JPS63236000A - 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] Industrial Application Field 3 -\1 The present invention relates to a voice recognition method for causing a machine to recognize a human voice.

2. Description of the Related Art Speech recognition technologies have been actively developed and commercialized in recent years, but most of these are for specific speakers whose voices are recognized only by those who have registered their voices.

Devices for specific speakers require time and effort to register the words to be recognized in the device in advance, which puts a heavy burden on the user unless the device is used continuously for a long time. In response to this, research has recently been actively conducted on recognition technology for non-specific speakers that is easy to use and does not require voice registration.

Generally speaking, a speech recognition method involves pattern matching between the input speech and standard speech stored in a dictionary (these are parameterized). This means that the voice of the user is output as the recognition result. In this case, there is no problem if the input voice and the voice in the dictionary are physically exactly the same, but in general, even if the input voice is the same voice, it may be voiced by different people or people who say it in different ways. Physically, differences in wording are expressed as differences in spectral characteristics and differences in temporal characteristics. In other words, since the shape of the articulatory organs (lo, tongue, throat, etc.) differs from person to person, different people will have different spectral shapes even when speaking the same word.

Furthermore, the temporal characteristics differ depending on whether the voice is spoken quickly or slowly.

Speaker-independent recognition techniques require such spectra and their temporal variations to be normalized and compared to standard patterns.

As an effective method for speaker-independent speech recognition, the present applicant has already proposed a method that uses parameter time series information and a statistical distance measure (Niyata et al., “Simple speaker-independent speech recognition”). Recognition method'', Proceedings of the Acoustical Society of Japan, 1-1-4
(March 1986)) A method for further improving the computational efficiency of recognition has been filed by one of the applicants (Japanese Patent Application No. 61-272475), and will be described below as an example of the prior art.

These methods use a pattern matching method to spot speech from noise, thereby making it possible to recognize speech and detect speech sections at the same time.

First, the distance measure (statistical distance measure) used in pattern matching will be explained.

If the input word audio length is linearly expanded or contracted to J frames and the parameter vector per frame is 1+J, the input vector x becomes as follows.

0 (moon, prisoner 2..., smell) Here, each &J is a p-dimensional vector.

As a standard pattern of word ωk (=1.2...,K), the mean value vector is /l' k s covariance matrix is I
WI(, then the word with the maximum posterior probability P(ωkl&) may be taken as the recognition result.

According to Bayes' theorem, P(ωk)-P(ωk)·P(Water 1ωk)/P(Me) The first term on the right side, P(ωk), can be regarded as a constant. Non-regular-
exp (-1/2 (/(-#k) ・%Wk・(
(X−yk)) (2) The denominator term P(to) can be regarded as a constant if the input parameters are the same, but it does not become a constant when mutual comparisons are made for different inputs. Here, P
〆) follows the normal distribution of the mean value 〃x1 covariance matrix wx and hypothetically -exp[1/2(&-I/1.x)・W)
'(JC-#x)) (3) If we take the logarithm of (1), omit the constant term, and set it as ILk, we get 1Lk-(A 7zk )・'w k' (x /1t
k) −(fi−#x), \wx・(m−/px)+
log IIWk IIWx l (4)
Here, lWk and WX are all set as 1w. That is, W=(lWI+lW2+−−−−−+lWk+1Wx)
/(K+1) (5) When formula (4) is expanded, Lk = Bk −/Ak − method
(6) However, Ak=2(vl/・/1tk−%W・/l1x)
(7) Bk=Ittk-SW −/
/zk -#x -SW ・〃lz
(8) To 7 Equation (6) is a first-order discriminant with a small amount of calculation. here,
Equation (6) is transformed as follows.

=-, calculated by adding the interest three times and subtracting it once.

Next, we will explain a method for spotting and recognizing speech in noise using the above distance measure, and then we will explain a method for reducing the amount of calculation.

