JPH0361957B2 - - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION Field of Industrial Application The present invention relates to a pattern comparison device that compares an input pattern with a plurality of types of standard patterns represented by a series of feature vectors and identifies the input pattern. In particular, the present invention relates to a pattern comparison device applicable to recognition of consecutively uttered word sounds.

Conventional configuration and its problems If voice, which is the most natural means of generating information for humans, could be used as an input means for a human-machine system, the effect would be very large. In this case, it is desirable for the speech recognition device to be able to recognize continuously uttered speech as a condition for recognizing more natural speech.

As a pattern comparison device that is effective in recognizing consecutively uttered word sounds, a pattern comparison device that uses the so-called two-stage DP method, which uses two steps of dynamic programming (hereinafter referred to as DP), is in practical use.
Various methods have been announced that are based on the DP method (therefore, the basic principle is the same as the two-stage DP), require significantly less calculation, and are therefore capable of high-speed calculation.

By the way, one of the problems when recognizing isolated word speech is how to correctly cut out speech sections. In other words, when uttering word speech or monosyllabic speech, the level and other conditions are unstable near the beginning and end of the word, making it difficult to correctly extract the speech section. is relatively high. One method for dealing with this problem is a method called endpoint free matching. This is a method that performs matching using the DP method by changing the range of the starting point and ending point of a standard pattern or input pattern under certain limiting conditions.

Even in the recognition of continuous word speech based on the DP method, there is a similar problem with correctly cutting out speech sections when registering standard patterns. Since it is unclear due to the influence of words, it is better to perform matching with free endpoints.

If the DP path (DP matching path with constraint conditions as shown in Figures 1 and 3) is asymmetric, end point free matching at the start and end points of the standard pattern can be easily realized. In contrast to one method of continuous word recognition that uses the DP method as its basic principle, there is a method that performs free endpoint matching at the start and end points of a standard pattern.
IEEETRANSACTION ON ACOUSTICS,
SPEECH, AND SIGNAL PROCESSING,
It is described in VOL.ASSP-29, No.2, P.294, APRIL1981. However, there is still nothing known about endpoint free matching for input patterns. In continuous word recognition based on the DP method, it is true that matching with a standard pattern in an arbitrary interval,
In the sense that the optimal one is adopted, the input pattern appears to have free endpoints, but in the sense that all frames of the input pattern are always matched with some part of some standard pattern. End points are not free.

OBJECTS OF THE INVENTION It is an object of the present invention to provide a pattern comparison device that can perform endpoint-free matching for input patterns and has significantly improved recognition accuracy.

Structure of the Invention The present invention provides a frame of an input pattern that corresponds to the start or end frame of the standard pattern to be matched when finding a standard pattern sequence that minimizes the matching distance between the combination of standard patterns and the input pattern using dynamic programming. Replace the distance between the vectors of the input pattern and the standard pattern in the vicinity δ frame with a constant estimated value, and 1δ with respect to the above δ
This pattern comparison device is equipped with a cumulative distance calculation means for calculating the matching distance while performing optimization within the range of Δ, and performs endpoint free matching for input patterns.

The basic idea of the present invention is that in the conventional method, the final frame of the matching of the immediately preceding word and the starting frame of the subsequent word are both consecutive on the corresponding input pattern. This is to avoid inconsistency when using the DP method even if the numbers are δ frames apart. In other words, in order for the DP method to be applicable, the normalization coefficient of the cumulative distance must be proportional to the input frame length, so the cumulative distance corresponding to the skipped δ frames can be supplemented with a reasonable estimate. be.

The principles utilized in the present invention will be explained below.

The pattern comparison device of the present invention can be used to recognize various input patterns. An example of an input pattern is, for example, continuous word speech that is continuously uttered. Next, the present invention will be explained using continuous word speech as an example.

The sounds uttered by humans change from person to person and from time to time, and are temporally expanded and contracted non-linearly with respect to a standard pattern. In order to recognize the input speech by comparing this non-linearly expanded/contracted input pattern with the standard pattern, the time axis is expanded/contracted non-linearly to associate each feature vector of the input pattern and the standard pattern. It is necessary to calculate which standard pattern the pattern is most similar to. However, even though this input voice expands and contracts nonlinearly, it never becomes abnormally long or short.

If we focus on the physical characteristics of such input patterns, when comparing input patterns and standard patterns, we do not have to compare all possible possibilities, but we can compare them with certain validity determined by the physical properties of the input patterns. The comparison should be made within the range that is considered to be.

