JP3425165B2 - Continuous sentence speech recognition device - Google Patents

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a continuous sentence voice recognition apparatus for recognizing continuously spoken sentence voices, and more particularly, the present invention uses a dependency relation, a semantic relation, and a degree of relation between clauses for faster and more efficient use. The present invention relates to a continuous sentence voice recognition device that can accurately recognize a sentence.

[0002]

2. Description of the Related Art Conventionally, as a sentence-speech recognition method for recognizing continuously-sentence-sentences, a sentence of context-free grammar is CYK.
Parser (method that divides sentence into components) and DP (Dynamic programming, etc.)
A method of recognizing a combination of collations has been used.

[0003]

By the way, the conventional methods described above are limited to context-free grammars that can be handled, and are unsuitable for expressing dependency relations and semantic relations in Japanese, and the degree of association between phrases. There was a problem. That is, the above-described method is suitable for recognizing sentence voices that have a meaning in the order of bunsetsu, such as English sentences, but as in Japanese, dependency relations (joint relations of clauses by particles), meaning It was not suitable for recognizing sentence-speech, which is highly dependent on the relation and the degree of association.

The present invention has been made in view of the above-mentioned problems of the prior art, and it is possible to efficiently recognize continuous sentence speech while considering the dependency relation, the semantic relation, and the degree of association between phrases. It is an object to provide a sentence voice recognition device .

[0005]

FIG . 1 is a block diagram showing the principle of the present invention.
FIG. In the figure, 1 is the input voice and phrase ten
The collating unit 2 that collates the plates and outputs the recognition score,
Refer to the clause group flag to make the next collation in collation unit 1.
The phrase selection section for selecting a phrase, 3 is a case governing grammar table
4 and 4 indicate the phrase group flag indicating the phrase to be selected next.
In the collation unit 1, the stored phrase group flag table and 5 are
The phrase table that contains the phrase template that matches the input speech.
Plate storage unit, 6 is a beam search unit for beam search.
Home search pruning unit, and 7 is a relevance score calculation unit.
It

In order to solve the above-mentioned problems, the invention according to claim 1 of the present invention proposes that a clause is defined in advance according to its case and meaning.
Classify them into clause groups and store them in the case governing grammar table 3,
Referring to the classification result above, the dependency relation and semantic relation of the bunsetsu
In the phrase of the input voice of phrase group flag table 4
Flag any clauses that may appear.

[0007] Then, the phrase group and the input
The matching section 1 matches the phrase of the voice, and a flag is set.
Included in the specified phrase group is actually in the phrase of the input speech
, Select a phrase based on the result, and
Based on the recognition score output by the collation means
To recognize .

According to a second aspect of the present invention, in the first aspect of the invention, the flagged phrase and the input voice phrase are included.
As a result of collating with the above collating means, a flag is set.
The phrase included in the phrase group can be matched in the input voice.
If it is, the clause group flag table corresponding to the matched clause
The flag of is reset . According to a third aspect of the present invention, in the first or second aspect of the invention, a bunsetsu candidate to be collated with a bunsetsu of the input voice is beamed.
Fixed by the beam search method in the search pruning unit 6
I tried to narrow it down to a number. The invention of claim 4 of the present invention is the same as the invention of claim 1, 2 or 3 ,
In the relevance score calculation unit 7, the relation is calculated from the relation between the phrases.
Calculate the degree of relevance score and input the calculated relevance score.
It is added to the result of voice verification .

According to a fifth aspect of the present invention, in the fourth aspect of the invention, co-occurrence relation data between clauses is used as the degree-of-relationship degree data. According to the invention of claim 6 of the present invention, in the invention of claim 4, the adjacency relation data between clauses is used as the degree-of-association degree data. In the invention of claim 7 of the present invention, in the invention of claim 4, as the degree- of- association data, the degree- of- association data between two clauses is used, and each time the term is matched with the clause of the input voice, The relevance score of is calculated and stored, and the relevance score at that time is calculated from the relevance score calculated last time and the relevance score calculated this time.

