JPS59124400A - Voice pattern matching system - 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 [°Technical Field of the Invention] The present invention relates to an improvement in a speech pattern matching method for a speech recognition device.

[Prior art and problems]

Conventional speech recognition devices convert input speech into patterns...
After converting into an input cover pattern for matching, the input cover pattern is pattern matched with a plurality of pre-stored standard patterns to obtain the most similar standard pattern as the recognition result. In this pattern matching, since the utterance time of word sounds varies considerably, it is necessary to correct this time variation and compare the incoming cover pattern and the standard pattern in a state in which they are most similar. Such time normalized pattern matching is typified by DP (dynamic programming) matching;
Matching requires a large amount of calculation and requires large hardware. In addition, route-limited matching, which is a modification of linear matching, predicts several matching routes, matches them, and then adopts the one that is the most similar.Although the amount of calculation is small, the matching route of the input cover pattern and standard pattern is the expected route. If there is a large difference, time normalization cannot be performed, and if the number of predicted routes is too large, the amount of calculation will increase.

[Purpose of the invention]

The present invention is based on the above consideration, and an object of the present invention is to provide a voice pattern matching method that requires a small amount of calculation and can be expected to produce results close to DP matching.

[Structure of the invention]

Therefore, in the voice pattern matching method of the present invention, the voice pattern matching method of the voice recognition device performs linear matching only on each side of the matching route predicted in advance in a net shape, and performs matching from the start end to the end of each side. This method is characterized in that the results of matching edges are summed for each route.

[Embodiments of the invention]

Hereinafter, the present invention will be explained with reference to the drawings.

Fig. 1 is a diagram showing an example of the mesh matching route in the present invention, Fig. 2 is a flowchart for determining the distance G by conventional route-limited matching, and Fig. 3 is a diagram showing frames at matching points on each side of the mesh. This is a flowchart according to the present invention in which the distance g of each route is determined by calculating the cumulative difference d between the two routes, and then calculating the distance G of each route.

In FIG. 1, A, B, C, . . . L indicate sides. Matching calculates the difference d between the input parameter registration parameters at the point on the route, and
The distance g of the route is determined by accumulating the distance g. The smallest distance among the distances g of each route is defined as the distance G between the input parameter and the registered parameter. Input parameters are determined as follows. The input audio is passed through a filter group consisting of, for example, 16 filters, the output of each filter is sampled every 10 m5, and the analog sampling signal is digitized. A collection of only a few digital outputs of 16 filters at a winter filtering point is called a frame. The input parameters consist of a collection of frames. The registered parameters have the same configuration as the input parameters. The difference d between the input parameter and the registered parameter at each point on the route is determined as follows. The difference d is determined by subtracting the registered parameter frame defined by the point on the route and the input parameter frame for each channel, and summing the differences for each channel.

Note that a channel corresponds to a filter. In FIG. 1, there are 1] switching routes. Each route consists of the following edges.

■ A-D- ■ 33- i ■ e-G-L ■ A-E-H- ■ A-E-i ■ A-E-J-L ■ B-H- ■ B-J-L ■ C-F-H-K [F] C-F-i G) C-F -J -L In addition, in No. 11d, the circled points indicate the edges of the sides, and the dots indicate the matching points. There is.

FIG. 2 is a flowchart for determining the distance G using conventional route-limited matching. First, a matching route is determined from input parameters and registered parameters. For example, since the number of input parameter frames is different when the same word is uttered for a short time and when it is uttered for a long time, it is necessary to determine a matching route by referring to the input parameters and registered parameters. Winter route■、■・
..., calculate the difference d at the matching point along ■, and calculate the cumulative g
is calculated from gl to gll for each route. gl is the result of calculating and accumulating the differences d at all matching points on the route ■, and gll is the calculating and accumulating the differences d at all matching points on the route @.

The distances g1 to g11 of each route are compared, and the minimum distance between the input parameter and the registered parameter is set as G. If you set the matching route as shown in Figure 1,
In order to obtain the distance G using conventional route-limited matching, it is necessary to calculate d for 451 locations.

FIG. 2 is a flowchart for calculating the distance G between an input parameter and a registered parameter according to the present invention. First, a matching route and each side are determined using input parameters and registered parameters. .

For each side A, B, ..., L, find the difference d at all the moat points that exist on the side, accumulate these differences d, and calculate the difference d for each side A, B, C, ...L. Find the cumulative gA or gL. Distance g1゜g2 for each route ■, ■, ..., ■.・
...Calculate 9g11 by combining the appropriate items from gmu to gL. The distances g1 to gll of each route are compared, and the smallest one is set as the distance G between the input parameter and the registered parameter. When the matching route and sides are set as shown in FIG. 1, in order to obtain the distance G according to the present invention, d only needs to be calculated at 163 locations, which is 1/2.7 of the conventional calculation amount.

〔Effect of the invention〕

As is clear from the above description, according to the present invention, even if the number of matching routes increases, the amount of calculation is much smaller than that of route-limited matching, and results similar to DP matching can be expected.

[Brief explanation of drawings]

Fig. 1 is a diagram showing an example of the mesh matching route 01 according to the present invention, Fig. 2 is a flowchart for determining distance by conventional route-limited matching, and Fig. 3 is a diagram showing the distance between frames at matching points on each side of the mesh. 2 is a flowchart according to the present invention in which the distance g of each route is determined by calculating the cumulative difference d between the two routes, and then calculating the distance G. A, B, C,...L,... Side Patent Applicant Fujitsu Ltd. Representative Patent Attorney Kyotani 4 Part Ya1 Figure': X3 Figure

Claims (1)

[Claims] In the speech pattern matching method of the speech recognition device, linear matching is performed only on each side of the matching route predicted in advance in a net shape, and the matching from the start end to the end is calculated by summing the matching results of each side for each route. A voice pattern matching method characterized by the following.