JPH06167992A - Voice pattern creating device and standard pattern registration device using the same - Google Patents

Abstract

(57) [Abstract] [Purpose] When registering a standard pattern of an arbitrary word for voice recognition, it is possible to significantly reduce the occurrence of erroneous recognition. [Structure] The standard pattern creation unit 6 includes a voice section detection unit 4
The registration unit 11 for registering the voice-related part of the feature conversion result output from the dictionary 5 as a standard pattern in the dictionary 5 as it is, and the plurality of standard patterns registered in the dictionary 5 as they are in the working area, for example, in the working area. Dividing unit 1 for reading and dividing each standard pattern into a plurality of portions
2 and a plurality of characteristic patterns are divided into a plurality of parts, respectively, these are combined and connected to generate a new standard pattern of voice, and the standard pattern registered in the dictionary 5 as it is by the registration unit 11. In addition, a pattern generation unit 13 that registers all or some of the new standard patterns in the dictionary 5 is provided.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a voice pattern creating apparatus for creating a voice pattern used for voice recognition and the like and a standard pattern registration apparatus using the same.

[0002]

2. Description of the Related Art Generally, in a voice recognition apparatus, standard feature patterns of various voices, that is, standard patterns are registered in a dictionary in advance before actual voice recognition is performed. The characteristic pattern of the unknown voice uttered by the person is collated with various standard patterns registered in the dictionary to find the similarity between them, and the standard that gives the largest similarity among the various standard patterns. A word (for example, a word) corresponding to the pattern is output as a recognition result.

By the way, in this type of voice recognition device, as a standard pattern, for example, "Adachi", "Inoue", "Uno", "Endo", "Ono", "Kato", "Kinoshita", "Kusenbe" , "Kobayashi", "Yamada" when the pattern corresponding to the words of 10 names is registered, the speaker is "Sato"
"Kato", which is the most similar to "Sato" among the above 10 standard patterns, is output as a recognition result.

In order to prevent such erroneous recognition, conventionally, Japanese Patent Laid-Open Nos. 61-133996 and 63-295 have been used.
As disclosed in Japanese Patent No. 394, a technique of setting a threshold value for the similarity has been proposed. According to this, even when the standard pattern that gives the highest similarity among various standard patterns is selected, if this similarity is less than or equal to the threshold value, it can be rejected, resulting in erroneous recognition. Can be prevented. Also, JP-A-58-86598
There is also proposed a technique for preventing the erroneous recognition as described above by weighting the standard pattern as disclosed in No.

[0005]

However, in the above-mentioned prior art, it is very difficult to set a threshold value for the similarity, and there is a limit to surely prevent erroneous recognition and the like. To be more specific, in the above example, in the case of “Kato” and “Sato”, / ka respectively
Both are / k because they are pronounced as too /, / satoo /
Everything else is the same except that / and / s / are different. As a result, there was no big difference between "Kato" and "Sato" originally,
There was a problem that it was difficult to determine the threshold. Further, in this case, if a strict threshold value is determined, it is rejected even though "Kato" is correctly uttered, and a recognition result cannot be obtained.

Further, there is a problem that the weighting to the standard pattern is limited when the words to be registered are predetermined or in the case of the specific speaker system. Specifically, in order to emphasize the difference between "Sato" and "Kato", it is possible to weight the / k / part of "Kato" in the standard pattern. It is enough to emphasize the difference between "and" Sato, but not enough for every word entered. That is, for example, when the name "Kano" is input, an abnormally high degree of similarity with "Kato" is displayed, resulting in erroneous recognition.

As described above, in the above-mentioned conventional technique, when a standard pattern of an arbitrary word is registered for voice recognition,
There is a limit to significantly reducing the occurrence of erroneous recognition, and it is very difficult to provide a general-purpose and practical voice recognition device. For example, when a user of voice recognition registers and uses an arbitrary word (especially, in an application that operates when a certain keyword is recognized), it is specified. It was not possible to determine whether or not the word was input, and therefore, there was a limit to the usable range.

The present invention can realize a general-purpose and practical voice recognition device capable of remarkably reducing the occurrence of erroneous recognition when a standard pattern of an arbitrary word is registered for voice recognition. The purpose of the present invention is to provide a voice pattern creating device and a standard pattern registration device.

[0009]

In order to achieve the above object, a voice pattern creating apparatus according to a first aspect of the present invention comprises a dividing means for dividing each of a plurality of voice characteristic patterns into a plurality of portions. , And a pattern creating means for creating a new voice characteristic pattern by combining and connecting the respective portions of the plurality of characteristic patterns divided by the dividing means. This makes it possible to easily create a voice characteristic pattern for another word.

A standard pattern registration device according to a second aspect is the voice pattern creation device according to the first aspect.
It further has a registration means for registering a plurality of voice characteristic patterns as a standard pattern in the dictionary as it is, the dividing means divides the standard pattern into a plurality of parts, and the pattern creating means divides the divided parts. A new characteristic pattern is generated by combining and connecting each part of a plurality of standard patterns, and the new characteristic pattern is used as a new standard pattern together with the standard pattern registered as it is by the registration means. It is characterized by registering in a dictionary. As a result, when a standard pattern of an arbitrary word is registered for voice recognition, the occurrence of erroneous recognition can be significantly reduced.