The invention of claim 8 of the present invention is such that, in the invention of claim 7 , the relevance score of the time is calculated from the sum of the relevance score calculated last time and the relevance score calculated this time. is there. According to the invention of claim 9 of the present invention, in the invention of claim 7 , the relevance score at that time is calculated from the product of the relevance score calculated last time and the relevance score calculated this time.

According to the invention of claim 10 of the present invention, in the invention of claim 7 , the larger one of the previously calculated relevance score and the relevance score calculated this time is set as the relevance score for the time. It is a thing. According to the invention of claim 11 of the present invention, in the invention of claim 7 , the smaller one of the relevance score calculated last time and the relevance score calculated this time is set as the relevance score of the time. .

According to a twelfth aspect of the present invention, in the seventh aspect of the invention, the relevance score for the current time is obtained from the average value of the relevance score calculated last time and the relevance score calculated this time. .

[0013]

According to the first and second aspects of the present invention , the clause
Are classified into bunsetsu groups according to their case and meaning, and the case governing grammar table
Bullet 3 and refer to the classification result above
It appears in the phrase of the input voice from the dependency relation and the semantic relation of
Flags a group of clauses that may
The phrase group of the input speech is matched with the phrase group that was input, and based on the result.
Recognition score output by the matching means
Recognize continuous sentence speech based on
The matching phrase and the phrase of the input voice were matched by the matching means.
As a result, the clauses included in the flagged clause group are actually
If the input voice can be matched, the matching phrase is found.
To reset the flags in the clause group flag table
Therefore , the continuous sentence voice input can be efficiently recognized in consideration of the dependency relation and the semantic relation, and the performance of the continuous sentence voice recognition can be improved.

In the invention of claim 3 of the present invention, since the bunsetsu candidates to be matched with the bunsetsu of the input voice are narrowed down to a fixed number by the beam search method , calculation for matching is performed. The amount can be reduced. In the inventions of claims 4 to 12 of the present invention, the relevance score is calculated from the relevance between clauses, and the calculated relevance score is added to the matching result of the input voice. It can recognize continuous sentence voice input.

[0015]

FIG. 2 is a diagram showing a first embodiment of the present invention. In FIG. 2, 11 is a DP collating unit for collating an input voice with a phrase template by dynamic programming and outputting a recognition score, 12 Is a phrase selection unit that selects the phrase to be matched next in the DP matching unit 11 by referring to the phrase group flag, 13 is a case governing grammar table that stores the case governing grammar, and 14 is a phrase that indicates the phrase to be selected next. Clause group flag table storing group flags, 15 is DP
It is a phrase template storage unit that stores a phrase template to be matched with the input voice in the matching unit 11.

FIG. 3 is a diagram showing an example of the structure of the case governing grammar table 13. As shown in FIG. 3, the case governing grammar table 13 has a dependency relation of verbs, noun phrases, etc., that is, verbs and nouns. The correspondence relation between a phrase such as a phrase and, when the phrase comes, which next phrase comes is stored. For example, the phrase "(person) ga" such as "I am" or "Taro ga" and the phrase "(place) such as" to school "or" to hospital "may be associated with the verb" go ".
It is described that there can be a clause like "he". Further, for example, it is also described that an adjective comes before a noun phrase.

In the sentence order, the verb "go" generally comes to the end of the sentence, but in the case of Japanese, the sentence that follows is determined by the dependency relation from the front to the rear. Can reflect the restrictions of grammar more clearly if you go from behind. Therefore, at the time of recognition, recognition starts from "go" at the end of the sentence, and "go to" is followed by "to a place" or "a person".

FIG. 4A shows a clause group flag table 14.
As shown in the figure, the bunsetsu group flag table 14 stores verbs, noun phrases, adjectives, and their bunsetsu group flags. These bunsetsu group flags are stored in the case dominance grammar table. As a result of referring to 13, the circles are given to the clauses predicted to come after the matched clauses, and the crosses are given to the other clauses and the matched clauses.

