JPH0441359B2 - - Google Patents

Description

[Detailed Description of the Invention] Technical Field of the Invention The present invention uses dynamic programming (hereinafter referred to as DP method) used in speech recognition to calculate the distance between an input feature vector time series and a dictionary feature vector time series. The present invention relates to a method for calculating distances between feature vector series to be calculated, and is designed to be able to deal with return states with small changes, especially when a feature is gradually changing from one state to another.

Problems with the conventional technology In the conventional feature vector sequence distance calculation circuit using the DP method, g _(i,j) = d _(i,j) +ming _(i,j-1) +d _(i,j-1) Distance calculations can be performed using the recurrence formula g _(i-1,j) +d _(i-1,j) g _(i-1,j-1) +2d _(i-1,j-1) or its transformations. I was gone.

Here, input feature vector sequence 〓=〓 ₁ ,
〓 ₂ …, 〓 _i …〓 _I , dictionary feature vector 〓=〓 ₁ ,
When 〓 ₂ …〓 _j …〓 _J , d _{(i, j)} represents the distance between 〓 _i and 〓 _j .

For example, if the feature vector of the number 3 (san) is schematically expressed as |SSSAANNN|, the partial expansion and contraction will be |SSAAAAANN|. For example, the dictionary feature vector has |SSSAANNN| as shown on the vertical axis in Figure 1, and the input feature vector |SSSAAANNN| is shown on the horizontal axis.
When performing distance calculations with SSSAAANNN, ideal distance calculations are performed because matching is achieved at the positions shown by the dots in Figure 1, but normally there is partial expansion and contraction as shown above, and the horizontal The input feature vector is as shown on the axis. Therefore, when calculating the distance between these feature vectors using the above formula, m ₁ ,
For m ₂ , m _{3 .} . . , when matching is performed based on the above equation, matching is obtained at a position where the direction has changed. That is, in the above equation, the three terms in the min term are as l ₁ to l ₃ in Figure 3,
It is oriented in only three directions: whether it is matched with an element diagonally above 45 degrees, with an element on the side, or with an element directly above.

Therefore, for example, as shown in FIG. 4, when there is a fine reversal of the change during the gradual change of the feature from one state to another (NAN part), from the point m ₀ to the above l ₁ The matching becomes very incomplete in the three directions ~l ₃ , and it is not possible to deal with the return that exists in the middle of such a change, so the distance calculation result is larger than the actual one, This has a problem in that accurate recognition results may not be obtained.

Purpose of the invention The purpose of the present invention is to solve this problem by:
In order to be able to deal with small reversals of changes that occur while a feature gradually changes from one state to another, the distance between feature vector series is added to the above recurrence formula with an additional term described later. This provides a distance calculation method for calculating .

Structure of the Invention In order to achieve this object, the feature vector series distance calculation method of the present invention calculates two feature vector time series = = = ₁ , = ₂ ... = _i = _{I ... 2}
... _{= j ...} , a recurrence formula calculation means is provided, and the feature vector time series 〓,
〓 is held in the first feature vector holding unit and _the second feature vector holding unit, and the distance d _{( i , )} , and by calculating the following recurrence formula or the following recurrence formula calculation means, the distance is calculated in response to the return of the change in the middle of the change of the feature from one state to another state. It is characterized by doing the following.

Recurrence formula g _(i,j) =d _(i,j) +ming _(i,j-1) +d _(i,j,1) g _(i-1,j) +d _(i-1,j) g _{( i-1,j-1)} +2d _(i-1,j-1) g _(i-1,j+1) +2d _(i-1,j+1) Recurrence formula g _(i,j) = d _{( i,j)} +ming _(i,j-1) +d _(i,j-1) g _(i-1,j) +d _(i-1,j) g _(i-1,j-1) +2d _{(i- 1,j-1)} g _(i+1,j-1) +2d _(i+1,j-1) Key Points of the Invention The key points of the present invention are to search for matching points M as shown in Figure 5 A or B. When doing so, diagonally downward to the right
Calculating means is provided so that matching points can also be calculated in the l ₄ or diagonally upward left direction l ₄ '. Therefore, the fifth
As shown in Figures A and B, compared to the conventional method of matching only in the three directions indicated by l ₁ to _{l 3} , matching can be performed in four directions. Distance can be calculated. As a result, as shown in FIG. 6, if there is a return state of change in the input feature vector, matching can be obtained in the direction of m ₀ →m _P. Conversely, as shown in FIG. 7, if there is a return state of change in the dictionary feature vector, matching can be obtained in the m ₀ '→m _P ' direction. Therefore, even if there is a return state of change, it is possible to perform distance calculations that are based on reality, and the recognition results are improved.

