JPS61102698A - Standard pattern adaption system - Google Patents

Abstract

(57) [Summary] This bulletin contains application data before electronic filing, so abstract data is not recorded.

Description

[Detailed Description of the Invention] (Field of Industrial Application) The present invention uses voice recognition! ! #! This invention relates to a method for adapting a standard pattern of a device to a user's voice quality.

(Prior art and its problems) Speech recognition technology has improved rapidly in recent years and is now widely used for inputting data to computers, inputting control commands to various mechanical systems, etc. However, its recognition performance is not ideal, and there are still problems such as the need for standard pattern registration by the user's own voice and the possibility of erroneous recognition.

All of these existing systems are designed to operate on the principle of pattern matching. The principle behind this is that a symbol pattern is prepared for each word, and when an unknown cover pattern is given, a determination is made by comparing these symbol patterns and searching for the most similar symbol pattern. Usually, the mark pattern is registered by the user using his or her own pronunciation. This is to deal with the problem of individual differences, where even the same word has significantly different speech characteristics.

However, even when the same word is uttered by the same person, the speech patterns vary. Regarding expansion and contraction in the time direction, a pattern matching method using dynamic programming (called DP matching) has been developed and effectively used, as described in the specification of Japanese Patent Application No. 45-53896. has been done.

In addition to this, there are changes in the frequency direction, such as vowel devoicing and nasal vowelization.
No effective measures such as matching have been found.
′ As a measure against this kind of variation (hereinafter simply referred to as variation), we prepare multiple standard patterns for the same word, and by doing so, the actual input pattern matches one of the standard patterns. A method is taken to try to do so. In other words, by uttering and registering each word multiple times and using these words with the expectation that they are representative examples of fluctuations in the frequency direction that can actually occur, errors caused by pattern changes can be avoided. It seeks to reduce awareness. However, in order to prepare a large number of standard patterns, a large capacity memory is required, and as a result, the speech recognition device becomes expensive. In addition, even if the user registers to speak many times,
Therefore, there is a problem in that there is no guarantee that all pattern variations that may actually occur can be covered. On the contrary, ordinary users often become nervous and repeat a certain, stiff utterance, and are often unable to register various fluctuating patterns.

(Objective of the Invention) The present invention improves the above-mentioned drawbacks of the conventional standard pattern adaptation method, improves the memo-reading efficiency of standard patterns, and recognizes speech patterns with high accuracy even with variations that have not been registered. (T quasi-pattern - realizes an adaptive method,
The purpose is to provide a high-performance and inexpensive voice recognition device.

(Structure of the Invention) According to the present invention, the standard pattern adaptation method expresses a mark pattern as a time-series network of features, and matches it with a voice pattern uttered by a user to find an optimal path in the network. A mark pattern adaptation method is obtained, which is characterized in that the characteristics on the optimum path are modified by the sound pattern.

Further, according to the present invention, an optimal path is determined in the network by matching a standard pattern expressed as a time-series network of features with a voice pattern uttered by the user, and the above-mentioned features on the optimal path and the A standard pattern adaptation method is obtained in which the standard pattern adaptation method is characterized in that the standard 4A turn feature is modified by the voice pattern feature only when the feature in the corresponding voice pattern is similar to or above a predetermined criterion.

Further, according to the present invention, an optimal path is determined in the network by matching a standard pattern expressed as a special time-series network with a voice pattern uttered by the user, and the above-mentioned characteristics on the optimal path and corresponding responses are determined. In a portion where the features in the voice pattern are similar to each other by more than a predetermined standard, the standard pattern feature is modified with the voice pattern feature, and in a portion where the feature is not similar by more than a predetermined standard, the standard pattern feature is modified in parallel with the portion. A standard pattern adaptation method is obtained in which a new path is created based on the voice pattern features.

(Operation) Japanese Patent Application No. 57-156413 discloses a method for recognition by expressing a quasi-culm pattern as a time-series network of features and matching it with a time-series standard pattern. Representing standard patterns in this way over a network has the advantage of high memory efficiency for storing standard patterns. Now, as an example, consider the case of the numeric sound "8" (/hatfi/).

This word is usually pronounced /hstfi/, but sometimes the initial vowels /, / are devoiced to become /a/ (the "Q" at the bottom means devoiced), or the final vowel /i/ is known to be devoiced to become /i/. Since these phenomena occur independently of each other, the result is (a) /hatfi /, (b) as shown in Figure 1.
/ h a t f i /, (c) / h a t
f i /, (d) / h at f r /,
A total of four types of fluctuations will occur. Fig. 1 is a so-called automaton representation, starting from the initial state 0, taking the phonemes, a(a), , t f , 1(i), and proceeding sequentially to states 1.2.3. , final state 4
and ends.

Moving from one state to the next in response to some input (in this case, phonology) is called a transition. In this case, the input may be a symbol indicating a phoneme, but generally, each input (for example, corresponding to /a/) may be a time series of quantities indicating characteristics (for example, ictor). .

(a), (b), (c), (d) in Figure 1
This example shows a case where an independent standard pattern is prepared for each type of transformation (combination of). In such a case, a method was filed on August 1, 1982 that allows recognition of variations in types that were not registered by the user's own voice at a high bGW degree, and the patent specification ``Standard Pattern Registration'' was filed on August 1, 1982. method”. However, such a method originally had the drawback that it required a large amount of memory because standard patterns had to be prepared independently for various variations.

. Therefore, the present invention is based on the premise that the standard pattern is expressed as a network-like (including branching) automaton as shown in FIG. This representation has the advantage that the amount of memory for storing the standard pattern is significantly reduced. In other words, the amount of memory is roughly proportional to the number of states (nodes) of the automaton and the number of transitions (branches) that control the states (nodes).

(a), (b), fc), fd) in Figure 1
In contrast to the total of 16 pieces in Figure 2, there are 6 pieces in Figure 2.

The expression of an automaton with branches in this way is called a node, and "→" (state transition) is called an edge, based on terminology in graph theory. Note that for each school ζ in FIG. 2, symbols such as /h/ and /a/ do not exist, but are similar to the quasi-pattern description in the specification of Japanese Patent Application No. 57-156413. The time series of voice features corresponds to

As an example, the features corresponding to the sound @/h/ are represented by vector sequences b1b, b. Similarly, for each phoneme /i/
Bla bts bt.

The vector sequences are associated with each other. Said patent application Sho 57-156
The specification of No. 413 describes such a standard pattern as B"" blbz b3 (b4 kg b6 b7
Tol/bg bz6 Toll ) btt bts
(btt ILLl bta b!q /b
ts bto bto ) (expressed as 21, entering Kabatan = as bz °°=' a A°゛゛a r
(31) is described. Using this pattern matching device, (2
) By preparing a standard pattern for each required word and performing the pattern matching method, stable recognition can be performed with a small amount of standard pattern memo IJit and even against pattern fluctuations.

The problem here is how to create standard pattern B. In contrast, in the present invention, the following two
The basic policy is to do this.

+11 The expected types of variations are incorporated in advance as techniques in the network and a standard standard pattern is prepared. This process can be performed visually and manually using a large number of sample utterance patterns.

(2) As it is, the recognition rate is expected to be low because the user's own speech pattern is not taken into account.

Therefore, this original standard pattern is modified based on the person's vocalization.

Of the above, (1) may be offline processing with human intervention, and the present invention does not relate to that method, so further explanation will be omitted.

Hereinafter, a method for executing 2) above will be described.

Now assume that the network in Figure 2, that is, B in equation (2), is given as the original standard pattern, and the person in equation (3) is the voice pattern uttered by the user himself/herself, and this is used to create the original standard pattern B. Let's discuss the case of modifying.

The voice pattern A and the original standard pattern B are matched to determine which feature of the standard pattern each feature 111 of the person entering the cover pattern is associated with.

As a method for determining this correspondence, the above-mentioned patent application No. 57-1
DP using the stack described in No. 56413
This is also possible by improving and using a matching method, but here we will show an example of a method that is easier to understand. Expanding the network in Figure 2, we get (a) and (b) in Figure 1.
−(c) * Convert to the form shown in (d). (ILf
2) When considered in the form of an equation, it is converted into the following four types of time series〇To show these in a unified manner, resubscripts are added as C:C
1C! It is expressed as ゛°”°゛ 1°”-” CJ (4).For example, this is the case of B4.

