22) Then, the average value μ _i can be obtained by the following equation (26), and the variance σ _i ² can be obtained by the following equation (28).

(Equation 23)

By modifying equation (28), the following equation (29) holds.

(Equation 24)

Here, the sum Ω of the statistical population

(Equation 25) In consideration of the following, the following Expressions (30) and (31) are obtained from Expression (26).
Is derived, and the following expressions (32) to (34) are derived from the expression (29).

(Equation 26)

In practice, the equations (31) and (34)
Is used after being multiplied by a coefficient as shown in the following equation.

[Equation 27] Here, a and b are coefficients for which an optimal value is determined by simulation.

The following equation (37) can also be employed. In the equation (37), only the variance σ _i ² is multiplied by the coefficient b.

[Equation 28]

Next, the operation of the speech recognition apparatus of FIG. 1 will be described.

The framing unit 2 receives the voice data collected by the microphone 1 (the uttered voice to be recognized including environmental noise), where the voice data is framed, and the voice data of each frame is , And the observation vector a are sequentially supplied to the noise observation section extraction unit 3 and the feature extraction unit 5. The noise observation interval extraction section 3, a timing t _b previous noise observation interval Tm and Tn of audio data speech switch 4 is turned on (environmental noise) is extracted, the feature extraction unit 5 and the silence acoustic model correction section 7 Supplied to

The silent acoustic model correcting section 7 calculates the noise based on the environmental noise as the voice data in the noise observation section Tm.
Silence feature distribution from each frame of noise observation section Tn PDF
Is required. Further, in the silent sound model correcting unit 7,
Based on the feature distribution PDF, the silence acoustic model is updated (adapted) by any one of the above-described first to fourth methods, and is supplied to the speech recognition unit 6. In the speech recognition unit 6, the identification function corresponding to the silence acoustic model that has been stored is updated by the identification function as the silence acoustic model supplied from the silence acoustic model correction unit 7. That is, adaptation of the silent acoustic model is performed.

On the other hand, in the feature extracting unit 5, the framing unit 2
The audio data as the observation vector a from the audio data is acoustically analyzed, and its feature vector y is obtained. Further, the feature extraction unit 5 calculates a feature distribution parameter Z representing a distribution in a feature vector space based on the obtained feature vector y, and supplies the calculated feature distribution parameter Z to the speech recognition unit 6. The speech recognition unit 6 calculates the values of the identification functions of the acoustic models corresponding to the silence and the predetermined number K of words using the feature distribution parameters from the feature extraction unit 5, and the acoustic model in which the function value is maximized. Is output as a speech recognition result.

As described above, the speech data as the observation vector a is converted into the feature distribution parameter Z representing the distribution in the feature vector space which is the space of the feature quantity, and the feature distribution parameter is converted into the speech data. The distribution function of the noise included is taken into consideration, and the identification function corresponding to the silent acoustic model for identifying (recognizing) silence is updated based on the voice data of the noise observation section Tn immediately before the utterance. Therefore, it is possible to greatly improve the speech recognition rate.

Next, FIG. 6 shows a silent period Ts from when the utterance switch 4 is turned on until the start of utterance (FIG. 2).
4 shows the results of an experiment (simulation) in which a change in the speech recognition rate when the value was changed was measured.

In FIG. 6, the curve a does not correct the silent acoustic model (no adaptation of the silent acoustic model).
Curve b shows the result of the first method, curve c shows the result of the second method, and curve d shows the result of the third method.
The curve e shows the result by the fourth method, and the curve e shows the result by the fourth method.

The experimental conditions are as follows. That is,
The voice data used for recognition is data collected in a vehicle traveling on a highway. The noise observation section Tn is about 0.2 seconds in 20 frames. The silent section Ts is 0.05
Seconds, 0.1 seconds, 0.2 seconds, 0.3 seconds, and 0.5 seconds.
MFCC (Mel-Frequency)
Analysis was performed in the Cepstral Coefficients domain (features were obtained by MFCC analysis). A total of eight speakers, four men and women, were to be recognized, and 303 words were uttered individually per person. The number of recognized words is 5000 words in Japanese. The acoustic model was an HMM, and learning was performed in advance using voice data prepared for learning. For voice recognition, Vite
The beam width was set to 3000 using the rbi search method.

In the first, second and fourth methods, the coefficient a was set to 1.0 and the coefficient b was set to 0.1.
In a third method, both coefficients a and b are:
1.0.