According to a ninth aspect of the present invention, a voice pattern creating apparatus shifts each of a plurality of voice characteristic patterns to a high frequency band and / or a low frequency band in a frequency axis direction to generate a new voice pattern. It is characterized by having a pattern creating means for creating a characteristic pattern. This allows
Although it is the same word as the original standard pattern, it is possible to easily create a standard pattern of voice corresponding to the case where the speakers are different.

The standard pattern registration apparatus according to a tenth aspect of the present invention is the voice pattern creation apparatus according to the ninth aspect, further comprising registration means for registering a plurality of voice characteristic patterns as standard patterns in the dictionary as they are. However, the pattern creating means shifts each of the plurality of standard patterns to a high band and / or a low band by a certain amount in the frequency axis direction, and when a new voice pattern is created, the registering means keeps the same form as it is. Along with the registered standard pattern, all or part of the new voice pattern is registered in the dictionary as a standard pattern. Accordingly, the dictionary for the unspecified speaker can be easily created, and the dictionary can be applied to the voice recognition device for the unspecified speaker.

[0013]

Embodiments of the present invention will be described below with reference to the drawings. FIG. 1 is a diagram showing a configuration example of a voice recognition device to which a voice pattern creation device according to the present invention is applied. Referring to FIG. 1, the voice recognition device includes a voice input unit 1 such as a microphone, an A / D converter 2 for analog-digital converting a voice signal input from the voice input unit 1, and a digitally converted voice. A feature conversion unit 3 that performs feature conversion on a signal, a voice section detection unit 4 that detects a portion related to voice from the feature conversion result as a voice section, a dictionary 5 in which various standard patterns are registered, and a standard pattern is created. Switch SW, which is switched between the case of registering and the case of actual voice recognition, and the time of creating and registering the standard pattern.
A standard pattern creating unit 6 that creates a standard pattern of a voice based on a part related to the voice of the feature conversion result output from the voice section detecting unit 4 and registers it in the dictionary 5,
A portion related to the voice of the feature conversion result output from the voice section detection unit 4, that is, an unknown pattern is collated with various standard patterns registered in the dictionary 5 to obtain a similarity,
The recognition unit 7 outputs a word (for example, a word) corresponding to the standard pattern that gives the highest similarity as a recognition result. The feature conversion unit 3 may be, for example, an FFT.
(Fast Fourier Transform), and in this case, the voice signal is transformed into the frequency domain, and the spectrum of the voice is obtained as a feature amount. In this case, the A / D converter 2 has 16 bits and 16 KH.
The one with a characteristic of about _Z is used.

Further, the standard pattern creating section 6 registers the part relating to the voice of the feature conversion result output from the voice section detecting section 4 as a standard pattern in the dictionary 5 as it is, and the registering section 11 as it is. When a plurality of standard patterns registered in the dictionary 5 are read into, for example, a working area and each standard pattern is divided into a plurality of parts, and a plurality of characteristic patterns are divided into a plurality of parts, respectively, Are combined and concatenated to generate a new standard pattern of voice, and all or a part of the new standard pattern is combined with the standard pattern registered in the dictionary 5 by the registration unit 11 as it is. And a pattern generation unit 13 that registers the pattern.

Next, the operation of the speech recognition apparatus having such a configuration will be described with reference to the flowchart of FIG. Prior to the actual voice recognition operation, the switch SW is switched so that the standard pattern creation process is selected. Then, the speaker utters this word in order to register a standard pattern of voice corresponding to a word, for example, the word in the dictionary 5. When the word voice uttered by the speaker is input from the voice input unit 1 (step S1), the input voice signal is converted into a feature amount (for example, spectrum) in the feature conversion unit 3 (step S2), and the voice section The detection unit 4 detects a voice-related portion of the feature conversion result as a voice section (step S3) and inputs it to the standard pattern creation unit 6 via the switch SW.

In the standard pattern creating section 6, first, in the registering section 11, the part relating to the voice of the feature conversion result output from the voice section detecting section 4 is made to correspond to a word (for example, a word) as a standard pattern as it is. And register it in the dictionary 5 (step S4). When the processes of steps S1 to S4 are repeated and, for example, a standard pattern of a predetermined type of word voice is registered in the dictionary 5 as it is (step S
5) Further, the dividing unit 12 divides each standard pattern registered in the dictionary 5 as it is into a plurality of portions (step S6). Next, in the pattern generation unit 13, the divided portions of the plurality of types of standard patterns are combined and connected to generate a new voice standard pattern (step S7), and the new voice standard pattern is also combined in the dictionary 5. To register (step S8). In this way, the new standard pattern is registered in the dictionary 5 together with the standard pattern output from the voice section detection unit 4 as it is.

After the standard pattern as described above is registered in the dictionary 5, the switch SW is switched so that the recognition process is selected. In this case, when the utterance of the speaker is input from the voice input unit 1, the input voice signal is converted into a feature amount (for example, spectrum) by the feature conversion unit 3, and the voice section detection unit 4 and the switch SW are turned on. It is input to the recognition unit 7 via the. In the recognition unit 7, the unknown voice pattern input via the switch SW is collated with various standard patterns registered in the dictionary 5 as described above to obtain the similarity, and the standard pattern giving the highest similarity is obtained. The word corresponding to is output as a recognition result. In this way, a series of voice recognition processing can be performed.

By the way, in this embodiment, as described above,
The standard pattern creation unit 6 is further provided with a division unit 12 and a pattern generation unit 13,
The mode of processing in the division unit 12 and the pattern generation unit 13 will be described in more detail.

