JPS6266300A - Voice recognition 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 The present invention relates to a voice recognition method. Long come technique! 11BTsP (Binary T
The present applicant has already proposed various types of voice recognition using the ime-3pectrum pattern) method. However, in this BTSP method, since the audio is subjected to binary processing, information on energy and power representing the loudness of the audio is lost, which may result in erroneous recognition. For example, /P/ is a plosive whose consonant part rises rapidly, and /P/ has a relatively slow rise.
The difference between / is difficult to distinguish. Therefore, it may be possible to binarize the audio power information using the usual method and store it together with the binary TSP (BTSP) pattern, but in this case,
Since the calculation method for the BTSP and the power part is different, a dedicated calculation unit for determining the similarity of the power part is required, which complicates the device and loses the high-speed calculation performance of the BTSP method. Purpose The present invention was made in view of the above-mentioned circumstances.
In particular, the purpose of this invention is to provide a recognition method capable of high-accuracy and high-speed recognition while adding voice energy or power information to a recognition system using binary patterns. Structure In order to achieve the above object, the present invention provides a speech recognition method in which a recognition result is determined by extracting a feature amount of a speech, retaining it as a standard pattern, and comparing it with a speech pattern of an unknown input speech. The standard is a sound pattern that is created by binarizing the position or its vicinity that indicates the energy or power shape of the sound with a different code from the rest of the sound, and then superimposing and adding one or more patterns created using the same procedure. The system is characterized in that the unknown input speech is similarly binarized as a pattern, and the similarity is determined by superimposing it on a standard pattern, and the standard pattern with the maximum similarity is taken as the recognition result. Hereinafter, the present invention will be explained based on examples. FIG. 1 is an electrical block diagram for explaining one embodiment of the present invention. In FIG. 2, 1 is a microphone, 2 is a power detection section,
3 is a voice section detection section, 4 is a binarization section, 5 is a register, 6
1 is an addition unit, 7 is a standard pattern, 8 is a superimposition unit, 9 is a similarity judgment unit, and 10 is a recognition result output unit. First, the power of the voice input from the microphone 1 is detected, the voice interval is determined, and the voice interval is Binarize the power only for the relevant part. The power may be detected, for example, by detecting the amplitude envelope of the speech waveform, and the speech section may be detected as a portion where the power determined by the power detection section 2 is equal to or greater than a certain value. In addition, the binarization unit converts the power signal shown in FIG. 2 (A) into (B)
As in the case of the binarized information shown in FIG. 1, a method may be used in which the part where the shape of the audio power is shown is represented by "1" and the other parts are represented by "0". In this example, the magnitude of power is quantized in five stages, and it is easy to associate the waveforms of (A) with (B). When creating a standard pattern, press switch S.

