JPS5936760B2 - Recognition method using nonlinear matching - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION The present invention is based on a recognition method based on non-linear matching, particularly based on the degree of similarity between each element constituting a standard pattern sequence and each element constituting an input speech pattern sequence. When performing nonlinear matching (determining the sum of similarities ΣR), the time series of Percoll coefficient time slope components extracted from the input speech is frozen, and the sum of similarities ΣR is determined in accordance with the peak occurrence timing in the time series. The present invention relates to a recognition method using nonlinear matching, which includes a matching processing means that sequentially changes the similarity R1j used for the purpose of the present invention.

As one method for pattern recognition of input speech, a method is being considered in which a time series of coefficients called Percoll coefficients is obtained and matched with a standard time series of Percoll coefficients.

The Percoll coefficient referred to here refers to the ``8th Tohoku University Research Institute of Electrical Communication Symposium Proceedings, Fumitada Itakura, Speech Feature Extraction Using Statistical Methods'' and ``1972 Institute of Electronics and Communication Engineers National Conference Lecture Proceedings 5. This is disclosed in ``PARCOR-type speech synthesis'' by Fumitada Itakura et al., 3-9. F. Like the time series of Percoll coefficients above,
So-called non-linear matching processing is widely used to find matching between a standard pattern sequence and an input audio pattern sequence extracted from the input audio (hereinafter sometimes abbreviated as input pattern sequence).

This can be thought of as the following 5. That is, the standard pattern series S-(Si)(i-1, 2, . . . , m
) and the input pattern series P ■(Pj) (j■1, 2,・
...n) and Wu, a matrix is created in which all the degrees of similarity Rlj between each element Si and Pj constituting the above two series are determined.

During the period from the similarity R,l corresponding to the start point to the similarity Rmn corresponding to the end point, for example, from Ri,j to Ri
,j+l<l! -Ri+1.j+1, the one with the larger similarity is selected, and the sum of these similarities is determined by sequentially tracing the selected similarities. Then, perform this process of calculating similarity for all standard pattern sequences, and select the one standard pattern sequence with the largest sum of similarities as a result, and assume that the input pattern sequence belongs to that standard pattern sequence. recognize. The nonlinear matching process is performed as described above, but as mentioned above, Ri,
In order to select the higher similarity from j to Ri,j+1 and Ri+I,j+l, i.e., to perform transition processing, a high similarity is partially taken for some reason. If a position exists in an undesired position, a transition occurs at an undesired position, and it may become impossible to determine the correct similarity sum.

The purpose of the present invention is to solve the above-mentioned problems, and the recognition method using nonlinear matching of the present invention combines each element constituting a standard pattern series prepared in advance with each element constituting a pattern series of input speech. In a recognition method using non-linear matching that performs a non-linear matching process between the two standard patterns and the pattern sequence of the input voice based on the degree of similarity between the husbands, the Percoll coefficient is determined based on the Percoll coefficient extracted from the input voice. a Perco ignition coefficient time slope component extraction unit that determines the time slope component of the Perco coefficient, a time series generation unit that generates a time series of the extracted Perco coefficient time slope component, and the standard pattern sequence and the pattern of the input audio. A matching processing means for obtaining matching with the series is provided, and the matching processing means is configured to calculate the similarity R used to obtain the similarity sum ΣR corresponding to the peak occurrence timing at 5 in the Percoll coefficient time slope component time series. is calculated from one similarity Rij between the first element constituting the standard pattern sequence and the j-th element constituting the pattern sequence of the input voice, the (1+1)th element constituting the four standard pattern sequences. The present invention is characterized in that a transition timing for transitioning to another degree of similarity Ri+1, j+l between the th element and the (j+1)th element constituting the input voice pattern sequence is given so that the transition is forcibly made.

This will be explained below with reference to the drawings. FIG. 1 shows the configuration of one embodiment of the recognition method according to the present invention, FIG. 2 shows the configuration of one embodiment of the time gradient processing section shown in FIG. 1, and FIG. 3 shows an example of the Q parameter extracted by the present invention. , FIG. 4 is an explanatory diagram for explaining the nonlinear matching process according to the present invention, and FIG. 5 shows an embodiment in the form of a flowchart of the process of the matching section according to the present invention.