As a method of division in the division unit 12, the standard pattern registered in the dictionary 5 is, for example, n in the time axis direction.
A mode of equally dividing is possible. More specifically, in the process of step S5, as shown in the dictionary 5 for example, as shown in FIG. (Corresponding to “.”) Is registered, the dividing unit 12 copies these three standard patterns A, B, and C into, for example, a working area, and thereafter, these three standard patterns A, B, and C are copied in the time axis direction.
Divide into equal parts (n = 3) and divide each part into A1, A2, A
3, B1, B2, B3, C1, C2, C3.

Further, the pattern generation unit 13 includes a division unit 12
A new standard pattern is generated by combining the respective parts of the plurality of types of standard patterns that have been divided in 1. At this time, a certain combination rule is set. That is, the pattern generation unit 13 has a rule that the portion at the beginning of the standard pattern before division is positioned at the beginning even after the connection. According to this rule, the pattern generation unit 13
When combining, for example, the portion at the beginning in the pattern before division is treated as a special group, and the portions other than the beginning are connected and then the beginning is added.

Further, the pattern generation unit 13 has a rule that the portion at the end of the standard pattern before division is positioned at the end even after connection. According to this rule, the pattern generation unit 13 treats the part at the end of the pattern before division as a special group, combines the parts other than the end, and adds this to the end when combining. Has become.

In other words, such a rule is that when combining the divided parts in the example of FIG.
It means to keep the order of the subscript n of n and Cn. As a result, it is possible to prevent a standard pattern that is not common sense from being created when the patterns are rearranged. That is, in the actual speech recognition process, as is clear from the examples of "Sato" and "Kato", the problem is that the sound has different beginnings and the same sound at the end. "," And "Fujika", such as words with the beginning and the end interchanged, are generally not a problem, and therefore only standard patterns created by combinations according to the above rules are important. Become.

In accordance with such a rule, three standard patterns A, B, C as shown in FIG. 4 are divided into three parts A1, A2, A3, B1, B2, B3, C1, C2, C.
When the combination and connection process is performed on No. 3, a new standard pattern D (A2 + B2 in this example) as shown in FIG.
+ B3) can be made, and this new standard pattern D
As shown in FIG. 3B, (A2 + B2 + B3) can be registered in the dictionary 5 together with the standard patterns A, B, and C. In this case, it is possible to create information such as the length of the new pattern created by the combination connection and add the information to register it. However, in the example of FIG.
At the time of registration in the dictionary 5, the standard patterns A, B, and C that are registered as they are are associated with words (words), that is, “Kusek”, “Kato”, and “Kobayashi”, respectively. The created standard pattern is not associated with a predetermined word (word).

For example, if the standard patterns of A, B, and C are "Kusekbe", "Kato", and "Kobayashi", the new standard pattern D (A2 + B2 + B3) is very similar to the pattern of "Sato". It will be close. Therefore, since a similar standard pattern is automatically created without weighting the standard pattern, it is possible to easily prevent erroneous recognition of an input voice having a similar sound. In other words, for similar voice input, the standard pattern newly created by the combination has greater similarity, so it is possible to output the recognition result as a reject or to output a predetermined message. it can. That is, in the above example, the speaker is "Sato".
, The input voice is the standard pattern A, B,
Of C and D, it is the most similar to the new standard pattern D, and thus is not erroneously recognized as "Kato". Further, in the example of FIG. 3B, even when the new standard pattern D is most similar,
Since this standard pattern D is not associated with a word, the recognition result is not a specific word but a reject. Thus, for example, in the case of an application that operates when a certain predetermined keyword is recognized, it is possible to determine with a simple method whether or not a specific word is uttered. . Note that, as shown in FIG. 3C, words are also included in the new standard pattern D.
It is also possible to correlate with and register in the dictionary 5, and in this case, a correct recognition result can be output instead of rejecting.

Further, in the above example, the division number n of the standard pattern in the time axis direction is "3", but n may be a division number other than three, and it is also necessary that the division is even division. Absent.

For example, normally, in the case of speech recognition of words, 1
A process for obtaining a spectrum is performed every 0 to 20 msec, and a short word speech is "2 / ni /",
It is composed of a single syllable like "5 / go /", and the duration is about 200 msec. Therefore, even if this is divided into a large number of parts, the time length of each part is short and the influence on the new pattern after connection is small. Further, in the case of a single syllable, the continuation part becomes a partial piece of the same vowel no matter how many times it is divided, so in consideration of the fact that there is no point in dividing it, a plurality of registered vowels as shown in FIG. Each of the standard patterns A, B, and C of the voice is divided into two parts near the beginning and the end with respect to the time axis, and the vicinity of the end of B is added to the beginning of A and registered as a new voice pattern. A new voice pattern is created and registered by adding the vicinity of the end of C to the vicinity of the beginning of B, and all or part of the created pattern is registered as a new standard pattern by repeating this process. You can also

FIG. 6 is a flow chart showing the flow of such processing. Referring to FIG. 6, the dividing unit 12 divides the m standard patterns registered in the dictionary 5 into two instead of dividing them into two, and gives the divided patterns to the pattern generating unit 13 (step S11). In the pattern generation unit 13, first, the m standard patterns registered as the original words are numbered (step S12), and the two-divided portions are used as the beginning and the end of each numbered pattern. Next, the counter i is initialized to "1" (step S13) and the beginning of the i-th pattern and the end of the (i + 1) -th pattern are connected to form a new standard pattern (step S13).
15). i is incremented by 1 (step S16),
The above process is repeated, and when the counter i reaches m (step S14), the end of the first 1st pattern is added to the beginning of the mth pattern (step S1).
7), the process ends. As a result, it is possible to perform pattern concatenation that is effective for short voices and has no waste. In the above processing example, the counter i is incremented by "1" and the pattern number is shifted by one, but the pattern number does not have to be one by one and may be randomly spliced. .

