JPH0419700A - Method for matching voice pattern - Google Patents

Abstract

PURPOSE:To execute correct collation by converting a specific part into a determined length, comparing the pattern of a part removed from the vicinity of an end part up to the end and its residual pattern with reference patterns and defining a result having higher similarity as a similar pattern. CONSTITUTION:The pattern of an obtained feature variable is expanded/contracted to the determined length, the expanded/contracted pattern is stored and registered in a reference pattern storing memory 6, and at the time of ending the registration, a switch 5 is turned to the recognition side to recognize the pattern. When there is a part whose power is lower than a fixed value, whether a frequency components are concentrated into a high range or not and whether the range is the head or the end of the pattern are checked respectively by comparing parts 13, 15, the whole length of the pattern is expanded/contracted by an expanding/contracting part 16 and the processed data are stored in a memory 17. A collating part 7 finds out the similarity or error of the obtained pattern with/from respective reference patterns and a maximum similarity detecting part 8 detects the reference pattern representing the maximum similarity and outputs a code or the like expressing the detected reference pattern as a recognized result. Thus, patterns can be correctly collated without expanding/contracting them at the time of collation.

Description

DETAILED DESCRIPTION OF THE INVENTION The present invention relates to a speech pattern matching method, and more particularly to a pattern matching method in speech recognition.

Bumien 4 Most current speech recognition devices use a pattern matching method, which compares a pre-registered standard pattern with an unknown input speech pattern to find the most similar standard pattern. This outputs categories as recognition results.

FIG. 3 is a diagram for explaining an example of a conventional voice pattern matching method. In the figure, 1 is a microphone, 2 is a microphone amplifier, 3 is a feature converter, 4 is an A/D converter, and 5 is a switch. Switch, 6 is standard pattern storage section, 7 is collation section, 8
9 is a maximum similarity detection unit, and 9 is a recognition result output unit. First, the switch 5 is set to the standard pattern registration side (a side), and a voice is input manually from the microphone 1. The voice converted into an electrical signal by the microphone 1 is amplified by the microphone amplifier 2, and then subjected to feature conversion by the feature converter 3. There are several known spectralizations as the feature amounts used. It is converted into discrete quantities and stored in the standard pattern storage section 6 as a standard pattern. At the time of recognition, the switch 5 is moved to the verification side (side b). In the same way as when registering, a voice pattern is created, compared with all standard patterns registered in advance, and the pattern with the highest similarity is determined as the recognition result.

Details of such a recognition method and feature amounts are described in, for example, "Speech Recognition J" by Niimi, and are well known, so a detailed explanation will be omitted here.

Among these, there is the issue of how to deal with pattern variations when matching patterns.

In particular, this variation is largely temporal, and is affected by the speed of speech, etc. There are two ways to counter this. One is non-linear matching, represented by DP matching, which dynamically matches two patterns to maximize the similarity while looking at the similarities between the two patterns. , the other is to insert data evenly over time without checking for similarity.

After matching by thinning, the two are compared for linear matching. Although the former requires a large amount of calculation, it has good accuracy, and the latter has the advantage of having a very small amount of calculation. In particular, in the latter case, by setting all patterns to a constant length, once the length of the input voice pattern is adjusted, there is no need to expand or contract the pattern during verification. This method is quite effective when the audio pattern is complete and there are no omissions or additions. If there are omissions or additions to the information, the matching accuracy will be extremely poor.

FIG. 4 is a diagram showing the temporal change in the energy of the voice.If you explain according to this diagram, as shown in the figure, the same r 5
When there is a sound pattern called taff J, when comparing the normal ones by linearly expanding and contracting them, the error between the two can be reduced as shown in (a), but (
As shown in b), if voice section detection fails and the /f/ in one pattern is missing, r sta J, the same pattern may have different sounds associated with each other near the end of the voice. As a result, the difference between the two patterns becomes significantly large.