In the pattern comparison device, the input audio signal is converted into a series of several sets of numerical values (feature vectors) by frequency analysis, LPC analysis, PARCO analysis, correlation analysis, etc., and is compared with the feature vector of this input pattern. Each vector is compared with the feature vector of the standard pattern. This comparison value for each vector, that is, the cumulative distance, which is the sum of distances between vectors, is used as a measure of pattern similarity. When calculating this cumulative distance, comparing each vector for all combinations requires an enormous amount of calculation, and cannot be put to practical use as a pattern comparison device.

Considering a plane with the input pattern on one axis and the standard pattern on the other axis (hereinafter referred to as the ij plane), the combination of each vector of the input pattern and the standard pattern is It can be indicated by each grid point (hereinafter simply referred to as a point). Therefore, calculating the distance between vectors for all the above combinations means calculating the distance between vectors at each point,
Calculating the cumulative distance means sequentially calculating and summing the distances between the feature vectors of the input pattern and the feature vectors of the standard pattern corresponding thereto. The correspondence between the feature vectors of the input pattern and the standard pattern, that is, the sequence of points selected in the process of calculating this cumulative distance, is called a path.

Limiting the range of comparison in consideration of the physical properties of the input patterns described above means that, in the present invention, constraint conditions are set for route selection.

Here, terms and symbols used in the following description will be explained.

A: Input pattern (A = a ₁ a ₂ ... a _i ... a _I ), a _i is the feature vector of the i-th frame, I is the number of frames of the input pattern R ⁿ : n-th standard pattern (R ⁿ = b ⁿ ₁ b ⁿ ₂ …b ⁿ _j …b ⁿ _Jo ) b ⁿ _j is the feature vector of the j-th frame of the n-th standard pattern J ⁿ is the number of frames of the n-th standard pattern, and when N is the total number of standard patterns, 1n
N d ⁿ (i, j): inter-vector distance between the feature vector b ⁿ _j of the j-th frame of the n-th standard pattern and the feature vector a _i of the i-th frame of the input pattern.

D(i): Inter-pattern distance between the input pattern from the first to i-th frames and the combined pattern of the optimal combination of each standard pattern (hereinafter referred to as the terminal cumulative distance) N(i): The first to i-th frames The number (hereinafter referred to as
(referred to as the last standard pattern name) A number indicating the frame one frame before the starting point frame of B(i), N(i) (hereinafter referred to as the back pointer) D ⁿ (i, j): i'th to i'th of the input pattern The distance between the partial pattern up to the i-th frame and the partial pattern from the first to j-th frame of R ⁿ (hereinafter referred to as partial cumulative distance D ⁿ _i ′ (i, j)), and D
The minimum value for i′ of the sum of (i′−1) (hereinafter,
) B ⁿ (i, j): Partial cumulative distance D ⁿ _i ′(i, j) and D(i′
−
1) i′ that minimizes the sum with i^′, i.e. i^′=
When argmin [D(i′-1)+D ⁿ _i ′(i, j)],
A number indicating the frame immediately before the i^' frame (hereinafter referred to as intermediate back pointer). However, argmin ^X [ ] indicates the value of X when the value in [ ] is minimized with respect to X.

D ⁿ (i): Intermediate cumulative distance when j = J ⁿ , D ⁿ
(i)=D ⁿ (i, J ⁿ ).

B ⁿ (i): An intermediate back pointer when j=J ⁿ , and B ⁿ (i)=B ⁿ (i, J ⁿ ).

Basically, the present invention assumes that the input pattern ends at the _i0th frame, and when the last pattern is R ⁿ , the intermediate cumulative distance D _o (i ₀ ) is calculated by using i′ as the starting point. This method performs DP matching between a partial pattern A (i'- ₁ , i ₀ ) of an input pattern and a standard pattern R ⁿ , with frame i 0 as the end frame, with a free start point and a fixed end point.
The initial value D ⁿ (i', 1) of the intermediate cumulative distance at i' and the initial value B ⁿ (i', O) of the intermediate back pointer are D ⁿ (i', 1) = minD (i'-1) + d ⁿ (i
′, 1) D(i′-2)+d ₂ +d ⁿ (i′, 1) D(i′-3)+2d ₃ +d ⁿ (i′, 1) 〓 〓 D(i′-δ)+(δ −1) d〓＋d ⁿ (i′, 1) 〓 〓 〓 ○△D(i′, △)＋(△−1)d△＋d ⁿ (i′, 1)…
...(1a) B ⁿ (i', O) = mini'-1...D ⁿ (i'-1)
= when i'-2...D ⁿ (i'-1) = i'-3...D ⁿ (i'-1) = 〓 〓 i'-δ...D ⁿ (i', 1) When = 〓 〓 i′−△…D ⁿ When (i′, 1)=○△…(1b) However, d〓=｜D(i′−δ) −D(i′−δ−1 )｜...(1c) By setting, up to the i'-1st frame
As a continuation of DP matching, partial pattern A(i′−
1, i ₀ ) and the standard pattern R ⁿ .

