JP2973805B2 - Standard pattern creation device - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a standard pattern creating device for creating a standard pattern used in a speech recognition device.

[0002]

2. Description of the Related Art In speech recognition, standard patterns such as phonemes and words to be recognized are prepared in advance, and the input speech is compared with the standard patterns. Alternatively, it is often determined that a word has been uttered. In such a method, in general, the more the number of standard patterns, the more various fluctuations in voice can be expressed, so that a good recognition rate is obtained.
However, on the other hand, it requires a large amount of memory and computation.

[0003] Clustering (A. Gersho an
dV. Cuperman, IEEE Commun,
Meg. 21,9, pp. 15-21, 1983, hereinafter referred to as Document 1) is known as a method for reducing the standard pattern in order to reduce the calculation time and the memory amount while maintaining the recognition performance. Among them, as a method for efficiently reducing the standard pattern, a separate clustering method in which a feature vector of a learning pattern is divided and clustering is performed for each of the divided feature vectors (Journal of the Acoustical Society of Japan, Vol. 44, No. 8, 1988, pp. 595-602).
"Normalization of spectrogram using separate vector quantization", hereinafter referred to as Document 2).
In Literature 2, the feature vector is composed of power and LPC parameters. Hereinafter, a conventional standard pattern creating apparatus will be described with reference to Document 2 as an example.

FIG. 2 is a block diagram showing an example of a conventional standard pattern creating apparatus. A voice is input to the voice input unit 200 and sent to the analysis unit 210. The transmitted voice waveform is analyzed in the analysis unit 210, and the power and LPC parameter feature vectors are extracted. The first standard pattern learned using the extracted feature vector is stored in the learning pattern storage unit 220. The power is sent from the learning pattern storage unit 220 to the power clustering unit 230 and clustered. The LPC parameters are sent from the learning pattern storage unit 220 to the LPC parameter clustering unit 240 and are clustered. Using the information clustered by the power clustering unit 230 and the LPC parameter clustering unit 240, a standard pattern is created in the pattern creation unit 250 from the learning pattern sent from the learning pattern storage unit 220. The standard pattern created by the pattern creation unit 250 is sent to the standard pattern output unit 260 and output.

[0005] As described above, by performing clustering of power and LPC parameters, a standard pattern with a smaller amount of memory and less quantization distortion was obtained than by clustering feature vectors collectively. Have been.

[0006]

In Reference 2, clustering is separately performed for each feature amount of power and LPC parameters. In this method, parameters having low correlation may be put together, and as a result, many clusters are required because quantization distortion increases and clustering efficiency decreases. An object of the present invention is to provide a standard pattern creating apparatus which solves this problem.

[0007]

According to the present invention, there is provided a standard pattern creating apparatus comprising: a voice input unit for inputting voice; an analysis unit for analyzing input voice data and extracting a feature vector; A learning unit that learns a standard pattern from a feature vector, a learning pattern storage unit that stores a learned first standard pattern, a correlation calculation unit that calculates a degree of correlation between the feature vector elements, A feature vector dividing unit that calculates the strength of correlation between feature vector elements from and a feature vector, and a distance calculating unit that calculates an inter-pattern distance from the feature vector.
A clustering unit that clusters a learning pattern for each divided feature vector based on the vector division information and the distance between patterns; a cluster center storage unit that stores a cluster center obtained as a result of the clustering; and a pattern that forms each cluster. It comprises a cluster member storage unit for storing and a standard pattern creation unit for creating a standard pattern based on the result of the clustering.

[0008]

The standard pattern creating apparatus according to the present invention calculates the strength of correlation between feature vector elements, divides the feature vector, and performs clustering for each of the divided feature vectors to reduce the number of clusters. Create a pattern.

The concept will be briefly described with reference to FIGS. In the figure, X1, X2, Y1, and Y2 are feature amount axes, distributions on the axes are distributions based on each axis, R1 to R5 are cluster center numbers, and portions surrounded by dotted lines are covered by cluster centers. The feature space to be covered, and the portion surrounded by the solid line is the feature space to be covered.

Referring to FIGS. 3 and 4, the distributions on each axis are equal. However, in the case of FIG. 3, since the feature space has a low correlation between parameters, many standard patterns are required to cover the entire space. In contrast, FIG.
When the correlation between the parameters is high as shown in (1), the entire space can be covered with a small number of standard patterns.
As described above, when the correlation between the parameters is high, the entire space can be expressed with fewer parameters, so that a standard pattern with a reduced number of patterns can be efficiently obtained.