FIG. 4B shows the clause group flag table 14.
Is a diagram showing how the bunsetsu group flag changes, and shows the state of the bunsetsu group flag when recognizing the sentence voice "I go to school". As an example, the addition of the phrase group flag when recognizing the sentence voice "I go to school" as described above will be described with reference to FIG. As described above, when recognizing a continuous sentence, recognition is performed from the end of the sentence voice, and in a Japanese sentence, the verb usually comes to the end of the sentence, so that the phrase group flag table 14 is
○ is given to all of the verbs. Then, when “go” is recognized as the verb, “x” is added to the clause group flag of the verb stored in the clause group flag table 14. By referring to the case dominance grammar table 13, the phrase that comes after the verb "go" is predicted, and as a result, in the case of "go", "(place) to" and "(person) ga" are relevant phrases. Since it is found, as shown in Fig. 4 (a), "to (place)" and "(person)" are marked with "(place)".
X is added to "de". As a result of matching the phrase that comes after "go", for example,
When the phrase “to school” is recognized, a cross is attached to the phrase “to (place)”. This uses the rule that a simple sentence rarely shows more than one case of the same type.

As a result, the state of the clause group flag in the clause group flag table 14 is as shown in FIG.
A circle is given to "(person) ga". As a result of collating the next phrase of "to school", if the next phrase is recognized as "I", an X is added to the flag of the phrase "(person) ga" in the phrase group flag table 14. The state of the added phrase group flag becomes the state shown in FIG.

As a result, x is added to all clause group flags, which means that the recognition of this sentence is completed. In the above example, all the bunsetsu group flags became x, and the recognition could be ended. However, depending on how the grammar is written, the end of the sentence may not be detected forever. Therefore, the above problem can be avoided by limiting the number of recognizable phrases.

FIG. 5 is a flow chart showing the processing in the phrase selecting unit 12, and the first embodiment shown in FIG. 2 will be described with reference to FIG. 3, FIG. 3 and FIG. The phrase selection unit 12 first puts the phrase group flag and the initial state of the DP (a state in which the collation result is not input) into the queue (step S1 in FIG. 5). In the initial state, the bunsetsu group flag has a circle attached to the verb bunsetsu as described above.

The DP matching result is a number string consisting of matching results up to each part of the phrase, and in the initial state, an infinite symbol is written in the number string. Then, in step S2, it is determined whether or not the queue is empty, and if the queue is empty, the process ends. If the queue is not empty (in the initial state, the queue has the phrase group flag input in step S1 and the initial state of DP), go to step S3, and obtain the result of the clause group flag and DP from the queue. Take it out. Further, I is set to I = 1 which is the initial state.

In step S4, it is determined whether or not the value of I is greater than the number of clause template registration clauses stored in the clause template storage unit 15 of FIG. 2, and the value of I becomes greater than the number of clause template registration clauses. If so, the process returns to step S2 on the assumption that the matching with the phrase template for the phrase is completed, and it is determined whether or not the queue is empty. If not, the process proceeds to step S3.

If the value of I is smaller than the number of registered phrase templates, the process goes to step S5, and the I-th phrase extracted from the phrase template is put in B.
Next, in step S6, with respect to the I-th bunsetsu extracted from the bunsetsu template, the bunsetsu group flag table 14 is referred to and whether or not the bunsetsu group flag is set (marked with a circle) is checked. To determine. If the phrase group flag is not set, the process goes to step S8, 1 is added to I, and the process returns to step S4.

If the phrase group flag is set, the process proceeds to step S7, where the DP collation unit 11 in FIG. 2 compares the input voice data with the I-th phrase B extracted from the phrase template. At the same time, the phrase group flag stored in the phrase group flag table 14 is updated. That is, as described above, for the matched clauses,
Change the clause group flag from ○ to ×.