As is clear from FIG. 6, in the conventional method (curve a), the speech recognition rate is remarkably reduced as the silent section Ts becomes longer, but the first to fourth methods (curve b) are used. In steps (e) to (e), even when the silent section Ts becomes longer, the speech recognition rate decreases only slightly. Accordingly, by adapting the silent acoustic model, it is possible to maintain the recognition performance even when the silent section Ts changes.

In each of the first to fourth methods described above, the normal distribution G (μ _s ,
The average value μ _s defining σ _s ² ) is a silent feature distribution G (μ _i ,
σ _i ² ) is the average value of μ _i . So, for example,
Now, the average value of the average value μ _i of the silent feature distribution G (μ _i , σ _i ² ) is represented as μ, and the normal distribution of the silent acoustic model obtained by the first to fourth methods is G, respectively.
^{_{s1 (μ, σ s1 2)}} , G s2 (μ, σ s2 2), G s3 (μ, σ s3
² ) and G _s4 (μ, σ _s4 ² ), these are centered on the average value μ (centroid) in the feature space, as shown in FIG.
Distribution.

By the way, the silent feature distribution G (μ _i , σ _i ² )
The adaptation of the silent acoustic model according to the above-described first to fourth methods based on the above can be defined by the following equation (38) using the mapping V. Hereinafter, G (μ _i ,
σ _i ² ) is described as G _i , and G (μ _s , σ _s ² ) is described as G _s .

[0079]

(Equation 29)

[0080] Further, here, as silence acoustic model G _s, and assuming a normal distribution, normal distribution, since is defined by the mean value and variance, the mean value for defining the normal distribution of the silence acoustic model G _s And variance are represented by μ _s and σ _s ² , as described above, the definition of equation (38) defines the average and variance mapping V _μ
And V _{σ 2} , respectively, using equations (39) and (4)
0).

[0081]

[Equation 30]

In the first to fourth methods represented by the above-described mappings V (V _μ and V _{σ 2} ), the time-series silence feature distribution G obtained from each of the M frames in the noise observation section Tn (FIG. 2). _1, G _2, ···, are handled equally G _M.

However, the environment in the speech recognition section
Strictly speaking, the noise observation just before the speech recognition section
It is not the same as the environmental noise in the section Tn.
Generally, environmental noise in the noise observation section Tn is:
Start time t of the speech recognition section ( _c)
It is expected that this will be different from the environmental noise in the recognition section.
Measured.

Therefore, the time-series silent feature distributions G ₁ , G ₂ ,..., G _M obtained from each of the M frames in the noise observation section Tn (FIG. 2) are not treated equally, but are not treated equally. Should be treated with weights closer to (the ones farther from the speech recognition section should be treated without weight), and by doing so, adaptation of a silent acoustic model that further improves speech recognition accuracy ( Correction and update) are possible.

[0085] Therefore, the silence feature obtained in the noise observation interval Tn distribution G _1, G _2, · · ·, for G _M, the freshness (here corresponds to proximity to the speech recognition section) fresh representing the A method of adapting a silent acoustic model in consideration of the freshness will be described.

FIG. 8 shows an example of the structure of the silent acoustic model correction unit 7 shown in FIG. 1 for adapting a silent acoustic model in consideration of freshness.

The freshness function storage section 31 stores a freshness function (a parameter defining the above) which is a function representing the freshness as described above.

The correction section 32 includes a noise observation section extraction section 3
Is output as a series of observation vectors (here, 2M-frame audio data) as audio data (noise) in the noise observation sections Tm and Tn, and the correction unit 32 from silence feature distribution G _1, G _2, · · ·, to obtain G _M, and these, on the basis of the freshness function stored in freshness function storage unit 31 performs adaptive silence acoustic model.

Here, the silent feature distributions G ₁ , G ₂ ,.
G _M is a discrete value observed in each of the M frames in the noise observation section Tn.
But if the system for processing discrete values, the silence feature distribution G _1, G ₂ is a discrete value, ..., it can be used as it is G _M. However, the silent acoustic model correction unit 7
But if a system for processing a continuous value, for example, as shown in FIG. 9, the silence feature distribution G _1, G ₂ is a discrete value, ..., a G _M, continuous values in continuous transducers , And then need to be processed by the silent acoustic model correction unit 7. As a method of converting a discrete value into a continuous value, for example,
There is a method of performing approximation using a spline function.