A simple example will be described.

[0012]

(Equation 1)

A parameter having three elements x, x,
Suppose y, z. Among these three parameters, two parameters x and y have a strong correlation and a correlation function of 1, but x and y, and y and z have no correlation and a correlation function of 0. I do. Under these conditions, x,
Consider a case where three parameters of y and z are divided into two sets.

First, the sum of x and y, and 2 of z
Consider the case of dividing into sets. Since x and y always take the same value, possible values are [-1, -1], [0, 0],
[1, 1]. There are three possible values of z, -1, 0 and 1. Therefore, x, y and z
, The number of parameters to be stored is 2 × 3 +
3 = 9. Next, let us consider a case where the image is divided into two sets of x and a set of y and z. The possible value of x is-
1, 0, and 1. If you put together y and z,
[-1, -1,], [-1,0], [-1,1,], [0,
-1], [0,0], [0,1], [1, -1],
It takes nine values of [1,0] and [1,1]. Therefore,
When divided into x, y, and z, the number of parameters to be stored is 3 + 2 × 9 = 21. In this case, by gathering parameters having high correlation, the entire space can be covered with 9/21 parameter numbers.

Although a simple example has been described here, the same applies when the number of parameters is increased.

As described above, by taking into account the strength of the correlation between parameters, it is possible to provide a standard pattern that can obtain better recognition performance with a small number of patterns.

[0017]

DESCRIPTION OF THE PREFERRED EMBODIMENTS Next, a standard pattern forming apparatus according to the present invention will be described with reference to the drawings.

FIG. 1 is a block diagram showing an embodiment of the present invention. A voice is input to the voice input unit 10 and sent to the analysis unit 20. The sent voice waveform is analyzed in the analysis unit 20 and a feature vector is extracted. Examples of the feature vector after the analysis include LPC mel-cepstrum and Δ-mel-cepstrum (“Speaker-independent”.
isolated word recognition
using dynamic features of
speech spectrum, "IEEE Tr
ans. Acoustic. , Speech Signal
Processing, vol. ASSP-34, p
p. 52-59, 1986. Hereinafter, this is referred to as Document 3), delta ² Mel cepstrum ( "Improved A
cosmetic Modeling with the
SPHINX Speech Recognition
n System, X. D. Huang, K .; F. Le
e, H .; W. Hon, and M.S. Y. Hwang, I
CASSP 91 pp. 345-348, 1991,
Hereinafter, this will be referred to as Reference 4.).

The extracted feature vector sequence is sent to the learning unit 30
Is used for learning standard patterns. The learning method depends on the recognition method. For example, the path cost DP (Watanabe, Kimura, Proceedings of the Acoustical Society of Japan, 2-5-9, 1962)
10, hereinafter referred to as Reference 5), as described in Reference 5, the average vector and the statistical path cost in each frame of the standard pattern are calculated.

Hereinafter, the path cost DP will be described as an example.

The learned pattern is input to a learning pattern storage unit 40. Next, the average vector

[0022]

(Equation 2)

(J = 1 to J: category number, n = 1 to
N _j : the number of feature vectors of the category j, p = 1 to P: the number of dimensions of the feature vector) are sent to the correlation degree calculation unit 50. Where the elements of the mean vector

[0024]

(Equation 3)

(P-dimensional element of the n-th feature vector of category j).

An embodiment of the correlation calculating section 50 will be described.

First, all learning patterns

[0028]

(Equation 4)

The average value μ (p) of each parameter of the feature vector over the range is determined. The average μ (p) is

[0030]

(Equation 5)

## EQU1 ##

Next, using the calculated parameter average value, the covariance matrix σ (p1, p
2) 1 <p1, p2 <P (p1, p2 are parameter numbers of the feature vector) are calculated.

[0033]

(Equation 6)

Next, the calculated covariance matrix σ (p1, p
2) and the correlation coefficient ρ (p1, p
2) is calculated.

[0035]

(Equation 7)

The correlation coefficient calculated by the correlation degree calculation section 50 is stored in the correlation degree storage section 60. Next, in the feature vector dividing unit 70, a determinant between the correlation coefficients is calculated based on the correlation coefficients calculated in 50, and parameters are collected.

Hereinafter, the feature vector dividing section 70 will be described. (1) First, each parameter is assumed to be independent, and is divided into P pieces so that each parameter is a partial vector.