When the above processing is completed, the collation result in the DP collation unit 11 and the updated clause group flag table are put in a queue. Then go to step S8 and set 1 for I
Is added and the process returns to step S4, and for the next phrase template stored in the phrase template storage unit 15, the DP collation and the phrase group flag updating process are performed in the same manner as above.

As described above, when all the phrase templates stored in the phrase template storage unit 15 have been collated with the input voice, the process returns from step S4 to step S2 to judge whether or not the queue is empty. If not empty,
Go to step S3. In step S3, the bunsetsu group flag and the result of the DP are taken out from the queue, I = 1 is set, and for the next bunsetsu of the input voice, based on the bunsetsu group flag taken from the queue and the result of the DP, the DP collation is performed in the same manner as above. And update the clause group flag.

As described above, in the processing of FIG. 5, among the templates stored in the clause template storage unit 15, the template of the clause to which the clause group flag is added and the first clause of the input speech (the clause at the end of the clause) ) And the DP collating unit 11 sequentially collates and updates the clause group flag,
The collation result and the updated clause group flag are stored in the queue.

Next, among the phrases of the next input speech and each template stored in the phrase template storage unit 15,
Similar to the above, the clause templates to which the clause group flag is added are collated, and the DP collation result is connected based on the DP collation result of the first clause stored in the queue. Then, the DP collation result and the updated clause group flag are stored in the queue as described above.

Similarly, each phrase of the input voice is sequentially collated with the phrase template. As described above, the collation ends when all the phrase group flags are in the x state. Then, the DP matching unit 11 obtains and outputs a recognition score from a plurality of DP matching results for continuous sentence speech obtained by connecting the DP matching results of each clause of the continuous sentence speech, and outputs the recognition score having the highest recognition score. And the recognition result.

FIG. 6 is a diagram showing a second embodiment of the present invention. In FIG. 6, the same parts as those of the first embodiment shown in FIG. In the embodiment, the beam search pruning unit 21 is the same as that of the first embodiment.
And the degree-of-association score processing unit 22 and the degree-of-association data storage unit 23 are added, and other configurations are the same as those in the first embodiment.

FIG. 7 is a diagram showing the concept of the beam search, and the beam search in this embodiment will be described with reference to the figure. Basically, the continuous sentence voice recognition can be realized by connecting the DP matching results (hereinafter, referred to as DP plane) of each clause of the input voice as described above, and FIG. The DP plane is illustrated using the sentence "I go to school" as an example.

As shown in FIG. 7, when recognizing a continuous sentence, a DP plane of the first appearing phrase (end of sentence) is created (in the figure, "go" as the first appearing phrase, The DP planes of “listen” and “see” are shown as examples), and the DP planes of the clauses that can follow after each clause are connected (in the figure, after the clause at the end of the clause, “to school”, “I By connecting the DP plane of "ga"), continuous recognition can be performed.
It should be noted that, as described above, it is possible to more clearly reflect the restriction of grammar when recognition is performed from the back, and in the example of FIG. 7, the verb such as “go” at the end of the sentence is used for matching.

By the way, in the first embodiment shown in FIG. 2, since the DP planes are created for all the matching results of each phrase of the inputted voice and the phrase template, and the DP planes are connected, DP matching is performed for all possibilities, and the amount of calculation becomes enormous. Therefore, at the stage of matching each clause, if the number of extending DP planes by pruning based on the matching result is limited to a fixed number, the above calculation amount can be suppressed.

Such a method is beam search,
In the example of FIG. 7, the number of extending DP planes is limited to one. That is, the first shown in FIG.
In the case of the above embodiment, all the DP planes of “go”, “listen”, and “see” are followed by D of the next clause of the input voice.
Since the P planes are connected, the number of branches in the DP plane increases and the amount of calculation becomes enormous, but as shown in FIG. 7, it is the most recognized among “go”, “listen”, and “see”. The amount of calculation can be suppressed by extending the DP plane only from “go” having a high score.

