JPS60254099A - Voice recognition system - Google Patents

Abstract

(57) [Summary] This bulletin contains application data before electronic filing, so abstract data is not recorded.

Description

DETAILED DESCRIPTION OF THE INVENTION (Technical Field) The present invention relates to a speech recognition system, and particularly to an improvement in a method for matching speech data in order to improve recognition performance.

(Background Art) A conventional speech recognition device is configured as shown in Figure 1.
1 is an input terminal, 2 is a frequency analysis section, 3 is a spectrum conversion section, 4 is a speech interval determination section, 5 is a dissimilarity calculation section, 6 is a standard speech skeletal/guterz memory, 7 is a judgment section, and 8 is a This is a recognition result output terminal.

In conventional speech recognition devices, the spectrally converted input speech spectrum pattern and the standard vector, k (k
In the dissimilarity calculation with -1 to K), the dissimilarity Dk is input, and the m-th channel element of the n-th time sample point of the cassette vector is set as A(m, n), and the image quality is The mth channel element of the nth time sample point in the vector pattern is S
When K (m, n), calculation is performed using equation (1), and the recognition result is the category of the standard spectrum pattern whose Dk is the highest lJ· among the K standard spectrum mi-9 turns. Here the weight W
There are many methods for calculating (m, -n), but they are not the purpose of this invention and will therefore be omitted.

In conventional recognition devices, the word information of the input voice is completely lost during vector conversion.As a result, for example, "ichi" may be mistakenly recognized as "two" or "co" may be misrecognized as "ro". ”
There are cases where it is misrecognized.

FIG. 2 shows examples of tunagrams for the speech patterns of "ichi", "two", "koi", and "roku". In FIG. 2, the horizontal direction is the frequency axis, and the vertical direction is the time axis.

In this way, ``ichi'' and ``two'' are converted by
``Go'' and ``10ri'' are similar to kanashi.
Although there is a large silent interval between ``'' and ``ri'', the power information is lost, so it is likely to be misrecognized as a result.
This resulted in a thick layer of decline in recognition rate.

(Problems to be solved by the invention) The purpose of the present invention is to solve these drawbacks and improve the recognition rate by using power information, especially information in silent sections, which was completely lost due to spectrum conversion in the past. The key point is to provide a voice recognition method that allows the number ``ichi'' to be ``two'' and the number ``roku'' to be ``ko゛J''.
By using the information on the presence or absence of power valleys (hereinafter referred to as power dizzos) in words in combination with conventional spectral distance information for recognition judgment, it is possible to eliminate confu-noyons and improve the phonology, especially in phone speech recognition. It is used to improve the recognition rate by using information that has deteriorated.

Figure 3 shows the situation of Gwardizo in terms of words.

(Structure and operation of the invention) FIG. 4 is a block diagram showing one embodiment of the invention. In FIG. 4, 100 is an input terminal, and 200 is a frequency analysis section. 300 is spectrum data, 400 is a voice section determining section, and 500 is a resampling section. Reference numeral 600 denotes a noyawardinop calculation unit, which includes: 9 information memo section 601, shift register 1602, shift register 2 603, register 3604 to register 3604, nostar 4605, addition/subtractor 606, . . . and word differentiator 607, Comparison section 1608, comparison section 2609, comparison section 3 610. DS determining unit 611, DE determining unit 612,
Power dizzo interval determination unit 613, PA calculator 614,
QA calculator 615, power dizzo evaluation value calculation unit 616
, power differential 0 distance calculation section 617, comparison section 4618
, latch 619, division divider 620, subtraction calculator 1 6
2], subtraction calculator 2622, subtraction calculator: 3623, subtraction calculator 4624, comparison section 5625, comparison section 6626,
Comparison section 7627, comparison section 8628, OR key-1-62
Consists of 9. 700 is a power dinozo constant table, 800 is a comprehensive distance calculation section, and 900 is a distance output terminal.