In FIG. 1, reference numeral 1 denotes a Percoll coefficient k-parameter extracting unit, and 2-1 to 2-10 each a time gradient processing unit, which calculates the average value of the above-mentioned k parameter within a predetermined short time and changes the average value over time. 3 is a time series generation unit that obtains the Percoll coefficient time slope component time series Vj, and 4 is an input voice pattern sequence generation unit that extracts the input voice based on the above k parameter, for example. 5 is a matching processing unit, and 6 is a standard pattern sequence group storage unit which obtains an input pattern sequence P corresponding to a plurality of predetermined standard pattern sequences s(0)...s(r).
It represents something that stores...

In addition, in the following examples, the standard pattern series s(
o)...... takes the same form as the input pattern series P, but compared to the input pattern series P, what is called thinned out in time series is stored in the storage unit 6. . Corresponding to the input voice, the Percoll coefficient k parameter extraction unit 1 using known means extracts the k parameter k1.
, K, and O are determined, and these k parameters k1 to KlO are led to the input pattern sequence generator 4.

Then, in the generation unit 4, for example, time period t=0,T,
is determined every 2T, . In the case of the present invention, when performing matching processing in the matching processing section 5, the above-mentioned Percoll coefficient time slope component time series Vj
, and the peak occurrence timing of the time series Vj is investigated and used.

FIG. 2 shows the configuration of an embodiment of the time slope processing section shown in FIG. It represents.

For example, when the k parameter K, is input, kl(T
i), kl(Ti+TO), and kl(Ti+2T0) are averaged by the summing operational amplifier 11 to obtain a value A.

The time delay circuit 10 delays the value A, and the differential amplifier 12
generates the difference between the delayed value A and the value A, that is, the Q parameter Ql. As shown in FIG. 3, the Q parameters obtained as described above correspond to each speaker A, B, . . . and include features corresponding to words such as "4" and "9". Since it is known that the above Percoll coefficient corresponds to the characteristics related to the vocal tract, the above Q parameter Ql is the ゜' vocal tract when the speaker pronounces a certain word. This corresponds to a change in the shape of the mouth. In other words, the above Q parameter Q
Vj(t←{Q,”(t}FQ
z(t}I-...+Q.

When the Percoll coefficient time slope component time series Vj given by ゛(t)}1/2-(1) is determined, the timing at which the time series Vj takes a peak is at a point where ``゜change in mouth shape'' is large. It can be thought of as the timing when the phonology changes. In other words, for example, in the case of an input corresponding to a word such as 1 Akai I, consider that this is the timing at which the phoneme 1 A I changes to the phoneme I power 1, or the timing at which the phoneme 1 power 1 changes to the phoneme l I 1. good. The time series generator shown in FIG. 1 generates the time series VJ according to the above equation (1).

FIG. 4 shows an explanatory diagram for explaining the nonlinear matching process of the matching section according to the present invention.

As in the case of conventionally known non-linear matching processing, each element (corresponding to the phoneme 1, 1, 1, etc.) configuring one standard pattern sequence S and the input pattern sequence P
The matrix R is obtained by arranging the elements P'' constituting the abscissa and ordinate, and determining the degree of similarity Rij between each element Sl (!:PJ). Using the Percoll coefficient time slope component time series Vj, the similarity R to obtain the similarity sum ΣR is extracted as follows: 1) Using the similarity R1l as a starting point, the above time series VJ is Under the timing that is less than the predetermined threshold EO, one extracted similarity R1j to Ri,
2) At the timing when the time series Vj exceeds a predetermined threshold EO, forcibly convert Ri+1,j+ from one extracted similarity Rij.
Make the transition to I.

The validity of the transition process as described above is easily confirmed by the fact that the Percoll coefficient time slope component corresponds to a change in the shape of the mouth, that is, a change in phoneme. be understood.

FIG. 5 shows, in the form of a flowchart, the process of calculating the similarity sum ΣR according to the transition described in connection with FIG. 4 in the matching processing section 5.

The process is as follows. That is, a) the similarity R11 corresponding to the elements S and PI is set in the register W at the starting point.

b) Next, extract the similarity R,2 and add it to the register W to see if it should be added to the register W. >Check whether it is EO or not. c) If V2>EO is not satisfied, the similarity R12 is added to the register W, and a process is started to check whether or not the similarity R13 is to be extracted next.

d) But V.

>EO, the similarity Rl2 is not extracted, the similarity R22 is added to the register W, and a process is started to check whether the similarity R23 is to be extracted next. The alignment processing section 5 shown in FIG. 1 performs the processing described in connection with FIGS. 4 and 5 above, but as described above, when Vj>E
Instead of checking O, the above transition can be performed while tracking changes in the peak level of the time series Vj.