For consonants produced with low energy, such as /f/ in rstaffJ given here as an example, detection of the vocal interval often fails, and the above problem occurs very often. Some pattern matching methods using nonlinear expansion/contraction make end points free, and can perform accurate matching even though /f/ is missing. However, as mentioned above, this method using nonlinear expansion and contraction still requires a large amount of calculation.

In addition, as one of the countermeasures, mark the standard patterns of unstable audio such as omissions, etc., and if there are unstable parts of the input audio, check the unstable standard patterns. If there are no unstable parts in the input audio with the parts left attached, there is a method that removes the unstable parts from all standard patterns and compares them. however,
Since this method changes the standard pattern depending on the input pattern, it has the disadvantage that the standard pattern must be corrected each time it is compared.

Purpose The present invention was made in view of the above-mentioned circumstances.
In particular, even if the detection of the voice section is not successful,
This was done with the aim of enabling accurate matching using a linear expansion/contraction method that requires less calculation.

Dark - 11 precepts.

In order to achieve the above object, the present invention provides a voice pattern matching method in which feature quantities are extracted from a voice signal and matched as a feature pattern with a constant length of time. When a specific part is found in which the energy of the voice is lower than that of vowels and the frequency spectrum components are concentrated in the high range, the entire pattern is converted to a predetermined length and the corresponding Converting the remaining pattern after removing the portion from near the end of the specific portion to the tip, or the portion from near the end of the specific portion to the end, to a predetermined length and retaining both; The feature is that both patterns are compared with a standard pattern, and the result in the direction of high similarity is defined as the similarity between the patterns. Hereinafter, the present invention will be explained based on examples.

FIG. 1 is a flowchart for explaining one embodiment of the present invention, and FIG. 2 is a block diagram for realizing the present invention shown in FIG. is a power calculation section, and 13 is a comparison section.

14 is a high frequency spectrum calculation section, 15 is a comparison section, 16 is an expansion/contraction section, 17 is a memory, and 18. 19 is a threshold value.
This method was developed by focusing on the fact that frequency components are concentrated on the higher side, and in particular, in the audio pattern matching method, which extracts feature quantities from audio signals and uses them as feature patterns to match them while keeping the length of time constant. When a partial portion is found at the beginning or end of the input unknown voice where the energy of the voice is lower than that of vowels and the frequency spectrum components are concentrated in the high range, the overall pattern is determined. At the same time, the remaining pattern after removing the part from near the end of the explanatory part to the tip, or the part from near the end of the explanatory part to the end, is converted to the specified length. Both are converted and retained, both are compared with a standard pattern, and the result with higher similarity is defined as the similarity between patterns.

First, an explanation will be given based on the flowchart shown in FIG. 1. First, in the voice registration flow (2), the entire voice input is set to a constant length and registered as a standard pattern. Next, in the speech recognition flow (■), the input speech is converted into a characteristic pattern using the same procedure as the standard pattern, and specific parts (in other words, the voice has relatively low energy and frequency Check whether there is a part where the components are concentrated in the high range.

Whether this voice energy is small or not can be checked by checking whether the voice energy drops below a certain value at the beginning or end.This constant value should be set at about 115 based on the energy value when the vowel is input. . Also, there are various ways to check whether frequencies are concentrated in the high range, but for example, divide the analysis frequency band into two and find that the high range contains components that are several times as large as the low range. This can be determined by drawing a fitted straight line in the direction of the frequency axis of the speckle distribution and determining if the slope is negative.

If there is no such part at the beginning or end of the voice where the energy is small and the frequency components are concentrated in the high range, that is,
If there is no specific part, the registration of this voice is finished, and if there is,
Depending on whether it is the beginning or the end, that is, /f/
First, find out where in the audio the sounds that are likely to be dropped, such as , are attached, and then create a pattern in which these sounds are dropped. In other words, it is determined that there is a sound that is likely to be lost at the beginning of the audio, and the part with low energy and frequency components concentrated in the high range is removed from the part to the end, and the remaining part is set to a certain length. It is retained in the buffer memory in the same way as the input cover pattern.

