JPH0311478B2 - - Google Patents

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 Fig. 1, where 1 is an input terminal, 2 is a frequency analysis section, 3 is a spectrum conversion section, 4 is a speech interval determination section, and 5 is a degree of dissimilarity. 6 is a standard speech spectrum pattern memory, 7 is a determination unit, and 8 is a recognition result output terminal. In a conventional speech recognition device, an input speech spectrum pattern and a standard spectrum pattern k (k=1 to
In the dissimilarity calculation with K), the dissimilarity D _k is defined as the m channel element of the nth time sample point of the input spectrum pattern as A(m, n), and the time sample point n of the standard spectrum pattern k. When the m-th channel element is S _k (m, n), D _k = _N 〓 ⁿ⁼¹ _M 〓 ^m=1 | A (m, n) − S _k (m, n) | × W (m, n)...(1) Calculated using equation (1), and the recognition result is the category of the standard spectrum pattern that minimizes D _k among the k standard spectrum patterns. There are many methods for calculating the weight W(m, n), but they are not the purpose of the present invention, so their description will be omitted. Conventional speech recognition devices frequency-analyze input speech, calculate a least squares approximation straight line of a speech spectrum, and use the slope of the least squares approximation straight line as a spectral slope value. If the spectral slope value is negative, the input speech is determined to be voiced, and a least square approximation straight line of the speech spectrum is drawn from the speech spectrum.
If the spectral slope value is positive, the input voice is determined to be unvoiced, and the vocal cord sound source characteristics and phonation intensity of the input voice are normalized by subtracting the average of the voice spectrum from the voice spectrum. Due to this conversion effect, the power information of the input voice is completely lost. As a result, "ichi" may be mistakenly recognized as "ni" or "go" may be mistakenly recognized as "roku". FIG. 2 shows an example of a sound pattern sonagram for "ichi", "ni", "go", and "roku". In FIG. 2, the horizontal direction is the frequency axis, and the vertical direction is the time axis. In this way, due to spectral conversion, "ichi" and "ni", "go" and "roku" have very similar patterns, and the difference is the silent interval between "i" and "chi", and the pattern between "ro" and "ro". Although the silent section between "ku" is large, the power information is lost, so it may be misrecognized as a result. Furthermore, even though voiced and unvoiced sounds are converted in very different ways, this information is lost, leading to misrecognition of "san" and "yon", or "ni" and "kiyuu". Misrecognition may occur. FIG. 3 shows examples of voice patterns for "san" and "yon" and "ni" and "kiyuu." In FIG. 3, the horizontal direction is the frequency axis, and the vertical direction is the time axis. In this way, as a result of the above conversion, "san" and "yon" and "ni" and "kiyuu" have very similar patterns, and the difference is that the first few frames of "san" are unvoiced, but "yon"'' is a voiced sound, and the leading class frame of ``kiyu'' is an unvoiced sound, but ``ni'' is a voiced sound, but the difference is very small, and due to the conversion effect described above, almost the same pattern As a result, it may be misrecognized. This 2
Misrecognition occurred due to two factors, and the recognition rate decreased. (Objective of the Invention) In order to solve these drawbacks, the present invention compares power patterns between a voice input and a standard voice and compares spectral slope value patterns between a voice input pattern and a standard pattern during dissimilarity calculation processing. This will be explained in detail below. (Structure of the Invention) FIG. 4 is a block diagram showing an example of a speech recognition device for implementing the present invention. In FIG. 4, 100 is an input terminal, and 200 is a frequency analysis section. 300 is a spectrum conversion unit, which includes a counter 301, a multiplication circuit 302, an addition circuit 303, a register 304, an addition circuit 305, and a register 30.
6. Multiplexers 307, 308, multiplication circuit 3
09, 310, subtraction/division circuit 311, register 3
12, subtraction/division circuit 313, register 314, counter 315, multiplication circuit 316, addition circuit 31
7. Consists of a delay circuit 318, a subtraction circuit 319, switching circuits 320, 321, and a division circuit 322. 400 is a voice section determining section. 500 is a dissimilarity calculation unit, which includes an input speech spectrum pattern memory 501, a subtraction circuit 502, and an absolute value circuit 50.
