JPH0199096A - Pattern generator - Google Patents

Abstract

PURPOSE: To improve the accuracy of recognition by reflecting dynamic features at the time of preparing a reference pattern for voice recognition or the like. CONSTITUTION: A feature extraction part 1 converts an input voice signal to be registered as a reference pattern into a feature vector sequence and stores the converted sequence in an input buffer memory 2. A mean value calculation part 5 successively reads out each partial section set up by a partial section setting part 3 from a memory 2, finds out a mean value and outputs the mean value to a least square approximate straight line calculating part 6. The calculating part 6 finds out the least square approximate straight line of feature vectors and a residual square sum (partial distance) based upon outputs from the calculation part 5 and the memory 2 and a partial residual square sum normalizing part 15 normalizes the residual square sum. A minimum accumulated residual square sum calculating part 7 divides a normalized residual square sum so as to minimize it based upon an output from the normalizing part 15, finds out the sum and stores the average vector and direction vector of a square approximate straight line in an average/direction vector storing part 13. Since the inclination of the straight line in the partial section indicates a dynamic feature and the average vector indicates a static feature, 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 essential to the present invention, the word distance will be used here to represent them. do. That is, for example, the distance is short,
A small distance corresponds to a high degree of similarity, and a large distance corresponds to a low degree of similarity. etc.

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. This is done by registering a pattern consisting of a series of feature vectors representing the word speech 3 to be recognized as a standard pattern in advance for each of the word speeches, and at the time of recognition, recognizing the same sequence of feature vectors. The input cover pattern to be used is compared with each of the standard patterns, the closest standard pattern 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 non-linearly expand and contract the time axes of patterns with different time lengths, but in order to do this efficiently, a method called DP matching uses dynamic programming, which is currently the best method. This is one of the ways to get results.

However, in this method, the time axis is expanded or contracted so that both patterns to be compared are closest in terms of distance, and features related to temporal changes in feature vectors such as slopes with respect to the time axis are used. (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.

In addition, we have a word dictionary in the form of symbol sequences representing phonemes and syllables (hereinafter referred to as speech segments), prepare standard patterns corresponding to each speech segment in advance, and use The kabataan is converted into a sequence of phonetic units, that is, a sequence of symbols representing each phonetic unit, based on the standard pattern, and matching is performed at the symbol level with each word in the word dictionary to find the closest word. In this case, since it is impossible to completely recognize the speech segments, the speech segment sequence converted from the input cover pattern is inserted or omitted.

Contains some errors such as substitutions. Therefore, in the symbol level matching, distances between speech segment sequences are determined by DP matching based on distances between speech segments 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.

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

Means for solving the problem Characteristics: A subsection setting means for setting a candidate section for the i-th subsection of a pattern consisting of a series of vector l-
)I/series and the pect)v sequence on a curve given by a time function that takes pect/L/value. Normalized partial distance (normalized partial similarity) normalized in relation to the length of the interval )
The normalized partial distance (normalized partial similarity) is calculated by optimal parameter calculation means for calculating the parameters of the time function such that a minimum cumulative distance (maximum cumulative similarity) calculating means for calculating the minimum (maximum) value of the sum of the degrees), and a parameter of the time function calculated corresponding to each optimally determined subinterval i The standard pattern storage means stores information related to the section number i.

The operation partial interval setting means sets a candidate interval for the i-th partial interval of the pattern consisting of the series of feature vectors I−/2, and the optimum parameter calculating means sets the i-th candidate interval of the set pattern. Normalized partial distance (normalized partial similarity) between a feature vector sequence of a candidate interval and a vector sequence on a curve given by a time function that takes vector values, normalized in relation to the length of the interval. The parameters of the time function are calculated so that the minimum (maximum)
optimally determine the number of sections and division positions of the partial sections in order to obtain the minimum (maximum) value of the total of normalized partial similarity),
The parameters of the time function calculated corresponding to each optimally determined subinterval i are stored in the standard pattern storage means in association with the interval number i, and this parameter string is used as a new standard for the pattern. It is a pattern.

7 to 7 Embodiment As the time function, an n-th order (n=1.2...) polynomial, a spline function, etc. may be used. Here, an embodiment of the present invention will be described using a linear function for the sake of simplicity and for the reason that it is sufficiently practical. Still, 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 becomes a so-called least squares approximation straight line, and the quantity corresponding to the distance becomes what is called the 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.