Specifically, in the flow chart of FIG. 5, when the DP collation result and the clause group flag are put in the queue, the DP collation result is shown in the beam search pruning unit 21 of FIG.
The beam search pruning unit 21 narrows the number of bunsetsu candidates within the beam width and sends the bunsetsu candidates to the queue of the bunsetsu selecting unit 12. That is, of the DP matching results, the beam search pruning unit 2
One or a plurality of phrases having a high recognition score selected in 1 are sent to the queue of the phrase selecting unit 12 as a phrase candidate (“go” is selected as a phrase candidate in the example of FIG. 7).

FIG. 8 is a diagram showing an example of the degree-of-association data stored in the degree-of-association data storage 23. In FIG. 8, there is a possibility that two clauses may appear in one sentence at the same time as the degree-of-association data. The indicated value (referred to as co-occurrence relation data) is shown. In the example of the figure, the probability that the phrase "today" appears before the phrase "go" is 0.3,
It is also shown that the probability that the phrase "today" appears twice is 0.

As the degree-of-association data, in addition to the above example, for example, data indicating the possibility that a certain phrase is adjacent to another certain phrase (referred to as adjacency relation data) can be used. 6 is a means for calculating a relevance score by referring to the relevance score data stored in the relevance score data storage unit 23, and the phrase selecting unit 12 selects a phrase. In this case, the degree of relevance to the already selected phrase in the sentence is calculated, and the degree of relevance is added to the collation result in the DP collation unit 11, or the product of the collation result and the degree of relevance is obtained and output as a recognition score. To do.

As a method of calculating the degree of association, for example, when the clause selecting unit 11 selects "to school" as the clause before "go", the degree of association is 0.8 from the degree-of-association data shown in FIG. Is required. Next, when “I am” is selected as the clause before “To school”, the degree of association between “To school” and “I” is 0.4 from the association degree data of FIG. Then, the relevance of 0.8 is added to the above-mentioned “go” and “to school” to obtain a relevance of 1.2.

Further, when "today" is selected as the next clause of "I am", the degree of association is 0.2 according to FIG. By adding 2, the degree of association becomes 1.4. That is, the degree of association is obtained by accumulating the values of the degree-of-association data shown in FIG. In the above example, an example of calculating the sum of the degree of association to obtain the degree of association of a plurality of clauses has been shown, but various other methods can be adopted as the method of calculating the degree of association. Products can be used instead. In this case, as in the above example, the degree of association when “go”, “to school”, and “I am” is selected is 0.8 × 0.4 = 0.
32, and the degree of association when “today” is selected is 0.32 × 0.2 = 0.064.

Further, as a method of obtaining the degree of association, other than this, the maximum value ma of the degree of association obtained from the degree-of-association data of FIG.
x is calculated (in this case, the degree of relevance when “go”, “to school” and “I am” is selected as described above is 0.
8) Alternatively, the degree of association can be calculated by obtaining the minimum value min, and further obtaining the average of the degree of association values.

Next, the second embodiment of FIG. 6 will be described with reference to FIGS. 7 and 8. Similar to the embodiment shown in FIG.
Among the templates stored in the phrase template storage unit 15, the template of the phrase to which the phrase group flag is added and the first phrase (the last sentence of the sentence) of the input voice are compared with the DP matching unit 1.
In step 1, the phrase group flag is updated. This matching result is sent from the DP matching unit 11 to the beam search pruning unit 21, and the beam search pruning unit 21 sets the number of phrase candidates obtained by the DP matching to a fixed number within the beam width by the recognition score. squeeze.

The phrase candidates narrowed down by the beam search pruning unit 21 are sent to the sentence selection unit 12 and put in a queue. Then, the phrase of the next input voice and the template of the phrase to which the phrase group flag is added among the templates stored in the phrase template storage unit 15 are collated in the same manner as above. Then, the DP matching result is connected to the DP plane of the phrase candidate of the previous phrase input to the queue.