Targeting a sufficiently long interval that definitely includes speech, we set various subintervals within this and calculate the degree of similarity with each word (9).
It is sufficient to obtain the word using a formula and use the word with the maximum degree of similarity across all subintervals as the recognition result. If this similarity calculation were to be performed as is, the amount of calculation would be enormous, but by limiting the partial interval length in consideration of the word duration, and by commonly using the partial similarity dk) during the calculation, it can be significantly reduced. "Calculated in 2019" can be reduced. FIG. 3 is an explanatory diagram of this method.

When matching input to words, the subinterval length n (n”
, '<n <nce)) to the standard pattern length J, and the similarity is calculated for each frame with the end fixed. The degree of similarity is calculated using equation (9) along a route starting from point T on QR and ending at P. Therefore, all similarity calculations for each frame are performed within APQR. By the way, since XJ in equation (9) is the j-th frame component after expanding and contracting the interval length n, a corresponding input frame co' exists.

Therefore, using the input vector, dkl can be expressed as follows.

arrow), k)' (i, j)-cho・Ryot (1
0) However, i'-re r, (J)+1 (11
) Here, r, (J) is a function that relates the word length n to the linear expansion and contraction of J. Therefore, if the partial similarity between each input frame and the turtle is determined in advance, equation (9) can be easily calculated by selecting and adding the partial similarities having the relationship l'. By the way, APQR is 1
(move to the right every frame) Then, calculate the machine on the PS, and / + [10 partial similarity,
Store it in memory for the amount equivalent to △PQS, and
・＼− If the structure is configured to shift each frame, all the necessary similarities are stored in the memory, so the duplication of calculations for obtaining partial similarities can be omitted and the amount of calculation can be reduced. However, the amount of calculations can be further reduced by using the value L'k given by the sequential calculations shown in equations (12) and (13) as the value 9 for Lk.

(transformation) Lk-dayk RJ
(12)Ro)=,wholesale+,,,ax(R(--l
) R(-1(J-1) p(r-7)>''J
J-1+ J-1 The meaning of this equation is not only for the set of straight lines PT in PQR in Figure 3, but also for the entire set of all monotonically increasing polygonal lines with slopes ranging from 1/m to IA. This corresponds to finding the optimal matching path. L' obtained in the above manner has the property of taking a value that is equal to or smaller than Lk, so although it takes a slightly different value, it does not pose a problem in practice, but it can actually yield better results. This has been confirmed by experiment. Furthermore, according to this method, the amount of calculation and memory required are only the amount required to perform the sequential calculations of equation (13), and the cost is about 1A compared to the method of directly calculating Lk. can be reduced to.

The method will be explained below using FIGS. 4 and 5. FIG. 4 is a functional block diagram thereof.

The inputted unknown input audio signal is converted to 8 by the AD converter 41.
It is sampled at KHz and converted into a 12-bit digital signal. The acoustic analysis unit 42 performs LPG analysis of the input signal every 10 m5ec (i-frame) to obtain the 10th-order linear prediction coefficient and residual power. The feature parameter extraction unit 43 uses the linear prediction coefficients and residual power to calculate L
The PC cepstrum coefficients 01 to C5 and the power ring co are determined as feature parameters.

Therefore, the feature vector X for each frame is gray'=(Co
, C1-.-C5) (14).

The frame synchronization signal generating section 44 is a part that generates a timing signal (frame signal) every 10 m5ec, and is performed in synchronization with the recognition processing frame signal.

The standard pattern selection unit 45 sets the word number stored in the standard pattern storage unit 46 during one frame to =1.
,2.・・・・・・Select one after another.

In the partial similarity calculation unit 47, the selected standard pattern -
and the partial classification degree c of the feature vector Xi of the first frame
Calculate i(k)(i+j).

do) (i, j) -&tkJ), X battle, -4,2501
,, ) (15) The similarity calculation unit 47 calculates the above −(i, j)
Otherwise, the value obtained by the method described below is obtained as L'.

