JP2845355B2 - Signal detection device - Google Patents

Abstract

PURPOSE:To provide a new spotting type pattern recognizing method by which high recognizing performance can be obtained. CONSTITUTION:A signal is inputted from a signal input part 1, the signal is converted into the time series of feature vector and a distance is calculated as a value showing the classification degree in each frame of an input continuous pattern series in a feature extraction part 2, and a recognition/learning processing part 6 decides whether the pattern of the class model based on the calculated class distance exists or not, decides that it exists if the class distance is smaller than a threshold 5 and it does not exist if not so. Further, the recognition/learning processing part 6 performs the learning update of a parameter.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a signal detection device,
For example, the present invention relates to a signal detection device that detects a pattern belonging to a specific class (class) from a pattern (signal sequence) that can be expressed numerically such as voice, text, and image.

[0002]

2. Description of the Related Art In pattern recognition based on spotting (hereinafter, referred to as spotting method), a pattern recognizer is composed of a pattern model of a class to be spotted, which is designed in advance, and a class membership threshold. Given an input continuous pattern sequence, the pattern recognizer assigns an attribute to each class at each position of the input continuous pattern sequence (the time point when the pattern is a time signal, and the spatial position when the pattern is a spatial signal). The degree of membership is measured using this class model, and the obtained degree of membership is compared with the above-mentioned threshold value. If the relationship between the measured membership level and the threshold value satisfies a preset condition, the pattern is positioned Is output, and when the relationship between the measured membership value and the threshold value does not satisfy the condition, the determination that the pattern does not exist at that position is output.

In the above spotting method, there are a recognition error (replacement) common to all pattern recognition tasks, and an error unique to the spotting method, ie, dropout and addition.
These three types of errors are hereinafter collectively referred to as spotting errors.

[0004]

In the conventional pattern recognition based on the spotting method described above, a class model realizing the maximum likelihood or the minimum distortion in each class, or a class model selected empirically, and an empirical model are used. , And their design or setting is performed only in the training stage preceding the recognition stage. These design techniques are inconsistent with the design goals for the original spotting pattern recognizer of minimizing spotting errors described above. For this reason, the conventional pattern recognition does not guarantee the optimality of the training result, that is, the minimum spotting error state, and cannot follow the change of the characteristics of the input continuous pattern sequence or the environment of the pattern recognition at the time of the input. There is a problem that performance (small spotting error) cannot be obtained.

[0005] Therefore, a main object of the present invention is to provide a signal detection device capable of obtaining high recognition performance.

[0006]

According to the present invention , a recognition unit is selected from a sequence of input signals such as voices, characters, and images.
Spotting whether a signal belongs to a class or not
Law determined using, a signal detection apparatus for extracting a desired signal sequence, any feature extraction means, in the feature sequence obtained by the feature extracting means for converting an input signal series FEATURES sequence May belong to the subsequence class of
The distance as a value representing the size, partial sequences and classes
And distance calculating means for calculating using the template, a detection means computed distance is less if partial sequence than class threshold is determined to belong to the class, it is determined not to belong to the class otherwise , part depending on the detection means
Even though the series belongs to the class,
Errors that have been determined to be
Judging that it does belong to Lass
The number of occurrences of errors with mistakes
And the first derivative defined for the threshold
And updating means for updating the template and the threshold value based on the gradient configured as described above.

[0007]

[Action] signal detection apparatus according to the present invention converts an input signal sequence to the feature sequence, the distance as a value representing the potential size belonging to any part series class in its features series Is calculated using the subsequence and the class template, and the distance is smaller than the class threshold.
Whether the subsequence belongs to the class depending on whether
Is detected. Whether the sub-series belongs to the class
Error and the sub-sequence
Belonging to a class that does not belong to the class
The number of occurrences of the error with the error
First derivative defined for template and threshold
And updates the template and the threshold based on the gradient formed by

[0008]

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS An embodiment of a specific adaptive learning rule for speech pattern recognition based on the spotting method will be described below. The optimal solution search method and the generalized stochastic descent method employed in the embodiment can be replaced with any other rational optimal solution search method without changing the definition of the spotting error criterion. In particular, when a batch search method such as the steepest gradient method is employed, the resulting learning rule is also of a batch type. Also, the embodiments described below are
Although related to one class, based on the fact that the spotting method is essentially performed independently for each class, the extension to the multi-class task is easily performed without changing the procedure described below.

