JPS6019199A - Voice recognition - 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 method designed to improve recognition performance.

(Prior 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 overlapping circuit memory, 7 is a judgment section, and 8 is a recognition section. This is a result output terminal.

In conventional speech recognition devices, the input speech spectrum tsukturn that has been subjected to space-coverage conversion and the standard spectral nosoturn k (
k=1~K), the dissimilarity Dk
Let the m-th channel element of the n-th time sample point of the standard spectrum pattern be A (m + n ), and let the m-th channel element of the n-th time sample point of the standard spectrum pattern be Sk (m, n ), Dk=f, 'l
, IA(m,n)-8k(m,n]XW(m,n) (
1) n = 1 m = 1 Calculated according to equation (1), and the category of the standard spectrum pattern that minimizes Dk among the K standard spectrum ie turns is set as the recognition result. Here, the weight W (m r
There are many methods for calculating n), but they are not the purpose of this invention, so their description will be omitted.

In conventional recognition devices, the power information of the input speech is completely lost due to spectral conversion. As a result, for example, "ichi" may be mistakenly recognized as "two" or "go" may be mistakenly recognized as "roku".

FIG. 2 shows examples of voice pattern tunagrams for ``ichi'', ``ni'', ``ko'', and ``roku''. In FIG. 2, the horizontal direction is the frequency axis, and the vertical direction is the time axis.

In this way, the spectral transformation allows the ``1'' and ``2''.

``Go'' and ``Roku'' have similar patterns, but the differences are the silent interval between ``i'' and ``chi'', and the ``mouth'' pattern.
Although there is a large silent section between "ri" and "ri", the power information is lost, resulting in misrecognition, which causes a drop in recognition rate.

(Objective of the Invention) An object of the present invention is to provide a speech recognition method capable of solving these drawbacks and improving the recognition rate.

(Summary of the Invention) The present invention is designed to compare power and turn between speech input and standard speech during dissimilarity calculation processing, and will be described in detail below.

(Embodiment of the invention) FIG. 3 is a block diagram showing one embodiment of the invention. In FIG. 3, 100 is an input terminal, and 2θO is a frequency analysis section. 300 is a spectrum conversion unit, counter 3011 multiplication circuit 302, addition circuit 303, register 3θ4, addition circuit 305, register 306, multiplexer 307.30B, multiplication circuits 309, 310, subtraction/division circuit 31j register 312, subtraction/division circuit 313,
register 314, counter 315, multiplication circuit 316, addition circuit 777, delay circuit 318, subtraction circuit 119,
It consists of switching circuits 320 and 321 and a division circuit 322.

400 is a voice section determining section. 50θ is a dissimilarity calculation unit, and input voice spectrum/, o-turn memory 50
1. Subtraction circuit 502, absolute value circuit 503, multiplication circuit 50
4. Weight determination circuit 505, constant generation circuit 506, accumulator, input audio nozzle and turn memory 508,
Addition circuit 509, register 510, division circuit 5 No. 1, standard audio averaging/warming memory 512, subtraction circuit 513, standard audio power pattern memory 514, addition circuit 515,
It consists of switching circuits 516, 517+'528.

600 is a standard speech overlap pattern memory, 7θO is a determination unit, and 800 is a recognition result output terminal.

The input audio signal inputted from the input terminal 1θO is inputted to the frequency analysis section 20θ, frequency-analyzed as a quantized signal corresponding to a plurality of frequency bands, and then transmitted to the spectrum conversion section 30.
'Sent to 0.

If the M pieces of data analyzed at a certain time n by the frequency analysis unit 20-0 are x(m, n) (m-1 to M), then the spectrum-converted input waste vector data A (
m + n ) + (m-1 to M) is given by equation (1).

A (m, n) = x (m, n') - (αn"m + β1) -
(1) In equation (1), αn and βn are each X (m
+ n ) means the slope and intercept of the least squares approximation straight line, and are determined by the following equations, respectively.

M−L r+rx(m, n)−L m, f:, x(
m, n ) m'''1 m=J m=1 number.

Self='! If we set l + m + C2 = missing m2, (2)
, Equation (3) is m= 1 m= 1 Scatter vector data A (m + n
) can be set. In FIG. 4, this input waste vector data A(m, n) is created as follows. first,
Input data x (m,
n) and a counter 301 that calculates in synchronization with input data.
The multiplication circuit 302 calculates the product with m generated by , and then the addition circuit 303 and register 304 calculate
By accumulating the values of in''X(m+n), the value of Σm''X(nl+n) can be set to -1 in the register 304. Further, the adder circuit 305 and the register 306 similarly output the signal as the register 306 power Pn.

Next, in the multiplexers 307 and 308, by selecting the value of M + CB, the multiplier circuit 32
Subtraction/division circuit by 1? The result, that is, the value of αn, is set in the register 312 by the subtraction/division circuit J I 1 connected to the I l side, and outputted to the dissimilarity calculation section.

Similarly, multiplexer 307. ．． In 30B, C1 and C2 are selected respectively, and a multiplication circuit, 709.310
and switching circuits 320 and 321 as a subtraction/division circuit 313
switch to the side, use the subtraction/division circuit 313,
The result, that is, the value of in, is set in the register 314.