Note that a discrete value is a finite number of observation values observed at discrete times in a certain finite observation section, and a continuous value is an arbitrary value in a certain finite (or infinite) observation section. It is an infinite number of observations observed at time, and is represented by a certain function.

When the silent feature distribution used for adapting the silent acoustic model is a discrete value, the freshness function is also a discrete value function. When the silent feature distribution is a continuous value, the freshness function is also a continuous value. Is a function of

Next, the freshness function and the adaptation of the silent acoustic model using the function will be described separately for a case where the freshness function is a discrete value and a case where the freshness function is a continuous value.

First, the freshness function g (t) can be defined, for example, as shown in equations (41) to (43).

[0094]

(Equation 31) Here, Ω _obs represents an observation section of the silent feature distribution, and corresponds to a noise observation section Tn in the present embodiment.

From equation (41), the freshness function g (t)
Is 0 outside the observation interval Ω _obs . Equation (4)
According to 2), the freshness function g (t) is a function that increases or does not change with the passage of time in the observation interval Ω _obs (referred to as a monotonically increasing function in the present specification). g (t) basically becomes larger as it approaches the voice recognition section (FIG. 2). Further, according to equation (43), the freshness function g (t) is a function whose integral value is 1 when integrated over the observation interval Ω _obs . From equations (41) to (43), the freshness function g (t) is, for example, as shown in FIG.

Here, in the present embodiment, the freshness function g
(T) is used as a multiplier for multiplying the silent feature distribution as described later. Therefore, the freshness function g (t) is
When its value is positive or negative, it acts as a weight for the silent feature distribution that is multiplied as a multiplier. When the value of the freshness function g (t) is 0, the silence feature distribution multiplied by the multiplier is invalidated, and acts so as not to affect the adaptation of the silence acoustic model.

The correction unit 32 shown in FIG.
Degree function g (t) and silence feature distribution G ₁, G_Two, ..., G
_MAnd basically, according to equation (44),
Silent acoustic model G after adaptation_sIs required.

[0098]

(Equation 32)

According to equation (44), the closer the silence feature distribution is to the speech recognition section, the higher the weight is treated, and the silence acoustic model is adapted. As a result, the speech recognition accuracy is further improved. It becomes possible.

Next, a specific example of the freshness function F (x) and adaptation of the silent acoustic model using the function will be described. In the following, the observation interval Ω _obs (in the present embodiment, the noise observation interval Tn) of the silent feature distribution is defined as t from 0 to t _M
The section up to. As a function value of the freshness function g (t), a value of only the observation section Ω _obs is considered (as shown in Expression (41), the function value of the freshness function g (t) is Since it is 0 outside the observation section Ω _{ob s} , that point will not be described below).

As the freshness function g (t), for example, a linear function can be used. When a continuous value is taken as the function value, the freshness function g (t) can be expressed by, for example, the equation (4)
5).

[0102]

[Equation 33]

Α in Equation (45) is a predetermined constant, and this constant α is 2 / t _M ² from the definition of the freshness function in Equation (43). Therefore, the freshness function g (t) in Expression (45) is expressed by Expression (46).

[0104]

(Equation 34)

Here, the freshness function g (t) represented by the equation (46) is shown in FIG.

In this case, the silence acoustic model G after the adaptation
_s is obtained according to equation (47).

[0107]

(Equation 35) Note that G _x (μ _x , σ _x ² ) represents a silent feature distribution at time x, and μ _x and σ _x ² are an average value and a variance defining a normal distribution representing the silent feature distribution, respectively. .

Next, as the freshness function F (x), for example, a function taking a linear, discrete value can be used. In this case, the freshness function F (x) can be obtained by, for example, the equation (48) It is represented by

[0109]

[Equation 36]

Α in Expression (48) is a predetermined constant, and this constant α is 2 / (t _M (t _M +1)) from the definition of the freshness function in Expression (43). Therefore, equation (4)
The freshness function g (t) of 8) is expressed by Expression (49).

[0111]

(37)

Here, the freshness function g (t) represented by the equation (49) is shown in FIG.

In this case, the silence acoustic model G after the adaptation
_s is obtained according to equation (50).

[0114]

(38) Gt represents a silent feature distribution at a sample point (sample time) t.

Next, as the freshness function g (t), for example, a non-linear function such as an exponential function, a higher-order binomial function, or a logarithmic function can be used. For example, when a quadratic function as a higher-order function that takes a continuous value is used as the freshness function g (t), the freshness function g (t) is expressed by, for example, Expression (51).