Equation (1a) expresses the initial value of the cumulative distance in the i'-th frame, and means that there may be a maximum of Δ-1 frame jump between words. The cumulative distance corresponding to the skipped (δ-1) frames is (δ-1) of d〓, which is given by formula (1c), which expresses the distance from the last frame of the word immediately before the last word.
By doing so, the normalization coefficient of the cumulative distance can be made proportional to the number of input frames, and the possibility of applying the DP method is maintained.

In this case, the cumulative distance corresponding to the skipped (δ-1) frames is calculated as (δ-1) of the inter-vector distance d ⁿ (i, 1) in the first frame of the last word.
It can also be doubled. At this time, equation (1a) becomes the following equation (2).

There is no change in equation (1b).

D ⁿ (i', 1) = minD (i'-1) + d ⁿ
(i', 1) D(i'-2)+2d ⁿ (i', 1) D(i'-3)+3d ⁿ (i', 1) 〓 〓 D(i'-δ)+δd ⁿ (i' , 1) 〓 〓 ○△D (i′−△)＋△d ⁿ (i′, 1) ……(2) Also, most simply, the distance between the vectors for the skipped (δ−1) frames is d ⁿ (i, 1) may be replaced with a predetermined constant value, and the cumulative distance corresponding to the (δ-1) frames may be set as (δ-1) times the inter-vector distance. The point is that the distance between the vectors of this skipped (δ-1) frame, and thus the cumulative distance corresponding to this (δ-1) frame, can be replaced with a reasonable estimated value.
The examples above illustrate a few examples of how this can be done.

By doing this, for example, under the constraint conditions for the matching path as shown in FIG. 1, D ⁿ (i, j) can be changed to D ⁿ (i , j)=d ⁿ (i, j) + minD ⁿ (i-2, j
-1)+d ⁿ (i-1, j) D ⁿ (i-1, j-1) D ⁿ (i-1, j-2) The possible matching path is the area P in Figure 2. and i ₀ −2J ⁿ +2−△ii ₀ −dJ ⁿ ,
The calculation of d ⁿ (i, j) and D ⁿ (i, j) required to obtain D ⁿ (i ₀ ), which is restricted to the line segment where j = 1, is performed for each point in the area P. You only need to do each one once. In FIG. 2, the horizontal axis is the input pattern, and the vertical axis is the last standard pattern R ⁿ . The region P is an area surrounded by a straight line _P1 with a slope of 1/2 and a straight line P2 with a slope of ₂ . To find D ⁿ (i, j), as is clear from Figure 1, D ⁿ (i-2, j-
1), D ⁿ (i-1, j-1), D ⁿ (i-1, j-
2), d ⁿ (i-1, j), d ⁿ (i, j) only need to be known, so the intermediate cumulative distance on the i-th frame
To obtain D ⁿ (i, j) (where j = 1, 2, ..., J ⁿ ), calculate the intermediate cumulative distance D ⁿ (i-1, j) on the i-1th frame and the i-2th frame. , D ⁿ (i-2,
j), the terminal cumulative distance D(i) and the i-1th frame,
Inter-vector distance d ⁿ (i-1,
j), d ⁿ (i, j) (where j = 1, 2,..., J ⁿ )
All you need to do is remember.

At this time, D ⁿ (i ₀ ) can be obtained as D ⁿ (i ₀ )=D ⁿ (i ₀ , J ⁿ ).

