JPH0199095A - Pattern generator - Google Patents

Abstract

PURPOSE: To improve the accuracy of recognition by calculating an optimum parameter so that a distance between a partial pattern and a vector sequence on a curve shown by a time function expressed by vector values is minimized (maximized). CONSTITUTION: A mean value calculation part 5 reads out a feature vector included in a section set up by a partial section setting part 4 from an input buffer memory 2 and finds out a mean value. Then a least square approximate straight line calculating part 6 finds out the residual square sum of the least square approximate straight line of feature vectors in a partial section and the feature vectors. An average vector direction vector obtained by the calculating part 6 is stored in a reference pattern storing part 7 based upon a section number applied from the setting part 4. Since a voice pattern consisting of a feature vector sequence is converted into a parameter string for expressing the least square approximate straight line of each partial section, the accuracy of recognition can be improved.

Description

DETAILED DESCRIPTION OF THE INVENTION Field of the Invention The present invention relates to an apparatus for creating standard patterns in speech recognition and the like.

BACKGROUND OF THE INVENTION A case in which word speech is recognized will be described below. In addition, differences between vectors or patterns can be expressed as similarity,
Words such as distance and error are used, and there are various measures for each, but since they are not relevant to the present invention, the word distance will be used here to represent them. That is, for example, short distance or small distance corresponds to high similarity or large similarity, while long distance or large distance corresponds to low similarity or low similarity. This corresponds to the fact that it is small.

A so-called DP matching method is often used as a method for recognizing a pattern consisting of a series of feature vectors, such as in speech recognition. In this method, a pattern consisting of a series of feature vectors representing the word sounds to be recognized is registered in advance for each of the word sounds to be recognized as a standard pattern. The input cover pattern is compared with each of the standard patterns 3', -, the standard pattern with the smallest distance is searched, and the word corresponding to the standard pattern is taken as the recognition result of the input cover pattern. At this time, it is necessary to nonlinearly expand and contract the time axis of patterns with different time lengths, but in order to do this efficiently, dynamic programming is used.
This is a method called P matching, and it is currently the method that provides the best results.

However, in this method, the time axis is expanded or contracted so that the two patterns to be compared are closest in terms of distance. (hereinafter referred to as dynamic features) may not be reflected appropriately. Therefore, this method alone is insufficient in terms of recognition accuracy for phonemes that are characterized by the way their spectra change.

A shortened word dictionary is held in the form of symbol sequences representing phonemes and syllables (hereinafter referred to as phoneme segments), and standard patterns corresponding to each phoneme segment are prepared in advance and recognized. The power input cover pattern is converted into a speech segment sequence, that is, a symbol series representing each speech segment, based on the standard pattern, and each word in the word dictionary is matched with the symbol series to find the one with the largest error. There are some that recognize small words. At this time, the speech segment sequence converted from the input cover pattern contains some errors such as insertions, omissions, and substitutions, since it is impossible to completely recognize speech segments as a practical matter. I'm here. Therefore, in matching the symbol sequences, distances between speech segment sequences are determined by DP matching based on distances between speech segment sequences that have been calculated and prepared in advance. In this case as well, it is important to reflect the dynamic characteristics when performing speech segment recognition on the input cover pattern or when calculating the distance between speech segments in order to improve recognition accuracy.

Problems to be Solved by the Invention In view of the drawbacks of the above-mentioned conventional examples, the present invention provides an appropriate method of expressing a standard pattern that takes temporal dynamic characteristics into account, and a pattern 5.7 creation device for creating a standard pattern based on the method. It is in realization.

Means for Solving the Problem A pattern dividing means divides a pattern R consisting of a series of feature vectors)/V to be registered as a standard pattern, and a partial pattern of each partial section of the divided pattern H. Optimal parameter calculation means for calculating parameters of a time function that takes a vector value so that the distance (similarity) between the partial pattern and a vector series on a curve given by the function is minimized (maximum); A constituent element is standard pattern storage means for storing parameters of the time function determined for each of the partial intervals calculated by the calculation means.