In this way, a plurality of types of registered standard patterns are divided into n in the time axis direction and combined.
I created a new standard pattern, with the above example,
For example, in the case of two divisions, division is performed for a fixed time length (for example, equally divided into two). However, with such division, strictly speaking, it is inevitable that a phoneme in a word is divided in the middle. Therefore, it is desirable to match the phoneme breaks and the pattern divisions as much as possible. For this purpose, for example, for each of the standard patterns A, B, and C of a plurality of voices, a portion with a large change in the pattern in the time axis direction is detected, and at this portion, it is divided near the end and near the beginning of A A new voice pattern is added by adding the vicinity of the end of B, and C is added near the beginning of B.
It is also possible to create and register a new voice pattern by adding the vicinity of the end of, and to register all or part of the created pattern as a new standard pattern by repeating this process.

FIG. 7 is a flow chart showing the flow of the above processing in the dividing unit 12. In the process of FIG. 7, when the time is i, the frequency is j, the standard voice pattern is a (i, j), and the total length of the standard pattern is n, first, the time counter i is initialized to "1". It is set (step S21). Then, with respect to the entire pattern at time i, the time change rate d of the pattern is expressed by the following equation.
_i is calculated (step S22).

[0030]

[Equation 1]

I is incremented by "1" (step S2
4) The above process is repeated to obtain the time change rate d _i of the pattern for each i, and when the counter i reaches (n-1) (step S23), d is selected from the total pattern length n
_{The i} that maximizes the i is obtained and is set as i _n (step S25). Then, the reference pattern 1 to i _n a i
It is divided into _{n + 1} to n (step S25). Such division can prevent a situation in which patterns are discontinuously connected at the center of the phoneme.

Alternatively, instead of the above processing, a minimum portion of the voice energy is detected in the time axis direction for each of a plurality of voice standard patterns A, B, and C, and this portion is divided into the vicinity of the tail end. Create a new voice pattern by adding the vicinity of the end of B to the beginning and registering it as a new voice pattern, and adding the vicinity of the end of C to the beginning of B to create and register a new voice pattern and repeating this process. It is also possible to register all or part of the created pattern as a new standard pattern.

FIG. 8 is a flow chart showing the flow of the above processing in the dividing section 12. In this process,
First, the time counter i is initialized to "1" (step S31), and then the energy p _i at the time i is obtained by summing a (i, j) for j (step S32). Then, the difference between p _i-1 and p _i is calculated,
This is designated as D _i (step S33). Next, i is incremented by "1" (step S35), the above process is repeated, and the energy p _{i and the} difference D _i are obtained for each i,
When the counter i reaches n (step S34), D
Based on _i , the minimum energy is detected. That is, in this detection process, the minimum value of p _i is D _i−1 and D _{i.
Since i is obtained when} the sign is different and D _i is positive, this point is obtained and set as i ′ (step S36). After that, the patterns 1 to i '
And i ′ to n (step S37). By such division, as in the processing example of FIG. 7, it is possible to prevent a situation in which the patterns are discontinuously connected at the center of the phoneme. When there are a plurality of i's in one voice, the one with a smaller D _i at that time is selected, the one near (or far from) the head is selected, and the one near the center is selected. You can take actions such as choosing.

As in each of the above-described examples, the process of dividing the standard pattern in the time axis direction and combining these can be intuitively understood to synthesize another word similar to the original voice. Although it is possible and very easy to think, considering that the patterns are similar to each other, the same effect can be obtained not only by dividing in the time axis direction but also by dividing in the frequency axis direction. Therefore, as the dividing unit 12, for example, FIG.
As shown in FIG. 3, each of the plurality of voice characteristic patterns A, B, and C is divided into a high-frequency neighborhood and a low-frequency neighborhood with respect to the frequency axis, and A1, A2, B1, B2, C1, respectively. C
As shown in FIG. 10, the high-frequency neighborhood A1 of A is added to the high-frequency neighborhood B2 of B to be registered as a new voice pattern D, and the high-frequency neighborhood C2 of C is connected to the low-frequency neighborhood B1 of B. Patterns D, E, created by repeating the process of creating and registering a new voice pattern E by adding
It is also possible to register all or part of ... as a new standard pattern.

Alternatively, instead of performing the division processing in the division unit 12, each of the plurality of standard patterns A, B and C of the voices is operated by shifting (shifting) to a high frequency range by a certain amount with respect to the frequency axis. It is also possible to register all or part of the created new voice pattern as a new standard pattern.