1 is a feature extraction unit which processes input audio signals to be registered as standard patterns through a filter bank, Fourier transform, LP
Several m5ec to several m5e are determined by well-known methods such as C analysis.
For each c (referred to as a frame), a series of feature vectors from multi-dimensional light to multi-dimensional light 1x (tl = 1x(1), X(2)
, =-, dl. x (t) is a feature vector at time t.

Reference numeral 2 denotes an input buffer memory, which temporarily stores a series of feature vectors obtained by the feature surface extraction section 1.

4 is a frame counter which is reset when there is no audio input and counts up every frame. That is,
Its contents will indicate the frame currently being processed.

Reference numeral 3 denotes a partial section setting section, which sets a partial section for the input cover turn. Now, when the content of the frame counter 4 is t, the partial section setting unit 3 sets r
= t -s to t -e are sequentially set as starting end candidate frames of the i-th partial section. Here, s and e are constants given in advance to limit the range allowed as a subinterval.

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, extracts them, and calculates the average value. . That is, if the subinterval t-τ+1 to t is the i-th subinterval and its average value is m(t, 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 least squares approximation straight line of the feature vector in the partial interval and the feature vector included in this partial interval. That's what I'm looking for. 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, i) (k=1-r) can be set if u(t, i) is its direction vector. At this time, x(t-r+k) and Q(k,i)
The partial residual sum of squares v(t-τ+1:t) between −1 and τ
is expressed as 10 ＼−/v(−τ+1:t). Therefore, the least squares approximation straight line can be obtained by determining u(t, i) in equation (2) so that the partial distance (partial residual sum of squares) v (-τ+1:t) is minimized.

That is, it is obtained by solving the equation regarding u(t, i) assuming that the partial differentiation of v(-τ+1:t) with respect to u(t, i) is equal to 0, and
1x (t-r+k) -Q (k, i) l=o
・From =-(a), it becomes 11 A-. However, m(t, i), u(t, i), Q(k
, i).

x(+-τ1,000), etc. are vertical vectors, and ' means transposition. Further, differentiation by a vector means to differentiate each element separately.

As described above, the average value calculation section 5 and the least squares approximation straight line calculation section 6 form the optimal parameter calculation means, and calculate the optimal parameter for the section set by the partial section setting section 3, that is, the section The mean value and direction vector are calculated as parameters of the least squares approximation straight line that best approximates the feature vector series of , and the partial residual sum of squares for that straight line is calculated.

Reference numeral 15 denotes a partial residual sum of squares normalization unit, which normalizes the partial residual sum of squares by a function of the number of frames in the partial interval. The purpose of normalization here is that, if normalization is not performed, the effect of partial sections with a large number of frames will have a greater influence on the division result than the effect of subintervals with a small number of frames. As the normalization coefficient, for example, τ, τ, etc. can be used, where τ is the number of frames in the partial section. By normalizing in this way, it becomes possible to compare different division numbers of the same section to see which one is more optimal. If, as in this example, the sum of the squares of the distances between vectors is used as the partial distance, in order to remove the influence of the number of frames τ in the partial interval, it is recommended to normalize it by an amount proportional to τ. When using the absolute value distance (urban area distance) as the distance between vectors, it is considered desirable to normalize by an amount proportional to τ in order to remove the influence of the number of frames τ in the partial section.

7 is a cumulative residual sum of squares calculation unit (cumulative distance calculation unit).

First, if the number of frames in the i-th subinterval is τ1 and the normalization coefficient is τ, then the 1st to T frames are the sum of the normalized residual sums of squares (normalized partial distances) in the 1st to 1st subintervals v(
1:f(1))/r+v(f(1)+1:f(2)
))/τ2+-・-+v(f (I-1)+1: f
(I ))/r1 is minimized by 13%.Hereafter, we will refer to the optimal I division), and the sum (hereinafter referred to as the minimum cumulative distance) V(T , I).

Here, f(i) (i=1 to 1) is the final frame of the divided i-th partial section. This calculation can be performed efficiently by using dynamic programming. That is, τ
Since 1 = tr, recurrence formula% formula%) (t≠0), and if we sequentially calculate t = 1 to T and i = 1 to I, we can obtain V(T, I). be.

