JPS628800B2 - - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION The present invention provides a pattern preprocessing device for globally normalizing or correcting variations in a feature pattern expressed as a time series of feature vectors due to individual differences in a speech recognition device. Regarding.

The feature parameters that represent the magnitude of each component of the feature vector of the audio pattern are (i) the output value of the filter bank corresponding to each channel when the audio frequency band is divided into several channels, and (ii) the self-phase value. Relationship coefficient (iii) Partial autocorrelation coefficient {……PARTIAL AUTO
−CORRELATION COEFFICIENT (hereinafter referred to as
It is possible to use various quantities such as PARCOR coefficient)}.

However, no matter which amount is used, various variations occur in the feature parameters based on individual differences in speech information, resulting in large errors in the extracted feature parameters, which poses the problem of making it difficult to correctly recognize speech patterns. It was hot.

In the case of speech patterns, typical examples of variations based on individual differences include: (1) variations in the absolute level of a feature parameter based on variations in input level; (2) variations in formant frequency when the output value of a filter bank is used as a feature parameter. There are fluctuations, etc., and the following methods have been considered to prevent the recognition accuracy from decreasing due to these fluctuations.

For (1), normalization in the frequency direction (relative value conversion of filter output) is performed at each time when, for example, the output value of the filter bank is used as a feature parameter.

Regarding (2), since the variation in formant frequency is due to the difference in the vocal tract length of each individual, the vocal tract length is estimated from speech information and this vocal tract length is normalized. .

However, in the conventional normalization method for (1), for example, the structure regarding the absolute magnitude relationship in the time direction of feature parameters is destroyed, and as a result, some of the essential information of the audio information is lost. There were flaws.

Furthermore, the conventional method for (2) has the drawback that it is difficult to accurately estimate the vocal tract length, and errors associated with the estimation make correct normalization impossible.

Therefore, the first object of the present invention is to provide pattern preprocessing that absorbs variations in characteristics based on individual differences in feature parameters extracted from voice patterns without losing the essential information of the voice patterns. The purpose is to provide a method.

Furthermore, a second object of the present invention is to provide a pattern preprocessing method that maintains the structure of the characteristic parameters in the time direction while also absorbing fluctuations in formant frequency as fluctuations in frequency characteristics based on individual differences. It is about providing.

In order to achieve the above object, in the present invention,
One of the coordinate axes constituting a two-dimensional plane is the time axis, and the other is the frequency axis. Regarding the feature parameters at a series of points on the time axis with respect to one point on this frequency axis, (a) The above series of feature parameters Perform division of each feature parameter by the maximum value.

(b) Perform nonlinear correction using a logarithm. (c) Calculate the magnitude relationship of the values of feature parameters in the time axis direction, such as by dividing a series of feature parameters that have undergone nonlinear correction using a logarithm by their maximum value. Perform retained pretreatment.

The principle of the present invention will be explained below with reference to FIG.

FIG. 1 shows the difference in filter bank output values due to the same voice content uttered by different speakers in a particular channel when the voice filter bank output values are used as feature parameters.

According to FIG. 1, it can be seen that although the positions on the time axis where the maximum points and minimum points are given are similar, the amplitude values differ widely. If the degree of similarity between speech patterns A and B by different speakers is calculated using the usual Euclidean distance, the variation based on the consensus of the speakers will be larger than the variation based on the difference in the speech patterns themselves. This results in the inability to correctly detect differences in voice patterns.

Therefore, it is necessary to normalize or correct the amplitude value to reduce variations due to speaker differences. In the present invention, in order to normalize or correct the amplitude value, one of the following two steps or a combination thereof is used.

Hereinafter, for example, a preprocessing method that combines two steps will be described, but a preprocessing method that uses only step 1 or a preprocessing method that omits the step is also possible.

Time i (i=1,
2. Feature vector a ₁ in ……I) and a ₁
The feature pattern A expressed as a time series is defined as follows.

〓〓=(a _i1 , a _i2 , ......, a _iJ ) A=a ₁ , a ₂ , ......, a Here, a _ij (j=1, 2, ......J) is the time i is the feature parameter amount corresponding to the filter bank output value of the j-th channel.

Step 1: Logarithmic correction To correct amplitude that approximates human hearing characteristics,
Non-linear correction of a _ij is performed using a logarithm having a base of 10 or e.