[0116]

[Equation 39]

Α in Expression (51) is a predetermined constant, and this constant α is 3 / t _M ³ from the definition of the freshness function in Expression (43). Therefore, the freshness function g (t) of Expression (51) is expressed by Expression (52).

[0118]

(Equation 40)

Here, the freshness function g (t) represented by the equation (52) is shown in FIG.

In this case, the silence acoustic model G after the adaptation
_s is obtained according to equation (53).

[0121]

[Equation 41]

Next, as the freshness function g (t), for example, a quadratic function as a higher-order function taking a discrete value can be used. In this case, the freshness function g (t) is, for example, , (54).

[0123]

(Equation 42)

In the equation (54), α is a predetermined constant. The constant α is 6 / (t _M (t _M +1) (2t _M +1)) from the definition of the freshness function in the equation (43). Becomes Therefore, the freshness function g (t) of the equation (54) is obtained by the equation (55).
Will be represented by

[0125]

[Equation 43]

Here, the freshness function g (t) represented by the equation (55) is shown in FIG.

In this case, the silence acoustic model G after the adaptation
_s is obtained according to equation (56).

[0128]

[Equation 44]

Next, when a logarithmic function having a continuous value is used as the freshness function g (t), the freshness function g (t) is expressed by, for example, equation (57).

[0130]

[Equation 45]

In the equation (57), α is a predetermined constant. From the definition of the freshness function in the equation (43), the constant α is 1 / ((t _M +1) log (t _M +1) -t _M ). Therefore, the freshness function g (t) of Expression (57) is expressed by Expression (58).

[0132]

[Equation 46]

Here, the freshness function g (t) represented by the equation (58) is shown in FIG.

In this case, the silence acoustic model G after the adaptation
_s is obtained according to equation (59).

[0135]

[Equation 47]

Next, as the freshness function g (t), for example, a logarithmic function having a discrete value can be used. In this case, the freshness function g (t) can be expressed by, for example, an equation (60). Is done.

[0137]

[Equation 48]

Α in Expression (60) is a predetermined constant, and this constant α is obtained from the definition of the freshness function in Expression (43). Therefore, the freshness function g (t) in Expression (60) is obtained.
Is represented by Expression (61).

[0139]

[Equation 49]

Here, the freshness function g (t) represented by the equation (61) is shown in FIG.

In this case, the silence acoustic model G after the adaptation
_s is obtained according to equation (62).

[0142]

[Equation 50]

Next, when a general higher-order function that takes a continuous value, for example, is used as the freshness function g (t), the freshness function g (t) is expressed by, for example, the equation (63). expressed.

[0144]

(Equation 51)

Α in equation (63) is a predetermined constant, and the order of the freshness function g (t) is determined by p.

The constant α can be obtained from the definition of the freshness function in the equation (43). Therefore, the freshness function g (t) in the equation (63) is represented by the equation (64). .

[0147]

(Equation 52)

In this case, the silence acoustic model G after the adaptation
_s is obtained according to equation (65).

[0149]

(Equation 53)

In the equation (64), for example, when p is 1 or 2, the freshness function g (t) is a linear function or a quadratic function that takes a continuous value. Each is represented as shown in (52).

In equation (64), for example, when p is 3, the freshness function g (t) is a cubic function having a continuous value and is expressed as shown in equation (66).

[0152]

(Equation 54)

Further, in the equation (64), for example, p
Is 4, the freshness function g (t) takes a continuous value 4
It becomes the following function and is expressed as shown in equation (67).

[0154]

[Equation 55]

Next, when a general higher-order function that takes a discrete value, for example, is used as the freshness function g (t), the freshness function g (t) can be expressed by, for example, the equation (68). expressed.

[0156]

[Equation 56]

Α in Expression (68) is a predetermined constant, and the order of the freshness function F (x) is determined by p.

The constant α can be obtained from the definition of the freshness function in the equation (43). Therefore, the freshness function g (t) in the equation (68) is represented by the equation (69). .

[0159]

[Equation 57]

In this case, the silence acoustic model G after the adaptation
_s is obtained according to equation (70).

[0161]

[Equation 58]

In equation (69), for example, when p is 1 or 2, the freshness function g (t) is a linear function or a quadratic function that takes a discrete value. Each is represented as shown in (55).

In equation (69), for example, if p is 3, the freshness function g (t) takes a discrete value 3
It becomes the following function and is expressed as shown in Expression (71).