Next, the counter 315 generates pm, and the multiplication circuit 3
The addition circuit 317 can be used to add αl'm and αn'm+βn to the adder circuit 317.

Next, the input data delayed by the delay circuit 318 )[(m
When the subtraction circuit 319 subtracts αn'm+βn obtained by the addition circuit 317 from be done.

Figure 4 shows input data X (m + n) r straight line Y=
αn−m+βn.

It is a diagram showing the relationship between input waste vector and mother turn data A(m, n). (n is a certain time lm=1~M) Y=
αn'rn+βn is the least squares approximation straight line of x(m, n), and A(m, n) is obtained by subtracting αn'm+βn from X(m+1).

The voice section detection section 40θ detects the start end and θm end of the voice section and sends a start end detection signal and an end detection signal to the dissimilarity calculation section.A simple detection method is frequency analysis for each sample period. There is a detection method in which the starting point is the point in time when the average value of M pieces of analysis data from a section first exceeds a preset threshold value, and the ending point is the point in time when the value finally falls below the threshold value.

When the voice section detection unit 400 detects the start of a voice, it writes the input voice sequence data A (m + n) into the input voice sequence pattern memory 501, and writes the voice information Pn of the input voice into the input voice. - Writing to the warp pattern memory 508 is started. When the end of the voice is detected, writing of the input voice to the vector pattern memory 50 and the input voice power pattern memory 508 is stopped, and dissimilarity calculation processing is started. The input audio switch and turn memory 501 is a two-dimensional memory, and its elements are input scrap data A (
m, n) (m-1 to M, n = 1 to N). The input voice power pattern memory 508 is a one-dimensional memory, and its elements are IP(n) (n = 1 to
N). The dissimilarity calculating unit 500 calculates the dissimilarity between the 5th standard speech and the input speech, and here, it is assumed that the dissimilarity with the 5th standard speech is calculated.

The degree of dissimilarity Dk is expressed by the following equation.

Here, Sk (m, n) is the 4-turn element of the spectrum of the second standard voice (m = 1~N, n = 1~
N). W (m, n) is the weight determined by the weight determination circuit 505, = P k (n) (n = 1 to N) is the element of the power pattern of the standard voice, and PP is the average of the input voice. APk is a weighting coefficient for setting the ratio of standard tones.

First, the dissimilarity calculation accumulator 505 is cleared to zero.

Next, the input audio element A (m +'n) is switched from the input audio power pattern memory 501 to the switching circuit 51G.
Also, the element Sk (m+n) of the standard voice from the standard voice spectrum A-turn memory coo is read through the switching circuit 517, and A(m', n) - Sk(m, n) is read out by the subtraction circuit 5θ2. is divided, the absolute value is obtained by the absolute value circuit 503, and the weighting coefficient W(m
, n). The weight coefficient W(m, n) is determined by the weight determination circuit 505. There are many methods for determining weights, for example, patent application 18/1982
416 "Voice Recognition Device" and is not the purpose of the present invention, so its explanation will be omitted. Further, the output of the multiplication circuit is added to the accumulator 505. m.

The above operation is repeated with m=1 to M and n=1 to N1 to calculate the first term of Dk.

Next, calculate the average of the input audio. The input voice power and turn IP (n), n = 1 to N are read from the input voice power/turn memo IJ s o s, and are accumulated by the adder circuit 509 and the reno star 510 to the reno star 51θ ΣIP (n) Set the value of This n=+ value is divided by 9N by the division circuit 51 to find the average of the input audio. PP can be expressed by the following formula.

ppm! -Kei IP (N) N. ==4 (7) Next, standard voice average power memo! The average power APk of the standard voice is read from J512, and APk is subtracted from SHPP by the subtraction circuit 513 to calculate the power correction value PP-APk.

Next, the ie power pattern pk(n) of the standard voice from the standard voice power pattern memory 514 is added to the power correction value '(pp-Apk) by the adding circuit 515.

The addition result is (Pk(n)+(PP-APk)).

On the other hand, the input voice power pattern IP(n) (n=1, N) is read from the input voice power pattern memory 50B through the switching circuit 517, and the switching circuit 51
7, selects the output of the adder circuit 515, and selects the output of the subtracter circuit 50.
2 and IP(n) (Pk(n) + (PP - APk)
), and the absolute value is taken by the absolute value circuit 503.

Next, a constant value is output from the pre-constant generation circuit 506, which is multiplied by the output of the absolute value circuit by the multiplier circuit 5θ4 through the switching circuit 51'8 and added to the accumulator 505.

When the addition to the accumulator is completed by changing n from 1 to N-j, the determination unit 7 uses the addition result as a dissimilarity calculation result.
Output to 00. The determination unit 700 takes the categorization of the standard speech with the smallest degree of dissimilarity as the recognition result. As a result of simulation, the optimum value of the constant wp is about 1/2 to 2.

