JPH0634192B2 - Voice recognizer - Google Patents

Description

Description: TECHNICAL FIELD The present invention relates to a voice recognition device, and more particularly to an improvement of a voice data matching method for improving recognition performance.

(Background Art) A conventional voice recognition device is configured as shown in FIG.
1 is an input terminal, 2 is a frequency analysis unit, 3 is a spectrum conversion unit, 4 is a voice section determination unit, 5 is a dissimilarity calculation unit, 6 is a standard voice spectrum pattern memory, 7 is a determination unit, and 8 is a recognition result output. It is a terminal.

In a conventional speech recognition device, a spectrum-converted input speech spectrum pattern and a standard spectrum pattern k (k =
1 to K), the dissimilarity Dk is the time sampling point of the input spectrum pattern, and the element of the n-th m-th channel is A (m, n), and the time sampling point of the standard spectrum pattern k. SK the element of the nth m-th channel
When (m, n), The recognition result is the category of the standard spectrum pattern that minimizes Dk among the K standard spectrum patterns calculated by the equation (1). Here, the weight Wk (m, n)
Although there are various methods for calculating, the description thereof is omitted because it is not the object of the present invention.

In the conventional recognition device, the power information of the input voice is completely lost due to the spectrum conversion. As a result, for example, "ichi" may be erroneously recognized as "d", or "go" may be erroneously recognized as "roku".

FIG. 2 shows an example of a sonargram of voice patterns of “ichi”, “d”, “go”, and “roku”. In FIG. 2, the horizontal direction is the frequency axis and the vertical direction is the time axis. In this way, by spectrum conversion, "ichi" and "ni", "go" and "roku"
Results in a very similar pattern, and the difference is that the silent section between “A” and “H” is large, but the power information is lost, so the result is It was sometimes mistakenly recognized, causing a decline in recognition rate.

(Problem of the invention) An object of the present invention is to solve these drawbacks and to improve the recognition rate by using the power information which has been completely lost by the conventional spectrum conversion, particularly the information of the silent section. It is to provide a recognition device, and the main point is whether there is a valley of power in the word (hereinafter referred to as power dip), such as "ni" for "ichi" and "go" for "roku" By using information for recognition judgment in combination with conventional spectral distance information, confusion is lost, and it is used especially for phone speech recognition and other phoneme information that has deteriorated to improve the recognition rate. Is.

Figure 3 shows the situation of power dip by word.

(Structure of the Invention) According to the present invention, means for frequency-analyzing an input voice signal, means for creating a power pattern of the analyzed input voice, and normalization with the spectral slope of the analyzed input voice A means for creating a spectrum pattern, a means for matching a spectrum pattern prepared in advance for a standard voice and the spectrum pattern for an input voice to calculate a spectrum dissimilarity, and a power in the voice section from the power pattern. Detecting a power dip section in which the power dip is minimum, and calculating the power dip dissimilarity for each recognition target category from the presence or absence of the power dip section and the size of the power dip, and the spectrum dissimilarity for each category. Add the power dip distance indicating the power dip dissimilarity to the spectrum matching distance indicating the degree There is provided a voice recognition device comprising means for calculating a total matching distance and means for recognizing and determining a category for which the calculated total matching distance is the smallest.

(Operation) The power dip section in which the power becomes minimum in the voice section is detected. By this detection, the characteristics of the presence or absence of the power dip section and the size of the power dip can be grasped for each recognition target category. In this way, since the dissimilarity can be calculated by focusing only on the sections having different powers between the recognition target categories, it is possible to significantly improve the recognition performance by calculating the accurate power dip dissimilarity for each recognition target category. You can

