JPH0634191B2 - Pattern feature normalization method - Google Patents

Description

Description: TECHNICAL FIELD The present invention relates to an improvement of a pattern normalization method in a pattern recognition apparatus, particularly for coping with fluctuations in a pattern to be recognized.

(Prior art) In the voice recognition by the pattern matching method of the standard pattern and the input pattern, the variation of the recognition target pattern due to the transmission line characteristic, individual difference, gender difference, etc. causes a reduction in the recognition rate. It is important to normalize the variation.

As a conventionally used normalization method, there is a method of obtaining a least-squares approximation line of the speech spectrum and subtracting it from the original speech spectrum. A Method of Word Speech Recognition by Matching ", October 1979.

(Problems to be Solved by the Invention) The above method is effective for normalization when the variation of the voice spectrum appears only in the variation of the slope, but the variation of the pattern due to the difference in the characteristics of the telephone and the microphone and the gender A normalization method that can deal with more complicated non-linear pattern fluctuations is required to normalize fluctuations due to gender differences.

An object of the present invention is to recursively calculate, from a variation pattern for learning, an operation having a normalizing action of a nonlinear variation component of a pattern that does not depend on a recognition target class and a type of variation that should be prepared as a standard pattern at the time of recognition. It is to provide a pattern feature normalization method that can be determined.

If the input pattern normalization method according to the present invention is used as a preprocessing unit of a recognition device, an input pattern that may have a plurality of fluctuation types can be accurately recognized by preparing a standard pattern for only one fluctuation type. It will be possible.

(Means for Solving Problems) In the pattern feature normalization method according to the present invention, in pattern recognition, an input pattern has a plurality of fluctuation types (for example, differences in characteristics of input devices, individual differences, gender differences, etc.). In the case of normalization of fluctuations in a case, a learning pattern for a small number of recognition target classes is used (however, patterns including all fluctuation types are included), and a standard fluctuation type and patterns of other fluctuation types are used as the standard. It is characterized in that a parametricized normalization operation group for normalization based on a variation type and an optimization means using the steepest descent method are recursively determined.

(Operation) The basic principle of the present invention is to recursively determine a normalization operation by using a typical fluctuation type pattern for a small number of classes in order to normalize a fluctuation component that does not depend on a pattern class of an input pattern. Is what you are trying to do.

A set of typical variation patterns of speech is {x ^(p) (a): p = 1 ,. ． ． , N _v , a = 1 ,. ． ． ，
N _c }. p represents a variation type (for example, in the case of gender difference, p = 1 (male) and p = 2 (female)). a represents a pattern class.

The normalization operation F from the fluctuation type p to the standard type pφ is y ^{(p, pφ)} (a) = F (x ^(p) (a):
c _m ^{(p, pφ)} ). Where {c _m ^{(p, pφ)} } (m =
1 ,. ． ． , M, p = 1 ,. ． ． N _v (≠ pφ)) is a parameter that characterizes the normalization function from the variation type p to the standard variation type pφ, and y ^{(p, pφ)} (a) is the pattern x of the variation type p with reference to the standard variation type pφ. ^(p) (a)
Is a normalized pattern. Below are standard fluctuation types p
A method for inductively determining φ and { _cm ^{(p, pφ)} } from the above variation pattern set will be shown.

Given the standard fluctuation types pφ and { _cm ^{(p, pφ)} }, the following quantity E ( _cm ^{(p, pφ)} ) is defined as an evaluation function that represents the error of the normalization.

Here, d [,] is a distance function between patterns. When pφ is given, { _cm
^{(P, pφ)} } can be obtained using the steepest descent method (see Iwanami Course Information Science 19 “Optimization” p.46). Therefore, if all the fluctuation types are provisional standard types pφ and E is minimized and the fluctuation type that minimizes the error amount is selected as the standard fluctuation type, the optimum pφ and { _cm
^{(P, pφ)} } is obtained.

When this method is applied to a pattern represented by time-series vectors having different lengths such as speech, the definition equation of the error amount E ( _cm ^{(p, pφ)} ) is as follows. You can change to.

Here, the subscript i represents the time axis of the time series pattern. x _i ^(pφ) (a) is a pattern in which the standard variation type pattern x _i ^(pφ) (a) is associated with the time axis by the DP matching method or the like.

A specific example of the case where the following functions are adopted as the normalization function F and the distance function d is shown below.