FIG. 11 shows one standard pattern a (i, j).
Is a flow chart showing the flow of processing for generating a new standard pattern by performing an operation for shifting the frequency axis by a certain amount to the high frequency band. In this process, first, the time counter i is initially set to "1" (step S41).
Next, the frequency counter j is set to "1" (step S42). Then, a (i, j) is shifted in the frequency direction by, for example, 2 samples to be b (i, j + 2) (step S43). When the range of j is 1 to J,
Stepping j by "1" (step S45), j = 1 to
This shift process is repeated in the range of J-2, and when j becomes J-2 (step S44), the b (i, j) created as a result of the shift is again converted into a (i, j). Map. That is, first, j is initially set to "1" (step S46), and then b (i, j) is set to a (i, j).
(Step S47). Stepping j by "1" (step S49), this mapping process is repeated within the range of j = 3 to J, and when j becomes J-2 (step S48), this mapping process ends. To do. As a result of this mapping processing, the original data remains in the part of a (i, j) where j = 1 and j = 2, and the part where j = 3 to J is from the original 1 to J.
The data up to -2 has been shifted. In this way, after the data shift to the high frequency band at one time i, i is incremented by "1" (step S5).
1) The same processing as described above is repeated, and when i becomes n (total pattern length) (step S50), the shift processing for one standard pattern is completed. In addition,
In this example, two samples are shifted in the frequency direction, but the shift amount is not limited to two samples and can be set arbitrarily, and this can be changed according to the sampling frequency.

Further, in the above example, the shift is performed in the high band direction, but it may be shifted in the low band direction. That is, all or one of the new voice patterns created by performing an operation of shifting (shifting) the standard patterns A, B, and C of a plurality of voices to the low frequency range by a certain amount with respect to the frequency axis. It is also possible to register a copy as a new standard pattern. This process is basically the same as the process of FIG. 11, and in step S43 of FIG. 11, only a process of shifting a (i, j + 2) by 2 samples in the frequency direction to obtain b (i, j). Is different.
In this way, when one standard pattern is shifted to the low frequency range by 2 samples with respect to the frequency axis, a
The original data remains in j = J-1 and J of (i, j), and the data of J is shifted from j = 3 in j = 1 to J-2.

It is also possible to perform both a shift to the high frequency band and a shift to the low frequency band. That is, all or a part of a new voice pattern created by performing an operation of shifting each of a plurality of voice standard patterns A, B, and C to a high frequency range by a certain amount with respect to the frequency axis. Registered as a standard pattern, and further, the original standard patterns A, B, C
All of the new voice patterns created by performing an operation that shifts each of the
Alternatively, a part of them may be registered as a standard pattern.

The above-described shift of the standard pattern to the high frequency range and the low frequency range is comparable to an operation for creating patterns of different utterers. Therefore, by registering a new standard pattern created by shifting to the high frequency region or the low frequency region in the dictionary 5, the voices higher or lower than the voice of the registrant (specific speaker) are registered. It is possible to prevent erroneous recognition from being made, and further, by making a predetermined word correspond to this new standard pattern, the dictionary 5 can be made for an unspecified speaker.

Further, by performing both the shift to the high frequency range and the shift to the low frequency range, both a voice higher than the voice of the registrant (specific speaker) and a voice lower than the voice of the registrant (specific speaker) can be obtained. It is possible to protect, and further, by making a predetermined word correspond to this new standard pattern, it is possible to make the dictionary 5 easy for an unspecified speaker.

In each of the above-described examples, the standard pattern is divided in the time axis direction or the frequency axis direction, and different patterns are connected to each other to create a new standard pattern.
It is possible to prevent erroneous recognition between similar words such as "Kato" and "Sato". However, for example, in the words to be registered, "Sato" When the beginning of the word is the same, or when the ending is the same, such as “Sakamoto” and “Hashimoto”, “Sato”,
In the example of "Saji", if the beginning and the end are divided and combined, the original "Sato" and "Saji" may be generated in the combined result. The same applies to "Sakamoto" and "Hashimoto". If the same original word as this is registered as a new standard pattern, the original correct recognition cannot be performed. Therefore, such a new standard pattern should not be registered in the dictionary 5.

Therefore, in order to avoid such a problem,
Further, the similarity between all the standard patterns of the original standard pattern and the new standard pattern registered in the dictionary 5 is calculated, and the standard pattern having the similarity larger than the determined value is deleted. Processing can also be performed.

FIG. 12 is a flow chart showing an example of such processing. In this processing example, a new standard pattern is created based on the original standard pattern, and when this new standard pattern is registered in the dictionary 5 together with the original standard pattern, it is first registered in the dictionary 5. All of the existing standard patterns (the number of original standard patterns W and the number of new standard patterns (L-W)) (total of L standard patterns) are represented by T ₁ , T ₂ , ..., T _L. Number them.
In this case, for example, the numbered sequentially from the original standard pattern, performs numbering of new reference pattern following the number (T ₁ ~T _W) original reference pattern (T W _{+ 1} ~T _L ).
Moreover, in order to obtain the similarity between two standard patterns,
For convenience, let us say that the number of one of the two standard patterns is r and the number of the other standard pattern is k.

After all standard patterns are numbered, r is initialized to "1" (step S6).
1). Then, k is set to "W + 1" (step S
62). After that, the standard pattern T _r and the standard pattern T _k
Then, the similarity S _rk between and is obtained (step S63). Then, it is determined whether r is smaller than k (step S
64). This judgment is made when r and k are the same (when the two standard patterns T _r and T _k are the same) and when r and k are opposite (in T _r and T _k) . (In the case of T _k and T _r ), it is done for the purpose of exclusion. That is, in the table as shown in FIG.
With respect to the new standard patterns T _{W + 1 to} T _L , this is done so as not to perform subsequent processing on the portion below the diagonal line DG, including on the diagonal line DG. When r is smaller than k by such a judgment, that is, when there is a relationship above the diagonal line DG in the table of FIG. 13, the similarity S _rk obtained in step S63 is smaller than a predetermined threshold S ′. It is determined whether or not (step S6)
5). When S _rk is smaller than S ′, it is determined that the similarity between the two standard patterns T _r and T _k is small, and the erasing process is not performed. On the other hand, when S _rk is not smaller than S ′, it is determined that the similarity between the two standard patterns T _r and T _k is large, and one standard pattern T _k is deleted from the dictionary 5 (step S66). .