R = PT (k), (1≤k≤P) (T (k) is the number of dimensions of the k-th partial vector) (k is the partial vector number) (2) Next, 1≤k , L ≦ r, parameters p1, p2, (1 ≦ p
1, p2 ≦ (T (k) + T (l)) correlation function ρ (p
1, p2) is read from the correlation degree storage unit 60, and (T
A correlation matrix C of (k) + T (l)) × (T (k) + T (l)) is created, and a determinant D (k, l) is obtained.

D (k, l) = det | C | (3) Next, the partial vectors k and l giving the minimum D (k, l) are combined into one partial vector.

T (k) = T (K) + T (l) (k <l) At this time, the numbers of the new partial vectors are
The smaller of the two.

Next, the partial vector numbers are renumbered based on the division information. At this stage, the number of divisions is reduced by one. (4) Next, r = r−1, and if r is larger than a predetermined threshold value K, the process returns to (2). This calculation is performed until r = K.

Finally, the partial vector number pv (p) to which the p-th element belongs and the number of dimensions T (k) of the k-th partial vector are obtained.

In the above procedure, the correlation sequence calculated from the covariance matrix of the feature vector as the degree of correlation has been described as an example, but other calculation methods are also possible.

Next, the clustering of the feature vectors is performed by the clustering section 80 for each partial vector.

For clustering, a method using the LBG algorithm (IEEE Trans.
un. , COM-28, 1PP. 84-95, Jan.
1980, hereinafter referred to as Document 6).

Hereinafter, an embodiment of the clustering unit will be described.

The controller 120 sends a partial vector number k (k
= 1 to K) and the cluster center number M of the partial vector number k
_k are sequentially sent to the clustering unit 80. The clustering unit 80 calculates the average vector stored in the learning storage unit 40.

[0048]

(Equation 8)

An element p for which pv (p) = k is extracted from among them, and is set as a T (k) -dimensional vector. The extracted T
(K) dimensional vector

[0050]

(Equation 9)

Assume that next,

[0052]

(Equation 10)

Then, M _k vectors are selected as the cluster centers. As this selection method, M _k numbers may be selected in numerical order or may be selected at random. M selected
_k cluster center values

[0054]

[Equation 11]

Is sent to the distance calculation 110.

The distance calculation unit 110 calculates each average vector stored in the learning pattern storage unit 40.

[0057]

(Equation 12)

And M _k sent from the clustering unit 80
Distance from each cluster center

[0059]

(Equation 13)

Is calculated and sent to the clustering unit 80.

As for the distance, the Euclidean distance can be used in the path cost DP.

The clustering unit 80 calculates the value of the cluster center.

[0063]

[Equation 14]

Is sent to the cluster center storage unit 100, and the cluster center storage unit 100 holds this. The clustering unit 80 calculates the Dc calculated by the distance calculation unit 110.
The cluster number h that takes the minimum value among l (j, n, k, h) is defined as member (j, n, k) = h (1 ≦ membe
r (j, n, k) ≦ M _k ) and sends it to the cluster member storage unit 90. member (j, n, k) is

[0065]

(Equation 15)

Indicates the number of the cluster to which each vector belongs. The cluster member storage unit 90 holds this.

Next, the clustering unit 80
er (j, n, k) is read from the cluster member storage unit 90, and the cluster center

[0068]

(Equation 16)

Belongs to the cluster of number h

[0070]

[Equation 17]

Update using the average value of Hereinafter, the above procedure is repeated until Dcl (j, n, k, h) calculated by the distance calculation unit 110 converges, and the final cluster center value is obtained.

[0072]

(Equation 18)

Is stored in the cluster center storage unit 100, and the cluster number membe to which each final average vector belongs
r (j, n, k) is stored in the cluster member storage unit 90.

Although the clustering at the partial vector number k has been described above, this operation is performed for k = 1 to K.

Next, the cluster number member to which each pattern belongs held in the cluster member storage unit 90
(J, n, k) and the cluster center value held in the cluster center storage unit 100

[0076]

[Equation 19]

Based on the information of the learning pattern storage unit 40,
The standard pattern is created in the pattern creating unit 130 using the learning pattern stored in the standard pattern.

