JP3004749B2 - Standard pattern registration method - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

The present invention relates to a reference pattern registered how, and more particularly it relates to a pattern matching speech recognition.

[0002]

BACKGROUND ART Current speech recognition apparatus includes a mainstream utilizes a pattern matching how, and the standard pattern registered in advance, by comparing the unknown speech pattern input, the most similar reference pattern The category is output as a recognition result.

FIG. 7 is a diagram for explaining an example of a conventional voice pattern matching method, in which 1 is a microphone,
2 is a microphone amplifier, 3 is a feature converter, 4 is an A / D converter, 5 is a changeover switch, 6 is a standard pattern storage, 7
Is a matching unit, 8 is a maximum similarity detection unit, 9 is a recognition result output unit, and first, the changeover switch 5 is set to the standard pattern registration side (a
Side), and input sound from the microphone 1. The sound converted into an electric signal by the microphone 1 is amplified by the microphone amplifier 2 and is subjected to characteristic conversion by the characteristic conversion unit 3, and some of the characteristic amounts to be used are known, such as a spectrum. It is converted into a discrete amount and stored in the standard pattern storage unit 6 as a standard pattern. At the time of recognition, the switch is moved to the collation side (b side). As in the case of registration, a voice pattern is created, collated with all standard patterns registered in advance, a pattern with the highest similarity is found, and that is used as a recognition result.

[0004] Details of such a recognition method and feature amounts are described in, for example, Niimi's "Speech Recognition".
Since it is well known, further description is omitted here.
Among them, there is a problem of how to cope with the fluctuation of the pattern at the time of pattern matching. In particular, this fluctuation is largely temporal, and is influenced by the speed of utterance and the like. There are two countermeasures, one is non-linear matching represented by DP matching, which looks at the similarity of the two patterns to be matched and dynamically associates the two patterns to maximize the similarity. And the other is to insert data evenly over time without checking for similarities,
Linear matching is performed in which the two are compared after being matched by thinning. The former has a merit that the calculation amount is large, but the accuracy is good, and the latter has a merit that the calculation amount is very small. In particular, in the latter case, by setting all patterns to have a fixed length, if the length of the input voice pattern is adjusted once, there is a characteristic that it is not necessary to expand or contract the pattern at the time of comparison. This method is very effective when the voice pattern is complete and there are no missing or added,
However, the voice is non-linearly expanded and contracted, and it is made up by linear expansion and contraction. Therefore, if the voice pattern is missing or added, the matching accuracy becomes extremely poor.

[0005] FIG. 8 is a diagram showing the time change of the energy of the voice. Referring to FIG. 8, as shown in FIG.
When there is the same voice pattern "staff", when normal ones are linearly expanded and contracted and compared, the error between both can be reduced as shown in (a).
As shown in (b), if voice section detection fails and “/ sta” of one pattern is missing,
Even though the patterns are the same, different sounds correspond to each other near the end of the voice, and the difference between the two patterns becomes significantly large. Consonants with low energy uttered like / f / of "staff" cited here as an example often fail to detect voice segments well, and the above problem occurs very often. In some pattern matching methods using non-linear expansion / contraction, end points are made free, and high-precision matching can be performed while / f / is missing. However, the method using the non-linear expansion and contraction still has a large amount of calculation as described above.

[0006] As one of the countermeasures, a mark is put on a standard pattern of an unstable voice such as a dropout or the like. When there is no unstable part in the input voice while the unstable part is added, there is a method in which the unstable part is removed from all the standard patterns for matching. However, in this method, since the standard pattern is changed depending on the input pattern, there is a drawback that the standard pattern must be corrected each time of collation.

[0007]

SUMMARY OF THE INVENTION The present invention has been made in view of the above-mentioned drawbacks of the prior art, and in particular, even when voice section detection is not successful, correct collation can be performed by a linear expansion / contraction method with a small amount of calculation. It is made for the purpose of being able to do it.

[0008]

[Configuration] The present invention, in order to achieve the above object, in the speech pattern matching how to match forms with feature pattern extraction a feature from the audio signal, than the voice energy of the vowel at the beginning or end of the speech Is low, and the part where the spectral components are concentrated in the high range is found,
The entire pattern is converted to a predetermined length, and the remaining pattern is removed to a predetermined length except for a portion where the energy is lower than that of the vowel and the spectral components are concentrated in the high frequency range. It is characterized by being converted into a standard pattern. Hereinafter, a description will be given based on an example of the present invention.

