JPH07146938A - Device and method for comparing signal waveform data - Google Patents

Abstract

PURPOSE:To compare two signal waveform data with high precision at a high speed by computing and outputting the similarity between fuzzy standard pattern data and after-normalization feature vector time-series data to be decided. CONSTITUTION:A fuzzy standard pattern data generating means 15 computes and stores the fuzzy standard pattern data, obtained by making respective feature quantity component values of plural after-normalization feature vector time-series data obtained on the basis of plural signal waveform data into fuzzy numbers, in a fuzzy standard pattern data storage means 13. A fuzzy relation arithmetic means 17 computes fuzzy relation with the standard pattern data as to feature quantity components of respective frames and dimensions of the after-normalization feature vector time-series data to be decided which are stored in an after-normalization time-series data storage means 11. On the basis of the obtained fuzzy relation, a similarity arithmetic means 19 calculates the similarity. Thus, a belonging degree can accurately be judged from the fuzzy standard pattern data.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a signal waveform data comparison device, and more particularly to high speed and high precision comparison.

[0002]

2. Description of the Related Art Frequency fluctuations become a problem when two audio signals are compared. Frequency variation means that the frequency of a spoken word varies from person to person. The multi-template method is known as one of the methods for absorbing frequency fluctuations.

The multi-template method will be described.
For example, if the input audio signal is 25.6 m per frame,
s, FFT cepstrum analysis is performed at a frame period of 15 ms. The cepstrum coefficients from the first dimension to the tenth dimension are used as feature quantities for word speech recognition, and further all speech signals are converted into, for example, DP (dynamic programmin).
g) Using the matching method, normalize to 22 frames.

As a result, the value of the cepstrum coefficient in the first dimension to the tenth dimension can be obtained for each frame for one word. Such feature vector time-series data is created by the number (i) of speakers Hi (i = 1 to I) (see FIG. 21). Next, i feature vector time-series data is classified into one or more groups, and a standard pattern of a certain word is created.

Specifically, among the feature vector time series data for the number of people, the cepstrum coefficient values are compared for each frame and each dimension, and the total deviation of each frame and each dimension is less than a predetermined value. If so, classify into the same group. Then, the average of the cepstrum coefficient values of the feature vector time series data classified into the same group is calculated for each frame and each dimension, and this is stored as a standard pattern of the group. In this way, for one word, a plurality of standard patterns are created from the feature vector time series data for the number of people. That is, a plurality of standard patterns exist for one word, which is called a multi-template method. The standard pattern is created for each word to be registered.

Next, comparison of audio signals will be described. When the audio signal to be compared is input, the values of the cepstrum coefficients of the first dimension to the tenth dimension are calculated for each frame, as described above. Next, the cepstrum coefficient value for each frame and each dimension is compared with all the stored standard patterns.

The deviation between the standard pattern and the audio signal to be compared for each frame and for each dimension is calculated for each standard pattern, and the standard pattern having the smallest total is calculated. Then, the audio signal to be compared is determined to be the audio signal represented by the standard pattern with the smallest total.

As described above, by the multi-template method, it is possible to absorb the frequency fluctuation and compare the audio signals with high accuracy.

[0009]

However, the conventional multi-template method has the following problems. In order to perform highly accurate comparison, it is necessary to store as many templates as possible. However, storing a large number of templates (that is, storing a large number of standard patterns) not only requires a large amount of storage area, but also requires the same amount of calculation processing as the number of templates, and the calculation time required to compare all the templates is increased. It costs.

Particularly, in the multi-template method, a plurality of standard patterns are stored for one word. Therefore, even if the number of words is increased by one, the stored amount is increased by the number of standard patterns.

On the other hand, in order to increase the calculation speed,
Reducing the number of standard patterns to be stored results in a less accurate comparison.

SUMMARY OF THE INVENTION It is an object of the present invention to solve the above problems and provide a signal waveform comparison apparatus or method capable of comparing signal waveforms with high accuracy and high speed.

[0013]

A signal waveform data comparison device according to a first aspect of the present invention is an input means for inputting signal waveform data,
Feature amount extraction means for dividing the signal waveform data into a plurality of frames and extracting a frequency component for each frame as a frame feature amount, extracting a desired feature amount component value from each frame feature amount, and the extracted feature amount component When calculating the feature vector time-series data that connects the element vectors of each frame while arranging the element vectors that have the extracted multiple feature quantity components as the components of each dimension based on the values in the multidimensional vector space A sequence data calculating means, a normalizing means for normalizing the feature vector time series data on a time axis and calculating the normalized feature vector time series data, and a normalizing means for storing the obtained normalized feature vector time series data Post-normalization time-series data storage means, and stores each feature amount component value of the plurality of normalized feature vector time-series data obtained based on the plurality of signal waveform data. Fuzzy standard pattern data storage means for storing in advance fuzzy standard pattern data obtained by digitizing, and each of the normalized feature vector time series data of the determination target stored in the normalized time series data storage means. For the frame and the feature amount component value of each dimension,
A fuzzy relationship calculation means for calculating a fuzzy relationship with the fuzzy standard pattern data stored in the fuzzy standard pattern storage means, and based on the calculated fuzzy relationship, the fuzzy standard pattern data and the normalized feature vector time-series data to be judged. Similarity calculation means for calculating the similarity,
It is characterized by having.

In the signal waveform data comparison device according to the present invention, the similarity calculation means 1) obtains the average value of the feature amount component values for each dimension of each frame based on the fuzzy standard pattern data for each word. , The weight value is calculated for each frame and each vector space according to the degree to which the feature amount component value is larger than the average value, 2) based on the obtained weight value and the fuzzy relationship, a fuzzy standard pattern The feature is that the degree of similarity between the data and the normalized feature vector time series data to be determined is calculated.