As described above, frame i of the input pattern is 1
Each time the frame advances, the intermediate cumulative distance D ⁿ (i, j) on that frame (where j = 1, 2, ..., J ⁿ ;
n = 1, 2, ..., N) as the intermediate cumulative distance D ⁿ (i-1, j) D ⁿ (i-2,
j) and the vector distances d ⁿ (i-1, j), d ⁿ (i, j) (where j
= 1, 2, ..., J ⁿ ; n = 1, 2, ..., N), and calculate the terminal cumulative distance to the i-th frame by setting D(i) = min [D ⁿ (i, J ⁿ )]. be able to. D ⁿ (i, j) obtained in this way (where j
= 1 ^, 2, . . . , J ⁿ , n = 1, 2, .

Further, the back pointer B(i) for D(i) (the value obtained by subtracting 1 from the starting point for D(i)) is determined as follows.

Set the intermediate back pointer for D ⁿ (i, j) to B ⁿ
(i, j), (1) D ⁿ (i, j)=D ⁿ (i-2, j-1)+d ⁿ (i-
1, j) + d ⁿ (i, j), then B ⁿ (i, j) = B ⁿ (i-2, j-1) (2) D ⁿ (i, j) = D ⁿ (i-1 , j-1)+d ⁿ (i,
j), then B ⁿ (i, j) = B ⁿ (i-1, j-1) (3) D ⁿ (i, j) = D ⁿ (i-1), j-2) + d ⁿ ( i
,
j), by setting B ⁿ (i, j) = B ⁿ (i-1, j-2), D(i) = D ⁿ (i, J ⁿ ), then B(i) = B ⁿ (i, J _o ). Therefore, also for B ⁿ (i, j), 1
Remember what happened a frame ago and what happened two frames ago. However, n^=argmin [D ⁿ (i, J ⁿ )].

In addition, in the case of the path as shown in Fig. 3, D ⁿ (i,
It is clear that it is only necessary to remember the values of B ⁿ (i, j) from one frame before.

DESCRIPTION OF EMBODIMENTS An embodiment of the pattern comparison device of the present invention using the principle explained above will be described. FIG. 4 is a block diagram showing an embodiment in which the pattern comparison device of the present invention is applied to continuous word recognition. In the figure, I _o is an input terminal for an audio signal, and 1 is a feature extraction unit composed of a filter bank, etc., which converts the input audio signal into a series A of feature vectors a _i . 2
is a word standard pattern storage unit, and each of the N words serving as a recognition vocabulary is a standard pattern R ⁿ =b ⁿ _i . . .
b ⁿ _j ...b ⁿ _Jo , (1nN) are registered in advance in the form of feature vectors. 3 is an inter-vector distance calculation unit which calculates the feature vector a _i in the i-th frame of the input pattern and the n-th word standard pattern R ⁿ
The distance d ⁿ (i, j) from the feature vector b _i of
Find 1, 2,..., J ⁿ and store it until it is no longer needed. In this embodiment, the vector distance between the frame immediately before the frame for which the intermediate cumulative distance is being calculated and the frame is stored until the intermediate cumulative distance of the frame is calculated. d ⁿ (i,
j) can be defined, for example, as the city distance between a _i and b _j . That is, let the dimension of the vector be 1, and a _i =
(a _i1 , a _i2 ,…, a _i1 ), b ⁿ _j = (b ⁿ _j1 , b ⁿ _j2 ,…, b ⁿ _j1
), then d ⁿ (i, j) = ₁ 〓 ^k=1 | a _ik + b ⁿ _jk |.

Reference numeral 4 denotes a cumulative distance calculation unit, which calculates an intermediate cumulative distance D ⁿ (i, j), an end cumulative distance D (i), an intermediate back pointer B ⁿ (i, j), and a back point B (i) for the i-th frame. j = 1, 2, ..., J ⁿ ; n = 1,
2. Find N for ..., N and indicate the last word.
Find (i). When the matching path constraint conditions shown in FIG. 1 are adopted, D ⁿ (i, j) and D(i) can be found from the following equations.

D ⁿ (i, j) = d ⁿ (i, j) + minD ⁿ (i-2, j-
1) + d ⁿ (i-1, j) D ⁿ (i-1, j-1) D ⁿ (i-1, j-2) ... (3a) Initial condition D ⁿ (i, 1) = minD ( i-1)+d ⁿ (i,
1) D(i-2)+d ₂ +d ⁿ (i, 1) 〓 D(i-δ)+(δ-1)d〓+d ⁿ (i, 1) 〓 〓 D(i-△)+(△ −1) d△＋d ⁿ (i, 1)……(3b
) However, d〓=D(i-δ)-D(i-δ-1) ……(3c) D(i)= min ⁿ [D ⁿ (i, J ⁿ )] ……(3d) B ⁿ (i, j) and B(i) can be found from the following equation.

