JPS63250698A - Voice recognition equipment - 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] (Industrial Application Field) The present invention calculates the similarity between an input speech and a dictionary phoneme sequence of each dictionary entry in a word dictionary in which phonemes are expressed using a phoneme likelihood matrix. The present invention relates to a speech recognition device that recognizes words by calculation.

(Prior Art) A first example of a conventional word speech recognition device will be explained based on FIG. FIG. 2 is a flowchart of the speech recognition device. In the figure, input speech is analyzed to extract parameters, and then phoneme recognition is performed. This recognized phoneme sequence is compared with the dictionary phoneme sequence of each dictionary item in the word dictionary, the likelihood between each corresponding phoneme is determined by referring to the likelihood matrix, and the phoneme sequence is calculated as the sum of the likelihoods of each phoneme. Find the sum of like-like likelihoods, divide this by the number of phonemes, and find the similarity between phoneme sequences.
The dictionary item with the maximum degree of similarity is taken as the recognition result.

Here, the word dictionary is written in phonemes, and phonemes are units almost similar to the Roman alphabet, and each item in the word dictionary in this conventional example is 5APPOR○, TOO
KY○0 (Tokyo).

It is expressed as 00SAKA (Osaka), etc.

Figure 3 shows the phoneme conflation matrix (hereinafter referred to as C
, M). In the figure, the vertical axis shows the phonemes in the dictionary phoneme series, and the horizontal axis shows the recognized phonemes. The probability that (i) is not recognized and is omitted (its existence is ignored) is 8%. This C,M
, are logarithmized to form the above likelihood matrix. C and M shown in FIG. 3 perform phoneme recognition on a large number of voice data whose words are known in advance.

It is created based on the phoneme recognition results. Sunahchi T
becomes A! ! The probability of being recognized is 0%, which means that C,
In the audio data used to create M, ■ is A.
There was nothing recognized as such. This C
When performing word recognition using ,M, if I is recognized as A in the phoneme recognition result of a certain input speech, the probability that such a phoneme recognition result will occur from the dictionary phoneme sequence of this input word is 0. Therefore, there is no possibility that the input word will be recognized correctly. In this way, if a correspondence relationship between phonemes that did not exist in the phoneme recognition of the audio data used to create C and M occurs, there is no possibility that the word will be recognized correctly. ,M, for each element.

A likelihood matrix used for recognition was created by assuming that everything below 0.1% is 0.1%.

However, the first conventional example described above had drawbacks as shown in the following example. The input voice is Nabeshima (dictionary phoneme series is N
ABESIMA), and the recognized phoneme sequence is N
When ABESIMAWA was obtained, the word recognition result was NAMERIGAWA (dictionary phoneme sequence is NAMERIGAWA). It is almost impossible to recognize the phoneme R as S as long as the pronunciation is normal, but in the conventional example, R Since the probability of recognizing S as S was assumed to be 0.1%, it was mistakenly recognized as Namerigawa, which has a very similar phoneme sequence in other parts.

There is a second conventional example described below as a speech recognition method that addresses the drawbacks of the first conventional example. The difference between the second conventional example and the first conventional example is that when C and M are created as shown in FIG.
%, instead of considering it as 0.1%, humans should judge in advance whether the correspondence between the phonemes is possible based on the nature of the speech, and if it is possible, the first step should be taken. Similar to the conventional example No. 1, the likelihood matrix of the phoneme is created by considering it as 0.1%, and if it is impossible, assuming it as 0.0001%. Here, the impossible case is 0
The reason why I set it to 0.0001% instead of % is because if I set it to 0% at all, it would become 1 ψ when logarithmized to create a likelihood matrix, so I set it as a value that can be considered as 0%. It is. According to this second conventional example, even in the case of erroneous recognition in the first conventional example, it is possible to use the information of the likelihood matrix that it is almost impossible to recognize R as a phoneme S. Recognized phoneme series NABESIM
AWA and Namerigawa dictionary phoneme series NAMERIG A
The degree of similarity with W A becomes small, and the word is correctly recognized. Furthermore, this second conventional example also deals with the following case. Human voice is symmetrical (dictionary phoneme series is S IN
MACI) and the recognized phoneme sequence is SINOMACI,
In the first conventional example, the word recognition result is Ishinomaki (I S
INOMAKI). Here, the probability of omission of the word-initial word is 8%, so it is mistaken for Ishinomaki, which resembles the part below S in the phoneme sequence. Dealing with this problem is not as simple as in the Nabeshima example, since the T at the beginning of a word can actually be dropped. Here, in the second conventional example, this problem was dealt with using the rules described below.