3. Multiplication circuit 504, weight determination circuit 505, constant generation circuit 506, accumulator, input audio power pattern memory 508, addition circuit 509, register 510, division circuit 511, standard audio average power memory 512, subtraction circuit 513, standard audio power pattern memory 514, addition circuit 515, input audio spectrum slope value pattern memory 516, standard audio spectrum slope value pattern memory 517,
It consists of switching circuits 518, 519, and 520. 600 is a standard audio spectrum pattern memory;
700 is a determination unit, and 800 is a recognition result output terminal. The input audio signal input from the input terminal 100 is input to the frequency analysis section 200, where it is frequency analyzed as a quantized signal corresponding to a plurality of frequency bands.
The signal is sent to the spectrum conversion section 300. If the M pieces of data analyzed at a certain time n by the frequency analysis unit 200 are x (m, n) (m=1 to M), the spectrum-converted input spectrum data A (m, n), (m =1 to M) is given by equation (1)'. A (m, n) = x (m, n) - (α _o m + β _o ) ...(1)' In equation (1)', α _o and β _o are the least square approximations of x (m, n), respectively. It means the slope and intercept of a straight line, and each is calculated by the following formula. If the number of data M is fixed in equations (2) and (3), _M 〓 ^m=1
m, _M 〓 ^m=1 m ² is a constant, so the denominators of equations (2) and (3) are also constants. If we set C ₁ = _M 〓 ^m=1 m, C ₂ = _M 〓 ^m=1 m ² , equations (2) and (3) become becomes. Here, C ₃ =M・_M 〓 ^m=1 m ² −( _M 〓 ^m=1 m) ² .
As is clear from equations (4) and (5), from the input data _M 〓 ^m=1
If we find m x (m, n) and _M 〓 ^m=1 x (m, n), we get
The values of α _o and β _o can be found using equations (4) and (5),
Furthermore, input spectrum data A(m,
n) can be obtained. In FIG. 4, this input spectrum data A(m,n) is created as follows. First, a multiplier circuit 302 calculates the product of input data x (m, n) inputted from the frequency analysis section 200 and m generated by a counter 301 that calculates in synchronization with the input data, and then an adder circuit 3
By accumulating the value of m x (m, n) using 03 and register 304, _M 〓 ^m=1 is stored in register 304.
The value of m.x(m,n) can be set.
Further, the value _M 〓 ^m=1 x (m, n) can be similarly set in the register 306 using the adder circuit 305 and the register 306. Further, the result is divided by M by the multiplication circuit 322 and output as the audio power P _o . However, P _o =1/M _M 〓 ^m=1 x (m, n). Next, in multiplexers 307 and 308,
By selecting the values of M and C ₁ respectively, the multiplier circuit 309 obtains M・_M 〓 ^m=1 m・x(m, n), and the multiplier circuit 310 obtains C ₁・_M 〓 ^m=1 x(m, n) is obtained,
The switching circuits 320 and 321 are connected to the subtraction/division circuit 311 side, and the subtraction/division circuit 311 is connected to the subtraction/division circuit 311 side.