The similarity comparison unit 410 uses the obtained Lk and the temporary memory 411.
, and the one with greater similarity (smaller distance) is recorded in temporary storage 411.

L'shi? (ma Then, with =2, compare L2 obtained in the range n"<n<n" with L, (max)S-e to find the maximum similarity, and in this way -
Repeat the same procedure up to K to obtain the maximum similarity L'A9 (m
ax ) and the word number at that time ′ is temporarily stored 41
Store in 1. Next, the same procedure is repeated as 1-10+Δi, and when the final frame I=I is reached, the word number temporarily stored in the memory is =km, which is the recognition result. In addition, if the frame number i = im and word length n = nm when the maximum similarity was obtained are stored in the temporary memory 411 and updated, the speech section at that time will be displayed as the result at the same time as the recognition result. You can ask for it. The voice interval is im-nm-im.

The speech recognition process is completed as described above.

Next, the method of calculating Lk by the similarity calculating section 47 in FIG. 4 will be explained in detail. Here, as an example, J-16, l=
1, the case where m=4 will be explained.

FIG. 5 is a diagram showing an actual calculation mechanism for one frame of the similarity calculation section.

First, before operation, the contents of all memory elements are set to -■. Next, after waiting for 64 frame synchronization signals to be input, the partial phoneme sequence determined from the unknown input speech is input. Also, each time one delay element frame synchronization signal is input, the delay element functions to transfer the signal on the input side of the element to the output side. The figure shows the same circuit consisting of 16 stages (the number of stages is equal to J)]
, L'sc is calculated by sequential calculations.

In other words, if the partial cumulative value of each stage is the numerical value at the position shown in the figure, the value of the first stage is (k-4) R, -1) (is) (j= 1, 2...., 16) 伽) It can be seen that it is given by. Therefore, the value R16 of the final stage is as follows. However, ξ(2)) is a set of all functions that simultaneously satisfy . By subtracting <4 obtained in this way from the constant Bk, we finally get L'
k is obtained. With the mechanism described above, 13 sequential operations can be realized, and the desired L'k can be obtained.

According to FIG. 5, the amount of calculation required for one word per frame in this method is 16. It only takes about 16 additions. This is formula 9 to 11
If Lk is directly calculated based on the formula, the number of candidate straight lines PT in APQR in Fig. 3, that is, the number of frames in the range of the start end (specifically, about 50), and the comparison calculation of the number of frames in the range of the start end, and the number of frames in the range of the start end. It requires about 16 calculations, and the amount of memory required is only the part corresponding to △PQ etc., specifically about 500, and the method explained later has a small amount of calculations. It turns out that this is a considerably advantageous method.

Problems to be Solved by the Invention In the prior art described above, in the calculation of equation (13), l and m are constant values for all frames (l = 1, m = 4 in the specific example). Although a certain degree of recognition performance can be maintained as is, the degree of freedom in time axis alignment in pattern matching is too wide, resulting in mismatching, which is a factor that degrades recognition performance.

It is an object of the present invention to solve these conventional problems, perform fine-grained matching, and improve the recognition rate.

Means for solving the problem The present invention achieves the above object, and its technical means are as follows:
In advance, a series of length parameters P1, P2 expressing each standard pattern of the sound of the word to be recognized is prepared.
, , , PL is created using the data of all the voices to be recognized and the surrounding information of all the voices, and on the other hand, the PL is created using the data of all the voices to be recognized and the surrounding information of all the voices, and on the other hand, the unit time interval is For each frame (referred to as frame), a parameter Qn (where n is the frame number) that expresses the data in that frame is calculated, and for each frame, each element P) (i = 1.2...・For L), the distance or similarity Rn, + (
i = 1.2゜3,...L), and then use the following recurrence formula 9 to calculate k and m in the recurrence formula dynamically according to the transition of n or for each target word. This speech recognition method is characterized in that, while changing the number of words, the words that give the optimum values of Sn and L calculated for all recognition target words or all target frames n are obtained as recognition results.

