JPH0223876B2 - - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION The present invention is an improvement of a device that compares two patterns expressed as a time series of feature vectors, such as speech, and calculates a measure for evaluating their similarity (hereinafter referred to as similarity). Regarding.

Although the scope of application of the present invention is not limited to the recognition of voice patterns, the following description will take the recognition of voice patterns as a typical case.

The method of recognizing speech patterns on a word-by-word basis is relatively simple and avoids problems such as articulatory combination, so it is considered a practical means of speech recognition. However, even when the same word is uttered by the same person, the utterance speed is not uniform, each phoneme that makes up the word expands and contracts, and the overall time axis exhibits nonlinear and complex fluctuations. For this reason, it is necessary to normalize such nonlinear expansion and contraction to perform comparison processing, and various methods have been devised.

Among these, the time-normalized matching method (hereinafter referred to as DP matching method) using dynamic programming is a relatively simple yet extremely effective method (Reference: Journal of the Acoustical Society of Japan, September 1971). Issue Vol.27
No.9P483-487, hereinafter referred to as cited document (1)). This method models fluctuations on the time axis by approximating them with a monotonically increasing and continuous nonlinear function (hereinafter referred to as time normalization function), and then calculates the nonlinear function that maximizes the similarity between the input pattern and the standard pattern and normalizes it. , and obtain the similarity normalized with respect to the speaking rate.

However, in such a DP matching method, since the time normalization function is determined by the degree of similarity of each feature of each time point, it also allows considerably unrealistic deformation of the time axis. A method to prevent this is to provide a time axis alignment window (cited document (1)),
A method of setting an inclination limit (Japanese Patent Application No. 49-2418, hereinafter referred to as cited document (2)) has been proposed. Generally, in a word that is uttered, the amount of expansion and contraction of the time axis is small in consonant parts, and large in vowel parts. Furthermore, in the consonant part, although the absolute amount of time axis expansion/contraction is small, the amount of change in features between time frames is large. If prevention of unrealistic distortion of the time axis is achieved by limiting the amount of time axis distortion, it is difficult to achieve correct correspondence between phonemes. There is a risk that you may not be able to perform to your full potential. on the other hand,
There are phoneme pairs that are distinguished by the rate of change of their features, but with the DP matching method, it is difficult to distinguish between these phonemes based on the difference in rate of change.

The present invention improves time axis normalization accuracy to realize precise correspondence between phonemes, and enables discrimination between phoneme pairs distinguished by change speed, which is difficult with general DP matching methods. The purpose is to make it easier to distinguish between two confusing words.

Next, before describing the configuration of the pattern-to-pattern similarity calculating device according to the present invention, its operating principle will be explained.
Generally, speech patterns are expressed as a time series of feature vectors.

Now, input pattern A = ₁ , ₂ , ..., _i , ..., _I standard pattern B = 1, _{2 ,} ..., _j , ..., _J
......(1) Consider two speech patterns.

Here, 〓 _i = (a _1i , a _2i , ..., a _oi , ...a _Ni ) ......(2) is a vector indicating the characteristics of the speech at time point i, and a _oi is, for example, N One can consider the output of a frequency analyzer in a channel.

Consider the problem of calculating the similarity between these two patterns A and B. Generally, since the time axis shows non-linear and complex fluctuations, it is necessary to match and compare the time axis fluctuations between both patterns. By using the DP matching method as described in document (1), it is possible to optimally determine a time normalization function that provides correspondence between time points i and j. As the degree of similarity between the feature vectors 〓 _i and 〓 _j , the following value is used in literature (2): d _1(i,j) =〓〓〓〓〓〓〓〓〔a _oi −b _oj 〕 ......(3). In the present invention, the similarity between the amount of change in the feature vector d _2(i,j) =〓〓〓〓〓〓〓〓〓〔a _oi /a _oi-1 −b _oj /b _oj-1 〕……(4) and the above ( The weighted sum d _{(i,j) =} kd _1(i,j) + d _2(i,j) (5) with equation d 1 in 3) is used instead of the similarity between vectors. It is a feature.

By changing the value of weight k in equation (5), it is possible to change the weight ratio between the similarity between feature vectors and the similarity between variation amounts.

DP matching is performed using this d _(i,j) . That is, based on the initial value g _(1,1) = d _(1,1) ......(6), the recurrence formula g _{(i, j) =} min [g _(i,j-1) +d _(i,j) g _(i-1,j-1) +2・d _(i,j) g _(i-1,j) +d _(i,j) 〕 ………… (7) is solved under the conditions ₁ ≦ _i ≦ _I , ₁ ≦ _j ≦ _J , _jr ≦ _i ≦ _j+r (8), and the resulting integral quantity is
Find the inter-pattern similarity S (A _, B) = g _{(i, j)} / _{I + J} (9) using g (i, j).