The quantity a′ _ij corresponding to the corrected feature parameter is given by the following equation (1).

a′ _ij = log(1+a _ij /A ₀ )A ₀ : Constant (1) 1 in equation (1) prevents steep fluctuations in a′ _ij when a _ij /A ₀ becomes a value close to 0. It was added in order to

Step 2: Normalization in the time axis direction Global normalization in the time axis direction is performed for each channel in the frequency axis direction.

The above a in the j-th channel in the frequency axis direction
When the maximum value of ′ _1j , a′ _2j , ......, a′ _Ij is M _j , the normalized feature parameter a″ _ij corresponding to the feature parameter a _ij is given by equation (2). It will be done.

a″ _ij =a′ _ij /M _j (2) However, M _j =Max(a′ _ij , a _2j ………, a′ _Ij ) (3) The above two-step normalization or correction method is This makes it possible to obtain new feature parameters a'' _ij that maintain the magnitude relationship of the feature parameters in the time direction and clearly express the essential features of the original audio information.

This a″ _ij is the characteristic parameter a _ij in the following points.
It can be said that it is a better feature parameter than .

(α) This is a mapping that converts the maximum value of the feature parameter in the time axis direction to the same level (for example, the maximum level 1 when normalizing to a level of 0 to 1) for each channel in the frequency axis direction. This means that the fluctuations in the absolute levels of the characteristic parameters, which had been a problem, have been absorbed.

In particular, according to the normalization using only Step 2 above, that is, a'' _ij = a _ij /Max (a _1j , a _2j , ......, a _Ij ), the absolute level will fluctuate in the range where the amplitude value of a _ij is small. There may be an inconvenience in that the variation is magnified when viewed in terms of the relative level after normalization even though it is small, but the correction in step 1 is effective in eliminating this inconvenience.

(β) The normalization in step 2 is a mapping that extracts characteristic parameters of speech with the same content more broadly and clearly than before even when speakers differ.

For example, by emphasizing the change in the characteristic parameter for each channel on the frequency axis, or conversely, in a channel having a characteristic parameter with a small value over the entire time, the small value becomes a characteristic, and the value of _a''ij is The conversion is to make it larger.

(γ) In particular, in the case where the output value of the filter bank is used as the feature parameter, the normalization in step 2 above is based on variations in the frequency axis direction based on differences between speakers (for example, variations in formant frequency between speakers to adjacent channels). ) has the effect of absorbing

This means that at time i that gives the maximum value of the feature parameter in the j-th channel on the frequency axis, the adjacent (j-1)th channel or the (j-th
+1) Since there is a high probability that the characteristic parameter in the channel will be the maximum, the characteristic parameter obtained in step 2 corresponds to lowering the Q (resonance sharpness) in frequency analysis. this is,
This corresponds to reducing the variation in feature parameters based on differences between speakers, and provides a solution to (2) above.

Although the preprocessing according to the present invention has been mainly explained so far in the case where it is applied to the filter bank output value of the audio pattern, it is also possible to apply it to the autocorrelation coefficient PARCOR coefficient. When applied to PARCOR coefficients, for example, the PARCOR coefficients may be subjected to adaptive inverse filtering processing to convert them into quantities corresponding to reflection coefficients, thereby ensuring linearity with respect to absolute level fluctuations.

Hereinafter, the present invention will be explained in detail with reference to Examples.

FIG _. 2 is a block diagram _showing an embodiment of a circuit implementing the preprocessing method according to _the present invention. ......, a _Ij ) is 1
The block data is read out from the input buffer 21 and sent to the logarithmic conversion section 2 of the logarithmic conversion/normalization section 22.
21. In the logarithmic conversion unit 221, the calculation of equation (1) is performed, and the resulting block data a' _1j , a' _2j , ......, a' _Ij
is input to the maximum value detection section 223 and the normalization section 224 through the output line 222. The maximum value detection section 223 executes the calculation of the above equation (3), and the resulting M _j is also input to the normalization section 224 . The normalization unit 224 executes the calculation of equation (2) above, and the resulting block data a
″ _1j ,a″ _2j ,……,a″ _Ij is output buffer 2
3.

For example, if the above calculation process is j=1, 2,...
. . , J are executed in accordance with control signals output from the control unit 24 in the order.