B ⁿ (i, j) is B ⁿ (i, O) = i-1 D ⁿ (i, 1) = D (i-1)
When +d(i, 1) B ⁿ (i, O) = i-1 D ⁿ (i, 1) = D (i-1)
+d(i, 1) then i-2 D ⁿ (i, 1)=D(i-2)+d ₂ +d(i, 1
) When 〓 〓 〓 〓 i−δ D ⁿ (i, 1)=D(i−δ)＋(δ−1)dδ＋d(i, 1) 〓 〓 〓 〓 i−△ D ⁿ (i, 1)=D(i-△)+(△-1)d△
When +d (i, 1)... (3e) As an initial condition, (1) D ⁿ (i, j) = D ⁿ (i-2, j-1) + d ⁿ (i, j) + d ⁿ
When (i-1, j), B ⁿ (i, j)=B ⁿ (i-2, j-1)...(3f) (2) D ⁿ (i, j)=D ⁿ (i- 1, j-1) + d ⁿ (i, j), then B ⁿ (i, j) = B ⁿ (i-1, j-1) ... (3g) (3) D ⁿ (i, j) = D ⁿ (i-1, j-2) + d ⁿ (i, j), then B ⁿ (i, j) = B ⁿ (i-1, j-2)
...(3h), and if n^ is the word number that satisfies equation (3d), then B(i) = B ⁿ (i, J ⁿ ) ... (3i). Also, N(i)=n^.

Terminal cumulative distance D(i) obtained as above
= D ⁿ ^ (i, J ⁿ ^) is stored in the terminal cumulative distance storage section 5, the back pointer B (i) = B ⁿ ^ (i, J ⁿ ^) is stored in the back point storage section 6, and the last word number N is stored. (i)=n^ is stored in the last word storage section 7.

Note that D ⁿ ^ (i, j), B ⁿ ^ (i, j) (however, j =
1, 2, . . . , J ⁿ ^; n = 1, 2, . . . , N) are temporarily stored in the cumulative distance calculation unit 14 until they are no longer needed. In this embodiment, the intermediate cumulative distances of the frames immediately before and two frames before the frame for which the intermediate cumulative distance is being calculated are stored until the intermediate cumulative distance of the frame is calculated.

Further, the terminal cumulative distance D(i) stored in the terminal cumulative distance storage unit 5 is necessary for determining the initial condition of Equation 1, and for D(i), D ⁿ ^(i+1,
It is sufficient that it is stored until 1) is obtained.

Reference numeral 8 denotes a voice section detecting section, which determines a voice section from the magnitude of the input signal. When the voice section detecting section 8 detects that voice input has started, the frame number counting section 9 starts counting for each frame. Although the above processing was for the i-th frame, the count value of the frame number counting unit 9 sets i. Therefore, the same processing as described above is performed every time the frame advances by one. The frame number counting section 9 starts counting when a voice section is detected, and is reset when the voice section ends. Therefore, in the last word storage unit 7 and back pointer storage unit 6, N(i) and B(i) are i=
1, 2, . . . , 1 will be stored.

The segmentation unit 10 issues a command to the back pointer storage unit 6 to read a predetermined back pointer. That is, when the segmentation section 10 issues the value i to the back pointer storage section 6, the back pointer B(i) is read out from the back pointer storage section 6. The segmentation unit 10 is connected to the back pointer storage unit 6 to B.
When the value (i) is received, the same value is sent to the back pointer storage section 6. Therefore, when the voice section detecting section 8 detects the end of the voice input, the final value I of the frame number counting section 9 is supplied to the segmentation section 10, and the segmentation section 10 first stores the value I in the back pointer. Issued to section 6. From then on,
According to the operation described above, B(I), B(B(I)), B(B(B(I))), . . . are stored from the back pointer storage unit 6.
, O
The following outputs will be obtained sequentially. These values are the frame at the end of the penultimate word,
The last frame of the fourth frame, the last frame of the fourth frame, etc., and since N(i) is a word that ends with the i frame, if this value is given to the last word storage unit 7, the last Recognition results are obtained in reverse order starting from the word. In order to prevent recognition results from being obtained in the reverse order, this order conversion may be performed on the output of the back pointer storage section 6 or the output of the last word storage section 7.