(Embodiment) FIG. 4 is a block diagram showing one embodiment of the present invention. In FIG. 4, 100 is an input terminal and 200 is a frequency analysis unit. 300 is a spectrum conversion unit, 400 is a voice section determination unit, and 500 is a re-sampling unit. Reference numeral 600 denotes a power dip calculation unit, which is a power information memory unit 601, shift register 1 602, shift register 2 603, shift register 3 604, shift register 4 605, addition subtractor 606, power differentiator 607, comparison unit 1 608. , Comparison unit 2 609, comparison unit 3 610, DS determination unit 611, DE determination unit 61
2. Power dip section determination unit 613, PA calculation unit 61
4, QA calculator 615, power dip evaluation value calculation unit 6
16, power dip distance calculation unit 617, comparison unit 46
18, latch 619, division calculator 620, subtraction calculator 1
621, subtraction calculator 2 622, subtraction calculator 3 62
3, subtraction calculator 4 624, comparison unit 5 625, comparison unit 6 626, comparison unit 7 627, comparison unit 8 628, O
It consists of an R gate 629. 700 is a power dip constant table, 800 is a total distance calculation unit, and 900 is a distance output terminal.

In such a configuration, the input voice signal input from the input terminal 100 is input to the frequency analysis unit 200, subjected to frequency analysis as a quantized signal corresponding to a plurality of frequency bands, and sent to the spectrum conversion unit 300. The data sent to the spectrum conversion unit 300 is subjected to spectrum conversion and normalized spectrum information for each frame,
It becomes voice power information and is sent to the voice section determination unit 400 and the re-sampling unit 500. Speech section determination unit 400
Determines the start and end of the voice section using the voice power information, and sends it to the re-sampling unit 500 and the power dip calculation unit 600. The spectrum data and the power data sent to the re-sampling unit 500 are subjected to time normalization at the 16 or 32 points in the extracted voice section, and only the power information of them is sent to the power dip computing unit 600. On the other hand, the spectral data is sent to another spectral information dissimilarity calculator (not shown here) by the same method as the conventional method described in FIG. 1, and the spectral distance is obtained. This result is input to 1000 in FIG.

Next, to calculate the distance using the power dip,
The resampled power information written in the power information memory unit 601 is sequentially transferred to the shift registers 602 to 605 every frame.

The present invention mainly aims to improve the recognition rate by converting the characteristics of the silent part that is different for each category into the distance of the matching calculation and using the information on the presence or absence of the power dip in the word.

FIG. 5 shows a power pattern of voice, and the voice is segmented into a start frame (STFR) and an end frame (EDFR). Now, the resampled voice power data j (j = STFR to ED) sequentially transferred to the shift registers 602 to 605.
FR) is sent to the adder / subtractor 606, the data is further sent to the power differentiator 607, and the value DFPW (j) is calculated from the equation (1).
To calculate.

DFPW (j) = POW (j + 2) + POW (j + 1) -POW
(J) -POW (j-1) ... (1) (j = STFR to EDFR) j starts from the start frame.

Next, in order to determine the power dip start point and the end point, comparison is performed in the following order.

The DFPW (j) calculated by the equation (1) is first compared with the comparison unit 1698 (THL1), and the point at which DFPW (j) is smaller than the threshold THL1 is j1. Next, starting from j1, the point where DFPW (j) is greater than or equal to the threshold THL2 is j2. Then, starting from j2, the point where DFPW (j) becomes smaller than the threshold THL3 is set as j3. Also, comparator 1 608-3
If it is already a power dip section candidate in 610,
The same operation is performed again with j1 = j3. If the power dip section candidate has not been obtained, j2 and j3 are newly obtained.

In FIG. 5, j1 = P1, j2 = P2, respectively.
j3 = P3, and the DS determination unit 611 and the DE determination unit 6
Sent to 12. The data written in the power information memory unit 601 is also sent to the comparison unit 4 618, and successive comparison is performed using the comparison unit 4 618 and the latch to obtain the maximum power value PMAX. Equation (2) shows the equation for obtaining the maximum power value PMAX.

PMAX = MAX (POW (j)) j = STFR to EDFR (2) j1, j2, j already obtained by the above-mentioned equation (1)
The values of 2-1 and j3 are subtracted / compared by the subtraction calculators 621 to 624 and the comparison unit 5 625 to the comparison unit 8 628, passed through the OR gate 629, and the power dip section is set only when the output result is “1”. Detect and output, DS determination unit 6
11 and the DE decision unit 612.