The logarithmic conversion section 221 in FIG. 2 can be configured by, for example, a read-only memory (ROM).

In this case, the output signal of the input buffer 21 is used as the address signal, and the addresses a _1j , a _2j , . . .
…, a _Ij is the data a′ _1j , a which is the calculation result of equation (1)
' _2j , ......, a' _Ij are written and the control unit 2
4
This may be read out in accordance with a readout signal given from.

Further, the maximum value detection unit 223 is composed of an arithmetic circuit and a register that stores the arithmetic results, and for example, data a' is stored in the _order of a' _1j , a' _2j .
_ij is input to the subtraction circuit, and the data R stored in the register is
Then, only when a′ _ij −R>0, the contents of the register are updated and a′ _ij is newly stored in the register as i=1, 2, ..., I.
All you have to do is run it.

The normalization unit 224 can be configured with a normal divider.

Note that the above calculation in the logarithmic transformation/normalization section 22 can also be executed by software.

FIG. 3 is a block diagram showing an embodiment of a speech recognition system including the circuit of the present invention shown in FIG. 2, in which the parts shown in FIG. 2 are given the same numbers.

The input speech is frequency-analyzed in the feature extraction unit 31, and the extracted feature vectors a ₁ , a ₂ ,
......, a are stored in the input buffer 21 in chronological order.

The normalized data stored in the output buffer 23 by executing the process described in FIG.
4 is input. On the other hand, normalized standard patterns of speech are sequentially read out from the standard pattern memory 32, and one of them is inputted to the recognition section 34 through the standard pattern buffer 33.

The recognition unit 34 calculates the degree of similarity between the normalized data corresponding to the input speech pattern and the normalized standard pattern, performs recognition, and outputs the recognition result to the terminal 35.

FIG. 4 shows experimental data regarding the difference in the degree of separation of speech recognition results when the preprocessing method of the present invention is used and when the conventional preprocessing method is used.

In Fig. 4, the horizontal axis represents _the amount of weight given during recognition, and _the vertical axis represents the ratio (S ₁ /
When the degree of separation is given by S ₂ ), () ~
() indicates the case where the following pre-processing was performed, and () indicates the case where the conventional method was used.

(): a″ _ij = log(1+a _ij /A ₀ )/M _j However, M _j =Max{log(1+a _1j /A ₀ ), log(1+a _2j /A ₀ ,……, log(1+a _Ij /A ₀ )} (): a″ _ij = log(1+a _ij /A ₀ ) (): a″ _ij = a _ij /Max(a _1j , a _2j , ………, a _Ij ) Experiment in Figure 4 From the results, it can be seen that according to the preprocessing method of the present invention, a greater degree of separation can be obtained than other methods by appropriately setting the weight amount in similarity calculation.

As described above, the preprocessing method according to the present invention makes it possible to convert the features of the above feature pattern into a new feature pattern that clearly extracts the features, and the feature parameters obtained by the above preprocessing method are separable. It is effective because it has good performance (intra-class characteristics are clear) and has the effect of improving recognition reliability. The above effect is also proven by the result that the recognition rate was improved in a speech recognition device incorporating the above preprocessing normalization method.

[Brief explanation of the drawing]

FIG. 1 is a diagram showing differences between speakers of the same voice using filter bank output values, FIG. 2 is a diagram showing an embodiment of a circuit that implements the preprocessing method of the present invention, and FIG. 3 is a diagram showing the difference between speakers of the same voice. FIG. 4 is a diagram showing an example of the configuration of a speech recognition device using the processing method, and FIG. 4 is a diagram showing experimental data regarding the degree of separation of input speech patterns. 21: input buffer, 22: logarithmic conversion/normalization section, 23: output buffer, 24: control section.

Claims (1)

[Claims] 1 Feature parameters a _ij (i = 1, 2, ......, I, j = 1, 2,
. . .J), and input (a _1j , a _2j , . . . a _Ij ) for each channel j from the means as one block data, and input a predetermined value for each parameter value in the block. a correction means for performing logarithmic conversion correction of the parameter; a maximum value detection means for detecting a large value of the parameter within one block data corrected by the correction means;
1. A preprocessing device for speech recognition, comprising normalizing means for normalizing the corrected one block data according to the maximum value of the parameter for each block.