As an example, both patterns A and B are the number "1"
Consider the case where the voice is (/it∫i/). In FIG. 1, the horizontal axis shows the time axis i of input pattern A, and the vertical axis shows the time axis j of standard pattern B. As shown in FIG. 1, the amount of expansion and contraction of the time axis of each phoneme, especially the vowel /i/, is large. Considering this expansion and contraction, pattern A,
A time normalization function 10 that optimally matches time points i and j of B is determined. The amounts of change in patterns A and B form curves as shown by reference numerals 11 and 12 in FIG. In this way, the amount of change generally takes maximum and minimum points at the points where the phoneme changes. Therefore, by considering the amount of change in this neighborhood, it is possible to accurately match the parts where the phoneme changes. Furthermore, since the value of the amount of change between feature vectors is also added to the similarity evaluation, it is possible to distinguish between phonemes that are distinguished by the rate of change of the feature amount. As described above, the present invention improves the accuracy of calculating similarity between patterns by adding a function to evaluate the amount of change in feature vectors to the time normalization algorithm of the DP matching method.

In formula (3), distance is used as a similarity evaluation measure, but other evaluation measures (for example, inner product) may be used. Further, although the ratio value is used as the amount of change in equation (4), the difference (a _oi −a _oi-1 ) may also be used. Also,
Other recurrence formulas shown in equations (6), (7), and (8) can also be used, for example, the equations shown in document (2).

The configuration of the inter-pattern similarity calculation device according to the present invention includes an input pattern buffer for holding an input pattern A expressed as a time series of feature vectors 〓 _i , and a standard pattern B expressed as a time series of feature values 〓 _j . A standard pattern memory for storing , and a change amount for calculating the amount of change between feature values 〓 _i and 〓 _i-1 and 〓 _j and 〓 _j-1 that are phases in a time series according to a predetermined order. A calculation unit, a similarity calculation unit 1 that calculates the similarity d _2(i,j) based on the distance between the changes, and the similarity d between the feature values 〓 _i and 〓 _j specified by the order. a similarity calculation unit 2 that calculates _1(i,j) ;
At least a similarity addition unit that calculates the weighted sum of the similarities to obtain the similarity d _(i,j) and an integral g _(i,j) calculated based on the similarity d _(i,j) are required. The weighted sum of the calculation register for holding the number, the output d _{(i, j)} from the similarity adder, and the integral quantity g _{(n, o)} [m≦i, n≦j] etc. in the calculation register. It consists of a recurrence formula calculation section that calculates multiple values and finds the minimum (or maximum) value as the integral quantity g _(i,j) .

The present invention will be explained using the drawings.

FIG. 2 is a block diagram showing an embodiment of the inter-pattern similarity calculating device of the present invention. In the figure 2
1 is an input pattern buffer that holds input pattern A; 22 is a standard pattern memory in which standard pattern B is stored in advance. Reference numeral 23 denotes a similarity calculation unit, which calculates the similarity d _(i,j) between 〓 _i and 〓 _j shown in equation (5). Reference numeral 24 denotes a recurrence formula calculation unit, which calculates a recurrence formula as shown in equation (7) from the output d of the similarity calculation unit 23 and the integral amount g calculated in advance. A register section 25 holds the value of g necessary for calculating the recurrence formula. 26 is a normalization unit, which normalizes the obtained g _{(I, J)} according to the number of frames to obtain inter-pattern similarity S (A, B). 2
Reference numeral 7 denotes a control section, which gives operation timing to each section.

FIG. 3 is a block diagram showing the circuit configuration of the similarity calculation section 23. 30 and 31 are shift registers,
It has the number of stages equal to the number of dimensions N of the feature vectors 〓, 〓, and holds one value of the vector. Divider circuits 32 and 33 calculate the ratio between two inputs.
34 and 35 are distance calculation circuits that calculate the absolute value of the difference between two inputs. 36 is a constant circuit whose input is multiplied by a constant k. 37 is an addition circuit;
Calculate the sum of 3 inputs. 38 is an integral register circuit which holds an intermediate value for calculating the sum of distance values. Also, 32, 33 division circuits, 34, 3
Each of the calculation circuits, ie, the distance calculation circuit 5, the constant circuit 36, and the addition circuit 37, each has an attached latch circuit for holding the output result.

Next, the operation will be explained based on the above configuration.