In such a configuration, an input audio signal input from the input terminal 100 is input to the frequency analysis section 200, frequency-analyzed as a quantized signal corresponding to a plurality of frequency bands, and sent to the spectrum conversion section 300. The signal is then sent to the vector converter 300 and then subjected to spectrum conversion, resulting in normalized spectrum information and audio power information for each frame, which are sent to the audio segment determining unit 400 and the resampling unit 500. Voice section determining unit 400
determines the start and end of the voice section using the voice/voice information and sends it to the resampling section 50'O and the power dip calculation section 600. The spectrum data and power data sent to the re-sanno unit 500 are subjected to time normalization using 16 points and 1 to 32 points for the extracted voice section.
Of these, only the p-word information is sent to the p-word dip calculation section 600. On the other hand, the spectrum data is sent to another spectrum information dissimilarity calculation unit (not shown here) using the same conventional method as explained in FIG. 1, and the spectrum distance is calculated. This result is input at 1000 in FIG.

Next, to calculate distance using power dip,
- The resampled 7F word information written in the program information memory section 601 is sequentially transferred to the 77 Trenostars 602 to 605 frame by frame.

The present invention aims to improve the recognition rate by converting the characteristics of silent parts, which differ for each category, into distances for matching calculations and using information on the presence or absence of power dips in words.

FIG. 5 shows the power pattern of the audio, in which the audio is cut out between the start frame (5TFR) and the end frame (EDFR). Now, soft register 602-605
Re-sampled audio and Wordake J (
j = 5TFR - EDFR) is sent to an adder/subtractor 606, and the data is further sent to a power differentiator 607,
Calculate the value DFPW (j) from equation (1)0 DFPW (j) = POW (j + 2) + POW (j + 1)
POW (j) POW (j 1) - (1) (j =
5TFR~EDFR) J starts from the start frame.

Next, in order to determine the power dip start and end points, comparisons are made in the following order.

The DFPW (j) calculated by the formula (1) is compared with the comparison unit 1 608 (THLI), and the DFPW
Let jl be the point where (j) has a smaller value than the threshold THL1. Next, start with jl-46 and DFP
Let 32 be the point where W(j) is greater than or equal to the threshold THL2. Next, starting from 32, the point where nFPV/(j) is smaller than the threshold THL 3 is set as j3. Also, comparison sections 1608 to 36
If it is already a power dip section candidate in 10, j
The same operation is performed again with 1=j3. Also, if the Power Day and Zo sections are not candidates, J2 and J3 will be added.

In Fig. 5, J is pl, j2=P2,
j 3=P3, and is sent to the DS determining section 611 and the DE determining section 612. The data written in the power information memory section 601 is also sent to the comparison section 4618, and the data written in the power information memory section 601 is also sent to the comparison section 4618.

Perform successive approximation using
Put an X.・The formula for calculating the maximum value PMAX of
) is shown in the formula.

PMAX = MAX (POW(j)) j = S
TFR-EDFR・(2) The values of j 1 , j 2゜j2-1 and J3 already requested by the above-mentioned equation (1) are calculated by the subtraction calculators 621 to 624 and the comparison units 5625 to
It is subtracted and compared by the comparison unit 8628, passed through the OR key 29', and only when the output result is "1", the power-day 7 section is detected and output, and the DS determination unit 611 and DE
It is sent to the determination unit 612.

After calculating the maximum power value PMAX using equation (2), if any of the following conditions (4) equations 1) to 4) are satisfied,
Let J1 to J3 be power dip section candidates.

1) POW(jl)-POW(j2))PMAX/8
2) Pow(jl)-POW(j2-1)>PMAX
/8 ...(4)3) POW(ko3)T P□W(
j2-1)) PMAX/84) POW (j3)-PO
W(j2-1)) PMAX/8 subtraction calculator 621-62
4, the left side of equation (4) is subtracted, and on the right side, the maximum power value PMAX, which has already been determined from equation (2) in the division calculator 620, is divided by 8. After both sides have been determined, they are compared in comparison sections 5625 to 8628, and ■)
4), it is considered as a power dip section candidate via OR key () 629, and D≦determination unit 6
The information is sent to DE determination section 11 and DE determination section 612. DS
The determination unit 611 stores the power dip starting point DS and the power dip starting point DS.
The determination unit 612 stores the end points DE of the power dips 7, respectively.