FIG. 1 is a block diagram of one embodiment of the present invention. In FIG. 1, an audio signal is input from a signal input unit 1 and supplied to a feature extraction unit 2. The feature extraction unit 2 converts the input speech signal into a time series of acoustic feature vectors. This acoustic feature vector is obtained by, for example, sampling an audio signal, which is sampled at a sampling frequency of 12 kHz and converted into a 16-bit digital signal, at an interval of 5 msec.
It is a six-point discrete Fourier transform, is divided into 16 channels based on the mel scale, and is converted into a 16-dimensional vector composed of the logarithm of the sum of the output of each channel. For each frame of each acoustic feature vector, seven consecutive vectors obtained by combining three vectors before and after each frame are set as one of the input continuous pattern sequences.
Frame.

Next, the distance calculation unit 3 calculates, as a value representing the degree of class membership in each frame of the input continuous pattern sequence, the distance between the partial sequence of the input continuous pattern sequence ending with that frame and the class model 4 ( Hereinafter, the distance is referred to as a class distance). The class model 4 is composed of a set of one or more of a sequence of vectors having the same dimension as the acoustic feature vector forming the input continuous pattern sequence (hereinafter, referred to as a template). The total number of templates constituting the class model 4 is represented by B, the template indicated by the index b is represented by λ _b , the class model 4 as a set of templates is represented by Α ^* = ｛Α _b ｝ ^B _b = 1, and input Assuming that the continuous pattern sequence is x, the class distance g in the i-th frame of the input continuous pattern sequence
_i (x; Α ^* ) is defined by the following equation (1). Sand
From equation (1), the subsequence of the input pattern and the
Distance to the class model, that is, the class distance is defined.
It is.

[0011]

(Equation 1)

Here, K is a natural number not less than 1 and not more than B, b (k) is an index indicating the k-th template in a certain permutation for extracting K out of B templates, and ζ is a positive number. Is a number. D _i (x; λ _b ) is the distance between the partial sequence ending with the i-th frame of the input continuous pattern sequence and the template of the index b (hereinafter referred to as the cumulative path distance), and ω _k is the index b It is a weight coefficient for the cumulative route distance from the template (k). Voice pattern
Since the length of a symbol varies in nature,
In order to measure the distance between various patterns,
Fluctuations need to be normalized. Therefore, the following
Approximation of secondary distances as shown in equations (2) to (4)
Just in case, the normalization procedure is implemented in it. Definition
The cumulative route distance among the distances obtained is defined by the following equation (2).

[0013]

(Equation 2)

Here, _Hib is the total number of time-axis expansion / contraction maps allowed between the subsequence ending with the i-th frame of the input continuous pattern sequence and the index b template.
M in a certain permutation is a natural number of 1 or more H below, h (m) is taken out M pieces from among H _ib number of time-scaling mapping
This is an index indicating the third time axis expansion / contraction map, and ξη is a positive integer. Further, D _ih (x; λ _b ) is a local distance δ _ihj (x; λ _b ) of all the frames constituting the template of the index b with respect to the frame of the input continuous pattern sequence corresponding to the index θ by time-axis expansion / contraction mapping. ) (Hereinafter referred to as path distance), and ρ _m
Is a weight coefficient for the path distance of the index h (m) by the time-axis expansion / contraction map. The path distance is defined by the following equation (3).

[0015]

(Equation 3)

Here, J _b is the total number of frames constituting the template of the index b, and w _bj is a weighting coefficient of the local distance corresponding to the j-th frame of the template of the index b. The local distance is defined by the following equation (4).