In the signal waveform data comparison device according to the present invention, the normalizing means is 1) fuzzy classifying means for fuzzy classifying the feature amount component values of the respective feature vector time series data, and 2) fuzzy classifying. Representative feature point time series data calculation means for calculating representative feature points based on each feature amount component value, concatenating the obtained representative feature points in time series, and calculating representative feature point time series data, 3) the representative A normalized time series data calculating means for calculating and outputting the normalized feature vector time series data based on the characteristic point time series data.

In the signal waveform data comparison method according to the fourth aspect, the input signal waveform data is divided into a plurality of frames, the frequency component of each frame is extracted as a frame characteristic amount, and a desired signal is extracted from the frame characteristic amounts. The feature amount component values are extracted, and the element vectors having the plurality of extracted feature amount components as the components of each dimension are arranged in the multidimensional vector space based on the extracted feature amount component values, and the element vector of each frame is set. The combined feature vector time series data is calculated, the feature vector time series data is normalized on the time axis, the normalized feature vector time series data is calculated, and the obtained normalized feature vector time series data is obtained. And fuzzy obtained by fuzzy numbering each feature amount component value of a plurality of normalized feature vector time series data obtained based on a plurality of signal waveform data. The standard pattern data is stored in advance, and the fuzzy relationship with the fuzzy standard pattern data is calculated for each frame and each dimension of the feature amount component values of the normalized feature vector time-series data to be determined. Calculating a similarity between the fuzzy standard pattern data and the normalized feature vector time-series data to be determined based on the fuzzy relationship, and outputting the similarity.
Is characterized by.

[0017]

According to the signal waveform data comparison device or the method thereof, the input signal waveform data is divided into a plurality of frames, and the frequency component of each frame is extracted as a frame feature amount. A desired feature amount component value is extracted from each frame feature amount, and an element vector having a plurality of extracted feature amount components as respective dimension components is arranged in a multidimensional vector space based on the extracted feature amount component values. To do. The feature vector time series data in which the element vectors of each frame are connected is calculated, the feature vector time series data is normalized on the time axis, and the normalized feature vector time series data is calculated and stored.

Further, fuzzy standard pattern data obtained by fuzzy numbering each feature amount component value of a plurality of normalized feature vector time-series data obtained based on a plurality of signal waveform data are stored in advance. .

For each frame and each dimension of the feature amount component value of the normalized feature vector time series data to be judged,
The fuzzy relation with the fuzzy standard pattern data is calculated, and based on the obtained fuzzy relation, the similarity between the fuzzy standard pattern data and the normalized feature vector time series data to be judged is calculated, and the similarity is output. To do. As described above, the fuzzy standard pattern data is obtained by fuzzy numbering each feature amount component value of the normalized feature vector time-series data, so that the degree of belonging can be accurately determined.

According to another aspect of the signal waveform data comparison device of the present invention, the similarity calculation means obtains an average value of the feature amount component values for each dimension of each frame based on the fuzzy standard pattern data for each word. The weight value is calculated for each frame and each vector space according to the degree to which the feature amount component value is larger than the average value, and based on the obtained weight value and the fuzzy relationship, fuzzy standard pattern data and judgment target The degree of similarity with the normalized feature vector time series data of is calculated. Therefore, it is possible to highly evaluate the degree of similarity between the frame and each vector space in which the feature amount component value is larger than the average value.

In the signal waveform data comparison device of the third aspect, the fuzzy classifying means fuzzy classifies the respective feature amount component values of the respective feature vector time series data. The representative feature point time-series data calculation means calculates the representative feature points based on the fuzzy classified feature amount component values, connects the obtained representative feature points in time series, and calculates the representative feature point time-series data. To do. The normalized time series data calculation means is
The normalized feature vector time series data is calculated and output based on the representative feature point time series data. Thereby, the normalization can be calculated at a higher speed.

[0022]

DESCRIPTION OF THE PREFERRED EMBODIMENTS An embodiment of the present invention will be described with reference to the drawings. The signal waveform data comparison device 1 includes an input unit 3, a feature amount extraction unit 5, a time series data calculation unit 7, a normalization unit 9,
The normalized time series data storage means 11, the fuzzy standard pattern data storage means 13, the fuzzy standard pattern data creation means 15, the fuzzy relation calculation means 17, and the similarity calculation means 19 are provided.

Signal waveform data is input to the input means 3. The feature amount extraction means 5 divides the signal waveform data into a plurality of frames and extracts the frequency component of each frame as a frame feature amount. The time-series data calculation means 7 extracts a desired feature amount component value from each of the frame feature amounts,
Based on the extracted feature amount component values, the element vector that has the extracted multiple feature amount components as the components of each dimension is arranged in the multidimensional vector space, and the feature vector time series in which the element vectors of each frame are connected Calculate the data.

The normalizing means 9 is a fuzzy classifying means 31,
A representative feature point time-series data calculating means 32 and a normalized time-series data calculating means 33 are provided, and the given feature vector time-series data is normalized and normalized on the time axis as follows. The post-feature vector time series data is calculated.

The fuzzy classifying unit 31 fuzzy classifies the feature amount component values of the feature vector time series data. The representative feature point time-series data calculation means 33 calculates a representative feature point based on the fuzzy classified feature amount component values. Further, based on the feature vector time-series data given to the fuzzy classifying means 31, the obtained representative feature points are connected in chronological order to calculate the representative feature point time-series data.
The normalized time series data calculation means 35 calculates and outputs the normalized feature vector time series data based on the representative feature point time series data.

Returning to FIG. 1, the normalized time series data storage means 11 stores the obtained normalized feature vector time series data. Fuzzy standard pattern data creation means 1
Reference numeral 5 calculates fuzzy standard pattern data obtained by fuzzy numbering the respective feature amount component values of the plurality of normalized feature vector time-series data obtained based on the plurality of signal waveform data. Fuzzy standard pattern data storage means 1
3 stores the fuzzy standard pattern data.