The working pattern dividing means divides the pattern R consisting of a series of feature vectors to be registered as a standard pattern, and the optimal parameter calculating means calculates vector values for the subpatterns in each subinterval of the divided pattern H. The parameters of the time function are calculated so that the distance (similarity) between the partial pattern and the vector series on the curve given by the function is minimum (maximum), and the partial interval calculated by this optimal parameter calculation means The parameters of the time function determined for each time are used as a standard pattern.

Example The above time function is n-th order (n=1.2...
・) Polynomials, spline functions, etc. can be used.

Here, an embodiment of the present invention will be described in which a linear function is used for the sake of simplicity and for the reason that it is sufficiently practical. Furthermore, the sum of squares of the Euclidean distances between the feature vector and its corresponding vector on the curve will be used as a quantity representing the difference between the curve and the actual feature vector corresponding thereto. In this case, the curve is a so-called least squares approximation straight line, and the quantity corresponding to the distance is called a residual sum of squares.

FIG. 1 shows an embodiment of the present invention, and describes a case where the standard pattern is created for an audio signal.

7.-1 is a feature extraction unit which processes the input audio signal to be registered as a standard pattern through a filter bank, Fourier transform, and LP.
By well-known methods such as C analysis, several m5ec to m5
For each eC (referred to as a frame), a series of several-dimensional to multi-dimensional feature vectors (x(1)) = (X (1), X(2)
), --...j x(t), --.-, x(T) ). t is the frame number.

Reference numeral 2 denotes an input Banoffer memory, which temporarily stores a series of feature vectors obtained by the feature extractor 1.

3 is a Hunningsbek)/l/ display unit which displays the running spectrum of the audio input.

Reference numeral 4 denotes a partial section setting section, which sets a partial section for the input cover turn. This partial section can be set by a person looking at the running vector (/l/) and setting the section that is considered to be the most appropriate.

Reference numeral 5 denotes an average value calculation unit, which reads the feature vectors included in the section of the τ=t-r frame set as the partial section from the input buffer memory 2, and calculates the average value thereof. Here, t and r are the final frame of the currently focused section and the final frame of the previous section, respectively.

That is, if the subinterval is the first subinterval from 1τ+1 to t, and the average value is m (i), then the following is obtained.

6 is a least squares approximation straight line calculation unit (partial distance calculation unit), which calculates the residual sum of squares (partial distance) of the small square approximation line of the feature vector in the partial interval and the feature vector included in this partial interval. This is what we seek. That is, the least squares approximation straight line and partial residual sum of squares (partial distance) for the feature vector in the partial interval i are determined as follows.

Least squares approximation straight line Q to be found for the partial interval i
(k, 1) (k=1 to r) can be set as 97, -1 if u(i) is its direction vector. At this time, x(t-f+lc) and Q(k, i
) and the partial residual sum of squares v(t-τ+1:t
) is expressed as v(−τ+1 :t) −X (x(t−T+k)−Q(k, i))′=1 (x(t−r+k)−Q(k, i)) . Therefore, the least squares approximation straight line can be obtained by defining u(1) in equation (2) so that the partial residual sum of squares v (-τ+1:t) is minimized.

That is, it is obtained by solving the equation regarding u (i) assuming that the partial differentiation of v (-τ + 1: t) with respect to u (i) is equal to ○, and is obtained by solving the equation regarding u (i),
r+k)-Q(k,i)) =0・・・・・・・・・(
3) From this, it becomes . However, m(t), u(i), Q(k, t
).

x(t-τ+1), etc. are vertical vectors, and ゛ means transposition. Further, differentiation by a vector means to differentiate each element separately.

Reference numeral 7 denotes a standard pattern storage unit which stores the average vector of the least squares approximation straight line of the i-th partial interval obtained by the least squares approximation straight line calculation unit 6 in relation to the interval number i given by the partial interval setting unit 4. m(i), direction vector) ) v u (i) is stored for the interval i = 1 to ■.