[0017]

(Equation 4)

Here, S is the number of dimensions of the frame, and r
_bjs is the s-th element of the j-th frame of the template of the index b, and _{q (i, b, h, j)} is the input continuous pattern corresponding to the j-th frame of the template of the index b by the time-axis expansion / contraction map of the index h Index of the frame of the sequence, x
_{q (i, b, h, j) s} is the s-th element of the _{q-th (i, b, h, j)} frame of the input continuous pattern sequence.

Next, in the recognition / learning processing unit 6, a judgment process by the spotting method and a process of updating the class model 4 and the threshold value 5 are performed. FIG. 2 is a diagram showing the procedure of these processes.

First, in the determination processing unit SP1 in FIG. 2, it is determined whether or not a pattern of a class to be spotted exists at each position of the input continuous pattern sequence. If the above-mentioned class distance is smaller than the threshold value 5, it is determined that a pattern of the class to be spotted exists, otherwise, it does not exist. Next, the update processing unit SP2 updates the system parameters. For the portion R of the input continuous pattern sequence given whether or not the pattern of the class to be spotted exists by the teacher signal, whether or not the pattern of the class to be spotted exists in the portion R A function that reflects the determination of, that is, a spotting measure, is defined by the following equation (5).

[0021]

(Equation 5)

Here, Λ represents the entire parameters of the present system including the class model 4 (Λ ^* ) and the threshold value 5 (z), and I _R is the total number of frames included in the portion R of the input continuous pattern sequence. , Μ are positive numbers. The fact that the spotting speed in the portion R of the input continuous pattern sequence is negative reflects the determination that the pattern of the class to be spotted exists in the portion R, and the determination that the pattern is not negative does not exist.

The expression (5) represents "belongs" and "belongs".
The distance between the threshold 5 and the formula (1)
It is realized by comparison with the modified L _p norm form distance
It can be said. The second term on the right side of equation (5) is
Is a distance in the form of L _p norm. Such a complex format
The reason for taking is that the equation (1)
Is preferable from the viewpoint of the stability of the judgment.
Without taking the average of the distances in the subsection R
To improve the stability. Where
In equation (5), the comparison result is expressed by the difference between two terms.
You. If the difference is negative and the spotting measure d is negative
If the average distance of the two terms on the right side is larger than the threshold
Become. In other words, if d is negative, it means that the class
Will determine that the turn is not in R,
If d is not negative (0 or positive), the two terms on the right
The average distance will be smaller than the threshold. Accordingly
It is natural to determine that the subsequence is close to the class model
In this case, the class pattern "exists"
judge. In the signal detection device according to the present invention,
If the pattern of that class exists in the series
Training input to which "whether does not exist" information is added in advance
By using the pattern, it is possible to make a determination using the equation (5).
A trainable threshold so that
Alas aims to optimize both with the model.

[0024]

(Equation 6) In the above equation (6), ifV (R) = T
Becomes a spotting target in the part R of the continuation pattern sequence
If the pattern of the class actually exists, ifV
(R) = F is the spotting target in the above part R
Class pattern does not actually exist
I have. The upper part of equation (6) shows the approximate value of the number of missing errors.
I do. In the case of the upper row, the sub-series is originally
Since it shows the situation where the lath pattern is included,
An omission error occurs only when it is determined that "exists".
In fact, the equation of the loss function in the upper part of this equation (6) is
The spotting measure in equation (5) is "generally non-negative"
Value (determining that it exists)
Indicates a value close to 0, in other words, evaluates to "no error"
And the spotting measure of equation (5) takes a negative value
A value close to 1 if (determined as "not present")
Is shown. In other words, "Even though it originally exists,
To prevent mistakes in dropping
It is made like. The lower part of equation (6) is the reverse of the upper part.
It is a source. In other words, the class pattern originally exists
The number of additional errors determined to be "exist"
This is the case. Here, the lower part of equation (6)
The number is given by the-sign before α.
A number close to 1 when the measure in equation (5) takes a large positive value
And when the measure of equation (5) takes a negative value, it approaches zero.
It is made to show a large number.