In other words, as long as Ishinomaki is pronounced clearly, the first I is a voiced sound, and the next S is a voiceless sound with a long duration.
It is not possible to include ■ in S at all and recognize the phoneme together with IS as S. If the pronunciation of (①) is a little weak, it may not be recognized as a ``work'', but in this case, it becomes a U or a nasal sound, or even if it is combined with an S, it is still voiced, so it is recognized as a Z, etc. In other words, a rule is set that IS may be recognized as US, NS, Z, etc., but not as S, and for phoneme correspondences that violate this rule, the likelihood obtained by the likelihood matrix is The likelihood is subtracted. As a result, in the case of the above-mentioned input example of symmetrical lines, when the dictionary entry is Ishinomaki, matching is performed such that the initial word is dropped and S is correctly recognized as S, but since this is against the above rule, the likelihood is points are deducted, and the degree of similarity naturally decreases compared to the case of the first conventional example. On the other hand, since dictionary item symmetry has no relation to the rules and the degree of similarity is the same as in the first conventional example, in this input example, correct word recognition results can be obtained in the second conventional example. This second conventional example has some of the above-mentioned rules, prevents high similarity from being obtained even when there is an unlikely phoneme correspondence due to the nature of speech, and achieves high recognition performance. However, this second conventional example has the disadvantage that the time required for similarity calculation increases significantly, since comparison with a large number of rules is performed each time the likelihood of each phoneme is calculated.

(Problems to be Solved by the Invention) In the first conventional example described above, when input speech is matched with a dictionary entry for a word different from the input, it contains phoneme correspondences that are obviously impossible due to the nature of the speech. However, a high degree of similarity was obtained, but the problem was that word recognition could be incorrect. Although the second conventional example solved the problems of the first conventional example, it had the disadvantage that the time required for recognition increased significantly.

SUMMARY OF THE INVENTION An object of the present invention is to provide an excellent speech recognition device that eliminates the conventional drawbacks, has good recognition performance, and requires less time for recognition.

(Means for Solving the Problems) The speech recognition device of the present invention includes a word dictionary in which phonemes are expressed,
In phoneme recognition, each phoneme in this word dictionary uses a likelihood matrix that indicates which phoneme is recognized, dropped, or added, a phoneme recognition means, and the likelihood of this phoneme. A means for calculating the similarity between a dictionary phoneme sequence and a recognized phoneme sequence using DP, and a table showing rules for determining whether the correspondence between a phoneme in a word dictionary and a recognized phoneme is appropriate based on the nature of speech. It has a table in the same format as the matrix, and uses a likelihood matrix to determine the likelihood of correspondence between each phoneme, and among the elements of this likelihood matrix, elements with low probability of occurrence.

If the element cannot occur due to the nature of speech, it gives a probability O or a value that can be considered as O. If the element can occur due to the nature of speech, it gives a finite value close to O but not 0. degree matrix.

A rule that not only determines whether or not it is possible due to the nature of speech, but also based on the results of phoneme recognition, omissions, or additions of phonemes between the phoneme and the previous phoneme. When recognizing input speech using the above configuration, first perform phoneme recognition of the input speech, calculate the similarity between the obtained recognized phoneme sequence and the dictionary phoneme sequence using DP, and then The sum of the likelihoods of the corresponding points of each phoneme is calculated from the beginning of the sequence, and the correspondence between the phonemes is determined as a sound, including the phoneme recognition results.

Check whether it is appropriate due to the nature of When a rule is used, the likelihood calculation is performed by applying the rule shown in the table.

(Function) According to the present invention, when calculating the likelihood, by referring to a table of rules having the same format as the likelihood matrix,
High Ly! with little increase in recognition time! A p'ra rate can be ensured.

(Example) An embodiment of the present invention will be described based on FIG.

FIG. 1 is a flowchart of the operation of the speech recognition device of the present invention. This embodiment is an improvement on the first and second conventional examples, and its operation is also based on the conventional examples.

In FIG. 1, input speech is analyzed, parameters are extracted, and first-order phoneme recognition is performed. This recognized phoneme sequence and the dictionary phoneme sequence of each dictionary entry in the word dictionary are referred to, the likelihood between each corresponding phoneme is calculated, and the sum of the likelihoods of each phoneme is calculated to measure the similarity between the phoneme sequences. The dictionary item with the highest degree of similarity is determined as the recognition result.

Here, the word dictionary is the same as in the first and second conventional examples, and the likelihood matrix is the same as in the second conventional example. That is, it has the same format as the first conventional example, and is created by logarithmizing the probability from the phonemes C and M as shown in FIG. 3, as in the first conventional example. M, probability O on
For things that are possible due to the nature of the voice, the probability is 0.1%, and for things that are impossible, the probability is 0.
．． It is created assuming that it is 0001%.

The differences from the conventional example will be described below. The rule table has a similar structure to a likelihood matrix, and each element indicates the correspondence between various phonemes, and is referred to in the same way as a likelihood matrix when calculating the likelihood. Each element has a pair of phonemes that it represents), i. A flag of 0 is set indicating that there is no need to use a rule. In other words, when calculating the likelihood of each phoneme, refer to this rule table, and if the value is O, simply use the likelihood shown in the likelihood matrix, and calculate if the element in the rule table is not O. In this case, the judgment as to whether the correspondence between the phonemes indicated by the element is possible due to the nature of speech depends on the recognition result of the previous phoneme, not on the phoneme alone. - For example, when the symmetry described in the explanation of the conventional example is input, the phoneme recognition result is IS INOMACI, and this is a rule to prevent it from being erroneously recognized by ISINOMAKI (dictionary phoneme series is ISINOMAKI). In other words, the rule is that, although it is possible for an incorrect initial I to be dropped, a word initial whose following phoneme is S will not be dropped, and S will not be recognized as S. This rule itself is stored in the rulebase.