[Mark] Turn to the side and say one sound, for example /P a / three times. First, the first utterance pattern is put into the register 5, superimposed and added with the second utterance pattern, and then put into the register 5 again. Then,
The third pattern and the contents of register 5 are added and registered as a standard pattern. In other words, (B) in Figure 2, (
By adding the patterns C) and (D), the standard pattern (E
) is created, and after repeating this for each voice to be registered, recognition begins. In the case of recognition, an unknown voice is input through the microphone 1, and the binarized pattern is compared with several previously created standard patterns through the same process as when creating standard patterns. In matching, the binarized pattern and one of the standard patterns are superimposed to calculate the degree of similarity. The binarized unknown voice pattern at this time is shown in Figure 2 (B).
If both have similar waveforms, by overlapping it with the standard pattern (E), the element of "1" in the pattern (B) will overlap with the element that shows a large value of (E). . Therefore, the degree of similarity may be defined as the sum of the products of elements that correspond to each other by overlapping them. In this way, the degree of similarity between all registered standard patterns and the unknown pattern is determined, and the one with the highest degree of similarity is output as the recognition result. This makes it possible to calculate similarity by adding power information to the binarized data. However, it is difficult to recognize speech based only on power information. FIG. 3 is an electrical block diagram showing another embodiment that eliminates the above drawbacks. This embodiment uses both the pattern created in the previous embodiment and the pattern created using other feature quantities, This method calculates the degree of similarity between the two patterns, and applies the degree of similarity of one to the degree of similarity of the other to determine the final degree of similarity. We chose a method using the TSP value. This is because it has been binarized together with the binary TSP power pattern and the same calculations can be performed, but of course it can be used in conjunction with other methods. After the speech section is cut out in the section 3, the power is detected in the power detection section 2, and on the other hand, the same signal is subjected to feature amount conversion in the feature amount conversion section 11.
What are the features? In this example, a spectrum is suitable. A binarization unit 4 binarizes the power and the shape of the spectrum. In this binarized pattern, later calculations are easier if the spectrum pattern and the power pattern are combined into one pattern. An example of a pattern in the binarization section 4 is as shown in FIG. 4, where F is the normal BTSP and G is the pattern corresponding to the pattern in FIG. 2(B). The similarity can be calculated using the same procedure as in the example shown in FIG. 1 to derive the results. In this case, if the similarity determination unit 9 considers that the size of the pattern has increased, there is no change in the procedure, and the result is determined based on the similarity of the sum of both patterns. This ultimately improves accuracy compared to the example of FIG. In this case, it is possible to place weight on either the power or spectrum pattern and use the other as an auxiliary means. FIG. 5 is an electrical block diagram for explaining an embodiment made from the above-mentioned viewpoint. According to this embodiment, when determining the degree of similarity, one degree of similarity satisfies a specific condition or Only when the similarity is not satisfied, the recognition result can be determined by calculating the other similarity. In this example, the third
Similar to the embodiment shown in the figure, a pattern as shown in FIG. 4 is created in which the spectrum and power are combined, and this pattern is overlapped several times and registered. During recognition, a portion of the unknown input power generated by the binarization section 4 is compared with the power pattern matching section 12 to check for similarity. If the similarity is significantly different, the determination unit 13 determines that the degree of similarity of the spectrum portion should not be calculated, and moves on to matching with the next standard pattern. If the determination unit 13 determines that the degree of similarity of spectral patterns should be calculated, the degree of similarity between the patterns will be determined in the same manner as in FIG. In this case, the degree of similarity may be calculated with or without including part of the power. Although the example here is a comparison of the power of the entire pattern, it goes without saying that this may be performed for each frame in the pattern rather than for the entire audio pattern. As is clear from the above explanation, according to the present invention, power information is added to the binarized spectral pattern, making it possible to improve the accuracy of speech recognition.

[Brief explanation of the drawing]

FIG. 1 is an electrical block diagram for explaining one embodiment of the present invention, FIG. 2 is a diagram showing a binarization pattern for explaining the operation of the present invention, and FIG. 3 is an electrical block diagram for explaining an embodiment of the present invention. An electrical block diagram for explaining another embodiment of the invention, FIG.
FIG. 5, which is a diagram showing an example of a value pattern, is an electrical block diagram showing another embodiment of the present invention. DESCRIPTION OF SYMBOLS 1... Microphone, 2... Power detection section, 3... Voice section detection section, 4... Binarization section, 5... Register, 6
... Addition section, 7. Standard pattern, 8. Overlay section, 9. Similarity determination section, 1o.. Recognition result output section, 11. Feature value conversion section, 12. ...Power pattern matching section, 13... Judgment section. Patent applicant: Ricoh Co., Ltd. Figure 1 Figure 3 Figure 4 FG

Claims (3)

[Claims] (1) In a speech recognition method that determines the recognition result by extracting the feature quantity of the speech and storing it as a standard pattern, and comparing it with the speech pattern of unknown input speech, the energy or power shape of the speech is used. A sound pattern is created by binarizing the position indicating or the vicinity thereof using a code different from the rest of the part, and one or more patterns created using the same procedure are superimposed and added to form a standard pattern. , a speech recognition method characterized in that unknown input speech is similarly binarized, superimposed on a standard pattern to determine similarity, and the standard pattern with maximum similarity is taken as the recognition result. (2) The above pattern is used together with patterns created using other features, the degree of similarity is determined between each type of pattern, and the degree of similarity of one type is applied to the degree of similarity of the other to determine the final similarity. The speech recognition method according to claim 1, wherein the recognition result is determined by determining the speech recognition method. (3) When using the above-mentioned pattern together with patterns created using other feature values to determine the degree of similarity between each type of pattern, only when one degree of similarity satisfies or does not satisfy a specific condition. The speech recognition method according to claim (1), wherein the recognition result is determined by also calculating the other similarity.