In this way, the degree of similarity S _rk between two certain standard patterns T _r and T _k is obtained, and a predetermined process is performed based on the degree of similarity, and then k is incremented by "1" (step S
68), the same processing is repeated. When k becomes L (step S67), r is incremented by "1" (step S70) and the same processing is repeated, r
When L becomes L-1 (step S69), the process ends.

In other words, the above-mentioned processing is performed by setting k to W.
Change from +1 to L in order, and r from 1 to L-1
13 to create a similarity table between T _k and T _r as shown in FIG. 13, in which one of the two standard patterns T having a large similarity is used. to eliminate _k . Further, in such a process (specifically, the process of step S64), since the number of the new standard pattern is always larger than the number of the original standard pattern, only the new standard pattern is to be erased. Yes, the original pattern can be left without being erased. The method of determining the threshold value S ′ at this time is not particularly limited, but can be determined as follows, for example. That is, when correct recognition is performed using this standard pattern, the threshold value S ′ can be set to about 90% of the similarity acquired by the standard pattern that is the correct answer. This method removes confusional standard patterns in misleading words.

The original purpose of the present invention is to prepare a similar standard pattern for distinguishing another similar word. Therefore, when a standard pattern having no similarity or a standard pattern having little similarity due to division is generated, it is not necessary to register these in the dictionary 5, and these should not be registered. By doing so, the dictionary 5 can be saved. That is, the similarity between the new standard pattern created by concatenation and the original standard pattern is checked, and the standard pattern whose similarity to all the original standard patterns is smaller than a predetermined value is deleted. You can A method similar to this is also disclosed in Japanese Patent Laid-Open No. 61-292696. According to the method disclosed in this publication, when a plurality of standard patterns are registered in advance for one voice, the degree of similarity between the plurality of patterns is obtained, and one of the similar patterns is deleted. In this method, either pattern may be erased at that time, and if this method is applied to the present invention, the original pattern may be erased, and as a result, correct recognition cannot be performed. Occurs. Therefore, in the application to the present invention, when erasing a standard pattern whose degree of similarity is lower than a predetermined threshold value, it is necessary to perform processing such that the original pattern is not erased and only a new pattern is erased. There is.

FIG. 14 is a flow chart showing an example of such processing. In this process, the number of original standard pattern L _1, the number of new reference pattern made by the connecting operation or the like and L _2, each of the original standard pattern, and the numbering of 1 to L _1, also , L ₁ +1 for each new feature pattern
Number them to L ₁ + L ₂ . The variable k represents the original standard pattern, and r represents the standard pattern created by the operation.

First, L + 1 is initialized to r (step S81). Next, k is set to "1" (step S
82), the maximum value mx of the similarity is initialized to "0" (step S83). Then, the similarity S _rk between the original standard pattern T _k with the number _k and the new standard pattern T _r with the number r is calculated (step S84), and this similarity S _rk is m.
It is determined whether it is larger than x (step S85).
If it is larger than mx, the value of this similarity S _rk is set in mx (step S86). Then, k is incremented by "1" (step S88), the same process is repeated, and when k becomes L ₁ (step S87), whether mx at this time is larger than a predetermined threshold value S '. It is determined whether or not (step S89). That is, the above-described processing obtains L ₁ similarity between each of the L ₁ original standard patterns T _k with numbers _{k 1} to L ₁ and the new standard pattern T _r with number r. Of these similarities, it is determined whether or not the largest similarity is larger than a predetermined threshold value S ′. As a result, when the largest similarity mx is not larger than the predetermined threshold value S ′, the new standard pattern T _{r with} the number r has a small similarity to any of the L ₁ original standard patterns T _k. Is erased (step S9
0). Then, r is incremented by "1" (step S9
2), the same processing as described above is repeated for each of the L ₂ new standard patterns T _r , and when r becomes L ₁ + L ₂ (step S91), the processing ends. In this way, among the L ₂ new standard patterns,
Any of the L ₁ original standard patterns having a small degree of similarity can be erased. In this process, the threshold value S ′ may be determined in the same degree as the degree of similarity when words having no similar sounds are compared with each other. As a result, it is possible to erase a word having a low similarity to all the words to be registered, that is, a word that becomes noise of the standard pattern. Therefore, the capacity of the dictionary 5 can be saved and the original effect can be obtained.

[0050] Further, in the above example was with the original standard pattern 1 to L _1, on the contrary, a 1 to L ₂ as a new reference pattern, based on the standard L ₁ + L ₂ from L ₂ +1 The same process can be performed by setting a pattern and erasing ₁ to L ₂ .

[0051]

As described above, according to the invention of claim 1, each of a plurality of voice characteristic patterns is divided into a plurality of parts, and each of the plurality of characteristic patterns divided by the dividing means is divided. Since the parts are combined and connected to create a new voice characteristic pattern, it is possible to easily create a voice characteristic pattern for another word.