In this case, the address in the rule base of the above rule is included in the element corresponding to the omission of the initial word in the rule table. Here, in calculating the similarity with Ishinomaki when inputting this symmetry, when calculating the likelihood of S: If S is matched with S, it means that the initial I before S has been dropped, so the initial Check the table of rules to see if there are any rules about dropping workers. Then, it is found that the above rule exists, and the address of that rule in the rule base is shown, so by referring to that rule, it is found that the above condition is met, and this is a phoneme recognition result that is impossible due to the nature of the speech. Since it can be judged that, a likelihood corresponding to a probability of 0.0001% is given and the likelihood is significantly reduced.

In this way, according to the above embodiment, when calculating the likelihood of each phoneme, the likelihood value is changed depending on whether or not such phoneme recognition, omission, or addition is possible due to the nature of the speech, and this determination is also performed based on the phoneme. It is possible to obtain high recognition performance by taking into consideration the recognition results of the previous phoneme, not just the recognition result alone, and ensuring that the likelihood is low for phoneme correspondences that are impossible due to the nature of speech.

Furthermore, as a method for determining whether or not a correspondence relationship between phonemes is possible due to the nature of speech, considering the recognition results of previous phonemes, which cannot be handled by simply referring to the likelihood matrix, we have developed a method similar to the likelihood matrix. In most cases where you refer to a table of rules with a table of rules and do not apply any special rules depending on the flags indicated by the elements of that table, you can simply take the likelihood value shown in the likelihood matrix as the likelihood. There is almost no increase in calculation time.

In addition, even when using special rules, the address of the rule base indicated by the flag shown in the table can be immediately referenced, compared to the second conventional example in which application of the rule was determined each time in every case. However, it is much more efficient, and the time required for recognition increases only slightly compared to the first conventional example.

In addition, in the claims, examples, etc. of this patent, the likelihood of representing proximity and the likelihood of similarity are used, but even if distance, which is the likelihood of representing distance, is used, the results are the same, only the magnitude is reversed. That's true.

The present invention also includes cases where distance is used in calculations.

In addition, although the above embodiment shows an example of isolated word recognition, 1
The present invention also applies to recognition of continuous speech.

(Effects of the Invention) According to the present invention, a table of rules is provided in a format similar to a likelihood matrix, and the table of rules is used to determine whether or not there is a correspondence relationship between a phoneme in a dictionary and a recognized phoneme due to the nature of speech. The system makes judgments by referring to the rules, and refers to the rule base as necessary to obtain an appropriate likelihood value.It has high recognition performance and shortens the time required for recognition. , its practical effects are great.

[Brief explanation of drawings]

FIG. 1 is a flowchart of the operation of a speech recognition device according to an embodiment of the present invention, and FIG. 2 is a flowchart of the operation of a speech recognition device in an embodiment of the present invention. ! FIG. 3 is a flowchart of the operation of the recognition device, which is a chart showing a portion of the phonemes C and M. Patent applicant: Matsushita Tsushin Kogyo Co., Ltd. Figure 1 Figure 2

Claims (1)

[Claims] A word dictionary with phoneme notation and each phoneme in the word dictionary,
In phoneme recognition, a likelihood matrix indicating which phoneme is recognized, dropped, or added, a phoneme recognition means, a dictionary phoneme sequence using the likelihood of the phoneme, and a recognized phoneme. A means for calculating the degree of similarity between sequences using DP, and a table indicating rules for determining whether the correspondence between a phoneme in the word dictionary and a recognized phoneme is appropriate based on the nature of speech, and is the same as the likelihood matrix. The likelihood of correspondence, addition, or omission of each phoneme is determined using the likelihood matrix, and for elements of the likelihood matrix that have a low probability of appearance, that is, an element that has a low likelihood, the likelihood of correspondence, addition, or omission of each phoneme is determined by In the case of something that cannot occur due to the nature of
Or, if the element can occur due to the nature of the voice, it is a likelihood matrix that gives a finite value that is close to 0 but not 0. This is not only done for each phoneme independently, but also based on the recognition results of the phoneme, its omission, or the addition of a phoneme between it and the phoneme (in the case of the beginning of a word, the addition of a phoneme to the beginning of the word). When recognizing input speech using the above configuration, first perform phoneme recognition of the input speech, calculate the degree of similarity between the obtained recognized phoneme sequence and the dictionary phoneme sequence using DP, and in the DP calculation, the phoneme sequence Find the sum of likelihoods from the beginning to the corresponding point of each phoneme, and check whether the correspondence between the phonemes is appropriate based on the nature of the speech, including the phoneme recognition results, by referring to the table of rules. , when no special rules are used, the likelihood values shown by referring to the likelihood matrix are used as they are; when special rules are used, the rules shown in the table are applied to calculate the likelihood. A voice recognition device characterized by: