JPS5830628B2 - pattern luigi dokeisan sochi - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION The present invention relates to a device that calculates the degree of similarity between time-series bangs, such as voice bangs.

Conventionally, for this type of device, a method has been considered in which the influence of the time axis is removed by approximating the time axis fluctuation of the time-series bang by nonlinear transformation and taking the maximum coincidence between the input button and the standard button. As a method for realizing this method, a DP matching method is considered that uses dynamic programming (DP) to calculate similarity with a normalized time axis.

The DP matching method is described in known documents (Journal of the Acoustical Society of Japan Vol.
27, /I69.1971, P483-490 "Continuous word recognition based on temporal normalization of speech using dynamic programming").

The voice button is A-1, L as a time series of feature vectors.
2, °°”°°°, Ui1°°°゛°゛Revised 1 (1) B=
Tol, Shi2,...Toi,...L
j (2).

The correspondence between time axes i and j is optimally determined, the sum of the distances between Ui and J is set as the minimum, and the degree of similarity between batons is defined as the minimum value of the average of the above distances over the batans.

This minimization problem can be handled by the following DP.

initial value This is a method of calculating as follows.

In this case, equation (5) limits the allowable range for (i, j), and is hereinafter referred to as a matching window and denoted by NW according to the literature.

Similarly, if j is constant and is included in the window NW (i,
The set of j) is called the j-th stage.

As a specific calculation method for formula (8), the recurrence formula for formula (4) is as follows:
Executes the (j+1)th stage, and after execution up to the 5th stage, (
g(1, J) and (J-γ≦l≦J+γ) in equation (8) are all determined, and the degree of similarity in equation (8) can be calculated.

The above is the algorithm for calculating similarity using the DP matching method.

When this method is executed by a concrete device, the value g(i, j) of the recurrence formula may overflow, especially in the recurrence formula part where there are many repeated operations because the device is configured with finite length bits. , occurs frequently, and in the minimum value detection process of equation (4), there is a problem that an erroneous state is generated in which the overflow value is considered as the minimum value.

An object of the present invention is to provide a slam similarity calculation device that eliminates the above-mentioned problems and allows the minimum value detection operation in the recurrence formula processing of equation (4) to be executed without contradiction.

A feature of the present invention is that the recurrence formula processing section is configured with an adder circuit having a state detection function of outputting the maximum value in the event of an overflow, and the minimum value detection operation is executed without contradiction.

Hereinafter, one embodiment of the present invention will be explained in detail with reference to the drawings.

The figure is a block diagram of an embodiment of the batan similarity calculation device according to the present invention, in which 1 is a distance calculation section, 2 is a recurrence formula calculation section, and 3 is a similarity calculation section.

When the input button is i, the distance calculation section 1 calculates
It consists of a buffer that stores the input button within the matching window from li+r, a memory that stores the standard pattern If>j, and an arithmetic circuit. The distance d(i, j) is calculated according to the equation, and outputted to the recurrence formula calculation section 2.

The recurrence formula calculation unit 2 includes a memory 4 that stores the value of the recurrence formula of the (j-1)th stage, and a recurrence formula calculation unit that stores the value of the recurrence formula of the (j-1)th stage according to the distance d(i, j) and equation (4). d(i,j)+g(
i-1, j), d(i, j)+g(i, j-1)・2d(i, j)+g(i-L j-1); When an overflow occurs and a carry occurs, the output E is set to a 1'' state, whereas when there is no overflow, the output E is output to the flip-flop 6 which indicates a state of 011, and when the state signal E is II, 71, a sufficiently large value G is output. When the signal E is °0'', a switching circuit 7 outputs the output F of the adder circuit 5, and a minimum value detection circuit 8.
and a memory 9 that stores the value of the recurrence formula of the j-th stage.The output of the distance calculation section 1 is used to calculate the formula (4) in the direction of increasing i, with j kept constant. When the value of the recurrence formula overflows during sequential execution, a sufficiently large value G is output to the minimum value detection circuit 8, so that the overflowed value is not output from the minimum value detection circuit 8.

Therefore, when the calculation of the recurrence formula within the matching window is completed, the value of the recurrence formula obtained by consistently performing the minimum value detection operation of the recurrence formula in equation (4) is obtained.

Further, after the above operation is completed, the second recurrence formula calculation unit calculates the (j−
1') Update the contents of stage j with the contents of stage j, and update the contents of stage j
+1) Prepare the calculation of the recurrence formula for the stage.

The similarity calculation section 30 is composed of a division circuit and a minimum value detection circuit, and normalizes the contents of the memory 9 that stores the results of the recurrence formula of the recurrence formula calculation section 2 according to equation (8), and calculates the minimum value. Output as similarity.

Each of the calculation units described above is operated sequentially from j=1 to j=J, and the minimum value detection operation in the recurrence formula process is executed without contradiction.
It is possible to realize a device that calculates the similarity between two batons.

As explained above, when calculating time-normalized similarity using the DP matching method, the present invention can execute the minimum value detection operation of the recurrence formula part without contradiction, and the baton similarity calculation device using this method can perform the minimum value detection operation without contradiction. This will have extremely effective effects in achieving this goal.

[Brief explanation of drawings]

The figure is a configuration diagram showing an embodiment of a batan similarity calculation device according to the present invention. 1 is a distance calculation unit that calculates the distance between vectors, 2 is a recurrence formula calculation unit, 3 is a similarity calculation unit, and 4 is a first storage circuit that stores the value of the recurrence formula of the (j-1)th stage. , 5 is an adder circuit, 6 is a flip-flop that detects an overflow state, 7
8 is a switching circuit that outputs two data by switching, 8 is a minimum detection circuit, and 9 is a second storage circuit that stores the value of the recurrence formula of the j-th stage.

Claims (1)

[Claims] 1. A distance calculating section that calculates the distance between a memory circuit that stores a certain baton A as a vector series over a finite interval and a vector at a certain point in the vector series of another baton B over the finite interval; a first storage circuit that stores past recurrence formula values, and an addition circuit that has a function of adding the recurrence formula value to the distance and outputting a status signal indicating an overflow state each time an addition process is performed. , in response to this state signal, a switching circuit that outputs a sufficiently large value in the case of an overflow event, and a switching circuit that outputs the output of the adder circuit as the addition result in other cases; It consists of a minimum value detection circuit that selects and outputs it as a new recurrence formula value, and a second storage circuit that stores this recurrence formula value within the finite interval. The contents of the first storage circuit are added in accordance with the conditions specified in a predetermined order, and the smallest of the multiple addition results is used as a new recurrence formula value and sequentially stored in the second storage circuit within the finite interval. It consists of a recursion formula calculation unit, which has a function of outputting data over a period of time, and updating the contents of the first storage circuit with the contents of the second storage circuit according to the state after the operation is completed, a division circuit, and a minimum value detection circuit, and a similarity calculation unit that divides the contents of the second storage circuit of the recurrence formula calculation unit by the normalization coefficient over the finite interval and outputs the minimum value within the interval:
A slam similarity calculation device, characterized in that the calculation units are operated according to the number of times set according to the length of the pattern B to calculate the similarity between the batons A and B expressed by a sequence of vectors.