Here, the function V (R) is a teacher signal indicating whether or not the pattern of the class to be spotted actually exists in the portion R of the input continuous pattern sequence.
It takes the value T when it exists, and takes the value F when it does not exist. α is a positive constant, β is a constant, γ _t is a weight for a drop error, and γ _f is a weight for an additional error. When the generalized stochastic descent method is used as the optimal solution search method, the system parameters are updated according to the following equation (7).
Equation (7) is a pattern recognition called the generalized stochastic descent method.
Parameters that can be trained according to the theorem proposed in the field of knowledge.
This is a meter update formula. Includes class patterns in advance
Training pattern to which information indicating whether or not
If a large number of symbols are given, the following equation (7) is used.
Has been repeated, and the nth update has just been performed.
Assume that The update here is basically a reference pattern
And threshold based on the gradient (first derivative)
Only a small amount is changed. Actual changes right
The second term of the right side, and the parameter at the end of the n-th update of the first term of the right side
By subtracting the second term on the right from the parameter state Α (n),
The (n + 1) -th update is performed and reaches Α (n + 1) on the left side.
You. ▽ y _R (x (n); Α (n)) in the second term on the right side is a loss
Let the gradient of the function, i.e. the trainable parameter be a variable
Vector of the first derivative of the loss function. In this formula
Is the loss of the judgment result registered at the time of the (n + 1) th update
Only used. In other words, one subsequence
Only the evaluation of the result of the spotting judgment is used
It is. Explain exactly the meaning of subtraction on the right side of equation (7)
And a loss represented by y _R (x (n); Α (n))
Basically approximates the number of errors, so the smaller is
desirable. This means that each y _R (x (n); Α (n))
Also wants to be closer to 0 than 1 as much as possible.
Therefore, we look at the form of the function y _R (x (n); Α (n))
And this monotonically connects 1 and 0, just 1
Like connecting a plateau to a lowland with a smooth slope
The slope, that is, the magnitude of the slope (▽ y
_R (x (n); Α (n)) corresponds to this
Is the direction of descending (-▽ y _R (x (n); Α (n))
Change the parameter to
You. This has been proven by stochastic descent,
Iteratively updates this loss function in the direction of going downhill
And eventually a loss function defined over the available samples
The fixed point of the averaging function
Will arrive.

[0026]

(Equation 7)

Here, x (n) and Λ (n) are n
The input continuous pattern sequence and the system parameters at the time of updating by the input continuous pattern sequence, ε _n is a learning coefficient taking a small value satisfying the condition required by the generalized stochastic descent method, and U is a positive definite value Is a matrix, and ▽ is an operator for calculating a gradient vector of the function. This update rule has been proved by the generalized stochastic descent method to realize the minimum solution search of the spotting error criterion which is the expected value of the above-mentioned loss function.

As the optimization learning method, a method other than the generalized stochastic descent method can be used. Depending on the selection, the spotting error criterion does not need to be defined by a continuous function as described above. The definition of the loss function in the example can be simplified to the expression by a discontinuous function, and the learning rule derived by the simplification is described in the second rule below.
It becomes a practical learning rule as shown in the embodiment.

The result output unit 7 shown in FIG. 1 outputs, based on the recognition result of the recognition / learning processing unit 6, whether or not there is a pattern of the class to be spotted at each position of the input continuous pattern sequence. I do.

The switches SW1 and SW2 shown in FIG.
Are closed, it is possible to update both the class model 4 and the threshold 5, and it is also possible to close only one and update only the class model 4 or only the threshold 5.

Next, a second embodiment will be described. In the second embodiment, an approximation using a discontinuous function at the infinite limit of μ, ξ, ξ of the loss function expressed as a continuous function in the definition of the update rule of the first embodiment as the loss function is performed. Use simplified update rules. The updating of the class model 4 and the threshold 5 in this example is performed based on the following equations (8) and (9).