According to a second aspect of the present invention, there is provided a standard pattern registration apparatus using the voice pattern creating apparatus according to the first aspect, wherein the dictionary is a feature pattern of a plurality of voices as a standard pattern as it is. The dividing means divides the standard pattern into a plurality of parts, and the pattern creating means combines and connects the respective parts of the plurality of divided standard patterns to create a new feature. Since the pattern is generated and the new feature pattern is registered in the dictionary as a new standard pattern together with the standard pattern registered in the same form by the registration means,
When a standard pattern of an arbitrary word is registered for voice recognition, it is possible to significantly reduce the occurrence of erroneous recognition.

According to the third to fifth aspects of the invention, the pattern creating means divides each of the plurality of standard patterns into a plurality of parts, and combines and connects them to create a new standard pattern. In this case, the part at the beginning of the standard pattern before division is positioned at the beginning of the new standard pattern after concatenation, and the part at the end of the standard pattern before division is the new standard after concatenation. Since the patterns are also combined so that they are positioned at the end, it is possible to prevent the generation and registration of uncommon patterns.

According to the sixth to eighth aspects of the invention, the dividing means divides each of the plurality of standard voice patterns in the time axis direction, and the pattern forming means divides in the time axis direction. Each part of each standard pattern is combined and connected according to a predetermined rule to create a new voice pattern, and all or part of the created voice pattern is registered as a standard pattern by repeating this process. Therefore, it is possible to easily create and register a voice characteristic pattern for another word.

According to the ninth aspect of the present invention, each of a plurality of voice characteristic patterns is shifted to the high frequency band and / or the low frequency band by a certain amount in the frequency axis direction to obtain a new voice characteristic pattern. Since it is created, it is the same word as the original standard pattern, but it is possible to easily create a standard pattern of voice corresponding to the case where the speakers are different.

According to a tenth aspect of the present invention, there is provided a standard pattern registration apparatus using the voice pattern creating apparatus according to the ninth aspect, wherein the characteristic patterns of a plurality of voices are used as standard patterns in the dictionary as they are. Further, the pattern creating means shifts each of the plurality of standard patterns to a high frequency band and / or a low frequency band in the frequency axis direction by a certain amount to register a new voice pattern. The whole or a part of the new voice pattern is registered as a standard pattern in the dictionary together with the standard pattern registered as it is by the means. It can be easily created and can be applied to a voice recognition device for an unspecified speaker.

According to the eleventh aspect of the invention, the similarity between the original standard pattern registered in the dictionary as it is and the new standard pattern created by the pattern creating means is calculated, For new standard patterns whose similarity to the original standard pattern is larger than a predetermined value, it is deleted from the dictionary, so that it may cause confusion in confusing words in speech recognition. You can remove various standard patterns.

According to the twelfth aspect of the invention, the similarity between the original standard pattern registered as it is and the new standard pattern is obtained, and the similarity with the original standard pattern is determined. The new standard pattern smaller than the given value is deleted from the dictionary, so that the standard pattern unnecessary for voice recognition can be removed.

According to the thirteenth aspect of the present invention, the registration means, when registering the standard patterns of a plurality of voices as they are, registers the standard patterns in the dictionary in association with predetermined words. On the other hand, the pattern creating means is designed to register a new standard pattern in the dictionary without associating it with a word, and thus is somewhat similar to the voice corresponding to the original standard pattern. When a different voice is input, the voice can be rejected when the voice is most similar to the new standard pattern, and it is possible to effectively prevent erroneous recognition.

According to the fourteenth aspect of the invention, the registration means, when registering the standard patterns of a plurality of voices as they are, registers the standard patterns in the dictionary in association with predetermined words. Also, the pattern creating means is also adapted to register a new standard pattern in a dictionary in association with a predetermined word, so that it is somewhat similar to the voice corresponding to the original standard pattern. When a different voice is input, the voice can be correctly recognized even if the voice is most similar to the new standard pattern.

[Brief description of drawings]

FIG. 1 is a diagram showing a configuration example of a voice recognition device to which a voice pattern creation device and a standard pattern registration device according to the present invention are applied.

FIG. 2 is a flowchart showing an operation of the voice recognition device of FIG.

FIG. 3 is a diagram showing an example of registration of standard patterns in a dictionary.

FIG. 4 is a diagram showing how standard patterns A, B, and C are each divided into three equal parts in the time axis direction.

5 is a diagram showing an example of a new standard pattern created by combining standard patterns divided into three equal parts in the time axis direction as shown in FIG.

FIG. 6 is a flowchart showing a flow of processing of dividing each of a plurality of standard patterns in the time axis direction to create a new standard pattern by combining them.

FIG. 7 is a flowchart showing a flow of a process of dividing each of a plurality of standard patterns in the time axis direction and creating a new standard pattern by combining them.

FIG. 8 is a flowchart showing a flow of a process of dividing each of a plurality of standard patterns in the time axis direction and combining them to create a new standard pattern.

FIG. 9 is a diagram showing how standard patterns A, B, and C are each divided into two equal parts in the frequency axis direction.

FIG. 10 is a diagram showing an example of a new standard pattern created by combining standard patterns bisected in the frequency axis direction as shown in FIG.

FIG. 11 is a flowchart showing a flow of processing for generating a new standard pattern by performing an operation of shifting one standard pattern a (i, j) to a high frequency band with respect to the frequency axis.

FIG. 12 is a flow chart showing the flow of processing for obtaining the similarity between all the standard patterns of the original standard pattern and the new standard pattern, and erasing the standard pattern whose similarity is larger than the determined value. is there.

FIG. 13 is a diagram showing an example of a similarity table between standard patterns.