The fuzzy relation calculating means 17 is a fuzzy standard pattern for each frame and each dimension of the feature quantity component value of the normalized feature vector time series data to be judged stored in the normalized time series data storage means 11. A fuzzy relationship with the fuzzy standard pattern data stored in the storage means 13 is calculated. The similarity calculation means 19 obtains an average value of the feature amount component values for each dimension of each frame based on the fuzzy standard pattern data for each word, and determines the degree of the feature amount component value being larger than the average value. The weight value is calculated for each frame and each vector space, and the similarity between the fuzzy standard pattern data and the normalized feature vector time series data to be determined is calculated based on the obtained weight value and the fuzzy relationship. To do.

FIG. 3 shows an example of a hardware configuration in which the signal waveform data comparison device according to the present invention is realized by using a CPU. The signal waveform data comparison device 21 includes the CPU 2
3, ROM25, RAM27, audio A / D conversion IF2
6, keyboard 28, CRT 29, and bus line 3
It has 0.

The ROM 25 stores a control program for the CPU 23 and the like, and the CPU 23 controls each unit via the bus line 30 in accordance with this control program. The keyboard 28 is a mode switching means for inputting a mode switching signal. The audio A / D conversion IF 26 is an input means for inputting signal waveform data.

The present apparatus has a mode for creating a fuzzy standard pattern (hereinafter referred to as a standard pattern creation mode) and a mode for comparing unknown signal waveform data with a fuzzy standard pattern stored in advance (referred to as a comparison mode).

[Operation of Standard Pattern Creation Mode] First,
The operation of the signal waveform data comparison device 21 in the standard pattern creation mode will be described with reference to FIG.

When signal waveform data is input from the audio A / D interface (IF) 26 (step ST in FIG. 3).
1), the CPU 23 divides the signal waveform data into a plurality of frames in chronological order and extracts the frequency component of each frame as a frame feature amount (step ST2). In the present embodiment, one frame is set to 25.6 ms, and the frame feature amount is calculated using the cepstrum analysis for obtaining the envelope information of the spectrum for each frame.

The CPU 23 calculates the frame feature amount
RAM 27 for finding characteristic points arranged in the multidimensional vector space
Remember. In this embodiment, a 10-dimensional vector space is adopted as the multi-dimensional vector space. FIG. 5 shows the state of arrangement in the 10-dimensional vector space. In this case, since the input signal waveform data has 30 frames, the characteristic points a1 to a30 are arranged. Note that these characteristic points a
1 to a30 are represented as vectors.

The above is expressed by the general formula as follows. The time series A of the feature vector representing the input signal waveform is A = a (o) a (1) ... a (i) ... a (I-1) (1) {a ( i) = (ai1, ai2, ..., aip, ..., aiP)}. Here, I is the input time length (the number of frames), and P is the number of dimensions of the vector.

Since this feature vector a (i) can be seen as one feature point in the P-dimensional space, by tracing the time series of this feature point, an approximate continuous curve in the P-dimensional space can be obtained. (Line) is obtained.

However, since there are fluctuations in the characteristic points, it is meaningless to obtain an approximate continuous curve in which the characteristic points a1 to a30 are directly tracked in time series. In the present embodiment, the CPU 23 fuzzy classifies the respective feature points to obtain representative feature points, connects the obtained representative feature points in time series, and obtains representative feature point time series line data.

The CPU 23 reads out the vector time series A1 of the characteristic points a1 to a30 from the RAM 27. The vector time series A1 is represented by A1 = a1, a2, ... A30.

The CPU 23 uses the input vector time series A
Fuzzy classification of each feature point vector a1 to a30 of 1,
Representative feature point vector (see a101 to a104 in FIG. 5)
Ask for. In this embodiment, the known theory, Fuzzy C-means, is adopted.

The concrete algorithm of the fuzzy C-means method is as follows. First, the degree (value from 0 to 1) that each individual Ai (Ai: vector; i = 1,2, ... N) belongs to each cluster g (g = 1,2, ..., G) is calculated as follows. Matrix (menbership matr
ix).

[0040]

[Equation 1]

Here, u (g, i) = [0,1] indicates that u (g, i) takes a value between 0 and 1.

Next, the number of clusters is set to G, and an initial division matrix U ^{(0) of} U and an appropriate convergence judgment value ε are given. Next, the average vector Vg of the initial cluster g (g = 1, 2, ..., G) is calculated by the equation (3).

[0043]

[Equation 2]

Next, U ⁽⁰⁾ is updated to U ⁽¹⁾ by the equation (4).

[0045]

[Equation 3]

The operator T is used to calculate U ⁽¹⁾ as T
⁽¹⁾ (U ⁽⁰⁾ ) and repeat this. That is, U ^{(k + 1)} = T ^{(k + 1)} (U ^(k) ) (k = 0,1,2, ...) (5).

The above iteration ends the calculation when | U ^{(k + 1)} −U ^(k) | ≦ ε. If | U ^{(k + 1)} −U ^(k) | ≦
If not ε, the above calculation is repeated.

The CPU 23 stores the representative feature point vector thus obtained in the RAM 27. In this state, the time series concept is removed from the obtained representative feature point vector. Therefore, the CPU 23 uses the RAM
The representative feature point vectors stored in 27 are connected in chronological order to obtain representative feature point time series line data.

Since each representative feature point vector is connected in time series, each representative feature point vector and each feature point vector a1.
The a30 is replaced with the nearest representative feature point vector in time series order and connected. For example, the feature point vector a1 is replaced with the closest representative feature point vector a101, and the feature point vector a2 is replaced with the closest representative feature point vector a1.
01, and the feature point vector a8 is replaced with the nearest representative feature point vector a102,
The feature point vector a30 is replaced with the nearest representative feature point vector a104. As a result, the representative feature point vectors a101 to a104 are connected in chronological order.