What the recurrence formula (5) means is that 1 to t frames are
The minimum cumulative distance V(t:i) when divided is 1 to
r (t-s≦r≦t-e) The minimum cumulative distance V (r, 1-1) when the frame is divided 1-1 and the normalized partial path of the i-th section @v (r:t)/ r = v(r:t)/(t
-r)2 and is determined as the minimum value of r. This means that T that satisfies equation (6) is r. p,
Then, when 1 to t frames are optimally divided into i, 1
~'opt The dividing point of each section in the frame is 1~
This is based on the so-called principle of optimality, which states that the partial process of the optimal process that corresponds to the division point of each interval when the ropt frame is optimally divided into i-1 parts is also an optimal process.

The present invention is characterized in that the calculation of equation (6) also optimizes the number of divisions. In this case, the recurrence formula (5
) to Noyo, the initial value V(o)-o, V(t)=
V(T), which is finally obtained by solving this recurrence formula as (t-4!O), is the minimum cumulative residual sum of squares that should be found, and the number of divisions is also optimized. ing.

In the end, the minimum cumulative residual sum of squares calculation unit 7 calculates
= 1 to T and find the value corresponding to V (t). Here, B(t) is (X(1),...
..., x(t)) is optimally divided in terms of the number of divisions and division points, and the final frame of the penultimate sub-interval of 15'..., and the initial value of B(t) is B) −〇.

Also, N(t) is (X(1), =-9x(t)
) is the number of divisions when optimally dividing N (
The initial value of t) is N(o)=.

It is.

8 is a minimum cumulative residual sum of squares storage unit which stores the result of the minimum cumulative residual sum of squares calculation unit γ, that is, the minimum cumulative residual sum of squares (t) when frames 1 to t are optimally divided. Remember. V[) is used in the subsequent calculation of the recurrence formula in the minimum cumulative residual sum of squares calculation unit 7.

Reference numeral 10 denotes a bank pointer storage unit, which stores the B (+) (hereinafter referred to as the back pointer of the t-th frame) calculated by the minimum cumulative residual sum of squares calculation unit 7 at t.
0B(t) stored for =1 to d is used to find the final result of the optimal division point by backtracking after performing the above processing for t-1 to T as described below.

Reference numeral 9 denotes a division section number storage section, which stores the N (t) calculated by the minimum cumulative residual sum of squares calculation section 7 for t=1 to T. After performing the above processing for t = 1 to T, N (t) is used to find the section number of each subsection when finding the final result of the backtrunk and optimal division points as described below. used.

Reference numeral 13 denotes an average/direction vector storage unit which stores the average vector m(t) of the least squares approximation straight line of the final section corresponding to v (g) calculated by the minimum cumulative residual sum of squares calculation unit 7 and the direction Bekulele u ( t) for t=1 to T. Here, m(t) and u(t) are m(t, i) and u(t,
i).

Reference numeral 14 denotes a voice section detecting section, which detects a voice section by detecting the level of the long voice from the output of the feature extracting section 1 using a well-known method. The frame counter 4 is controlled by the output of the voice section detecting section 14, and sequentially counts the section in which voice exists starting from one frame. 11 is a bank pointer read control section, which inputs the frame number to the back pointer storage section 10. Immediately after the end of the voice section, the value T of the frame counter 4 at 17 \-7 is output to the back pointer storage section 10, and from then on, the back is continued until the value of the back pointer becomes 0. The output of the pointer storage section 10 is fed back to the back pointer storage section 1o. The back pointer storage unit 1° outputs the final frame B (t) of the previous partial section for a given frame number t. Therefore, first, when the end of a voice section is detected by the voice section detection section 14, the bank pointer read control section 11 outputs the value T of the frame counter 4 at that time to the back pointer storage section 10, and thereafter uses the back pointer. Bank pointer B (t) read from storage unit 1o is ○
The frame number is fed back to the back pointer storage unit 1o as a new frame number until the frame number is reached. As a result, the back pointer readout control unit 11 outputs the final frame of each divided section in the reverse order from T when the T frame is optimally divided. In addition, the already calculated average vector in the last divided section when frames 1 to t are optimally divided) /L/ m (t) and the direction vector) / 18 ... - u (t), the number of divisions In order to read out N (t), the frame number t, which is the output of the bank pointer read control section 11 obtained in this way, is supplied to the average/direction pect/l/ storage section 13 and the division section number storage section 9. Ru.