In the manner described above, a speech pattern consisting of a series of feature vectors can be converted into a parameter sequence that provides a least squares approximation straight line in each subinterval.

11.

FIG. 2a explains the concept of the method for creating a standard pattern in the present invention, assuming a pattern represented by a one-dimensional feature vector series.

The vertical axis represents the feature amount of the feature vector, the horizontal axis represents the frame, and * represents the coordinate position of the feature vector at each time point. This example shows the case of 3 divisions, and line segments 100 to 102
is the least squares approximation straight line for each section. Similarly, the figure shows the case where the direction vector (L/) of the least squares approximation straight line is always ○, and each interval is one representative value, which is the average value of the feature vectors included therein. It will be expressed as a vector. Comparing a and b, a clearly expresses the temporal change of the feature vector) /L/ l~
ing. Furthermore, it is confirmed that small changes are smoothed out and global temporal dynamic characteristics are expressed.

As is clear from the above description, in the present invention,
Divide the input cover turn appropriately, find the least squares approximation straight line in each divided section, and calculate the vector) /L/ that represents the average value for each section, and the slope (direction) of the least squares approximation straight line passing through it. We will have a vector representing this as a standard pattern.

Effects of the Invention According to the present invention, the slope of the straight line of the partial section represents the dynamic feature of the partial section, and the average vector represents the static feature. In the present invention, by having these as standard patterns, the dynamic characteristics thereof are reflected, and the drawbacks of the above-mentioned conventional examples can be eliminated.

In addition, the present invention has a feature that the only parameters to be stored as a standard pattern are a vector representing the average value for each subinterval and a vector representing the slope (direction) of the least squares approximation straight line passing through it. This also eliminates the drawback of requiring a large storage capacity when the series of feature vectors output from the extraction unit itself is held as a standard pattern.

Furthermore, when the present invention targets unspecified speakers, points on the least squares approximation straight line are expressed as the average value of /L/ (specifically, a normal This method has features that are not found in the conventional example, such as being able to implement the method by providing a distribution type and dispersion such as a distribution.

In addition, in this embodiment, the case where the approximate curve is a straight line has been explained, but as stated at the beginning of the explanation of this embodiment,
Approximations can also be made with various curves using a similar method,
Needless to say, it is possible to express the dynamic characteristics of a recognition unit with great precision, and the patterns are not limited to voice patterns.

Further, as a measure of the distance between vectors, the sum of absolute values of the differences of each component, that is, the distance between urban areas, or various distances or similarities can be used.

Furthermore, in this embodiment, a case has been described in which the running spectrum is used to divide into subintervals, but it is of course possible to use other display methods such as soundsvector diagram. It is possible to display the temporal transition of the feature amount of the audio signal using the output of .

[Brief explanation of the drawing]

FIG. 1 is a block diagram showing one embodiment of the present invention, and FIG. 14 and FIG. 2 are conceptual diagrams explaining the present invention in detail. 1...Feature extraction unit, 2...Input buffer memory, 3...Running spectrum display unit,
4... Partial interval setting unit, 6... Average value calculation unit, 6... Least square approximation straight line calculation unit, 7...
・Standard pattern storage section. Name of agent: Patent attorney Toshio Nakao and 1 other person No. 1
Diagram Human Power Voice Running Spectrum
″″ Possessive extraction part? , 4 Human power'<7 y Partial section setting section memory Figure 2 ((IJ frame frame

Claims (1)

[Claims] A pattern dividing means that divides a pattern R consisting of a series of feature vectors to be registered as a standard pattern, and a parameter of a time function that takes a vector value for a partial pattern in each partial section of the divided pattern R. optimal parameter calculation means for calculating the distance (similarity) between the pattern and the vector series on the curve given by the function to be minimum (maximum); and for each of the partial intervals calculated by this optimal parameter calculation means. A pattern creation device comprising: standard pattern storage means for storing parameters of the determined time function.