(M・_M 〓 ^m=1 m・x(m, n)−C ₁・_M 〓 ^m=1 x(m,
n))/ _C3 is performed, and the result, that is, the value of _αo , is set in the register 312, and is output to the dissimilarity calculating section. Similarly, multiplexers 307 and 308 select C ₁ and C ₂ respectively, and multiplier circuits 309 and
310 and switching circuits 320 and 321 to the subtraction/division circuit 313 side, and the subtraction/division circuit 313
Using (C ₂・_M 〓 ^m=1 x (m, n)−C _1M 〓 ^m=1 m・x
(m, n))/C ₃ is performed, and the result, ie, the value of β _o , is set in the register 314. Subsequently, the counter 315 generates m, the multiplier circuit 316 calculates α _o ·m, and the adder circuit 317 calculates α _o ·m+β _o . Next, the input data x(m, n) delayed by the delay circuit 318 and α _o・m obtained by the addition circuit 317
If +β _o is subtracted by the subtraction circuit 319,
Spectrum-converted input spectrum data A
(m, n) is output to the input audio spectrum pattern memory 501. Figure 5 shows input x (m, n),
Straight line Y = α _o・m + β _o Spectrum conversion data A (m,
It is a figure showing the relationship of (n). (n is a certain time,
m=1~M) Y= _αo・m+ _βo is the least squares approximation straight line of x(m,n), and _αo・m from x(m,n)
The value obtained by subtracting +β _o is A(m, n). The voice section detection section 400 detects the start and end of a voice section and sends a start detection signal and an end detection signal to the dissimilarity calculation section.As a simple detection method, the frequency analysis section 200 detects the beginning and end of a voice section, and sends a start detection signal and an end detection signal to the dissimilarity calculation section. There is a detection method in which the average value of M pieces of analysis data is determined, and the starting point is the point in time when the value first exceeds a preset threshold value, and the ending point is the point in time when it finally falls below the threshold value. When the voice section detection unit 400 detects the beginning of the voice, input spectrum data A(m,n)
is written into the input audio spectrum pattern memory 501, the power information P _o of the input audio is written into the input audio power pattern memory 508, and the spectral slope value α _o of the input audio is written into the input audio slope value pattern memory 516. Begins. Furthermore, when the end of the audio is detected, writing to the input audio spectrum pattern memory 501, input audio power pattern memory 508, and input audio spectrum slope value pattern memory 516 is terminated, and the input spectrum pattern, input power pattern, and input spectrum are discontinued. Dissimilarity calculation processing is started based on the gradient value pattern. The input audio spectrum pattern memory 501 is a two-dimensional memory whose elements are input spectrum data A.
It is represented by (m, n) (m=1 to M, n=1 to N). The input voice power pattern memory 508 is a one-dimensional memory whose elements are IP(n), (n=1
~N). The input audio spectrum slope value pattern memory 516 is a one-dimensional memory, and its elements are represented by IA(n), (n=1 to N). The dissimilarity calculation unit 500 calculates the dissimilarity between the K standard voices and the input voice, and here we will consider calculating the dissimilarity with the k-th standard voice. The degree of dissimilarity D _k is expressed by the following equation. D _k = _N 〓 ⁿ⁼¹ _M 〓 ^m=1 | A (m, n) − S _k (m, n) | × W (m, n) + _N 〓 ⁿ⁼¹ | IP (n) − P _k (n) −PP+AP _k |×WP+ _N 〓 ⁿ⁼¹ |IA(n)−SA _k (n)|×WA …(6) Here, S _k (m, n) is the spectrum of the k-th standard speech Pattern elements (m=1~M, n=1~
M). W (m, n) is the weight determined by the weight determination circuit 505, P _k (n) (n = 1 to N) is an element of the power pattern of the standard voice k, PP is the average power of the input voice, AP _k is the average power of standard speech k, and can be expressed as AP _k =1/N _N 〓 ⁿ⁼¹ P _k (n). IA (n) is the element of the spectral slope value pattern of the input voice, SA _k (n) is the element of the spectral slope value pattern of standard voice k, and WA is the proportion of dissimilarity due to the spectral slope value among the dissimilarities. This is a weighting coefficient for WP is a weighting coefficient for setting the proportion of dissimilarity due to power pattern among dissimilarities. First, the dissimilarity calculation accumulator 507 is cleared to zero. Next, from the input speech spectrum pattern memory 501 through the switching circuit 517, the input speech element A (m, n) and the standard speech element S _k of the standard speech k are sent from the standard speech spectrum pattern memory 600.