(The superscripts m and n of x and y represent the components of the pattern vector.) D [x, y] = Σ _m (x ^m −y ^m ) ² This normalization function is x ⁿ _i ^(p) Considering (a) to be the n-th component of the short-time spectrum of the i-th frame of speech, it is effective in normalizing the difference in the position of the spectrum peak (formant frequency) on the frequency axis due to gender differences. Is. Parameter c ^m _n by the steepest descent method in this case
_{^{(P, pφ) c m n}} (p, pφ) as minute constant ε updating the ^{_{← c m n (p, pφ}} ) + Δc m n
^{(P, pφ)} And the flow of processing is as follows.

1) Set the initial values of pφ, { ^cm _n ^{(p, pφ)} }.

2) the error amount _{^{E (c m n (p,}} pφ)) and the correction amount {.DELTA.c ^m _n
^{(P, pφ)} } is calculated.

_{^{3) {c m n (p}} , pφ)} Update.

_{^{c m n (p, pφ)}} ← c m n (p, pφ) + Δc m n
^{(P, pφ)} 4) ^{Repeat steps} (2) to (3) until convergence.

5) Update the center pφ and return to (1).

6) The center of fluctuation is pφ which minimizes the error amount E ( ^cm _n ^{(p, pφ)} ).

(Embodiment) FIG. 1 is a block diagram showing an embodiment of an apparatus realizing the present invention. The learning data storage unit 1 stores typical variation pattern data for a small number of classes. The parameter initialization unit 2 uses pφ, { ^cm _n ^{(p, pφ)} },
E ( ^cm _n ^{(p, pφ)} ) is initialized to the parameter
Store in buffer 3. 4 is a steepest descent calculator, the error amount E by the steepest descent method using the parameters and data of the learning data storage unit 1 of the normalization function read from the parameter buffer 3 _{^{(c m n (p, pφ}} )) performs minimization, the resulting error amount E (c ^m _n
^{(P, pφ)} ) and the parameter {c ^m _n ^{(p, pφ)} } are written back to the parameter buffer 3. This operation is repeated with all the fluctuations as the temporary center pφ, and finally the output unit 5 selects pφ and {c ^m _n ^{(p, pφ)} } that minimize the error amount E from the parameter buffer 3. , Output as the learning result of the normalization function.

Normalization function group of thus obtained {pφ, c ^m _n
FIG. 2 shows an example in which ^{(p, pφ)} } is applied to voice recognition. In FIG. 2, in the standard pattern storage unit 23,
Only the standard variation type pattern {x _i ^(pφ) (a)} is stored. The normalization parameter storage unit 22 stores the parameters of the normalization function determined by the above method. When the input pattern is input, the normalization calculation unit 21 normalizes the pattern using the parameters read from the normalization parameter storage unit. In this case, since the variation type of the input pattern is unknown, the normalization operation from all variation types to the standard variation type is applied in parallel to the input pattern and sent to the matting unit 24. The matching unit performs parallel matching between the standard pattern stored in the standard pattern storage unit and the normalized input pattern group sent from the normalization unit, and the best matching pattern as a recognition result. Output.

(Effects of the Invention) As described above, according to the present invention, a conversion function group that accurately normalizes a predicted variation of an input pattern is recursively and adaptively determined from a small number of recognition target class data. be able to. Since the resulting normalization function group does not depend on the recognition target class, it is effective for normalizing the variation of the pattern of the class not used for learning, and it becomes unnecessary to redo learning every time the recognition target class changes. That is, when adding a recognition target class, it is sufficient to collect only the standard variation type patterns and add them to the standard pattern, and the normalization operation unit can be used without changing.

[Brief description of drawings]

FIG. 1 is a block diagram showing an embodiment of the present invention, and FIG. 2 is a block diagram showing an application example of the present invention to a voice recognition device. In the figure, 1 is a learning data storage unit, 2 is a parameter initialization unit, 3 is a parameter buffer, 4 is a steepest descent calculation unit, 5 is an output unit, 21 is a normalization calculation unit, and 22 is a normalized parameter storage unit. , 23 is a standard pattern storage unit, and 24 is a matching unit.

Claims (1)

[Claims] 1. In pattern recognition, when a variation is normalized when an input pattern has a plurality of variation types, a learning pattern for a small number of classes to be recognized is used to determine a standard variation type and other variation types. A pattern feature normalization characterized in that a parameterized normalization operation group for normalizing a variation type pattern on the basis of the standard variation type is recursively determined by an optimizing means using a steepest descent method. Method.