[0164]

[Equation 59]

Further, in the equation (69), for example, p
Is 4, the freshness function g (t) takes a discrete value 4
It becomes the following function and is expressed as shown in equation (72).

[0166]

[Equation 60]

Note that the concept of the freshness function g (t) can be applied not only to the adaptation of a silent acoustic model, but also to the application of a speaker under a noise environment and the adaptation of acoustic models other than the silent acoustic model. Furthermore, it can be applied to voice detection and non-stationary noise detection. Also, in the field of sound signal processing, image signal processing, and communication, the freshness function F
By using the concept of (x), it is possible to improve robustness (robustness) against environmental noise and improve system performance.

The speech recognition apparatus to which the present invention is applied has been described above. Such a speech recognition apparatus, for example,
The present invention is applicable to a car navigation device capable of voice input and other various devices.

In the present embodiment, the characteristic distribution parameter is determined in consideration of the noise distribution characteristic. However, this noise includes, for example, external noise in an environment where speech is made, and When recognizing voice transmitted via a telephone line or other communication lines, the characteristics of the communication line are also included.

The present invention can be applied to the case of performing image recognition and other pattern recognition in addition to voice recognition.

Further, in the present embodiment, the silence acoustic model is adapted by using the silence feature distribution represented as the distribution in the feature space. However, the adaptation of the silence acoustic model is performed as points in the feature space. It is also possible to perform this by using the feature amount of the expressed noise.

The present invention can also be used for adapting acoustic models other than the silent acoustic model.

Next, the above-described series of processing can be performed by hardware or software. When a series of processing is performed by software, a program constituting the software is
Installed on a general-purpose computer.

FIG. 17 shows an example of the configuration of an embodiment of a computer in which a program for executing the above-described series of processing is installed.

The program is stored in a hard disk 105 or a ROM 1 as a recording medium built in the computer.
03 can be recorded in advance.

Alternatively, the program may be a floppy (registered trademark) disk, a CD-ROM (Compact Disc Read Onl
y Memory), MO (Magneto optical) disc, DVD (Digita
l Versatile Disc), a magnetic disk, a semiconductor memory, etc., can be temporarily or permanently stored (recorded) in a removable recording medium 111. Such a removable recording medium 111 can be provided as so-called package software.

The program can be installed in the computer from the removable recording medium 111 as described above, can be wirelessly transferred from the download site to the computer via a digital satellite broadcasting artificial satellite, or can be transmitted to a LAN (Local Area Area Network). Network), the Internet, and the like, and can be transferred to a computer by wire. In the computer, the transferred program can be received by the communication unit 108 and installed on the built-in hard disk 105.

The computer has a CPU (Central Processing).
Unit) 102. The CPU 102 has a bus 1
01, an input / output interface 110 is connected. The CPU 102 receives a command via the input / output interface 110 by operating the input unit 107 including a keyboard, a mouse, and the like. Then, according to it, ROM (Read Only Memory)
The program stored in 103 is executed. Alternatively, the CPU 102 transmits the program stored in the hard disk 105, a satellite, or a network, receives the program by the communication unit 108, and
The program installed in the hard disk 105 is read from the removable recording medium 111 installed in the drive 109 and loaded into the RAM (Random Access Memory) 104 and executed. As a result, the CPU 102 performs processing performed by the configuration of the above-described block diagram. Then, the CPU 102 transmits the processing result to an LCD (Liquor
output from an output unit 106 composed of an id CryStal Display) or a speaker, or transmitted from the communication unit 108, and further recorded on the hard disk 105.

Here, in this specification, processing steps for writing a program for causing a computer to perform various processes do not necessarily have to be processed in chronological order in the order described in the flowchart, and may be performed in parallel. Alternatively, it also includes processing executed individually (for example, parallel processing or processing by an object).

The program may be processed by one computer or may be processed by a plurality of computers in a distributed manner. Further, the program may be transferred to a remote computer and executed.

[0181]

According to the model adapting apparatus, the model adapting method, the recording medium, and the pattern recognition apparatus of the present invention,
A predetermined model is adapted based on extracted data in a predetermined section and freshness indicating the freshness of the extracted data. Therefore, by performing pattern recognition using the model, it is possible to improve recognition performance.

[Brief description of the drawings]

FIG. 1 is a block diagram illustrating a configuration example of a speech recognition device to which the present invention has been applied.