After obtaining the maximum power value PMAX from the equation (2),
(4) When any of the expressions 1) to 4) is satisfied, j1 to j3 are set as power dip section candidates.

1) POW (j1) -POW (j2)> PMAX / 8 2) POW (j1) -POW (j2-1)> PMAX / 8
(4) 3) POW (j3) -POW (j2-1)> PMAX / 8 4) POW (j3) -POW (j2-1)> PMAX / 8 In the subtraction calculators 621 to 624, the expression (4) is used. On the right side of the division calculator 620.
The maximum power value PMAX already obtained from equation (2) is divided by 8. After both sides are obtained, the comparison unit 5 625 to the comparison unit 8
When compared in 628 and any one of 1) to 4) is satisfied,
It is considered as a power dip section candidate via the OR gate 629, and the information is sent to the DS determination unit 611 and the DE determination unit 612. A power dip start point DS is stored in the DS determination unit 611, and a power dip end point DE is stored in the DE determination unit 612.

J sent to the DS determination unit 611 and the DE determination unit 612
1, j3 and the power dip section detection data are further passed through the power dip section determination unit 613 to the PA calculator 6
14 and QA calculator 615. In the power dip section determination unit 613, when two power dip section candidates are created, for example, P1 to P3 and P3 to P5 are candidates in FIG. 5, but the following Expression (5) 1) to 3) is used. When all are satisfied, P1 to P5 are set as the power dip section, and 1) POW (P1) -POW (P3)> 0 2) POW (P5) -POW (P3)> 0 ...... (5) 3 ) POW (P1) + POW (P5)> 3 * POW (P3) In FIG. 5, two of D1 to D2 and D2 to D3 are power dip sections.

The data determined by the power dip section determination unit 613 is sent to the PA calculator 614 and the QA calculator 615,
The slope and intercept of a straight line connecting the start and end values of the power dip section shown in (i) of FIG. 5 are calculated.

First, (DS, POW (DS)), (DE, POW (DE))
The equation of the straight line connecting the two points of is calculated from the equation (9).

(j) = PA * j + QA (7) where PA is the slope of the line and QA is the intercept.

Equation (8) is used to find the slope PA of the straight line, and equation (9) is used to intercept Q.
An equation for obtaining A is shown.

After obtaining the intercept and the slope of the straight line in this way, the power dip evaluation value calculation unit 616 calculates the evaluation function value of the size of the power dip. The evaluation function value is defined as being normalized by the equation (10).

In the power dip evaluation function value PDV of the equation (10), C
Is a constant and C = 2. SS represents the area of the power dip and is calculated from the equation (11).

WW represents the width of the power dip and is calculated from equation (12).

WW = DE-DS + 1 (12) That is, the larger the value of WW, the less the possibility of power dip.

AA represents the slope of the power dip and is calculated by the equation (13).

AA = PA ² +1 (13) The power dip magnitude PDV calculated from the equation (10) as described above is sent to the power dip distance calculation unit 617. The power dip distance calculation unit 617 calculates the distance due to the power dip according to the equation (14), and calculates the distance value according to the presence or absence of the power dip.

(14-1) is the case with power dip, and (14-2) is the case without power dip. C1, DBMAX, C2 are constants, and C1 = 3, C2 = 5, DBMAX = 300, respectively.
And The values of CCP and CCN are given from the power dip constant table 700.

Here, Fig. 6 shows some of the set values of the CCP and CCN values according to each category.

In this way, CCP and CCN are determined in advance by the presence or absence of power dip and the size of power dip for each category, and the value of CCP is determined in the range of 0 to 8,
For categories with power dips, such as "Ichi", "Roku", "Hachi", "Horyu", CCP is "0",
On the contrary, the category of power dip such as “Young”, “Go”, “High” is “8”.

The value of CCN is “0” for the category that is the reverse of CCP and has no power dip, and is “30” for the category with power dip.