j1 sent to the DS determining unit 611 and DE determining unit 612
The +j3 and power dip section detection data are further passed through the worddy 7 section determining section 613 to the PA calculator 6.
14 and QA calculator 615. The power dip section determination unit 613 determines that the mother Wardizzo section candidate is 2.
If completed, 41 For example, in FIG. 5, P] to P3. ? 3~P
5 is a candidate, but when all of the following equations (5)]) to 3) are satisfied, P1 to P5 are set as a word dip interval, and 1) POW (PI) - POW (P3)
) 02) POW(P5) -POW(P3) )
O-(5)3) POW(PL)+POW(P5)
:> 3 10 POW (P3) In Figure 5, D1 to D2.
Two sections, D2 and D3, are power dip sections.

(Left below) The data determined by the Pal No Dizzo interval determination unit 613 is sent to the PA calculator 614 and the QA rule calculator 615.
・Guwadino shown in f(i) of FIG.

The slope and intercept of the straight line connecting the values at the start and end of the section are calculated.

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

f(j)=PA*,+−+−qA (7
) Here, PA is the slope of the straight line and QA is the intercept.

The formula for inserting the straight line parameter PA into the second equation (8) and the equation for inserting the intercept QA into the second equation are shown.

After determining the straight line intercept and consideration, the power dip evaluation value calculation unit 616 calculates the evaluation function value of the magnitude of the power dip. The evaluation function value is defined as having been normalized by Equation 04.

-5S 1? DV engineering, □ ...... θ0) WW-AA 'PMAX In the □ 0 evaluation function value PDV of equation (10), C is a constant and C=2. SS is Power DeW
W is calculated from the trench style.

WW=DE-DS+1 (one, that is,
The larger the first shift of WW, the less the possibility of] or Wardino.

The short AA is calculated according to Equation 03, where the slope of Wardig is calculated by I-).

AA=PA +1 ......The power di calculated from the square hole is 7.
) is determined by the power dip distance calculation unit 617.
sent to. Power dip distance calculation section 61.7 Tel
"l:, □ The distance based on the 7-tipe is calculated using the formula (U→), and the distance value is calculated depending on the presence or absence of a power dip.

(14-1) is... with wardip, (1-4
, -2) is the case without nogwardinog. C1
, DBMAX, and C2 are constants, each with C1=3
．． Let C2=5°DBMAX=300. The CCP and CCN0 values are given from the Nogwardig constant table 00.

Therefore, Figure 6 shows each color code'l) CCP according to VC.
, shows a part of the setting value of CCN0 value.

In this way, CCP and CCN are predetermined in each category based on the presence or absence of power dips, the size of power dip, etc., and the CCP value is set in the range of 0 to 8. For example, “
"Ichi", "Roku", "Hachi".

In words such as "horieu", ccp is "O" and reverse r is "yon".

The category of /guwadizo such as "go" and "hai" is "8".

The CCN0 value is the opposite of CCP, and is set to "0" for categories without power dips, and "30" for categories with power dips.

If the result calculated by the αth → formula satisfies the condition of formula (14-1), it is power di 7, and (14.-2)
If the condition of the expression is satisfied, it is determined that there is no power dip, and the determination result DBC is sent to the total distance calculation unit 800. The total distance N1 calculation unit 800 calculates the distance between the conventional distance distance value sent from the distance information 1000 and the Paadizzo judgment result DBC obtained from formula 04, and outputs the distance as a result. terminal 90
Outputs 0.

As described above, in the present invention, in addition to the vector sowing distance in addition to the normal speed, by incorporating information on the number of each category and the presence or absence of warp f'yip, it is possible to perform a category with no dip. 1”, 10ri”.

Categories that do not have a power dip, such as ``Hachi'',
It is possible to reduce confusions such as ``2'', ``Yon'', and ``Go'', thereby improving the recognition rate.

In order to prove the effectiveness of the present invention described above, we will explain the results of a recognition experiment.