Patent Application No. 1982-62782 The correspondence between the features al and (J is determined as mapping j = j (i) between the time series A and C such as Kohoko, and the correspondence is established. A method for minimizing the distance between features as a whole is described.The resulting mapping j = j
According to (i), the sounds @/h/, /, /, /lf/.

It is said that the vector 1i belonging to /i/ and cI are accurately correlated. For this purpose, the patent application No. 46-6278
In specification No. 2, u21. A so-called DP matching process is executed in which calculation is performed based on the dynamic programming recurrence formula % formula % on the u31 formula. Here, d(1,j) is the distance II −1−s It between al and cj. In parallel to the calculation of equation (6), W (i, j) = 0゜1.2 depending on each case where the minimum value in the right side [] is the first equation and the second equation.
Store the values in a table. Based on this, mapping j(i) is determined by the following recursive process.

j (I) = J j (i-1) = j (i) - W (i, j (i))
(8) Also, g(1,
J) is known to be the distance D(person, C) between patterns A and C.

Now, as pattern C, pattern Bl is developed from the above B.
m Bm @ g, e Substitute B4 and get (6) (7
By repeatedly performing DP matching using equation 1, the distance D(
person, B1).

D (person, B, ), D (person = B, ), D (A-84)
Find Brn where these distances are the minimum, then
This voice pattern person has the same tendency of fluctuation as Bff. Therefore, B in the original standard pattern B
The features on the branch sequence (referred to as the optimal path) corresponding to m are modified using the features of the voice pattern person. (Then, as long as the fluctuations have the same tendency, the user's own characteristics will be used for matching, making it possible to perform accurate recognition operations.

The operation for modifying the features on the optimal path described above will be explained using a specific example. Assume that the development pattern with the minimum distance D (person, Bm) is B4. In this case, the feature on the path, ie, t+ in equation (5), will be modified. Mapping j = j obtained by equation (8) (1) should be associated with lees in time series C by (i) ,l
Replace by. For example...j(31=4.j(4)=4.j(5
) = 6..., then make the correction by replacing 1.

Such modification is preferably carried out by repeatedly applying a plurality of voice patterns (naturally, of the same word) to one original standard pattern. In this way, the adaptation of the standard pattern (to the user) is completed.

The standard pattern B obtained in this way requires a small amount of memory because it is expressed in a network. Adaptation has been completed to the types of fluctuations that appear in the user's own utterances. Furthermore, for types of variations that do not happen to appear in the user's own utterances, branches given as the original standard pattern are left. Therefore, such standard pattern adaptation processing is performed for each word, and using these standard patterns B, pattern matching with the input cover pattern is performed by the means described in the above-mentioned Japanese Patent Application No. 57-156413. This makes it possible to perform compact and highly accurate speech recognition.

(Embodiment) FIG. 3 is a block diagram for explaining a first embodiment of the present invention. The control unit 10 receives control signals fi, I'fl
generates and controls the entire operation. This recognition device operates in an adaptive mode to adapt the standard pattern to the user's own characteristics, and in actual recognition.

It is divided into recognition mode and mode. More and more misfires
: Describe the adaptation mode in which Ming is directly involved.

Now, the numbers “O”, “1”...n, are the words to be recognized.
...Think of "9". Original standard pattern B for each chapter
1 is stored in the storage unit 14 in the format illustrated in equation (2). These are designated and read out by the control signal n from the control section 10, and sent to the buffer 15 for adaptive processing. An example will now be shown in which a standard pattern Ba of the number 8 is adaptively processed with n=8. The subscript 8 will be omitted and the description will be made using the form B of equation (2) corresponding to FIG. 2.

The voice of "8" uttered by the user is inputted from the microphone 11, and analyzed by the preprocessing unit 12, (3)
The voice pattern of the expression is converted into a human form and input into the buffer 13. After this point, the control signal m from the control unit 10 is changed to take values of 1, 2, 3, and 4°5. In response, the buffer 15 generates patterns B", B", BS, and number B, which are developed from the original pattern, and each of them is sent to the buffer 16 as pattern C in equation (4), and the input cover pattern is The M1 matching unit 17 performs matching processing with the person. The first matching section 17 is the
It is configured similarly to the figure, and the above f6), (7), (
A distance n (person, C) and a mapping j (i) are output by processing such as the equation s).