FIG. 5 is a flowchart when the functions of the apparatus of the embodiment are realized by software, and will be explained below in connection with the operations of each part of the apparatus of the embodiment.

Steps 100 to 104 are the parts for initializing the cumulative distance D(i), the intermediate cumulative distance D ⁿ ^(i, j), the back pointer B(i), and the intermediate back pointer B ⁿ ^(i, j). .

Steps 106 to 111 are processes executed at the i-th frame. Step 107 is a part for giving initial values of the intermediate cumulative distance and intermediate back pointer in the i-th frame. The processes of steps 109 to 111 are mainly performed by the cumulative distance calculating section 4. The notation n=argmin ⁿ [D ⁿ (i, J ⁿ )] in step 110 means that n that minimizes D ⁿ (i, J ⁿ ) is set to n^. Step 109 calculates the intermediate cumulative distance D ⁿ (i,
j), the intermediate back pointer B ⁿ (i, j) is determined. Step 110 performs the calculation of step 109 with n=
When the i-th frame of the result input for 1, 2, ..., N is the end of the input, the cumulative distance D ⁿ
This is a process to find the last word n^ that minimizes (i, J ⁿ ). In step 111, for the optimal word n^ found in step 110, N(i)=n^, D(i)=D ⁿ ^(i,
J ⁿ ^), B(i) = B ⁿ (i, J ⁿ ^), and these memories are stored in the last word storage unit 7, cumulative distance storage unit 5, and back pointer. It corresponds to the storage section 6.

Steps 112 to 115 are steps for obtaining word recognition results in the reverse order from N(i) and B(i) obtained in the above manner. This corresponds to the processing performed between the word storage units 7.

In this embodiment, formulas (1a) to (1c) are used as the initial value of the intermediate cumulative distance, but of course, formula (2) can also be used as described above.

As described above, according to the present embodiment, it is possible to appropriately deal with unstable parts of word combinations of input word speech.

Next, a second embodiment will be described in which the starting and ending ends of the standard pattern are also free. In this second embodiment, the calculation of the recurrence formula in the cumulative distance calculating section 4 is changed as follows.

Let the end point free sections at the beginning and end of the standard pattern be δ ₁ frame and δ ₂ frame, respectively. That is, the starting frame for standard pattern matching is set to be a frame between the first to δ ₁ frames, the matching end frame is set to a frame between J ⁿ -δ ₂ to ^{J n} frames, and the optimal frames are selected for each.

To do this, the intermediate cumulative distance D ⁿ (i, 2)
..., how to determine D ⁿ (i, δ ₁ ) and the terminal cumulative distance D(i)
It is sufficient to simply change the method for determining , as in the following equations (4a) to (4f). In this example as well, the limiting conditions shown in FIG. 1 are provided for the route.

D ⁿ (i, 2) = minD (i-1) D ⁿ (i-2, 1) + d ⁿ (i-1, 2) D ⁿ (i-1, 1) + d ⁿ (i, 2) ... (4a) D ⁿ (i, 3) = minD (i-1) D ⁿ (i-2, 2) + d ⁿ (i-1, 3) D ⁿ (i-1, 2) D ⁿ (i-1 , 1) + d ⁿ (i, 2) 〓 ... (4b) D ⁿ (i, δ ₁ ) = minD (i-1) D ⁿ (i-2, δ ₁ -1) + d ⁿ (i-1, δ ₁ ) D ⁿ (i-1, δ ₁ -1) D ⁿ (i-1, δ ₁ -2) 〓 D ⁿ (i-1, 1) + d ⁿ (i, δ ₁ ) ...... (4c) D(i)=D ⁿ ^ (i, j ⁿ ^) ... (4d) However, j ⁿ ^ = argmin [D ⁿ (i, j)] J ⁿ −δ ₂ ≦j≦J ⁿ n^= argmin ⁿ D ⁿ ^ (i, j ⁿ ^) Also, the intermediate back pointer B ⁿ (i, j) is D ⁿ (i, j) = D (i-1) + X ₁ , then B ⁿ (i, j) = i - 1 ... (4e) When D ⁿ (i, j) = D ⁿ (α, β) + X ₂ , B ⁿ (i, j) = B ⁿ (α, β) ... (4f) However , X ₁ , X ₂ are intermediate cumulative distances D ⁿ (i, j)
represents the distance between vectors that should be added to the intermediate cumulative distance D ⁿ (α, β) when calculating .