Effect: In the above configuration, the present invention dynamically changes β and m as j in the above equation (13) changes, and by controlling the degree of freedom of time axis alignment in matching. It aims to improve recognition performance.

Example Embodiments of the present invention will be described below with reference to the drawings.

FIG. 1 is a functional block diagram embodying a speech recognition method according to an embodiment of the present invention.

The input unknown input audio signal is sent to the AD converter 11 by 8
The signal is sampled at Hz and converted into a 12-pit digital signal. Acoustic analysis section 12 is 10 m5ec
Performs LPC analysis of the input signal for each frame (1 frame),
Find the 10th order linear prediction coefficient and residual power. The feature parameter extraction unit 13 uses the linear prediction coefficients and the residual power to extract the LPC cepstral coefficients 01 to C5 and the power ring co.
is obtained as a feature parameter.

Therefore, the feature vector X for each frame is Xt−(Co
C1-Cs) (19).

The frame synchronization signal generator 14 is a part that generates a timing signal (frame signal) every 10 m5ec, and recognition processing is performed in synchronization with the frame signal.

The standard pattern selection unit 15 selects the word number stored in the standard pattern storage unit 16 during one frame period, k=
1.2. ...select one after another.

The partial similarity calculation unit 17 calculates the partial similarity d(k'(I, j)) between the selected standard baton aj and the feature vector of the first frame×1.

d(k)(I,j)=Ei)t-XI(J""t2t...J) In the similarity calculation unit 17, from the above d(k)(I,j), the conventional example will be described later. The value obtained by the method is obtained as L'.

The similarity comparison unit 110 calculates the obtained Lk and the temporary memory 11.
1, and the one with greater similarity (smaller distance) is recorded in temporary memory 11.

In this way, starting from frame I ``'' +6, the standard pattern has n(1) < n < n♀) for =1
Find the maximum similarity L1 (max) in the range of , then -2 and find the bow 0 and L in the range of ns<, ≦ne■:
' (max) ■ to find the maximum similarity, and in this way =
Repeat the same procedure up to K until the maximum similarity is 14? (ma
x) and the word number at that time is stored in the temporary memory 111. Next, the same procedure is repeated with +=io+Δi, and when the final frame i=l is reached, the word number temporarily stored in memory is =km, which is the recognition result. Also, if the maximum similarity is obtained, the frame number i at the time of lettering is
==im and word length n-nm are stored in temporary memory 111,
If it is updated, the speech section at that time can be obtained as a result at the same time as the recognition result. The voice interval is im-nm-im.

The speech recognition process is completed as described above.

Finally, the configuration and operation of the similarity calculation unit, which is the part in which the present invention is implemented, will be explained. FIG. 2 is a diagram for explaining this. Figure 2 is shown in Figure 5 of the conventional example (1
3) It basically operates in the same way as the mechanism that calculates equations, but
The improvement lies in the provision of a new mechanism for controlling the degree of freedom in aligning time axes during pattern matching.

First, before operation, the contents of all memory elements are set to -■. Next, after waiting for 64 frame synchronization signals to be input, the partial phoneme sequence determined from the unknown input speech is input. Further, the delay element functions to transfer the signal on the input side of the element to the output side every time one frame synchronization signal is input. In addition, the square frame labeled max in the figure is a square frame that has the function of detecting the maximum value from among a plurality of input numerical data, and also a frame for inputting the control signal SJ and a plurality of numerical data. Elements that output multiple numerical data are
, is a data selection element that validates only the data set specified by . By introducing the above data selection element in addition to the conventional configuration, the values of f, R1, R2,...R16, which are the partial cumulative values of each stage, can be expressed as in equation (16). Instead, it is given by . Here, the symbol S is an operator that specifies one of the subsets in the set consisting of the elements in the symbol parentheses as 21-. means value. Finally, L' is obtained as (temporarily) Lk=Bk-R16 (21), and this value becomes the final similarity.