The minimum value detection unit 18 determines m:m' where the distance D (A, C) calculated for each C=Bm is the minimum. The mapping j(i) corresponding to this m:m' is held in the mapping buffer 19. Assume that m'=4 as in the previous explanation of the principle. The conversion storage unit 20 stores the correspondence between tl and bk (correspondence as shown in equation (5)) for generating B4. This correspondence and the mapping buffer 1
As illustrated in equation (9) using the mapping held in 9, the process of modifying the standard pattern is modified by determining the vector Tok corresponding to -j(+) and replacing it with 11. It is executed in section 21. The standard pattern modified in this way is stored in the storage section 14 as a number 4.
'' is stored as a standard pattern Ba.

After performing such processing for numbers 0 to 9, the adaptive mode ends.

Next, the recognition mode will be briefly explained. In this mode, the pattern input to buffer 13 is an unknown input pattern. This pattern is guided to the second matching section 22, where the adapted standard pattern Bn (n=0.1.--9) held in the storage section 14 is used.
compared to The structure of this matching section is the same as that shown in FIG. 7 of the above-mentioned specification of Tokuton Sho 57-156413. However, this document exemplifies the application to online character recognition, and in the case of the present invention targeting speech recognition, the distance calculation unit 26 in FIG. 7 is used to calculate the distance between the feature al and the feature bl. Change the format to

This matching section calculates the distance D (A, B") between the incoming cover turner and each standard pattern Brl. These are compared with each other by the minimum value detection section 23, and the minimum value n"n is determined. . This n is output as the recognition result.

Although the first embodiment has been described above, the following problems remain in the first embodiment. To describe the example of FIG. 2, the original standard pattern B has two variations for the vowels / and /: normal utterances /, / and devoiced utterances /, /. In addition to this, it is known that there is a phenomenon called nasalization in the variation of vowels, /? A variation pattern of / (~ indicates nasalization) can be uttered. Now, if there is a variation of /7/ in speech pattern A, if we perform adaptation using this, either branch of /, / or /, / (
Here /, /) are replaced by the feature of /7/. In this case, the normal vowel /, / branches disappear, /h, t f i/ or / ha
Recognition of the utterance t f I / will deteriorate.

The second embodiment described below is a standard pattern that does not cause deterioration in the original standard pattern even if adaptive processing is performed using a speech pattern person having a type of variation that did not exist in the original standard pattern B as described above. It provides an adaptive method.

A second embodiment of the present invention determines an optimal path in a network by matching an obstacle pattern expressed as a time-series network of features with a voice pattern uttered by a user, and and the corresponding feature in the voice pattern are similar to each other by more than a predetermined criterion, the standard pattern feature is modified by the voice pattern feature.

If we decide to adapt in this way, the recorded vowel /7/ will be applied to both the normal vowel /a/ and the devoiced vowel /, /.
are patterns that differ by more than a certain degree, so the similarity is below the reference value, and the /, /, /, / branches will not disappear due to modification, and therefore the phenomenon of deterioration of the original standard pattern can be avoided.

In FIG. 3, the first matching section 17 (6)
In order to execute the formula, the distance d(i, D) between 21 and C3 is calculated. When the correction section corrects the vector 6k corresponding to c+(1), the distance d(i, D) between 21 and C3 is calculated.

If j(i)) is less than a predetermined reference value, the process of replacing bk with 21 is omitted.

In this way, even if adaptation is performed using a voice pattern that is not included in the original standard pattern, the standard pattern will not deteriorate.

According to the second embodiment described above, the standard pattern deteriorates even if adaptive processing is performed using a speech pattern that has a type of variation that is not included in the original standard pattern that was left in the embodiment of tJI,l. This problem could also be prevented. However, even in this second embodiment, variations that are not incorporated as branches of the original standard pattern network will remain permanently in the standard pattern even if adaptive processing is performed using a pattern person with such variations. The drawback remains that it is not incorporated into the system.