First, the cluster center value stored in the cluster center storage unit 100

[0079]

(Equation 20)

Is read out and held. Next, the member (j, n, k) is read from the cluster member storage unit 90.
Read out the value of and keep it. What should be remembered for creating the average vector is

[0081]

(Equation 21)

Cluster center value and N × K mem
ber (j, n, k), and a standard pattern with a smaller amount of memory can be created. Path cost

[0083]

(Equation 22)

As for the value, the value stored in the learning pattern storage section 40 is used as it is, and one pattern is combined with the average vector calculated above. In the above example, only the average vector is targeted for clustering, but the path cost can also be targeted for clustering.

The standard pattern created here is sent to the standard pattern output unit 140 and output. The application of this method is not limited to the path cost DP. For example, a continuous HMM (B
-H. Juang, IEEE Trans. Acous
t. , Speech & Signal Processes
s. , ASSP-33,6, pp. 1404-141
3, 1985, hereafter referred to as Document 4), the average vector of the distribution and the like can be clustered by the method described here.

To use the created standard pattern for speech recognition, for example, the SPLIT method (Sugamura, Furui, "Large vocabulary word speech recognition using onomatopoeia standard pattern", IEICE
65-D, 8, pp 1014-1048, 1982, hereinafter referred to as Reference 7.). When the standard pattern created above is used for speech recognition, a recognition device with a small amount of memory and a small amount of calculation can be realized.

[0087]

According to the present invention, a standard pattern creating apparatus capable of creating a standard pattern having a higher recognition rate with fewer patterns than the conventional standard pattern creating apparatus can be obtained.

[Brief description of the drawings]

FIG. 1 is a block diagram showing an embodiment of a standard pattern creation device according to the present invention.

FIG. 2 is a block diagram showing one embodiment of a conventional standard pattern creation device.

FIG. 3 is a diagram illustrating a difference in a covering space depending on a level of a correlation between feature amounts;

FIG. 4 is a diagram illustrating a difference in a covering space depending on a level of a correlation between feature amounts;

[Explanation of symbols]

 Reference Signs List 10 voice input unit 20 analysis unit 30 learning unit 40 learning pattern storage unit 50 correlation degree calculation unit 60 correlation degree storage unit 70 feature vector division unit 80 clustering unit 90 cluster member storage unit 100 cluster center storage unit 110 distance calculation unit 120 control unit 130 pattern creation unit 140 standard pattern output unit 200 voice input unit 210 analysis unit 220 learning pattern storage unit 230 power clustering unit 240 LPC parameter clustering unit 250 pattern creation unit 260 standard pattern output unit

Continuation of the front page (56) References JP-A-4-363000 (JP, A) JP-A-1-233499 (JP, A) JP-A-5-119790 (JP, A) JP-A-4-111189 (JP) JP-A-58-147062 (JP, U) Patent 2800618 (JP, B2) JP 673345 (JP, B2) JP 6-7344 (JP, B2) IEEE Transactions on Communications Vol. COM-28, No. 1, January 1980, "An Analogism for Vector Quantifier Design", p. 84-95 IEEE Communication Magazines, Vol. 21, No. 9, December 1983, "Vector Quantitation: A Pattern-Matching Technique for Speech Coding", p. 15-21 Journal of the Acoustical Society of Japan, Vol. 44, No. 8, 1988, “Normalization of spectrogram using separate vector quantization” p. 595-602 (Published on August 1, 1988) Proceedings of the Fall Meeting of the Acoustical Society of Japan in 1987 2-5-16 "Spectrogram Normalization Using Vector Quantization" p. 81-82 (August 10, 1987) (58) Field surveyed (Int. Cl. ⁶ , DB name) G10L 3/00 515 G10L 9/18 H03M 7/30 JICST file (JOIS)

Claims (1)

(57) [Claims] A voice input unit for inputting voice, an analysis unit for analyzing input voice data and extracting a feature vector, a learning unit for learning a first standard pattern from the extracted feature vector, and a learning unit. A learning pattern storage unit that stores the obtained first standard pattern, a correlation degree calculation unit that calculates the degree of correlation between the feature vector elements, and calculates a correlation strength between the feature vector elements from the correlation degree. A feature vector dividing unit that divides a feature vector, a distance calculating unit that calculates a distance between patterns from the feature vector,
A clustering unit that clusters a learning pattern for each divided feature vector based on the vector division information and the inter-pattern distance; a cluster center storage unit that stores a cluster center obtained as a result of the clustering; and a pattern that forms each cluster. A standard pattern creation device, comprising: a cluster member storage unit for storing; and a standard pattern creation unit for creating a standard pattern based on the result of the clustering.