Here, to describe the effect of this embodiment in comparison with the conventional method, the conventional method removes the function of the symbol Sj in equation (20), and all the functions in the parentheses after the symbol max are Element Rj-1 (n-1, 2, 3, 4) is the target of maximum value calculation.

When a pattern matching operation is performed according to the conventional method (i.e., equation (16)), in an extreme case, for example, as mentioned above, the partial similarity value (k□1) is adopted in a certain frame j, and in the next frame J+1, J1 (k-4), R was adopted, and there was a high possibility that matching would be performed that did not match reality, and there was a risk that problems would arise due to too large a degree of freedom during matching. On the other hand, according to the method according to the present embodiment, the degree of freedom in selecting the partial similarity in the matching calculation is controlled according to the transition of the recognition target word k and the frame number j, as shown in equation (20). Problems in the method can be solved, and higher recognition rates can be achieved.

Effects of the Invention In short, the present invention enables fine-grained matching by controlling the degree of freedom in selecting partial similarities in matching calculations according to changes in recognition target words and frame numbers, resulting in a higher recognition rate. This has the advantage that it is possible to realize the following.

[Brief explanation of the drawing]

FIG. 1 is a functional block diagram for embodying the speech recognition method in one embodiment of the present invention, FIG. 2 is a detailed configuration diagram of the similarity calculation section of this embodiment, and FIG. A conceptual diagram for explaining the matching method, Figure 4 is a functional block diagram for explaining an example of a configuration of a speech recognition method based on a conventional method, and Figure 5 is a functional block diagram for explaining an example of a configuration of a speech recognition method based on a conventional method. FIG. 3 is a detailed configuration diagram of a similarity calculation section. 11... AD conversion section, 12... Acoustic analysis section, 13... Feature parameter extraction section, 14.
...Frame synchronization signal generation section, 15...
Standard pattern selection unit, 16...Standard pattern storage unit, 17...Partial similarity calculation unit, 18...
...Section candidate setting unit, 19...Similarity calculation unit, 110...Similarity comparison unit, 111...
・Temporary storage.

Claims (3)

[Claims] (1) In advance, a series of parameters P_1, P of length L expressing each standard pattern of the sound of the word to be recognized.
_2,...P_L is created using the data of all the voices to be recognized and the surrounding information of all the voices, and on the other hand, a unit time interval ( For each frame, a parameter Q_n (where n is the frame number) that expresses the data in that frame is calculated, and for each frame, each element P_i (i=1, 2,...L) of the parameter series is calculated. , the distance or similarity R_n, _ between P_i and Q_n
i (i=1, 2, 3,...L), and further the following recurrence formula S_n, _o=constant S_n, _j=R_n, _j+opt(S_n_-_k
, _j_-_1, S_n_-_k_-_1, _j_-_
1, ..., S_n_-_k_-_m, _j_-_1) (j=1, 2, ..., L) {However, k and m are predetermined positive constants, and the symbol o
pt means to adopt the optimal one in parentheses. } S_n is calculated for all recognition target words or all target frames n while dynamically changing k and m of the recurrence formula according to the transition of n or for each target word, using A speech recognition method characterized in that a word giving an optimal value of L is obtained as a recognition result. (2) The speech recognition method according to claim 1, characterized in that the degree of similarity or distance between the characteristic parameters of the unknown input signal and the standard pattern of each speech is calculated using a statistical distance measure. (3) A patent claim characterized in that the statistical distance measure is any one of a measure based on posterior probability, a linear discriminant function, a quadratic discriminant function, Mahalanobis distance, Bayesian judgment, and a measure based on composite similarity. The speech recognition method according to item 2 of the scope.