FIG. 1 is a flow chart for explaining an embodiment of the present invention, and FIG. 2 is a block diagram for realizing the flow chart shown in FIG. 1. In FIG. Calculation unit, 12 is a comparison unit, 13 is a high-band spectrum calculation unit, 14 is a comparison unit, 15 is a stretching unit, 16 is a memory, 17, 18 are thresholds, 19 is a pattern shaping unit,
Reference numeral 20 denotes an expansion / contraction portion, and other portions that operate in the same manner as in the prior art shown in FIG. 6 are denoted by the same reference numerals as in FIG. In the present invention, consonant energies that are difficult to detect a voice section are relatively small, and attention is paid to the fact that frequency components are easily concentrated in a high frequency range. in the speech pattern matching how to match No, the beginning of the voice,
Or, at the end, find a part where the energy of the voice is lower than that of the vowel and the spectral components are concentrated in the high frequency range, convert the entire pattern to a predetermined length, and convert the energy to that of the vowel. The remaining pattern, which is lower than that and where the spectral components are concentrated in the high frequency band, is removed and converted into a predetermined length to become a standard pattern.

[0010] Figure 1 is an illustration in flow chart how the present invention will now be described with reference to this flowchart. First, the registration of voice will be described. As usual, the entire voice is set to a fixed length, registered as a standard pattern, and then a specific part (that is, low energy and low spectral component) is added at the beginning or end of the voice. Is concentrated in the high frequency range).
The energy of the voice is determined depending on whether it falls below a certain value, and this constant value is determined to be about one fifth of the energy value when the vowel is input. In addition, various methods of examining the concentration of the frequency can be considered. For example, the analysis frequency band is divided into two, and when there are components several times higher in the high frequency than in the low frequency, the frequency distribution of the spectrum distribution is considered. It is possible to determine whether the inclination is negative by drawing a fit straight line in the axial direction. If there is no part where the energy is small and the frequency components are concentrated in the high frequency band at the beginning or end of such a sound, the registration of this sound is finished, and in some cases, depending on whether it is the beginning or end, It is checked where a sound that is easily lost such as / f / is attached to the voice. Next, a pattern in which this is missing is also created in advance. In other words, when a sound that is likely to be missing is added to the beginning of the voice, the remaining energy is removed from the portion where the energy is small and the frequency components are concentrated in the high frequency band to the end, and registered with a fixed length.

On the other hand, at the time of recognition, the input voice is made to have a predetermined length and is collated with all the registered standard patterns. If the beginning and end consonants of the input speech are dropped, it can be compared with the previously registered missing speech pattern, so that the recognition accuracy can be improved.

FIG. 2 is a block diagram for realizing the present invention, and is the same as the prior art shown in FIG. 6 up to the point where the voice from the microphone 1 is converted into a discrete amount by feature conversion. First, registration will be described. The switch 5 is set to the registration side (a side), and is supplied to the power calculator 11 for calculating the power of the audio signal. Here, it is checked whether there is a portion where the power is lower than a certain value, and if there is, whether or not the frequency component is concentrated in a high frequency range, and whether the position is the beginning or end. In the next expansion / contraction section, the entire length of the pattern is expanded / contracted to a certain length and registered in the standard pattern storage section (memory) 16. If there is a portion where the power is lower than a certain value and the frequency components are concentrated in a high frequency range, the pattern shaping unit removes the leading end or the trailing end as shown in the flowchart of FIG.
After making the pattern shaped by the elastic part again a certain length,
It is registered in the standard pattern storage. When the voice to be registered has been registered in the standard pattern storage unit, the switch 5 is moved to the recognition side (b side) to perform recognition. In the recognition, after the feature pattern is converted into a characteristic pattern in the same manner as the registration, the data is collated with a fixed length in the expansion and contraction unit 10. The expansion / contraction section may have the same function as that at the time of registration, and is separately shown in the figure, but may be the same. The collation unit 7 does not particularly limit the collation method, and may be a method of calculating a difference between patterns based on an urban area distance, or a method of calculating similarity by an inner product between vectors. The similarity or error between the unknown input pattern and each of the standard patterns is determined in advance. The maximum similarity detection unit 8 finds a standard pattern showing the greatest similarity, and outputs its name or a symbol representing the same as a recognition result. According to this method, a voice pattern in which a part of the voice is missing is also registered in advance with a fixed length, so that when the beginning and end consonants of the input voice are dropped, it can be compared with this pattern so that it expands and contracts. The amount of calculation is smaller than that of, and the accuracy of recognition can be improved.