(m, n) is read, and A(m, n) - S _k (m, n) is obtained by the subtraction circuit 502 through the switching circuit 518.
is calculated, the absolute value is taken by the absolute value circuit 503, and the multiplication circuit 504 passes through the switching circuit 519 and multiplies the weighting coefficient W(m,n). Weighting factor W
(m, n) is determined by the weight determining circuit 505. There are many methods for determining weights,
An example of this is the patent application No. 56-184416 "Voice recognition device"
The description thereof will be omitted since it is not the purpose of the present invention. Furthermore, the output of the multiplication circuit 504 is cumulatively added by an accumulator 507. m, n as m=
Repeat the above operations until 1 to M, n = 1, n.
The first term of D _k will be calculated. Next, calculate the average power PP of the input voice. The input audio power pattern IP(n), n=1 to N, is read from the input audio power pattern memory 508, accumulated by the adder circuit 509 and register 510, and stored in the register 510 as _N 〓 ⁿ⁼¹ IP(n) value. Set. This value is divided by N by the division circuit 511 to obtain the average power PP of the input voice. PP can be expressed by the following formula. PP=1/N _N 〓 ⁿ⁼¹ IP(N) ...(7) Next, the average power AP _k of the standard voice k is read from the standard voice average power memory 512, and the subtraction circuit 5
13, subtract AP _k from PP and get the power correction value PP
−Calculate AP _k . Next, the power pattern P _k (n) of the standard voice k is read from the standard voice power pattern memory 514, and the power correction value (PP-
AP _k ). The addition result is (P _k (n) + (PP−
AP _k )). On the other hand, from the input audio power pattern memory 508, the input audio power pattern IP
(n) (n=1, N) is read out through the switching circuit 518, and the switching circuit 519 reads out the addition circuit 5.
15 output is selected, and the subtraction circuit 502 calculates IP(n).
-(P _k (n) + (PP - AP _k )) is calculated, and the absolute value is taken by the absolute value circuit 503. Next, constant generation circuit 5
The constant WP is output from 06, and the switching circuit 520
The output of the absolute value circuit is multiplied by the multiplier circuit 504 and added to the accumulator 507.
When the addition to the accumulator is completed by changing the value up to ~N, it means that the second term of equation (6) has been calculated. Next, the input audio spectrum slope value pattern IA(n), (n=1~
N) in order, while the standard audio spectrum slope value pattern memory 517 is read out from the switching circuit 5.
19, the spectral slope values SA _k (n) (n = 1 to N) of the standard spectral pattern k are sequentially read out, the subtraction circuit 502 calculates IA (n) - SA _k (n), and the absolute value circuit 503 calculates the absolute Takes a value. Next, a constant WA is output from the constant generator circuit 506, multiplied by the output of the absolute value circuit by the multiplier circuit 504 through the switching circuit 520, and added to the accumulator 507. After changing n from 1 to N and completing the addition to the accumulator, the addition result is output to the determination unit 700 as a dissimilarity calculation result. The determination unit 700 determines the standard speech category with the smallest degree of dissimilarity as the recognition result. The value of the constant WP is the simulation result 1/2 ~
The optimum value for the constant WA is 2 to 8. However, if the recognized words are only words that do not include silent sections, it is also possible to set WP=0. Table 1 shows the results of a simulation conducted to compare the conventional dissimilarity calculation unit and the dissimilarity calculation unit according to the present invention. Learn a total of 12 words, including 10 numbers, hi, and yes, uttered by approximately 400 people.
A standard voice was created and the voices of another 100 people were recognized and evaluated. At this time, the standard number of voices was 192 patterns.