In this way, the vector time series A101 obtained by vector-quantizing the vector time series A1 is obtained.

The vector time series A101 is represented by A101 = a101, a102, a103, a104. A representative feature point time series line L1 representing the vector time series A101 is shown in FIG.

The CPU 23 normalizes the obtained representative feature point time series line. When the mode switching signal is applied to the keyboard 28, the normalized feature point time series line data is calculated based on the mode switching signal based on the representative feature point time series line data stored in the RAM 27.

The reason why such normalization is performed is as follows. First, the same phoneme (stationary part), for example, "O"
Even in the case of
This is because there is a variation in the distribution of characteristic points called "o" and the center point of the variation also fluctuates.

Further, the characteristic points in the vector space corresponding to the transition (non-stationary part: for example, "o" and "m") sections between different phonemes are different depending on the sampling time points on the time axis of the transition sections. , Their position is different. Therefore, those variations affect the length of the spatial curve. That is,
This is because even the characteristic point time series line data representing the same word have different lengths and shapes in the vector space.

The outline of the calculation of the normalized feature point time series line data performed by the CPU 23 is as follows. First, the length of the trajectory of the representative feature point time-series line data is calculated, and this is calculated as (N
-1) Divide into equal parts. A new locus is obtained based on this division point, and N feature points resampled along the obtained locus are used as the normalized feature points.

Next, a method for normalizing an approximate continuous curve in the P-dimensional vector space will be specifically described with reference to FIGS. 6 to 8.

First, the relationship between the input feature points and the interpolation points in the P-dimensional space will be described with reference to FIG. As shown in the figure, when there are input feature points a (i-1), a (i), and a (i + 1), the interpolation point x is represented by the following equation (6).

X = a (i) + (a (i + 1) -a (i)) · t; (0 ≦ t ≦ 1; i = 0, ..., I-2) (6 ) Further, when the length of the locus of A is represented by L (A), it is defined as in Expression (3).

[0059]

[Equation 4]

Incidentally,

[0061]

[Equation 5]

It is

Next, the locus length L (A) is divided into N-1 equal parts and resampled along the locus X to use N feature points as normalized feature points.

Here, the normalized characteristic points of N points to be resampled are B = b (0) b (1) ... b (n) ... b (N-1). 8) When represented by {b (n) = (bn1, bn2, ..., bnp, ..., bnP)}, the normalized feature points are the following (9) to (1
It is calculated by the formula 2). Note that b (0) = a (0)
Let (N-1) = a (I-1).

However, I, P, i, n, A and d shown in FIG.
(i), L (A), N-1, ΔL, B, D (A), D (B), S (n)
Indicates the following.

I: input time length (number of frames) P: vector dimension number i: input feature point number on the trajectory; i = 0, ..., I-1 n: normalized feature point on the trajectory Number; n = 0, ..., N-1 A: Time series of input speech feature vector; A = a (0) a (1) ... a (i) ... a (I) {a ( i) = ai1, ai2, ..., aip,
..., aiP)} d (i): distance between neighboring feature points;

[0067]

[Equation 6]

L (A): locus length of A;

[0069]

[Equation 7]

N-1: number of equal divisions on the locus ΔL: length of line segment of equal division on the locus B: time series of normalized feature vector; B = b (0) b (1) ... b (n) ・ ・ ・ b (N-1) {b (n) = (bn1, bn2, ・ ・ ・, bnp, ・ ・ ・, bnP)} D (A): A feature point a ( trajectory length up to i);

[0071]

[Equation 8]

D (B): Normal feature point b (n) along the locus
D (B) = n · ΔL S (n): Local trajectory length at resampling time b (n),
Specifically, as shown in FIG. 8, it is the trajectory length from the adjacent input feature points.

ΔL = L (A) / (N−1) (9)

[0074]

[Equation 9]

Here, i = number {k | (·
..)} means that the number when "k""satisfies the condition in (...) is taken as i.

T (n) = S (n) / d (i + 1); (0 ≦ t (n) ≦ 1) (11) b (n) = a (i) + (a (i +1) -a (i)) · t (n) (12) This calculation algorithm will be described with reference to FIG. In step ST11 of FIG. 9, initialization is first performed.

Next, steps ST12 to ST
At 15, S (n) is calculated. In step ST12, the distance d (i + 1) to the next input feature point is added to D (A) (here, D (A) = 0). In step ST13,
It is determined whether D (A)> D (B). That is, it is determined whether or not the distance d (i + 1) to the next input feature point is larger than ΔL.

If the conditions are satisfied, step ST14
Then, D (A) = D (A) -d (i + 1). And S (n)
= D (B) -D (A) (step ST15). Thereby, the trajectory length S (n) from the input feature point a (i) can be obtained.

In step ST13, D (A)> D
If (B) is not established, the process proceeds to step ST16 and i = i +
It is set to 1, and the distance to the next input feature point is added to D (A) (step ST17). Return to step ST12,
Steps ST13 to ST17 are repeated.

Next, the interpolation point b (n) to be resampled is obtained based on the obtained trajectory length S (n). The interpolation point b (n) is obtained based on the relationship between the input feature point and the interpolation point already described (step ST18). In step ST19, n =
When N-2 is reached, the process ends. Step ST19
If n = N-2 is not reached, step S
Proceeding to T20, n = n + 1 and D (B) = n · ΔL are set, and step ST12 and subsequent steps are repeated to obtain the trajectory length S (n +
Ask for 1).

The normalized feature point time series line SL1 thus obtained is shown in FIG. In the figure, feature point a101
To a104 are input feature points, and feature points b101 to b11
0 is the calculated interpolation point.