FIG. 14 is a flowchart showing the flow of processing for checking the similarity between the original standard pattern and the new standard pattern and erasing the standard pattern whose similarity to all the original standard patterns is smaller than a predetermined value. Is.

[Explanation of symbols]

 1 voice input unit 2 A / D converter 3 feature conversion unit 4 voice section detection unit 5 dictionary 6 standard pattern creation unit 7 recognition unit 11 registration unit 12 division unit 13 pattern generation unit SW switch

Claims (14)

[Claims] 1. A new voice feature pattern is created by combining and connecting a dividing unit that divides each of a plurality of voice feature patterns into a plurality of parts, and the respective portions of the plurality of feature patterns that are divided by the dividing unit. And a pattern creating means for performing the same. 2. A standard pattern registration device using the voice pattern creation device according to claim 1, further comprising registration means for registering a plurality of voice feature patterns as standard patterns in the dictionary as they are, The dividing means divides the standard pattern into a plurality of parts, and the pattern creating means combines and connects the respective parts of the divided plurality of standard patterns to generate a new characteristic pattern, which is directly registered by the registering means. A standard pattern registration device characterized in that the new feature pattern is registered in the dictionary as a new standard pattern together with the standard pattern registered in the form. 3. The standard pattern registration apparatus according to claim 2, wherein the pattern creating means divides each of the plurality of standard patterns into a plurality of parts, and combines and connects these parts to create a new standard pattern. In addition, the standard pattern registration device is characterized in that the portion at the beginning of the standard pattern before division is combined so as to be positioned at the beginning of the new standard pattern after connection. 4. The standard pattern registration device according to claim 2, wherein the pattern creating means divides each of the plurality of standard patterns into a plurality of parts, and combines and connects them to create a new standard pattern. In addition, the standard pattern registration device is characterized in that the part at the end of the standard pattern before division is combined so as to be positioned at the end of the new standard pattern after connection. 5. The standard pattern registration device according to claim 2, wherein the pattern creating means divides each of the plurality of standard patterns into a plurality of parts, and combines and connects them to create a new standard pattern. In addition, the part at the beginning of the standard pattern before division is positioned at the beginning of the new standard pattern after concatenation, and the part at the end of the standard pattern before division is in the new standard pattern after concatenation. A standard pattern registration device characterized in that it is also combined so that it is located at the end. 6. The standard pattern registration device according to claim 2, wherein said dividing means divides each of a plurality of standard patterns of voice in a time axis direction, and said pattern creating means is divided in a time axis direction. By combining and connecting each part of each standard pattern according to a predetermined rule to create a new voice pattern, by repeating this process, all or part of the created voice pattern can be registered as a standard pattern. A standard pattern registration device characterized in that 7. The standard pattern registration device according to claim 6, wherein the dividing means divides the standard pattern into equal parts in the time axis direction, or detects a portion in which the standard pattern changes largely in the time axis direction. A standard pattern registration device characterized in that the standard pattern is divided at this portion, or a minimum portion of voice energy in the time axis direction is detected, and the standard pattern is divided at this portion. 8. The standard pattern registration device according to claim 2, wherein the dividing unit divides each of the plurality of standard patterns of voice in the frequency axis direction, and the pattern creating unit divides in the frequency axis direction. By combining and connecting each part of each standard pattern according to a predetermined rule to create a new voice pattern, by repeating this process, all or part of the created voice pattern can be registered as a standard pattern. A standard pattern registration device characterized in that 9. A pattern creating means for creating a new audio feature pattern by shifting each of a plurality of audio feature patterns to a high frequency band and / or a low frequency band in the frequency axis direction by a fixed amount. A voice pattern creating device characterized by the above. 10. A standard pattern registration device using the voice pattern creation device according to claim 9, further comprising registration means for registering a plurality of voice characteristic patterns as standard patterns in the dictionary as they are, The pattern creating means shifts each of the plurality of standard patterns to a high frequency band and / or a low frequency band in the frequency axis direction by a certain amount, and when a new voice pattern is created, is registered in the same form by the registration means. In addition to the standard pattern that is
A standard pattern registration device characterized in that all or part of the new voice pattern is registered in a dictionary as a standard pattern. 11. The standard pattern registration device according to claim 2 or 10, wherein an original standard pattern registered in the dictionary as it is and a new standard pattern created by the pattern creating means are provided. Seeking similarity,
A standard pattern registration device characterized in that a new standard pattern whose similarity to the original standard pattern is larger than a predetermined value is deleted from the dictionary. 12. The standard pattern registration device according to claim 2 or 10, wherein the similarity between the original standard pattern registered as it is and the new standard pattern is calculated to obtain the original standard pattern. The standard pattern registration device is characterized in that a new standard pattern whose similarity is smaller than a predetermined value is deleted from the dictionary. 13. The standard pattern registration apparatus according to claim 2 or 10, wherein said registration means associates the standard pattern with a predetermined word when registering the standard patterns of a plurality of voices as they are. The standard pattern registration device is characterized in that, while registering in a dictionary, the pattern creating means registers a new standard pattern in the dictionary without associating it with a word. 14. The standard pattern registration device according to claim 2 or 10, wherein said registration means associates the standard pattern with a predetermined word when registering the standard patterns of a plurality of voices as they are. A standard pattern registration device, characterized in that the standard pattern registration means registers in a dictionary, and the pattern creating means also registers a new standard pattern in the dictionary in association with a predetermined word.