FIG. 3 is a diagram for performing the operation of FIG. 1 by hardware using a microcomputer, in which a standard pattern of voice normalized to several lengths is previously loaded in a register 28. It will be described as. An unknown voice to be recognized is input from the microphone 1, amplified by the microphone amplifier 2, and then analyzed by the bandpass filter bank 3 into several frequencies (for example, 15 frequencies). The result is quantized by the A / D converter 4 to about 12 bits, and voice data is detected 21 using the data. A part related to the detected voice is stored in the register 22. How to detect voice sections is by Niimi, "Speech Recognition" (Kyoritsu Shuppan) P68
Is shown in Subsequent operations including the detection of the voice section are performed by software of the microcomputer, and therefore, there are many cases where no hardware other than the register is provided. Several kinds of frame lengths to be normalized are registered in the register 23. The frame length of the input voice generated when the voice section is detected is sent to the comparator 24 and compared with the contents of the register 23. One or two closest frame lengths are selected from the register 23, and the register 22 and the comparator 27 are selected.
As a frame length signal. The register 22 sets the frame length determined by the copy in the register based on the transmitted signal. FIG. 4 shows the operation of the register at the time of copying.

In FIG. 4, if the unknown voice length is 1
_{At 1} , it is assumed that this is stored in the register 22. If this must be expanded to 1 ₁ +2 frames (a), the input frame length is increased by the number of inserted frames + 1.
It is easy to determine the insertion part by dividing by. in this case,
Then the integer 1 _1/3 is assumed to be 1 ₁ '. First,
The 1 ₁ th data to the ₁ 1 +2 th, Yuku repeat the copy to the ₁ 1 +1 th ₁ 1 -1-th data. However,
21 ₁ 'data of 21 _1' copy to both +1 and 21 ₁ '+ 2-th frame. Then, ₁ 1 '-1st ₁ 1'
To the repeated copy th, it operates in was copied 1 ₁ '1 _1' to +1 th ends. Next, a case of 1 ₁ to ₁ 1 -2 frame (b). Just reverse to start from the younger of the number and, first of all, to copy 1 ₁ '+1 th data 1 _1' to the second. 'The + n ₁ 1' 1 ₁ to + n-1 and repeat copy, 'when it reaches, 21 _1' 21 ₁ Copying +2. Thereafter, copying is repeated from 21 ₁ ′ + n to 21 ₁ ′ + n−2, and the process is completed when ₁₁ minutes have passed. The above operation has been described by adjusting two frames, but the same applies to other frame lengths.

When the pattern length reaches a predetermined length, binarization is performed next. However, there are many general methods that do not require binarization. This is for recognition by the method described in the literature (Introduction to Ohm's Applied Fuzzy System). The binarization is performed by the comparator 25 for each frame. The sum of all data of one frame is shifted by 3 bits, that is, reduced to 1/8 from the register 22 and sent to the threshold. After that, the threshold value is compared with each value of the frame. When the value is larger than the threshold value, 1 is set, and the other values are set to 0, binarized and stored in the register 22 again. The contents of the register 28 are cleared to 0 in advance, and the pattern in the register 22 and the contents of the register 28 are added by the adder 27, and the result is returned to the register 28. This is for registering an averaged pattern as a standard pattern by uttering several times for one word. If it is not necessary to average, register the contents of the register 22 as a standard pattern in the register 28 as it is. Just do it. Here, description will be made assuming that the utterance made three times is registered. First, the pattern of all zeros and the pattern created by the first utterance are added and stored in the register 28, and the second utterance forms the same pattern as the first utterance. (1st utterance pattern) and register 2
Return to 8. The third utterance is similarly added and returned to the register 28, and the result is written to the register 28 as a dictionary unit. At the same time, the present invention is applied, and in the manner described with reference to FIG. 3, it is checked whether or not there is a portion having a small energy and having only a high frequency component near the leading or trailing end of the voice. If there is, remove that part and convert the remaining part to the determined length. When addition is necessary, the removed parts are added together, and the parts not removed are added. If the number of patterns is insufficient, further vocalization may be prompted, or a specific pattern may be weighted and averaged. After all necessary words have been registered in this way, the standard pattern of the register 28 is written to the floppy disk 42 or the like so that the contents can be preserved even when the power is turned off.