As described above, in this embodiment, the fuzzy C- is used to normalize the feature vector of the input signal waveform.
Since the means method is used, high-speed calculation is possible compared with the conventional DP matching method. This is because, unlike the DP matching method, it is not necessary to compare each one on the time axis plane, and the amount of calculation can be reduced. In the standard pattern creation mode, after performing such normalization, the process proceeds to step ST4 in FIG. 4 to create a fuzzy standard pattern. Specifically, the feature amount component value of the feature vector is calculated for each frame for each frame, and
A table showing the feature amount component values of the normalized feature vector in which the speaker i generated the word W1 J times as shown in A is created for each frame and each vector space.

In the present embodiment, the voice which is A / D converted at 10 kHz and 10 bits is 25.6 ms per frame,
The FFT cepstrum analysis was performed at a frame period of 15 ms, the cepstrum coefficients from the 1st to the 10th order were used as feature parameters for word speech recognition, and all the word speeches were normalized to 22 frames (that is, P = 10, L = 22).

Next, a table as shown in FIG. 10A is prepared for each word for a plurality of persons, and the feature amount component value of the feature vector is fuzzy numbered for each frame and vector space as shown in FIG. . In this embodiment, fuzzy numbers are obtained by the method shown in FIG. For example, I × J feature vectors F ⁿ _ij obtained by different I speakers speaking the word W _n J times and time-axis normalization are F ⁿ _ij = (f ⁿ _ij (L )) (I = 1, ..., I; j = 1, ..., J; L = 1, ...,
L). Here, f ⁿ _ij (L) is represented by the following formula.

F ⁿ _ij (L) = (f ⁿ _ij (L, 1), ..., f ⁿ _ij (L, p), ..., f ⁿ _ij (L, P)) where L : Frame number (L = 1, ..., L), the number of frames of the normalized word W _n , p: vector dimension number, feature vector dimension number, i: story Number of the speaker, j: number of utterances, I: number of speakers, J: number of utterances.

[0086] Next, the average feature pattern ^a f ⁿ of the word W _n
Ask for. The average feature pattern ^a f ⁿ is represented by the following formula.

[0087] ^{a f n = ((a f} n p (L), f n p, min (L), f n p, max (L)) (p = 1, ··
・, P; L = 1, ・ ・ ・, L)

[0088]

[Equation 10]

Here, ^a f ⁿ _p (L): average value of p-dimensional component (cepstral value) of the L-th frame, f ⁿ _{p, min} (L): L-th
The minimum value of the p-dimensional component of the frame, f ⁿ _{p, max} (L): The maximum value of the p-dimensional component of the L-th frame.

Specifically, as shown in the mathematical expression of FIG.
Different I speakers obtain the average value ^a f ⁿ _p (L) of the feature amount component values of the feature vector obtained by uttering the word W _n J times for each frame and each vector space. Further, among the values for each frame and vector space, the minimum value is f ⁿ _p , min and the maximum value is f ⁿ _p , max. Then, the average value ^a f ⁿ _p (L) is used as the vertex coordinates of the triangle representing the fuzzy number, and the minimum value f ⁿ _p , min and the maximum value f ⁿ _p , max are used as the base coordinates of the triangle.

That is, _f f ⁿ _p (L) = ( _f f ⁿ _p (L), _f f ⁿ _p (L)
-f ⁿ _{p, min} (L), f ⁿ _{p, max} (L) _-f f ⁿ _p (L)), and _f f ⁿ _p (L)
Is the right and left ambiguity _f f ⁿ _p (L) ー f ⁿ _{p, min} (L), f
^{_{n p, ma x (L)}} - is _f f ⁿ regular triangular fuzzy numbers become _p (L) (see FIG. 12).

As a result, the fuzzy number _f f ⁿ _p (L) shown in FIG. 12 is obtained. Such a standard pattern represented by a fuzzy number for each dimension and each frame is called a fuzzy standard pattern.

FIG. 16 shows fuzzy standard patterns of the words W ₁ "Tokyo" and W ₂ "Aichi" (both are the fuzzy feature vector of the 9th and 14th frames of the first dimension and the 3rd and 11th frames of the second dimension). (A fuzzy eye feature vector is displayed as an example).

Next, the case where the fuzzy standard patterns are created using the words "Tokyo" and "Aichi" as the words W1 and W2 will be specifically described.

Five speakers (I = 5) uttered five times (J = 5) for the word "Tokyo", and time-axis normalization (L
= 22), the 25 feature vectors (I ×
14A and 14B, the feature amount component values of the first and second dimensional components obtained from J) are arranged in time series.

In FIG. 14A, for the first dimensional component of the word W 1 "Tokyo", the feature amount component value of the 9th frame is the fuzzy number f f 1 1 (9) = ( a f 1 1 (9) , f 1 1, min (9),
Since f 1 1, represented by m ax (9)), fuzzy numbers f f 1 1 (9) is, f f 1 1 (9) = (0.07, -2.49, represented by 2,71). The feature amount component value of the 14th frame is the fuzzy number f f 1 1 (14) = ( a f 1 1 (14), f 1 1, min (14),
Since f 1 1, max (14)), the fuzzy number f f 1 1 (14) is represented by f f 1 1 (14) = (4.51, 0.52, 7.04).

Similarly, referring to FIG. 14B, the word W
₁ For the second dimension component of “Tokyo”, the feature amount component value of the third frame is the fuzzy number _f f ¹ ₂ (3) = ( ^a f ¹ ₂ (3), f ¹
_2, min (3), f ¹ _{2, max} (3)) ＝ (0.50, -1.95, 2.82), the first
The feature amount component value of the first frame is the fuzzy number _f f ¹ ₂ (11) = (
^a f ¹ ₂ (11), f ¹ _{2, min} (11), f ¹ _{2, max} (11)) ＝ (-3.23, -5.4
8, -1.19).