First, pattern A is input to the input pattern buffer 21 and held. A standard pattern B is stored in the standard pattern memory 22 in advance. Control part 2
According to control signals _{i and j} from 7, patterns A and B
The feature vectors 〓 _i , 〓 _j are sequentially output. The similarity calculation unit 23 calculates the similarity d _(i,j) between 〓 _i and 〓 _j . This operation will be explained with reference to timing charts shown in FIGS. 3 and 4. In addition, the fourth
In the figure, the shaded area indicates calculation time. First, the integration register 38 is reset at timing _T0 . Subsequently, at timing _T1 , the input pattern buffer 21 outputs one element a _oi of the vector 〓 _i shown in equation (2). Following this, at timing L ₁ , the shift register 30 is shifted by one stage, so that the a _o,i is written to the first stage of the shift register 30, and the value N stages before, that is _{, a o,i-1} is written to the last stage. It is output from Therefore, at this point, the division circuit 32 is given one element a _oi of the vector 〓 _i and one element _{a o, i-1} of the vector 〓 i-1. Based on these values, the division circuit 32 calculates the amount of change a _o,1 /a _o,i-1 at timing _T2 , and the result is held in the attached latch circuit at timing _L2 . Similarly, vector 〓 one element of _j b _o,j
Also, at the same timing as a _oi , the change amount b _oj /b _oj-1 is determined by the shift register 31 and the division circuit 33.
is calculated. In the distance calculation circuit 32, the timing
At T ₃ , the above two changes a _oi /a _oi-1 and b _oj /b _oj-1
The distance is calculated and held in the attached latch circuit at timing _L3 . In addition, in the distance calculation circuit 35, the
The distance between a _oi and b _oj is calculated at timing T ₄ and held in the attached latch circuit at timing L ₄ . Further, in the constant circuit 36, the output from the distance calculation circuit 35 is multiplied by a constant k at timing _T5 , and the result is held at timing _L5 . Addition circuit 37
Then, at timing _T6 , the output of the integral register 38, the output of the distance calculation circuit 34, and the output of the constant circuit 36 are added, and the result is held in the attached latch circuit at timing _L6 . This output result is held in the integral register 38 at timing _L7 . Sequentially, through these operations, the value of n in equation (2) is set to n=1,
_{2 , .} When the calculation of the similarity d _(i,j) between the vectors 〓 _i and 〓 _j is completed, the similarity calculation section 23 in FIG _.
is output.

In the recurrence formula calculation unit 24, the similarity degree d _{(i, j)} and the integral quantity g _{(i, j-1)} held in the register unit 25,
The recurrence formula (7) is calculated by inputting g _(i-1,j-1) and g _(i-1,j) , and the calculation result g _(i,j) is stored in the register section 25. Retained. The register unit 25 holds at least as many g values as are necessary for asymptotic equation calculation. This recurrence formula calculation is performed according to control signals tw and tr from the control section 27, but since these operations are common to a pattern matching device using a general DP matching method, detailed explanations will be omitted. As an example, the values at time point _{(i, j) are} (1, 1), (2, 1),
..., (i, 1), ..., (I, 1), (1, 2), (2,
2), ..., (I, 2), ..., (1, j), (2, j),
..., (I, j), ..., (I, j), g(I, J) can be obtained.

g(I, J) obtained by the recurrence formula calculation unit 24
On the other hand, the normalization unit 26 normalizes it by dividing it by the sum of the number of frames as shown in equation (9) to find the inter-pattern similarity S.

Although the principle of the present invention has been explained above using a simple example, the formula and circuit form are not limited to this example.

Furthermore, although the description has been made using voice pattern recognition as an example, it goes without saying that the present invention is not limited to this, and can also be applied to pattern recognition of characters, graphics, fingerprints, and the like.

[Brief explanation of drawings]

Figure 1 is a diagram of the principle of time axis normalization of audio patterns.
FIG. 2 is a block diagram showing an embodiment of the present invention, FIG. 3 is a block diagram showing an embodiment of the similarity calculation section of FIG. 2, and FIG. 4 is a signal for operating the circuit shown in FIG. 3. FIG. 3 is a waveform diagram showing an example of the occurrence time of the . In the figure, 10 is the time axis normalization function of pattern A and pattern B, and 11 and 12 are patterns A and B.
21 is an input pattern buffer, 22 is a standard pattern memory, 23 is a similarity calculation section, 24 is a recurrence formula calculation section, 25 is a register section, 26 is the normalization part, 3
0 and 31 are shift registers, 32 and 33 are division circuits, 34 and 35 are distance calculation circuits, 36 is a constant circuit, 37 is an addition circuit, and 38 is an integration register, respectively.

Claims (1)

[Claims] 1. An input pattern buffer for holding an input pattern A expressed as a time series of feature quantity _i ;
Feature value = Standard pattern B expressed as a series of _j
Changes that calculate the amount of change between adjacent features 〓 _i and 〓 _i-1 and 〓 _j and 〓 _j-1 in the time series using a standard pattern memory for storing , and a predetermined order. a similarity calculation unit 1 that calculates a similarity d _{2 (i, j)} based on a measure such as a distance between the amounts of change;
A similarity calculation unit 2 calculates the similarity d _1(i,j) between the feature quantities 〓 _i and 〓 _j specified by the order, and calculates the weighted sum of the similarities and calculates the two feature quantities 〓 _i , 〓 Similarity d _(i,j) between _j
, a calculation register for holding at least the required number of integral quantities g _(i,j) calculated based on the similarity d, and an output from the similarity addition unit.
d _(i,j) and the integral amount g _(n,o) in the calculation register [m≦i,
n≦j], etc., and calculates the minimum value (or maximum value) of the weighted sums as the integral quantity g _(i,j) . Similarity calculation device.