[Table] As shown above, the recognition rate has clearly improved compared to the conventional method. As explained above, in the first embodiment, in addition to normal pattern matching, the power pattern of the voice and the spectrum slope value pattern are compared. FIG. 6 is a diagram comparing the power of the sounds of "ichi" and "ni". Since "chi" is a voiceless plosive, there is no sound between "i" and "chi." On the other hand, “ni”
Since the power is continuous, for example, if the pattern of the input voice uttered ``ichi'' and the standard voice pattern uttered ``ni'' are compared using the dissimilarity calculation unit of the present invention, the dissimilarity calculation unit of the present invention will find that the input voice pattern uttered is more dissimilar than the conventional one. The degree of similarity increases. Furthermore, if we compare the input speech pattern of "ni" uttered with the standard speech pattern of "ni", we find that there is no silent section within the word in both cases, and even if the voice volume is different, the Since it is normalized using the average power, the degree of dissimilarity does not increase. Therefore, the standard speech pattern for "ni" and "ichi"
The degree of dissimilarity between the voice uttered as ``'' becomes larger, and the degree of dissimilarity between the voice uttered ``ni'' and the voice uttered hardly changes, so that the number of misrecognitions is reduced. These relationships are "go" and "roku", "hai" and "hachi"
It also holds true between. Moreover, FIG. 7 is a diagram comparing the spectral slope values of "San" and "Yon". The first few frames of "san" are unvoiced sounds and have positive spectral slope values, but the first few frames of "yon" are voiced sounds and have negative spectral slope values.
For this reason, if the input speech pattern uttered, for example, "san" and the standard speech pattern uttered "yon" are compared by the dissimilarity calculating section according to the present invention, the dissimilarity will be greater than that of the conventional method. Therefore, the degree of dissimilarity between the standard pattern of "yon" and the voice pronounced "san" becomes larger,
Since the degree of dissimilarity of the voice uttered as "Yon" hardly changes, misrecognition is reduced. These relationships also hold true between ``ni'' and ``kiyu.'' There is an advantage that the recognition rate is improved due to the above two factors. (Effects of the Invention) In addition to normal pattern matching, the present invention calculates dissimilarity by comparing voice power patterns and voice spectral slope patterns. , "Go" and "Roku",
It can be used in voice recognition response systems because there are fewer misrecognitions between ``san'' and ``yong,'' and the recognition rate is improved.

[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 pattern, FIG. 3 is an example of a speech pattern, and FIG. 4 is an example of a speech recognition device for implementing the present invention. Figure 5 shows the input data x
(m, n) and input spectrum pattern data A
(m, n), FIG. 6 is an example of a power pattern, and FIG. 7 is an example of a spectral slope value pattern. 100...Input terminal, 200...Frequency analysis section, 300...Spectrum conversion section, 400...Speech interval determination section, 500...Dissimilarity calculation section, 50
1...Input audio spectrum pattern memory, 50
2... Subtraction circuit, 503... Absolute value circuit, 504
... Multiplication circuit, 505 ... Weight determination circuit, 506
... Constant generation circuit, 507 ... Accumulator,
508...Input audio power pattern memory, 50
9...Addition circuit, 510...Register, 511...
...Division circuit, 512...Standard voice average power memory, 513...Subtraction circuit, 514...Standard voice power pattern memory, 515...Addition circuit, 51
6... Input speech spectrum slope value pattern memory, 517... Standard speech spectrum gradient value pattern memory, 518, 519, 520... Switching circuit, 600... Standard speech spectrum pattern memory, 700... Judgment unit.

Claims (1)

[Claims] 1. A process of creating a spectral slope value pattern of input audio, a process of creating a power pattern of input audio, and a process of creating a spectral pattern normalized by the spectral slope of input audio, A process of performing pattern matching between a pre-prepared spectral pattern of standard speech and the spectral pattern of input speech to calculate a first dissimilarity, _N 〓 ⁿ⁼¹ | IP (n) − P _k ( n) - PP + AP _k | (where N is the total number of power data in the voice section, IP (n) is the n-th power data of the input voice, P _k (n) is the n-th power data of the standard voice Power data, PP is the average power data of the input audio,
AP _k is the average power data of the standard voice), a process of performing pattern matching between a pre-prepared power pattern of the standard voice and the power pattern of the input voice to calculate a second degree of dissimilarity; performing pattern matching between a pre-prepared spectral slope value pattern and the spectral slope value pattern of input speech to calculate a third degree of dissimilarity; A speech recognition method comprising the step of adding weights to each of the input speech, and recognizing the input speech as a degree of dissimilarity between the input speech and the standard speech using the added value.