Further, FIGS. 15A and 15B are time series examples of the feature amount component values of the first and second dimensional components regarding the word W2 "Aichi" obtained in the same manner. The same applies to each dimension component of the word "Aichi".

In this way, a fuzzy standard pattern in which the feature amount component value for each dimension and for each frame is represented by the fuzzy number is created for each word (see FIG. 11).

Such a fuzzy standard pattern is created for each word W1 to WN, and a fuzzy standard pattern group as shown in FIG. 13 is obtained. In FIG. 13, for each template, the number of fuzzy numbers for each dimension for each of the words W1 to WN is equal to the number of normalized frames.

In FIG. 13, _f f ⁿ (L) means the fuzzy feature quantity component value of the word WN in the L-th frame and is represented by a vector. Note that _f f ⁿ (L) is _f f ⁿ (L) = ( _f f ⁿ , ₁ (L), _f f ⁿ , ₂ (L), ..., _f f ⁿ ,
_p (L), ... _f f ⁿ , _P (L)) (n = 1, 2, ...
・ ・ N)

Further, _f f ⁿ _p (L) is the P-th element of the fuzzy feature vector _f f ⁿ (L) and is a fuzzy number (see FIG. 12). Also, the ^a f _p (L), for the entire set of words the mean of fuzzy feature vector of the L-th frame is represented by the following equation (11).

[0103]

[Equation 11]

Further, ^a f (L) is expressed by the following equation.

[0105] _{^{a f p (L) = (}} a f 1 (L), a f 2 (L), ···, a
f _p (L), ... ^a f _P (L)) In this way, the fuzzy standard pattern corresponding to the voice signal of the word to be registered is created by the number of words, and the RAM 27
(Step ST5 in FIG. 4).

[Comparison Mode] The comparison mode for comparing the unknown signal waveform data with the fuzzy standard pattern stored in advance will now be outlined. When the unknown signal waveform is input (step ST1 in FIG. 4), the normalization is performed in steps ST2 to ST3, and the feature amount component value for each frame and each dimension (hereinafter, unknown signal waveform feature) shown in FIG. 10B. The quantity component value) is obtained. Here, if the mode switching signal given from the keyboard 28 is the comparison mode, the CPU 23 proceeds to step ST6 to obtain a fuzzy relationship between the unknown signal waveform characteristic amount component value and the fuzzy standard patterns of all registered words.

The CPU 23 then proceeds to step ST in FIG.
The degree of similarity between the fuzzy standard pattern stored in 5 and the input unknown signal waveform is determined (step ST8). Next,
In step ST9, the cumulative similarity is determined, and the input unknown signal waveform is determined to be the signal waveform corresponding to the fuzzy standard pattern having the maximum similarity (step S).
T10).

Each process will be described in detail below. First, the feature vector of the unknown signal waveform is f ^x _p (L) (p = 1, ...
,, 10; L = 1, ..., 22). Then, referring to the fuzzy standard pattern group shown in FIG. 13, a fuzzy similarity relationship matrix between the unknown signal waveform data and each word W _n is created.

Specifically, for the unknown signal waveform feature amount component value, the degree of conformity (0 to 1.0) with the fuzzy number stored in advance is obtained for each dimension and for each frame. As a result, the goodness of fit μ _L (P, W _n ) (p = 1 ... P, L = 1 ... L, W _n = 1 for each dimension and frame as shown in FIG. 20.
... N) can be obtained.

The above fuzzy relationship is shown in FIG.
As shown in 18. FIG. 17B shows time-series data of the first-dimensional component of the feature amount component value of the unknown input audio signal. 17A and 17C respectively show the components of the first and second frames of the first-dimensional component of the unknown signal waveform feature amount component value and the two fuzzy standard patterns (Tokyo,
Goodness of fit with 2 words in Aichi μ L (P, W n ) (p = 1, L = 9,14, WN = W1
= Tokyo). From FIG. 17B, for example, the feature amount component value of the ninth frame of the unknown input signal is “−0.5.
17 ", and this value is the fuzzy number in the 9th frame of the fuzzy standard pattern" Tokyo "in FIG. 17A.
f f 1, 1 (9) fit mu 9 with (1, W1) = found to be 0.78. Similarly, it can be seen that the compatibility μ 9 (1, W 2) with the fuzzy number f f 2 , 1 (9) in the ninth frame of the standard pattern “Aichi” is 0. Similarly, from FIG. 17C, the feature amount component value of the 14th frame and the standard pattern “Tokyo” are shown.
The conformity with “Aichi” is μ 14 (1, W1) = 0.80, μ, respectively
It can be seen that 14 (1, W2) = 0.

As for the second dimension component, as shown in FIGS.
Similarly, using C, μ ₃ (2, w ₁ ) = 0.57, μ ₃ (2, W2) =
It can be seen that 0, μ ₁₁ (2, W1) = 0.37, μ ₁₁ (2, W2) = 0.16. With respect to the goodness of fit thus obtained, the total goodness of fit is found for each fuzzy standard pattern. Specifically, the goodness of fit for each frame and each dimension may be summed up for each standard pattern. For example, in FIG. 19, the word W1
The cumulative similarity V _{1 with} is expressed by the following equation.

V ¹ = μ ₁ (1,1) + μ ₁ (2,1) + ... + μ ₁ (p, 1) + ...
・ + Μ ₁ (P, 1) + μ ₂ (1,1) + μ ₂ (2,1) + ・ ・ ・ + μ _L (P, 1) In this way, the cumulative similarity V _N of each standard pattern And the word WNA corresponding to the maximum cumulative similarity is output as the recognition result.