When we conducted a recognition experiment using a standard button created from approximately 5,800 buttons, compared to the male data (approximately 2,900) buttons, we found that the conventional recognition rate without adding power dip was 96.
The recognition rate improved by about 1 degree from 61% to 97.80%? lRarenoko. Simultaneously (C) The difference in distance between the 1st and 2nd positions widens, and the stability of recognition improves.

(Effects of the Invention) The present invention allows for more accurate discrimination between words by converting speech power information into a matching distance in addition to the usual turn matching distance. , which is effective in improving the recognition performance of speech recognition devices.

[Brief explanation of the drawing]

Fig. 1 is a block diagram of a conventional speech recognition device, Fig. 2 is a diagram showing an example of a speech /, o-turn, Fig. 3 is a diagram showing an example of a word-based situation, and Fig. 4 is a diagram showing an example of a word-based situation. 5 is a block diagram of an embodiment of the speech recognition device according to the present invention, FIG. 5 is a diagram showing the setting range of the wordinno, and FIG. 6 is a diagram showing a power delag constant table. 100...Input terminal, 200...Frequency analysis section, 3
00...Speech overlap conversion section, 400...Speech interval determination section, 500...Resampling section, 600...Matching calculation section, 601...Power information memory section, 6o2
...Shift register (J+2), 603...Shift register (J+1), 604...Shift register (J
), 605...Soft register (J-1), 606...
・Addition subtractor, 607 ・Power differentiator, 60・
・Comparison part 1.609... Comparison part 2.6], O...
・Comparison section 3.611...DS determination section, 612...D
E determining section, 6 E3... Power dip section determining section, 6
14...PA calculator, 615...QA calculator, 6
16...Nogwardizzo evaluation value calculation section, 617...
- Adder, 618... Comparison unit 4. 619... Latch, 62o... Division calculator, 621... Subtraction calculator 1
．． 622...Subtraction calculator 2.623...Subtraction calculator 3.624...Subtraction calculator 4.625...Comparison section 5
．． 626...Comparison part 6.627...Comparison part 7.62
8... Comparison part 8.629 ・OR dirt, 7o. ...Power delag constant table, 8oo...Comprehensive distance calculation unit, 900...Distance output terminal. Patent applicant: Oki Electric Industry Co., Ltd. Patent application agent: Patent attorney Megumi Yamamoto - Figures 1 to 6 Figure 2 (a) (b) (c) (d) (e) (f) Figure 5j No. 6 Figure CCP-CCN Direct Procedural Amendment (Voluntary) July 31, 1980 Commissioner of the Patent Office Gakudono Shiga1 Indication of the Case 1986 Patent Application No. 1086672 Name of the Invention Speech Recognition Method 3 Person making the amendment Relationship to the case Patent applicant name (029) Oki Electric Industry Co., Ltd. 4 Agent 5, Subject of amendment 11. Column 6 of Detailed Explanation of the Invention in Book IIl, Contents of Amendment (1) Page 3 of the Specification Column No.
) J Tob>Le's r XV1'K (m, n
) Correct as J. (2) "'W (rr3.n)" on page 3, line 13
Correct J to mouth IVK (ITI. 1 piece). -1d

Claims (1)

[Claims] Frequency analysis of the input audio signal, normalization of the analyzed spectrum slope, resampling to a certain data length to create an input audio pattern, calculation of the distance between the corresponding button and the standard button, and the distance In a speech recognition method that recognizes and determines the minimum recognition category, means for creating a power q-turn of input speech; means for creating a spectral pattern normalized by the spectral slope of input speech; means for matching the spectral pattern of the input voice with the spectral pattern of the input voice to calculate a first degree of dissimilarity; After normalization based on the average power, a second dissimilarity is calculated by performing an evaluation operation between the normalized ``Gwaa'' pattern and the above-mentioned ``ewordy'' and ``zoo'' information of the human voice. Means for producing a total matching distance by adding the distance lost points related to power D and Zo to the normal spectrum matching distance for each category; Recognizing the category name that gives the minimum total matching distance. 1. A voice recognition method comprising: means for determining.