In this way, all registered standard patterns are matched. If the standard pattern is likely to have consonants at the beginning or end, two introductory turns can be created, so even if a part of the standard pattern is missing, it can be matched, improving recognition accuracy. can.

FIG. 2 is a block diagram for realizing the present invention as described above. In this case, the conventional method shown in FIG. It's the same with technology. First, registration will be explained. Turn the switch 5 to the registration side (a side), expand or contract the obtained feature pattern (feature pattern) to a predetermined length, store it in the standard pattern storage memory 6, and register it. When the voice to be registered has been registered in the standard pattern storage section, the switch 5 is turned to the recognition side (side b) to be recognized. For recognition, the audio signal is input into the power calculation unit 12 for power calculation after or before being converted into a characteristic pattern in the same manner as registration. Here, the comparison units 13 and 15 check whether there is a part where the power is lower than a certain value, and if so, whether the frequency components are concentrated in the high range, and whether the position is at the beginning or the end. Then, the length of the entire pattern is expanded or contracted to a certain length by the expansion/contraction unit 16 and stored in the memory 17.
If there is a part at the beginning or end of the audio where energy is low and frequency components are concentrated in the high range, remove that part and reshape the pattern using the expansion/contraction section as shown in the flowchart in Figure 1. After making it a certain length, it is also stored in the memory 17. Matching calculates the similarity between the previously stored pattern in memory and the standard pattern. If two patterns are stored in memory, the similarity is calculated twice for one standard pattern, and the one with the higher similarity is adopted as the similarity between the input and standard pattern. In this case, there are two telescopic parts, but they only need to have the same function, and they can be the same.

The matching method of the matching unit 7 is not particularly limited, and it may be a method of finding differences between patterns based on the distance between urban areas, or a method of calculating similarity based on an inner product between vectors. The similarity or error between the unknown input pattern and each standard pattern is determined in advance. Maximum similarity detection unit 8
Then, the standard pattern that shows the greatest similarity is found, and its name or symbol representing it is output as the recognition result.

According to this method, a voice pattern in which a part of the voice is missing is registered in advance as a fixed length, so if a consonant at the beginning or end of the input voice is missing, it can be compared with this pattern, so it can be expanded or contracted. The amount of calculation is small compared to , and recognition accuracy can be improved.

Effects As is clear from the above explanation, according to the present invention, even if voice section detection is unsuccessful, correct verification can be performed without expansion or contraction during verification.

[Brief explanation of the drawing]

FIG. 1 is a flowchart for explaining one embodiment of the present invention, FIG. 2 is a diagram showing an example of a block diagram used to implement the present invention, and FIG. 3 is a general pattern matching block diagram. 4 are diagrams for explaining the correspondence when a weak consonant is detected and the correspondence when a weak consonant cannot be detected. 1...Microphone, 2...Mic amplifier, 3.
... Feature conversion section, 4... A/D conversion section, 5... Changeover switch, 6... Standard pattern storage section, 7... Matching section, 8... Maximum similarity detection section, 9 . . . Recognition result output section, 11 . . . Expansion/contraction section, 12 . . Power calculation section, 1
3... Comparison section, 14... High frequency spectrum calculation section, 1
5... Comparison section, 16... Expansion/contraction section, 17... Memory, 18, 19... Threshold section. Figure Figure

Claims (1)

[Claims] 1. In a voice pattern matching method that extracts feature quantities from a voice signal and uses them as a feature pattern to match a constant length of time, there is a possibility that the voice has more energy than a vowel at the beginning or end of the input unknown voice. low and
When a specific part where frequency spectrum components are concentrated in the high range is found, the entire pattern is converted to a predetermined length, and the part from near the end of the specific part to the tip, or , the remaining pattern after removing the part from near the end of the specific part to the end is converted to a predetermined length and both are retained, and both are compared with the standard pattern to find the one with high similarity. A voice pattern matching method characterized in that a direction result is defined as a similarity between patterns.