As described above, in this embodiment, unlike the conventional multi-template method, it is not necessary to store many templates, and one fuzzy standard pattern may be stored for each word. Therefore, the storage area can be reduced and the number of calculation processes can be reduced. At that time, since the standard pattern is stored in a fuzzy number, it is possible to perform highly accurate comparison by only storing one standard pattern for each word. Surawa,
The signal waveforms can be compared with high accuracy and high speed.

[Regarding Weighting] In the above calculation method, the similarity is evaluated for each frame and each vector space regardless of the magnitude of the feature amount component value.
However, by weighting so that the degree of similarity in the frame having a large feature amount component value and each vector space is highly evaluated, the degree of similarity can be evaluated more accurately. This is because the feature of the signal waveform appears more strongly in the frame having a large feature amount component value and each vector space.

The weight of the characteristic amount component value is obtained as follows. Each element of fuzzy relation matrix μ _L (p, WN) (L =
1, ・ ・ ・, L; p = 1, ・ ・ ・, P; N = 1 ・ ・ ・ N) as W _p (L)

[0116]

[Equation 12]

It is calculated by

For example, W ₁ (9) is obtained as follows. From the above formula (14), ^a f ₁ (9) = ( ^a f ¹ ₁ (9) + ^a f ²
_{It is 1} (9)) / 2. Here, the ^a f ^{₁ 1} (9) and ^a f ² ₁ (9), from Fig. 16A, each 0.07, because it is + 5.06,
Substituting gives ^a f ₁ (9) = 2.57. In addition, (13) above
From the formula, W ′ ₁ (9) = f ^X ₁ (9) / ^a f ₁ (9), so W ′
₁ (9) =-0.57 / 2.57 = -0.22. The obtained _W'p (L) is normalized as shown in the equation (12). That is, W
₁ (9) is represented by W ₁ (9) = (-0.22) / (-0.22 + W ' ₂ (9) + ... + W' ₁₀ (9)).

In this way, after the fitness is multiplied by the weight obtained for each dimension and for each frame (see FIG. 20), the cumulative similarity may be obtained as described above.

In this way, by weighting each dimension and each frame, more accurate similarity determination can be performed.

[Other Application Examples] In this embodiment, since the fuzzy C-means method is used to normalize the input signal waveform, high-speed calculation is possible as compared with the conventional DP matching method. Is. However, the present invention is not limited to this, and any method may be used as long as the feature vector of the input signal waveform can be normalized, and for example, a conventional DP matching method or the like may be used for the portion.

In this embodiment, the audio signal is used as the input signal waveform data, but any signal waveform data capable of extracting the characteristic amount of the signal may be used. For example, it can be applied to collation of handwriting signals.

Although the mode switching signal is input from the keyboard 28 in this embodiment, it may be supplied from another device (not shown) via the bus line 30.

In this embodiment, a plurality of people utter a plurality of times to create a fuzzy standard pattern.
The number of times is not limited to this, and the same person may be used. In the above embodiment, in order to realize the function shown in FIG.
This is realized by software using the CPU 23. However, some or all of them may be realized by hardware such as a logic circuit.

[0125]

According to the signal waveform data comparison apparatus or the method of the present invention, the input signal waveform data is divided into a plurality of frames, and the frequency component of each frame is extracted as a frame feature amount. . A desired feature amount component value is extracted from each frame feature amount, and an element vector having a plurality of extracted feature amount components as respective dimension components is arranged in a multidimensional vector space based on the extracted feature amount component values. To do. The feature vector time series data in which the element vectors of each frame are connected is calculated, the feature vector time series data is normalized on the time axis, and the normalized feature vector time series data is calculated and stored.

In addition, fuzzy standard pattern data obtained by fuzzy numbering each feature amount component value of a plurality of normalized feature vector time-series data obtained based on a plurality of signal waveform data are stored in advance. .

For each frame and each dimension feature value component value of the normalized feature vector time series data to be determined,
The fuzzy relationship with the fuzzy standard pattern data is calculated, and the similarity between the fuzzy standard pattern data and the normalized feature vector time series data to be judged is calculated based on the obtained fuzzy relationship, and the similarity is output. To do. As described above, since the standard pattern data is obtained by fuzzy numbering each feature amount component value of the normalized feature vector time-series data, the degree of belonging can be accurately determined. Accordingly, it is possible to provide a signal waveform data comparison device or a method thereof that can compare two signal waveform data at high speed and with high accuracy.

In the signal waveform data comparison device of the second aspect, the similarity calculation means obtains the average value of the feature amount component values for each dimension of each frame based on the fuzzy standard pattern data for each word, and The weight value is calculated for each frame and each vector space according to the degree to which the feature amount component value is larger than the average value, and based on the obtained weight value and the fuzzy relationship, fuzzy standard pattern data and judgment target The degree of similarity with the normalized feature vector time series data of is calculated. Therefore, it is possible to highly evaluate the degree of similarity between the frame and each vector space in which the feature amount component value is larger than the average value.

As a result, it is possible to provide a signal waveform data comparison device capable of comparing two signal waveform data with higher accuracy.

In the signal waveform data comparison device of the third aspect, the fuzzy classifying means fuzzy classifies the respective feature amount component values of the respective feature vector time series data. The feature point time series data calculating means calculates representative feature points based on the fuzzy classified feature amount component values, and also connects the obtained representative feature points in time series to calculate feature point time series data. The normalized time series data calculating means calculates and outputs the normalized feature vector time series data based on the characteristic point time series data. Thereby, the normalization can be calculated at a higher speed.

Therefore, it is possible to provide a signal waveform data comparison device capable of comparing two signal waveform data at higher speed.

[Brief description of drawings]

FIG. 1 is a functional block diagram of a signal waveform data comparison device 1 according to the present invention.

FIG. 2 is a block diagram showing a configuration of a normalizing means 9.