FIG. 2 is a diagram for explaining an operation of a noise observation section extraction unit 3 of FIG.

FIG. 3 is a block diagram illustrating a detailed configuration example of a feature extraction unit 5 of FIG. 1;

FIG. 4 is a block diagram illustrating a detailed configuration example of a speech recognition unit 6 in FIG. 1;

FIG. 5 is a diagram showing an HMM.

FIG. 6 is a diagram showing a simulation result.

FIG. 7 is a diagram showing a normal distribution of a silent acoustic model.

FIG. 8 is a block diagram illustrating a configuration example of a silent acoustic model correction unit 7 in FIG. 1;

FIG. 9 is a diagram showing how discrete values are converted into continuous values.

FIG. 10 is a diagram showing a general freshness function g (t).

FIG. 11 is a diagram illustrating a first example of a freshness function g (t).

FIG. 12 is a diagram illustrating a second example of the freshness function g (t).

FIG. 13 is a diagram illustrating a third example of the freshness function g (t).

FIG. 14 is a diagram illustrating a fourth example of the freshness function g (t).

FIG. 15 is a diagram illustrating a fifth example of the freshness function g (t).

FIG. 16 is a diagram illustrating a sixth example of the freshness function g (t).

And FIG. 17 is a block diagram illustrating a configuration example of a computer according to an embodiment of the present invention.

[Explanation of symbols]

1 microphone, 2 framing unit, 3 noise observation section extraction unit, 4 speech switch, 5 feature extraction unit, 6 speech recognition unit, 7 silent acoustic model correction unit,
31 freshness function storage unit, 32 correction unit, 101
Bus, 102 CPU, 103 ROM, 104 RA
M, 105 hard disk, 106 output unit,
107 input unit, 108 communication unit, 109 drive, 110 input / output interface, 111 removable recording medium

Claims (13)

[Claims] 1. A model adapting apparatus for adapting a model used for pattern recognition for classifying time-series input data into one of a predetermined number of models, comprising: Data extraction means for extracting the input data observed in the section and outputting the extracted data as extracted data; and adapting the predetermined model based on the extracted data in the predetermined section and freshness indicating the freshness of the extracted data. And a model adapting means for performing the following. 2. The model adaptation apparatus according to claim 1, wherein the pattern recognition is performed based on a feature distribution of the input data in a feature space. 3. The method according to claim 2, wherein the model adapting means adapts the predetermined model using a function whose value changes in accordance with a temporal position of the extracted data in the predetermined section as the freshness. The model adaptation apparatus according to claim 1, wherein 4. The model adaptation apparatus according to claim 3, wherein the function is a monotonically increasing function that increases with time. 5. The model adaptation apparatus according to claim 4, wherein the function is a linear or non-linear function. 6. The model adaptation apparatus according to claim 4, wherein said function takes a discrete value or a continuous value. 7. The model adaptation apparatus according to claim 4, wherein the function is a quadratic function or a higher-order function of third or higher order. 8. The apparatus according to claim 4, wherein the function is a logarithmic function. 9. The model adaptation apparatus according to claim 1, wherein the input data is audio data. 10. The model adaptation apparatus according to claim 9, wherein the predetermined model is an acoustic model representing noise in a section other than a speech section. 11. A model adaptation method for adapting a model used for pattern recognition for classifying time-series input data into one of a predetermined number of models, the method comprising: A data extraction step of extracting the input data observed in the section and outputting the extracted data as extracted data; and adapting the predetermined model based on the extracted data in the predetermined section and the freshness indicating the freshness of the extracted data. Performing a model adaptation step. 12. A recording medium storing a program for causing a computer to adapt the model used for pattern recognition for classifying time-series input data into one of a predetermined number of models. A data extraction step of extracting the input data observed in a predetermined section corresponding to a predetermined model and outputting the extracted data as extracted data; and extracting data in the predetermined section and freshness representing the freshness of the extracted data. And a model adaptation step of adapting the predetermined model based on the program. 13. A pattern recognition apparatus for classifying time-series input data into one of a predetermined number of models, wherein: a feature extraction unit for extracting a feature amount of the input data; A classifying unit that classifies a feature amount of the input data into one of the predetermined number of models; and extracts the input data observed in a predetermined section corresponding to a predetermined model. Data extracting means for outputting as extracted data, and model adapting means for adapting the predetermined model based on the extracted data in the predetermined section and freshness indicating the freshness of the extracted data. Characteristic pattern recognition device.