When actually performing calculations, if matching is to be found in the conventional directions l ₁ to _{l 3} , it is sufficient to find the more minimal one of the following equations (1) to (3).

g _(i,j-1) +d _(i,j-1) ……(1) g _(i-1,j) +d _(i-1,j) ……(2) g _{(i-1,j- 1)} +2d _(i-1,j-1) ...(3) In addition to these three formulas (1) to (3), this invention also provides the following (4)
Equation (when there is a return state of change in the input feature vector) g _(i-1,j+1) +2d _(i-1,j+1) ...(4) or (5) equation (when the input feature vector has a change return state) When there is a return state of change) g _(i+1,j-1) +2d _(i+1,j-1) ...(5) is added to the gradual Find the distance using the formula.

g _(i,j) =d _(i,j) +ming _(i,j-1) +d _(i,j-1) g _(i-1,j) +d _(i-1,j) g _{(i-1 ,j-1)} +2d _(i-1,j-1) g _(i-1,j+1) +2d _(i-1,j+1) ...[] g _(i,j) =d _{(i, j)} +ming _(i,j-1) +d _(i,j-1) g _(i-1,j) +d _(i-1,j) g _(i-1,j-1) +2d _{(i-1, j-1)} g _(i+1,j-1) +2d _(i+1,j-1) ...[] Embodiment of the invention An embodiment of the invention will be described based on FIG. 8 and FIG. A case where the input feature vector has a change return state will be described with reference to the figure.

FIG. 8 is a configuration diagram of an embodiment of the present invention, and FIG.
The figure is an explanatory diagram of the output state of the address and write signal output from the controller when there is a return state of change in the input feature vector, that is, the control sequence of the controller.

In the figure, 1 is a controller, 2 is an input feature memory, 3 is a dictionary memory, 4 is an intermediate result holding memory,
5 is an intermediate distance holding memory, 6 is an inter-vector distance calculating section, and 7 is a recurrence formula calculating section.

The controller 1 outputs addresses i, j, i', j' and a write signal W in the order shown in FIG.

The input feature memory 2 stores the input feature vector series (〓
i), and is also a dictionary memory 3
is a memory in which the dictionary feature vector (〓i) is stored.

The intermediate result holding memory 4 is a memory that holds g _(i , _j) as an intermediate result of the operation according to the above formula [], and this is displayed as g _(i ', _j ' ₎ . The intermediate distance holding memory 5 is a memory that holds distances of feature vectors calculated in the past, and holds intermediate results of calculations so that they can be used to improve the efficiency of calculations.

The inter-vector distance calculation unit 6 calculates the inter-vector distance d _(i , _j) between the input feature vector and the dictionary feature vector, and when the feature vector is 16-dimensional, _k=16 〓 ^k=1 | It calculates a _ik −b _jk |.

The recurrence formula calculation section 7 calculates the above formula [], but in this case, it calculates the minimum value of the formulas (1) to (4) above that make up the formula []. and add this to d _(i,j) .

Next, the calculation control state shown in FIG. 8 will be explained.

(1) First, as shown in FIG. 9, the controller 1 outputs i=1 and j=1 as addresses, and the write signal W is set to "0". As a result, the input feature vector from input feature memory 2
a _{1k (K=1 to 16)} is output, and the dictionary memory 3 outputs a dictionary feature vector b _1k . Then, these are calculated by the vector distance calculation unit 6 as ₁₆ 〓 ^k=1 | a _1k
−b _1k | is performed to calculate d _(1,1) , which is output to the recurrence formula calculation unit 7. On the other hand, controller 1
Separately, i' = 0, j' = 1 are output as addresses, and g _(0,1) and d _(0,1) are output from the intermediate result holding memory 4 and intermediate distance holding memory 5, and the recurrence formula calculation section 7, and the following address (i'=1, j'=0),
(i'=0, j'=0), (i'=2, j'=0), (i'
=0,
j′=1) and output g ₍₁ , ₀₎ , d _(1,0) ; g _(0,0) ,
d _(0,0) ; g _(2,0) , d _(2,0) ; g _(1,1) , d _(1,1) are sent to the recurrence formula calculation unit 7, and the above [] The calculation of the formula is performed, and when the controller 1 outputs W = "1", the minimum value is written in the intermediate result holding memory 4 as g _(1,1) , and the above d _(1,1) is written in the intermediate result holding memory 4. It is written in the intermediate distance holding memory 5.