FIG. 5 shows a flowchart for implementing the above processing using software when the standard pattern is also free of end points. In this figure, steps given the same numbers as in FIG. 4 perform almost the same processing. The difference is step 116
- Steps 118 and 120 have been inserted. In steps 116 to 118, the starting point is set free between the first to _δ1 frames, and the intermediate cumulative distance D ⁿ (i,
j), is the part that calculates the intermediate back pointer B ⁿ (i, j), and steps 119 and 109 are δ ₁ +1≦j≦J ⁿ
This is the part for finding the intermediate cumulative distance D ⁿ (i, j) and the intermediate back pointer B ⁿ (i, j). In step 120, the optimal frame j ⁿ ^ is determined with the end point set as a free frame from J ⁿ -δ ₂ to ^{J n} frames. step
110 and 111 are J ⁿ and J ⁿ ^ at the same step in Figure 4.
The only difference is that j ⁿ ^ and j ⁿ ^ are used instead of .

As described above, in this embodiment, since the input pattern and the standard pattern are matched at the end points, continuously uttered speech can be recognized with higher accuracy than the conventional device, and it is highly practical. It has become possible to realize a continuous speech recognition device.

In addition, in the first and second embodiments, the case where continuously uttered words are recognized is described, but instead of words, monosyllables may be recognized, and other continuous patterns may be recognized. It is also applicable.

In addition, although the constraints shown in Fig. 1 are used as the constraint conditions for the route, other constraints, such as those shown in Fig. Can be set.

Furthermore, the present invention is of course applicable to recognition of any continuous pattern based on the DP method.

Effects of the Invention The pattern comparison device of the present invention is configured to match input patterns without end points, so that boundary parts of continuously generated patterns, that is, ambiguous parts affected by the preceding and succeeding patterns, can be removed. It can be matched, the recognition rate can be improved, and its practical value is high.

[Brief explanation of drawings]

FIG. 1 is a diagram showing the constraint conditions of the matching calculation path, FIG. 2 is a diagram showing the area where matching calculation is performed, FIG. 3 is a diagram showing another constraint condition of the matching calculation path, and FIG. A block diagram of a speech recognition device according to one embodiment, FIG. 5 is a flowchart when the functions of the same embodiment device are realized by software, and FIG. 6 is a flowchart when the functions of the second embodiment are realized by software. This is a flowchart. 1... Feature extraction section, 2... Word standard pattern storage section, 3... Inter-vector distance calculation section, 4... Cumulative distance calculation section, 5... End cumulative distance storage section, 6...
Back pointer storage section, 7... Last word storage section, 8... Speech section detection section, 9... Frame number counting section, 10... Segmentation section.

Claims (1)

[Claims] 1 Standard pattern storage means for storing a standard pattern R ⁿ (where n=1, 2..., N) consisting of a sequence of vectors b ⁿ ₁ b ⁿ ₂ ...b ⁿ _Jo , and Standard pattern combination R ^q(1) R ^q(2) …R ^q(X) =b ^q(1) ₁ b ^q(1) ₂ …b ^q(1) _Jq
(1) b ^q(2) ₁
b ^q(2) ₂ …b ^q(2) _Jq(2) …b ^q(X) ₁ b ^q(X) ₂ …b ^q(X) _Jq(X) and vector sequence
When finding the standard pattern sequence R q( ₁ ⁾ R ^q(2) ...R _q ^(X) that minimizes the cumulative matching distance with the input pattern A consisting of a 1 a ₂ ... a 1 by dynamic programming, the matching The frame i of the input pattern corresponding to the starting frame of the standard pattern R ^q(X) and the frame i−δ δ traced back from there
(1≦δ≦∆, where ∆ is a predetermined constant value) is the standard pattern that was matched immediately before the standard pattern.
Let the frame of the input pattern A correspond to the terminal frame of R ^q(x-1) , replace the partial cumulative matching distance between the frame i-δ and frame i with an estimated value determined depending on δ, and with respect to δ and cumulative distance calculating means for calculating the cumulative matching distance so as to minimize the cumulative matching distance. 2 The cumulative distance calculation means sets the matching start frame of the standard pattern to the first frame for the nth standard pattern to be matched with the partial pattern of the input pattern.
A frame to the δ ₁ frame, a matching end frame to the (J ⁿ - δ ₂ ) frame to a J ⁿ th frame, and optimization is performed with respect to the matching start frame and the matching end frame. The pattern comparison device according to item 1.