[0032]

(Equation 8)

[0033]

(Equation 9)

However,

[0035]

(Equation 10)

[0036]

[Equation 11]

[0037]

(Equation 12)

[0038] Here, the index b of the template are assigned in ascending order of cumulative path distance in the i ^* frame consecutive input pattern sequence, the index m in the time-scaling mapping in the i ^* frame consecutive input pattern sequence R _bjs (n) and z (n) are the s-th elements of the j-th frame of the template of the index b at the time of updating with the n-th input continuous pattern sequence, respectively. And the threshold z, and ν is the derivative of the loss function l _R (d _R (x; Λ)).

In the second embodiment, since a high-speed calculation method such as the DP method or the A ^* search method can be used in the recognition / learning processing unit 6 in FIG. 1, the efficiency is far higher than in the first embodiment. Well practical.

Next, the effect of the present invention was confirmed by a spotting experiment of phonemes / k / from the text. The voice pattern sample used for training was composed of 50 sentences in which the number of phonemes by one male speaker was almost equal, and the phoneme / k determined by inspection
V (R) = T in the portion consisting of the frame at the end of / and three frames before and after it, and V (R) = T in other portions
F, K = M = 1, ω _k = ρ _m = 1, ω _bj = 1, α
= 0.05, β = 10.0, γ _t = 1.0, γ _f = 0.
As No. 5, the training according to the second embodiment of the present invention was performed. According to the conventional class model and threshold value designing method, the dropout error 33 and the addition error 170 were found for the total number 134 of phonemes / k /, but when the training according to the second embodiment of the present invention was performed, the dropout error was found. Error 18 and additional error 40 could be reduced.

[0041]

As described above, according to the present invention, the sports
Signal detection device using the spotting method.
Templates and thresholds to reduce the number of
By updating, the recognition performance can be improved.

[Brief description of the drawings]

FIG. 1 is a schematic block diagram showing an example of a speech recognition system to which an embodiment of the present invention has been applied.

FIG. 2 is a diagram showing a processing procedure of a recognition / learning processing unit 6 shown in FIG.

[Explanation of symbols]

 DESCRIPTION OF SYMBOLS 1 Signal input part 2 Feature extraction part 3 Distance calculation part 4 Class model 5 Threshold value 6 Recognition learning processing part 7 Result output part

 ────────────────────────────────────────────────── ─── Continuing on the front page (72) Inventor Shigeru Katagiri Kyoto, Soraku-gun, Seika-cho, 5th, Inaniya, 5th, Sanraya, ATR Co., Ltd. (JP, A) JP-A-4-124784 (JP, A) JP-A-4-148384 (JP, A) "Pattern recognition" by Kyoya Iijima, first edition issued on November 10, 1973, March 1979 15th Third edition issued, Corona Co., Ltd. p. 225-255

Claims (1)

(57) [Claims] Is 1. A voice, text, signal sequence belonging to the class is the recognition unit from among a series of input signals such as an image
Is determined using the spotting method, and the desired signal is determined.
A signal detection apparatus for extracting a sequence, feature extraction means for converting the input signal sequence FEATURES sequence, it can belong to any part series of classes in the feature sequence obtained by said feature extracting means Distance as a value representing a certain size of the subsequence and the class
Distance calculating means for calculating with the rate, the distance the partial sequence is smaller than the distance calculated by the calculation means the class <br/> threshold is determined to belong to the class <br/> , detecting means for determining not to belong to the class otherwise, and the class in which the partial sequence corresponding by said detecting means
I judge that I belong but do not belong
An error and the subsequence does not belong to the corresponding class
Despite the mistakes that he judged to belong,
The number of occurrences of errors in the template and the
Constituted by the first derivative defined with respect to the threshold
A signal detection device comprising: an updating unit configured to update the template and the threshold based on a gradient to be obtained.