(2) Next, controller 1 uses the address i=
1, j=2, and extracts the input feature vector a _1k from the input feature memory 2 in the same manner as in (1) above, but extracts the dictionary feature vector b _2k from the dictionary memory 3.
The above calculation is performed to find the inter-vector distance d _(1,2) . And controller 1 has addresses (i'=0, j'=2), (i'=1, j'=1), (i'
=0,
j'=1), (i'=2, j'=1), (i'=1, j'=
2) are sequentially output from the intermediate result holding memory 4 and the intermediate distance holding memory 5 as g ₍₀ , ₂₎ , d _(0,2) ; g _(1,1) ,
d _(1,1) ; g _(0,1) , d _(0,1) ; g _(2,1) , d _(2,1) ; g _(1,2) , d _{(
1, 2)} , and calculate the above formula [] using these. When W = "1" is output from the controller 1, the minimum value is set as g _{(1, 2)} and stored in the intermediate result holding memory 4. and also writes the above d _(1,2) into the intermediate distance holding memory 5.

(3) The above operations are performed according to the output of the controller 1 as shown in FIG. When all calculations are completed, the intermediate result holding memory 4
The distance between the input feature vector and the dictionary feature vector can be found using g _{(I, J)} held in .

Further, as shown in FIG. 7, when there is a return of change in the dictionary feature vector, the controller 1 in FIG. 8 may change its control sequence so that the control is performed in the order shown in FIG. . This makes it possible to perform calculations based on the above formula [], and even when there is a return of change in the dictionary feature vector, accurate matching can be performed.

Effects of the Invention According to the present invention, even if there is a return of change in the input feature vector or the dictionary feature vector, it is possible to calculate an accurate distance using the DP method. Therefore, it becomes possible to greatly improve the recognition rate of speech recognition.

[Brief explanation of the drawing]

Figures 1 to 3 are explanatory diagrams of the DP method, Figure 4 is an illustration of problems with the conventional DP method, Figures 5 to 7 are illustrations of the principle of the present invention, and Figure 8 is an illustration of the present invention. The configuration diagram of one embodiment, FIGS. 9 and 10, are explanatory diagrams of control sequences in the controller. In the figure, 1 is a controller, 2 is an input feature memory, 3 is a dictionary memory, 4 is an intermediate result holding memory,
5 is an intermediate distance holding memory, 6 is an inter-vector distance calculating section, and 7 is a recurrence formula calculating section.

Claims (1)

[Claims] 1. Two feature vector time series = = ₁ , = ₂ ... = _i
...〓 _I ,〓=〓 ₁ ,〓 ₂ ...〓 _j ...〓 In a feature vector sequence distance calculation method that evaluates the distance between _J by dynamic programming, a controller, a first feature vector holding unit, and a second A feature vector holding section, an inter-vector distance calculation means, and a recurrence formula calculation means are provided, the feature vector time series 〓, 〓 are held in the first feature vector holding section and the second feature vector holding section, and the distance between the vectors is Using distance calculation means, the distance d _(i,j) between 〓 _i and 〓 _j
By using this distance d _{(i, j)} and calculating the following recurrence formula using the recurrence formula calculation means, it is possible to calculate A distance calculation method between feature vector series, characterized in that distance calculation is performed in response to a return of a change in . Recurrence formula g _(i,j) =d _(i,j) +ming _(i,j-1) +d _(i,j-1) g _(i-1,j) +d _(i-1,j) g _{( i-1,j-1)} +2d _(i-1,j-1) g _(i-1,j+1) +2d _(i-1,j+1) 2 Two feature vector time series = = = ₁ , 〓 ₂ …〓 _i
...〓 _I ,〓=〓 ₁ ,〓 ₂ ...〓 _j ...〓 In a feature vector sequence distance calculation method that evaluates the distance between _J by dynamic programming, a controller, a first feature vector holding unit, and a second A feature vector holding section, an inter-vector distance calculation means, and a recurrence formula calculation means are provided, the feature vector time series 〓, 〓 are held in the first feature vector holding section and the second feature vector holding section, and the distance between the vectors is Using distance calculation means, the distance d _(i,j) between 〓 _i and 〓 _j
By using this distance d _{(i, j)} and calculating the following recurrence formula using the recurrence formula calculation means, it is possible to calculate A distance calculation method between feature vector series, characterized in that distance calculation is performed in response to a return of a change in . Recurrence formula g _(i,j) =d _(i,j) +ming _(i,j-1) +d _(i,j-1) g _(i-1,j) +d _(i-1,j) g _{( i-1,j-1)} +2d _(i-1,j-1) g _(i+1,j-1) +2d _(i+1,j-1)