If the result calculated by the formula (14) satisfies the condition of the formula (14-1), it is a power dip, and if the condition of the formula (14-2) is satisfied, the judgment without the power dip is performed. The determination result DBC is sent to the total distance calculation unit 800. The total distance calculation unit 800 adds the distance between the conventional spectral distance value sent from the spectral distance information 1000 and the power dip determination result DBC obtained from the equation (14), and outputs the result from the distance output terminal 900. To do.

As described above, according to the present invention, in addition to the normal spectrum matching distance, the information on the presence or absence of the power dip of each category is taken in, so that the category having the power dip, such as “ichi”, “roku”, “bee”, etc. And, it is possible to reduce the confusion of categories such as “d”, “yeon”, “go” that do not have a power dip, and it is possible to improve the recognition rate.

In order to prove the effectiveness of the present invention described above, the results of recognition experiments will be described.

A recognition experiment was performed using a standard pattern created from about 5800 patterns for about 2900 male data, and the conventional recognition rate without power dip was 96.61%.
On the other hand, the recognition rate was improved to 97.80%, which is a little over 1%. At the same time, the difference in the distance between the first and second places widens, and the stability of recognition is improved.

(Effect of the Invention) As described in detail above, in the present invention, in addition to the normal spectral dissimilarity, the power dip section in which the power becomes the minimum in the voice section is detected, and the presence or absence of the power dip section and the power dip section are detected. The power dip dissimilarity is calculated for each recognition target category from the size. Therefore, since the dissimilarity can be calculated by focusing only on the sections having different powers between the recognition target categories, the accurate power dip dissimilarity can be calculated for each recognition target category, and as a result, the recognition performance is significantly improved. Can be made.

[Brief description of drawings]

FIG. 1 is a block diagram of a conventional voice recognition device, FIG. 2 is a diagram showing an example of a voice pattern, FIG. 3 is a diagram showing an example of a power dip situation by words, and FIG. 4 is according to the present invention. FIG. 5 is a block diagram of an embodiment of a voice recognition device, FIG. 5 is a diagram showing a power dip setting range, and FIG. 6 is a diagram showing a power dip constant table. 100: input terminal, 200: frequency analysis unit, 300
...... Spectrum conversion unit, 400 ...... Voice section determination unit, 5
00 ... Re-sampling unit, 600 ... Matching calculation unit,
601 ... Power information memory unit, 602 ... Shift register (J + 2), 603 ... Shift register (J +
1), 604 ... Shift register (J), 605 ... Shift register (J-1), 606 ... Adder / subtractor, 60
7 ... Power differentiator, 608 ... Comparison unit 1, 609 ...
Comparing unit 2, 610 ... Comparing unit 3, 611 ... DS determining unit, 612 ... DE determining unit, 613 ... Power dip section determining unit, 614 ... PA calculator, 615 ... QA calculator, 616 ... ... Power dip evaluation value calculation unit, 617
...... Adder, 618 …… Comparison unit 4, 619 …… Latch,
620 ... Division calculator, 621 ... Subtraction calculator, 622
... Subtraction calculator 2, 623 ... Subtraction calculator 3, 624 ...
... subtraction calculator 4, 625 ... comparison unit 5, 626 ... comparison unit 6, 627 ... comparison unit 7, 628 ... comparison unit 8, 62
9 ... OR gate, 700 ... power dip constant table, 800 ... total distance calculation unit, 900 ... distance output terminal.

Claims (1)

[Claims] 1. A means for frequency-analyzing an input voice signal, a means for producing a power pattern of the analyzed input voice, and a means for producing a spectral pattern normalized by a spectral slope of the analyzed input voice. And a means for calculating a spectral dissimilarity by matching a previously prepared spectrum pattern of the standard voice with the spectrum pattern of the input voice, and a power dip that minimizes the power in the voice section from the power pattern. Means for detecting a section, calculating a power dip dissimilarity for each recognition target category from the presence or absence of the power dip section and the size of the power dip, and a spectrum matching distance representing the spectrum dissimilarity for each category To the power dip distance, which indicates the power dip dissimilarity, Means for calculating the distance, the speech recognition apparatus characterized by total matching distance thus calculated is a recognized means for determining the category to which the minimum.