For example, the original standard pattern B in Figure 2 does not include the branch /7/ in which / and / are nasalized, so /h'7tf
Even if adaptive processing is performed using a speech pattern person with i/, the processing of the second invention does not generate a branch with the characteristics of /7/, and therefore /h7*fx/ or /h7tfi
/ speech pattern recognition does not work well.

The third embodiment described below also improves on the above drawbacks and is capable of learning and adapting to types of variations that were not incorporated into the original standard pattern network.

A third embodiment of the present invention determines an optimal path in the network by matching a standard pattern expressed as a time-series network of features with a voice pattern uttered by a user, and and the corresponding features in the voice pattern are similar to each other by more than a predetermined standard, the standard pattern features are modified by the voice pattern features, and in parts where they are not similar by more than a predetermined standard, The method is characterized in that a new path parallel to the part is created based on the voice pattern feature.

Consider the original standard pattern B shown in Figure @2 or equation (2). On the other hand, it is assumed that adaptation processing is performed using voice pattern A of /h7tfi/.

The correction unit 21 that refers to the mapping Ni) obtained in the first matching unit in FIG. 3 and the distance d(i, j) calculates the distance d (
i, j(i)) is /a/ or /, / and /7/
It can be detected that the value increases due to the discrepancy between the two. In these parts, as shown in FIG. 4, the features of /7/ are extracted from the voice pattern and arranged, and new branches are added. (2
) For example, the formula is as follows: OB = bt bt bs (h4 b5 b
e b ding bs / be also 10”11/as
ah as ) bat b+3 (1114
its bts bst /1itsbtu b
to ) OCj Here, 23242g is extracted from the speech pattern human as a feature corresponding to the nasalized vowel /, /. In this way, it has become possible to automatically learn and adapt to variations that are not incorporated in the original standard pattern B in advance.

Although the principle of the present invention has been explained above based on examples, these descriptions do not limit the scope of the rights of the present invention.

In particular, as a method to correct standard pattern feature J by al (E5), replace vf of bj by al with +""1""pa"
2゛(1) (tt:t')"F-14icr6
.

You can also do that. In addition, there are other methods for expressing and handling network standard patterns (for example, Japanese Patent Application No. 1982-
131438) may also be used. Furthermore, the unit of sound is not a word such as a number, but may be another unit such as a consonant, a vowel syllable, a consonant, a vowel, or a consonant sound.

(Effects of the invention) When such an adaptation method is used, the standard pattern memory is small, the speaker adaptation is good, and the speaker adaptation can correctly recognize type 1 variations that were not input in the adaptation mode. It has become possible.

[Brief explanation of the drawing]

1, 2, and 4 are principle explanatory diagrams, and FIG. 3 is a block diagram of an embodiment. In the figure, 10 control unit, 11...microphone, 12...
・Preprocessing unit, 13...Buffer, 14...Storage unit, 15.Buffer, 16...Buffer, 1
7... First matching section, 18 - Minimum value detection section, 1
9...Photo'<? Buffer, 20... Conversion storage unit, 2
1...<6 positive part, 22-2 matching part, 23... Minimum value detection part.

Claims (3)

[Claims] (1) An optimal path is determined in the network by matching a standard pattern expressed as a time-series network of features with a voice pattern uttered by the user, and the above-mentioned features on this optimal path are modified by the above-mentioned voice pattern. A standard pattern adaptation method featuring: (2) An optimal path is determined in the network by matching a standard pattern expressed as a time-series network of features with a voice pattern uttered by the user, and the above features on the optimal path and the corresponding voice pattern are determined. A standard pattern adaptation method characterized in that the standard pattern feature is modified by the voice pattern feature only when the features in the middle are similar to each other by a predetermined criterion or more. (3) An optimal path is determined in the network by matching a standard pattern expressed as a time-series network of features with a voice pattern uttered by the user, and the above features on the optimal path and the corresponding voice pattern are determined. In a part where the features are similar to each other by more than a predetermined standard, the standard pattern feature is modified with the voice pattern feature, and in a part where the features are not similar by more than a predetermined standard, a new path parallel to the part is created. , a standard pattern adaptation method characterized in that it is created based on the voice pattern characteristics.