JPH06105399B2 - Voice recognition system - Google Patents

Description

TECHNICAL FIELD The present invention relates to a speech recognition system, and in particular, pattern matching is performed while shifting an input pattern end point in the time direction at each time, and a distance from a standard pattern is calculated, and at each time. The present invention relates to a voice recognition method that evaluates the distance between the two and determines the recognition result.

2. Description of the Related Art Conventionally, the evaluation distance threshold used in the recognition result determination unit of this type of voice recognition method is a fixed value. Thus, in the recognition result determination unit of the conventional voice recognition method, the threshold value suitable for the speaker cannot be set because the evaluation distance threshold value is a fixed value. Therefore, there is a drawback that incorrect recognition may occur and a correct recognition result cannot be obtained.

OBJECT OF THE INVENTION It is an object of the present invention to provide a voice recognition method capable of obtaining an accurate recognition result by setting an evaluation distance threshold suitable for a speaker.

According to the present invention, a method of determining a voice recognition result by providing an evaluation distance threshold value and a duration threshold value for the evaluation distance that is time-series data, and from a frame determined from the start of input , By calculating the average value of the evaluation distance to the other defined frame and adaptively calculating the value obtained by subtracting the coefficient value defined from the average value of the evaluation distance as the evaluation distance threshold, It is possible to obtain a voice recognition method capable of setting a suitable evaluation distance threshold.

Example Next, an example of the present invention will be described with reference to the drawings.

FIG. 1 is a block diagram of an embodiment of the present invention. The embodiment of the present invention includes a voice analysis unit 1, a standard pattern unit 2, a pattern matching unit 3, and a recognition result determination unit 4.

The input voice 10 is input to the voice analysis unit 1, analyzed into characteristic patterns, and input to the pattern matching unit 3 as an input pattern 11. The pattern matching unit 3 reads the standard pattern 12 from the standard pattern unit 2, performs pattern matching with the input pattern 11 for each frame, calculates the inter-pattern distance, and outputs it as the evaluation distance 13 to the recognition result determination unit 4.

The pattern matching is performed after shifting the end point of the input pattern for each frame by one frame behind the end point of the previous frame. The pattern matching method is not particularly limited, but in this embodiment, an example using the start end free DP matching will be described. Fig. 2 shows free start
FIG. 4 is a diagram showing a DP path allowable area of pattern matching when DP matching is used, and the DP path shown in FIG. 3 is used. 2 and 3, the horizontal axis is the frame of the input pattern, the vertical axis is the frame of the standard pattern, i is the frame of the input pattern, and j is the frame of the standard pattern.

The start-end free DP matching method is a method for finding a DP matching path that gives a minimum inter-pattern distance to an end point without limiting the end point. In the example shown in FIG. When I is the end point,
It indicates that one of the frames from IS1 to IS2 is the starting point. Further, shifting the end point of the input pattern for each frame to perform pattern matching means that the allowable area of the DP path shown in FIG. 2 has slid in the direction of the horizontal axis i.

Next, the recognition result determination unit 4 will be described with reference to FIG. Since the evaluation distance 13 is calculated for each input frame,
It becomes time-series data for the input frame as shown in FIG. The horizontal axis of FIG. 4 is the input frame, and the vertical axis is the evaluation distance. 4 (a), (b), (c), and (d) are graphs of the evaluation distances of the speakers A, B, C, and D, respectively.

The recognition result determination unit 4 defines an evaluation distance average value calculation start frame 19, an evaluation distance average value calculation end frame 20 and an evaluation distance threshold calculation coefficient 21 as shown in FIG. Calculate the threshold. In this embodiment, the evaluation distance average value calculation start frame is F = 20, the evaluation distance average value calculation end frame is F = 25, and the evaluation distance threshold calculation coefficient is 100.

DIST (F); Evaluation distance of frame F Fstr; Start frame of calculation of evaluation distance average value Fend; End frame of calculation of evaluation distance average value Adaptive evaluation distance threshold = evaluation distance average value−evaluation distance threshold calculation coefficient The distance threshold value is calculated and becomes an optimum value according to the speaker, as indicated by 22,23,24,25 in FIGS. 4 (a), (b), (c) and (d).

The continuity threshold is set as the threshold for continuity, and the evaluation distance
It is determined whether the time of the section in which 13 is continuously below the adaptive evaluation distance thresholds 22, 23, 24, 25 is equal to or longer than the duration threshold, and the section which is equal to or longer than the duration threshold is set as a valley section. In this embodiment, the evaluation distance at the lowest point of the valley section is used as the judgment score. However, what is used as the determination score is not particularly limited. In this way, the judgment score with all the standard patterns is obtained, and the standard pattern with the smallest judgment score is selected and output as the recognition result 14.

As is clear from the examples shown in FIGS. 4 (a), (b), (c), and (d), the evaluation distance moves differently depending on the speaker, and when the evaluation distance threshold is a fixed value, all It becomes difficult to find a value suitable for the speaker. For example, speaker A
When the speaker D is compared with the speaker D, the evaluation distance threshold suitable for the speaker D is about 650, but the value is too large for the speaker A, resulting in erroneous recognition.

On the other hand, the adaptive evaluation distance thresholds 22, 23, shown in the above-mentioned embodiment,
By using 24 and 25,
As is clear from (c) and (d), the evaluation distance threshold suitable for each speaker is set, and the incorrect recognition generated in the conventional method can be removed to obtain a correct recognition result.

EFFECTS OF THE INVENTION As described above, according to the present invention, an accurate recognition result is obtained by using the adaptive evaluation distance threshold for determining the evaluation distance, instead of using the fixed evaluation distance threshold for determining the evaluation distance. Therefore, there is an effect that false recognition can be reduced.

[Brief description of drawings]

FIG. 1 is a block diagram of an embodiment of the present invention, FIG. 2 is a diagram showing a permissible region of starting end free DP matching, FIG. 3 is a diagram showing a DP path of starting end free DP matching, and FIG.
(D) is a graph of the evaluation distance of each of the speakers A to D. Description of main part code 1 …… Speech analysis section 2 …… Standard pattern section 3 …… Pattern matching section 4 …… Recognition result judging section

Claims (1)

[Claims] 1. A standard pattern portion for storing a characteristic pattern of a word or a phoneme to be recognized as a standard pattern, and an input pattern which is a characteristic pattern of the input speech by analyzing the input speech at a predetermined frame period. The pattern matching unit includes a voice analysis unit for outputting, a pattern matching unit for matching a standard pattern and an input pattern to calculate an evaluation distance, and a recognition result determination unit for determining a recognition result based on the evaluation distance. Then, the end point of the input pattern is shifted by one frame after each frame of the input pattern to perform pattern matching, the inter-pattern distance is calculated for each frame as the evaluation distance, and the evaluation distance is output to the recognition result determination unit as time series data. However, in the recognition result determination unit, the evaluation distance threshold value and the duration threshold value for the evaluation distance that is time series data. Is a speech recognition method for determining a recognition result, wherein an average value of evaluation distances from a frame frame determined from the start of input to another predetermined frame frame is determined, and the average value of the evaluation distances is determined. A voice recognition method, wherein a value obtained by subtracting the obtained coefficient value is adaptively calculated as the evaluation distance threshold.