In this case, the following transition processing is performed.

That is, (3) If the similarity R extracted to calculate the similarity sum ΣR is the similarity Ri, then the above similarity Rij is divided by the Percoll coefficient time slope component Vj at the corresponding point in time. The value (Rij/Vj) is calculated in advance.

(4) - Find the value (Ri+1, j+1/K) obtained by multiplying the degree of similarity Ri+1, j+1 that may be extracted next at the relevant point in time by a constant (1/K).

(5) One above f direction (Rij/Vj) and (Ri+!, j+
Check the magnitude relationship with 1A0, and if the former is large, then the similarity R is extracted as the similarity Ri,,+! is extracted, and the same value (Ri, j+1/Vj+!) as above is obtained. Then compare with the value (Ri+1,j+2A0.(6)
If the latter is large as a result of the investigation according to (5) above, it is assumed that there is a phonological change, and the similarity Ri+1,j+1
Extract. Then, the value (Ri+1, j+1 /Vj
+! ) starts the comparison process with (Ri+j+V'K). In the above transition process, when the peak of the time series VJ appears, the value (Rij/Vj) becomes a small value and the value (Ri+
1,j+1/K) increases, the similarity Rij is still made to transition to Ri+1,j+1 in response to the peak generation timing of the Percoll coefficient time slope component.

In the above explanation, it has been shown that all of the Q parameters Q1 to QlO are used according to the above equation (1) to obtain the Percoll coefficient time slope component time series Vj, but the peak in the above time series Vj The above Q parameter Q is set so that the value appears smoothly as shown in FIG.
1 to some of QlO can be selected to obtain the time series Vj.

According to simulations conducted by the inventors, it was confirmed that employing this method yields more favorable results. Needless to say, as mentioned above, in the conventionally known nonlinear matching process, from Ri,j to Ri,j+
Since the path is traced by selecting the one with a greater degree of similarity between 1 and Ri+1, j+1, there is a risk that a transition will occur in an undesired manner. On the other hand, in the case of Honmeiba, since a sufficiently reliable transition timing is given,
Originally Ri+1,j+! The point where the transition should be made is getting stronger and Ri+
The possibility of proceeding to 1,j becomes stronger. As explained above, according to the present invention, the time gradient component time series Vj of the Percoll coefficient is utilized, and the transition processing is performed using the peak occurrence timing. Therefore, undesired transitions do not occur due to the undesired appearance of partially high degrees of similarity as in conventionally known methods.

[Brief explanation of drawings]

FIG. 1 shows the configuration of one embodiment of the recognition method according to the present invention, FIG. 2 shows the configuration of one embodiment of the time gradient processing section shown in FIG. 1, and FIG. 3 shows an example of the Q parameter extracted by the present invention. , 4th
The figure is an explanatory diagram for explaining the nonlinear matching process according to the present invention, and FIG. 5 shows an embodiment in which the process of the matching unit according to the present invention is expressed in the form of a flowchart. In the figure, 1 is a Percoll coefficient k parameter extraction unit, 2-1
Reference numerals 2-10 to 2-10 respectively represent time gradient processing units, 3 a time series generation unit, 4 an input voice pattern sequence generation unit, 5 a matching processing unit, and 6 a standard pattern sequence group storage unit.

Claims (1)

[Claims] 1 Based on the degree of similarity between each element constituting the standard pattern series prepared in advance and each element constituting the input voice pattern series, non-linearity between the standard pattern series and the input voice pattern series is determined. In a recognition method using non-linear matching that performs matching processing, a Percoll coefficient time slope component extraction unit that determines a time slope component of the Percoll coefficient based on the Percoll coefficient extracted from the input voice; It has a time series generation section that generates a time series, and a matching processing means for matching the standard pattern series and the pattern series of the input audio, and the matching processing means includes a time series generation section that generates a time series, and a matching processing means that detects a peak in the Percoll coefficient time slope component time series. Corresponding to the occurrence timing, the similarity R used to calculate the similarity sum ΣR is calculated by calculating the similarity R between the i-th element constituting the standard pattern sequence and the j-th element constituting the pattern sequence of the input voice. Transition from one similarity Rij to another similarity Ri+_1, j+1 between the (i+1)th element constituting the standard pattern sequence and the (j+1)th element constituting the pattern sequence of the input voice. A recognition method using non-linear matching, characterized in that a transition timing is given and the transition is forcibly made.