FIG. 3 is a diagram showing a hardware configuration in which the signal waveform data comparison device 1 is realized by a CPU.

FIG. 4 is a flowchart of overall processing of the signal waveform data comparison device 1.

FIG. 5 is a diagram showing a state in which feature points are arranged in a P-dimensional vector space.

FIG. 6 is a diagram showing a normalized feature point time series line SL1.

FIG. 7 is a diagram showing a relationship between input feature points and interpolation points in a P-dimensional vector space.

FIG. 8 is a diagram for explaining N-point normalized feature points to be resampled in a P-dimensional vector space.

FIG. 9 is a diagram showing a calculation algorithm for normalization.

FIG. 10 is a diagram showing feature amount component values of a signal waveform uttered J times by a speaker i.

FIG. 11 is a diagram showing fuzzy standard pattern data of a word Wn obtained based on a plurality of feature amount component values.

FIG. 12 is a diagram for explaining a method of performing fuzzy number conversion based on a plurality of characteristic component values.

FIG. 13 is a diagram showing a fuzzy standard pattern data group of words W1 to WN.

FIG. 14 is a diagram showing fuzzy feature amount component values of the word “Tokyo” in frame order (time series) for each dimension.

FIG. 15 is a diagram showing fuzzy feature amount component values of the word “Aichi” in frame order (time series) for each dimension.

FIG. 16 is a diagram showing an example of a fuzzy number for each frame of the first-dimensional component of the feature amount component values of words W1 and W2.

FIG. 17 is a diagram showing a fuzzy relationship between a normalized feature vector of an unknown voice signal and a fuzzy standard pattern in the first-dimensional component.

FIG. 18 is a diagram showing a fuzzy relationship between a normalized feature vector of an unknown voice signal and a fuzzy standard pattern in a second-dimensional component.

FIG. 19 is a diagram showing a fuzzy relationship with a fuzzy standard pattern for each dimension of an unknown voice signal and a feature amount component value in each frame.

FIG. 20 is a diagram for explaining a case where weighting is given to feature amount component values.

FIG. 21 is a diagram for explaining a standard pattern in the conventional multi-template method.

[Explanation of symbols]

 3 ... Input means 5 ... Feature amount extraction means 7 ... Time series data calculation means 9 ... Normalization means 11 ... Normalized time series data storage Means 13 ... Fuzzy standard pattern data storage means 17 ... Fuzzy relation calculation means 19 ... Similarity calculation means

Claims (4)

[Claims] 1. Input means for inputting signal waveform data, feature amount extracting means for dividing the signal waveform data into a plurality of frames, and extracting a frequency component for each frame as a frame feature amount, from each frame feature amount. A desired feature amount component value is extracted, and based on the extracted feature amount component value, an element vector having a plurality of extracted feature amount components as components of each dimension is arranged in a multidimensional vector space, and Time-series data calculation means for calculating feature vector time-series data in which element vectors are connected, the feature vector time-series data is normalized on a time axis, and normalization means for calculating the normalized feature vector time-series data, Normalized time series data storage means for storing the obtained normalized feature vector time series data, and a plurality of obtained plurality of signal waveform data based on the plurality of signal waveform data. Fuzzy standard pattern data storage means for pre-storing fuzzy standard pattern data obtained by fuzzy numbering each feature amount component value of the normalized feature vector time series data, and stored in the normalized time series data storage means A fuzzy relation calculating means for calculating a fuzzy relation with the fuzzy standard pattern data stored in the fuzzy standard pattern storage means for each frame and each dimension of the feature amount component values of the normalized feature vector time series data to be judged. A signal waveform data comparison device, comprising: similarity calculation means for calculating the similarity between the fuzzy standard pattern data and the normalized feature vector time-series data to be determined based on the calculated fuzzy relationship. 2. The signal waveform data comparison device according to claim 1, wherein the similarity calculation means 1) calculates an average value of feature amount component values for each dimension of each frame based on fuzzy standard pattern data for each word. The weight value is calculated for each frame and each vector space according to the degree to which the feature component value is larger than the average value, and 2) the fuzzy standard is calculated based on the obtained weight value and the fuzzy relationship. A signal waveform data comparison device characterized in that the degree of similarity between the pattern data and the normalized feature vector time series data to be judged is calculated. 3. The signal waveform data comparison device according to claim 1 or 2, wherein the normalizing means 1) fuzzy classifying means for fuzzy classifying the respective feature amount component values of the respective feature vector time series data, 2 ) Representative feature point time series data calculating means for calculating representative feature points based on the fuzzy classified feature amount component values, concatenating the obtained representative feature points in time series, and calculating representative feature point time series data , 3) A signal waveform data comparison device characterized by comprising normalized time series data operation means for calculating and outputting normalized feature vector time series data based on the representative characteristic point time series data. . 4. The input signal waveform data is divided into a plurality of frames, the frequency component of each frame is extracted as a frame feature amount, and a desired feature amount component value is extracted from each frame feature amount and extracted. Based on the feature amount component values, the element vectors that have the extracted multiple feature amount components as the components of each dimension are arranged in the multidimensional vector space, and the feature vector time-series data that connects the element vectors of each frame is calculated. Then, the feature vector time-series data is normalized on the time axis, the normalized feature vector time-series data is calculated, and the obtained normalized feature vector time-series data is stored and converted into a plurality of signal waveform data. Fuzzy standard pattern data obtained by fuzzy numbering each feature amount component value of the plurality of normalized feature vector time-series data obtained based on For each frame and dimension feature value of the normalized feature vector time-series data to be judged, a fuzzy relation with the fuzzy standard pattern data is calculated, and based on the obtained fuzzy relation, fuzzy standard pattern data is obtained. A signal waveform data comparison method characterized in that a similarity with the normalized feature vector time-series data to be judged is calculated and the similarity is output.