FIG. 5 shows a standard speech pattern created and evaluated by learning the steps 1 and 2 performed in order to determine the value of the weighting coefficient when incorporating power pattern comparison into dissimilarity. At this time, the number of standard voice patterns is 192 patterns, and ■ is varied from 0 to 41.

In this way, it can be seen that the recognition rate is clearly improved compared to the conventional dissimilarity calculation unit (corresponding to Vt/P=O), and the optimum value of wp is 1/2 to 2.

As explained above, in the first embodiment, in addition to normal pattern matching, audio power patterns are compared.

Figure 6 is a diagram comparing the sounds of ``ichi'' and ``ni''. Since “chi” is a voiceless plosive, “i” and “
There will be no sound during "ch". On the other hand, the power of "two" is continuous, so for example, the input voice pattern of "ichi" and the standard voice i? of "two" are uttered. When the turns are compared using the dissimilarity calculation unit according to the present invention, the dissimilarity becomes larger than that of the conventional method.

Also, if you compare the input speech pattern generated as "2" with the standard speech pattern for "2", you will find that there is no silent section within the word in both cases, and even if the voice volume is different, the average of the speech is Since the powers are normalized somewhat so that the powers are equal, the degree of dissimilarity does not increase.

Therefore, the degree of dissimilarity between the standard speech pattern of ``two'' and the voice uttered as ``ichi'' is yoshiinu kikunashi, and the degree of dissimilarity between the voice uttered as ``two'' hardly changes. The recognition rate is improved with fewer misrecognitions. These relationships also exist between ``go'' and ``roku,'' and between ``hai'' and ``hachi.''

(Effect of the invention) In addition to normal pattern matching, this invention compares the four turns of speech in a power-normalized form and calculates the degree of dissimilarity. , ``go'' and ``roku'', ``hai'' and ``hachi'', etc. are less likely to be misrecognized, and the recognition rate is improved, so it can be used in voice recognition response systems.

[Brief explanation of the drawing]

Fig. 1 is a block diagram of a conventional speech recognition device, Fig. 2 is an example of a speech ie turn, Fig. 3 is a diagram showing an embodiment of a speech recognition device according to the present invention, and Fig. 4 is a diagram showing input data. X (m
+ n ) and incoming scrap overlap pattern data A, (
m, n), FIG. 5 is a diagram showing the sea milling result for determining the weighting coefficient fraud, and FIG. 6 is an example of the power pattern. 100...Input terminal, 200...Frequency analysis section. 300... Spectrum conversion section, 400... Speech interval determination section, 500... Dissimilarity calculation section, 501... Input speech spectrum pattern memory, 5o2... Atsushi, arithmetic circuit, 503... Absolute value Circuit, 5o4... Multiplication circuit, 5
05... Weight determination circuit, 506... Constant generation circuit,
5oz...Accumulator, 5o8...Input audio power ie turn memory, 509...Addition circuit, 510
'・Register, 511...Division circuit, 512...Standard audio average power memory, 513...Subtraction circuit, 51
4...Standard audio power pattern memory, 515...
Addition jfE circuit, 516, 517, 518... Switching circuit, 600... Standard audio overlap pattern memory, 700... Judgment unit. Patent Applicant: Oki Electric Industry Co., Ltd. Figure 1 Figure 2 Figure 2 Dongichi t + 7 Eight calls → Number of carvings ↓ i18 questions (bl Figure 5 P Figure 6 Procedural Amendments Transfer Showa -9°11゛ Month 1 Commissioner of the Japan Patent Office 1. Indication of the case Patent Application No. 126237 of 1982 2. Name of the invention Speech recognition method 3. Relationship with the person making the amendment case Patent applicant's office (105) Minato-ku, Tokyo Toranomon 1-7-12
No. 4 Agent address (105) 1-7-12 Toranomon, Minato-ku, Tokyo
Item 6. Contents of amendment Attachment. Toshiro, Contents of amendment (1) Page 9, lines 16, 17, and 18 of the specification,
"Detection" on page 10, lines 2, 4, 5, and 9
I corrected it to ``decision''. (2) Formula (6) on page 11 of the same book is amended as follows. +Σl IP(n) −(Pk(n) +(PP AP
k)) lxwpi1=1 ・・・・・・(6) (3) Kaimei line 14, line 2, line 6, page 16, number 11
In the first line, [Simulation j is corrected to ``Simulation.'' (4) Amend the drawing “Figure 2” as shown in the attached sheet.

Claims (1)

[Claims] A process of creating a power no P turn of an input voice, a process of creating a swictor pattern of the input voice normalized by a swictor slope, and a spectral A' turn prepared in advance of a standard voice. A process of calculating a first degree of dissimilarity by performing tsukturn matching with the spectrum/mother turn of the input speech, and a step of matching the average power, power and After normalizing the input voice based on the average number of words, the normalized iQ word/e turn is matched with the power number and turn of the input voice, and a second dissimilarity is calculated. and a step of adding a weight of (i to 2) to the first dissimilarity and then adding it to the second dissimilarity, and the added value is added to the input speech and the standard speech. A speech recognition method characterized by recognizing input speech as a degree of dissimilarity between the input speech and the input speech.