Reference numeral 12 denotes a standard pattern storage section, which stores the output t of the back pointer readout control section 11, the accompanying output of the average/direction vector storage section 13, and the divided section number storage section 9, which are given as described above. It stores the output.

As described above, a least squares approximation straight line in each partial section is obtained when the audio pattern consisting of a series of feature vectors is optimally divided in terms of the number of divisions and the division position in the above-mentioned sense.

FIG. 2(a) illustrates the concept of the division method in the present invention assuming a pattern expressed as a one-dimensional feature vector (yV 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. In this example, the case of 3 divisions is shown.
is the least squares approximating straight line 19''-/ in each section. Similarly, Figure (b) shows the case where the direction vector )/ of the least squares approximating straight line is always Q, and this corresponds to the case where each interval is divided to minimize the sum of squares of errors of the vectors included in that interval with respect to the average vector of that interval, and each interval is one representative vector, which is the average value of the feature vectors included in it.) It will be expressed as /L/.

As is clear from the above description, in the present invention,
Divide the input cover turn by the optimal number of divisions and optimal division points, find the least squares approximation straight line in each division section, and find the vector representing the average value and the minimum 2 lines passing through it.
A vector representing the slope (direction) of the power approximation straight line is included 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. By having these as standard patterns, the present invention is able to reflect the dynamic characteristics of the patterns and eliminate the drawbacks of the prior art example described above.

However, in the present invention, 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 is aimed at unspecified speakers, points on the least squares approximation straight line are set as the average value of the feature vector at the corresponding point in time to form a distribution (specifically, a distribution such as a normal distribution). It has features not found in the conventional example, such as being able to realize this by providing different types and distributions.

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,
It goes without saying that not only is it possible to express the dynamic characteristics of the recognition unit very precisely, but the patterns are not limited to speech patterns.

Furthermore, as a measure of the difference between vectors, the sum of absolute values of the differences of each component, that is, various distances or similarities in addition to the urban area distance, can be used.In addition, in the embodiment, the voice interval detection is automatically performed. However, by looking at the speech analysis output, we manually extract sections corresponding to the speech segments such as syllables and phonemes, and then divide them into smaller sections using the method described in the present invention. Of course, it is also possible to express the pattern of each section by the parameters of the approximate curve corresponding to those sections, and to make the standard pattern of the speech segment by concatenating them. The output is applied to a method for recognizing speech segments.

[Brief explanation of the drawing]

FIG. 1 is a block diagram showing an embodiment of the present invention, and FIG. 2 is a conceptual diagram explaining the present invention in detail.
Feature extraction unit, 2...Input cover sofa memory, 22
・ 3... Partial section setting unit, 4... Frame counter, 5... Average value calculation unit, 6...
... Least squares approximation straight line calculation section, 7 ... Minimum cumulative residual sum of squares calculation section, 8 ... Minimum cumulative residual sum of squares storage section, 9 ... Division Section number storage unit, 1o...
... Back pointer storage section, 11 ... Bank pointer read control section, 12 ... Standard pattern storage section, 13 ... Average/direction vector)/L/ storage section , 14... Voice section detection unit, 15...
Partial residual sum of squares storage.

Claims (1)

[Claims] a subinterval setting means for setting a candidate interval for the i-th subinterval of a pattern consisting of a series of feature vectors, a feature vector series of the set candidate interval for the i-th interval of the pattern, and a time function that takes a vector value; The normalized partial distance (normalized partial similarity) normalized in relation to the length of the interval with the vector sequence on the given curve is the minimum (maximum)
The minimum total of the normalized partial distances (normalized partial similarities) is determined by optimal parameter calculation means for calculating the parameters of the time function such that a minimum cumulative distance (maximum cumulative similarity) calculation means for calculating a (maximum) value, and a parameter of the time function calculated corresponding to each optimally determined subinterval i in relation to the interval number i. 1. A pattern creation device comprising: standard pattern storage means for storing standard patterns.