Here, processing relating to the present invention is performed. FIG.
Is looking for a part of the voice where the energy is small and the spectrum is concentrated in the high band. The band-pass filter 3 has 15 filters, and the lowest frequency is 250H.
z, the maximum is 1/3 oct. It is assumed that they are lined up. After A / D conversion, 11
The sum is $ 31 and the twelfth to fifteenth are $ 32, and the two sums are compared. If the output of # 32 is large, the comparator 33 outputs 1. On the other hand, if the sum of # 31 and # 32, that is, the energy of the sound output from the adder 34 is smaller than the value determined as the threshold 35, the comparator 36 is also 1
Is output, otherwise 0 is output. Comparator 33,
When the product of the signal of 36 is 1, the unknown voice pattern temporarily stored in the register 37 is transferred to the register 22, and thereafter, the same operation as in FIG. 7 is performed to obtain a recognition result.

The comparator 24 and the register 22 perform the same processing as the previously collated pattern. After collation, the similarity calculated this time is successively written after the similarity previously stored in the register 43. Finally, as in the previous example, the word with the highest similarity is checked and its name is output. However,
At this time, the maximum value is obtained for all the similarity values written in the register 43.

FIG. 5 is a diagram showing a flow of speech recognition, and FIG.
Is a diagram for performing the flow shown in FIG. 5 by hardware using a computer, and only the portions different from FIG. 3 will be described below. This is the same until the binarization pattern is created by the register 22. The method of normal collation is as follows. Comparator 4
In the case of 0, the frame length of each word sent from the dictionary unit of the register 41 is compared with the frame length of the input voice, and only when the value is the same, the dictionary pattern is reloaded into the matching unit. Then, the similarity is calculated and written into the register 43. However, it is assumed that the register 43 has been cleared to 0 in advance, and that those which have not been collated due to a difference in frame length have a similarity of 0. This is repeated until an end signal indicating the end of the pattern registered in the dictionary section is output. When this is completed, the first similarity value of the register 43 is transferred to the register 45, and the second and subsequent similarity values of the register 43 are compared with the value of the register 45 by the comparator 44. Then, this value is written to the register 45. Thereafter, the newly written value is compared with the similarity value of the register 43 in order, and this is repeated. However, as in FIG. 3, all of these operations are controlled by a program and are operated by a microcomputer.

[0019]

As is clear from the above description, according to the present invention, even when the detection of a voice section is not successful, correct collation can be performed without expansion / contraction in collation.

[Brief description of the drawings]

FIG. 1 is a flowchart illustrating an embodiment of voice registration according to the present invention.

FIG. 2 is a block diagram for realizing the voice registration shown in FIG.

FIG. 3 is a diagram for realizing FIG. 1 by hardware using a microcomputer;

FIG. 4 is a diagram for explaining expansion and contraction of a frame.

FIG. 5 is a diagram showing a flowchart of voice recognition.

FIG. 6 is a diagram for realizing FIG. 5 by hardware using a microcomputer;

FIG. 7 is an explanatory diagram of general pattern matching.

FIG. 8 is a diagram for explaining the association when a weak consonant is detected and the association when a weak consonant is not detected;

[Explanation of symbols]

DESCRIPTION OF SYMBOLS 1 ... Microphone, 2 ... Microphone amplifier, 3 ... Feature conversion part, 4 ... A / D conversion part, 5 ... Switch, 6 ... Standard pattern storage part, 7 ... Collation part, 8 ... Maximum similarity detection part, 9 ... Recognition Result output unit, 10 ... Expansion unit, 11 ... Power calculation unit, 1
2 ... Comparing unit, 13 ... High-frequency spectrum calculating unit, 14 ... Comparing unit, 15 ... Expandable unit, 16 ... Memory, 17, 18 ... Threshold,
19: pattern shaping section, 20: expandable section, 21: voice section detector, 22, 23, 28, 30, 32, 39, 41 ...
Registers, 24, 25, 27, 31, 35, 38 ... comparators, 26, 37 ... thresholds, 29 ... collating units, 36, 40 ... adding units, 42 ... floppy disks.

──────────────────────────────────────────────────続 き Continued on the front page (58) Field surveyed (Int. Cl. ⁷ , DB name) G10L 3/00 521 G10L 3/00 531 G10L 7/08

Claims (1)

(57) [Claims] 1. A generates a characteristic pattern is taken out a feature from the audio signal, the audio pattern matching how to match the standard pattern of the feature pattern, than that of speech energy vowel at the beginning or end of the speech Find a part where the spectral components are low and concentrated in the high frequency range, convert the entire pattern to a predetermined length, and the energy is lower than that of the vowel and the spectral components are concentrated in the high frequency range standard pattern registration how to standard pattern is converted to a length defined the remaining pattern Remove the part.