Similarly, each phrase of the input voice is sequentially collated with the phrase template, and as described above, the collation ends when all the phrase group flags are in the x state. Further, when the phrase is selected by the sentence selection unit 12, the relevance score processing unit 22 refers to the relevance data stored in the relevance data storage unit 23 to calculate the relevance by the above-described method, The degree of relevance is added to the matching result of the matching unit 11 (for example, the sum or the product is calculated), and the result is output as a recognition score.

FIG. 9 is a diagram showing a third embodiment of the present invention. In FIG. 9, the same parts as those of the second embodiment shown in FIG. In the embodiment, the relevance score processing unit 22 of the second embodiment is composed of a relevance score addition unit 24 and a relevance score calculation unit 25, and other configurations are the same as those of the embodiment of FIG. is there.

In the embodiment shown in FIG. 9, the relevance score calculator 25 refers to the relevance data stored in the relevance data storage 23 and determines the relevance from the sum, product, etc. of the relevance data as described above. The score is calculated and the relevance score adding unit 24
Output to. The relevance score adding unit 24 uses the relevance score obtained by the relevance score calculation unit 25 as the DP matching unit 1.
1 is added to the recognition score output (for example, as described above, the recognition score is added to or multiplied by the relevance score) to output the recognition score.

[0048]

As is apparent from the above description,
In the present invention, bunsetsus are classified into bunsetsu groups according to their case and meaning, and with reference to the above classification result, bunsetsu groups that may appear in the bunsetsu of the input speech from the dependency relation and semantic relation of the bunsetsu. Flag, match the flagged phrase group with the phrase of the input voice, and flagged phrase
Does the phrase contained in the group actually appear in the phrase of the input speech?
And select a phrase based on the result,
Since the continuous sentence speech is recognized based on the recognition score output by, the continuous sentence speech input can be efficiently recognized in consideration of the dependency relation and the semantic relation. The performance can be improved.

Further, by limiting the number of phrase candidates to be matched with the phrase of the input voice to a certain number by beam search, the amount of calculation for matching can be reduced. Furthermore, by calculating the relevance score from the relevance between phrases and adding the calculated relevance score to the matching result of the input voice, it is possible to recognize the continuous sentence voice input more accurately.

[Brief description of drawings]

FIG. 1 is a principle block diagram of the present invention.

FIG. 2 is a diagram showing a first embodiment of the present invention.

FIG. 3 is a diagram showing an example of a case dominance grammar table.

FIG. 4 is a diagram showing an example of a clause group flag table.

FIG. 5 is a flowchart showing processing in a phrase selecting unit.

FIG. 6 is a diagram showing a second embodiment of the present invention.

FIG. 7 is a diagram showing the concept of beam search.

FIG. 8 is a diagram showing an example of association degree data.

FIG. 9 is a diagram showing a third embodiment of the present invention.

[Explanation of symbols]

11 DP collator 12 clause selection section 13 Case rule grammar table 14 clause group flag table 15 clause template storage 21 Beam search pruning unit 22 Relevance score processing unit 23 Relevance data 24 Relevance score adder 25 Relevance score calculator

Continuation of the front page (56) References JP 63-128467 (JP, A) JP 63-311398 (JP, A) JP 3-180899 (JP, A) Hideki Kojima et al. Recognizing Japanese Sentences by Dependency Analysis, Proceedings of the 1992 Autumn Meeting of the Acoustical Society of Japan, Japan, October 5, 1992, 2-Q-9, p. 189-190 Hideki Kojima et al., Continuous speech recognition technology, FUJITSU, Japan, January 12, 1998, Vol. 49 No. 1, p. 71-75 Yoshiya Suzuki, Japanese voice recognition using case structure, The Acoustical Society of Japan 1992 Spring Lecture Proceedings, Japan, March 1992, 3-1-7, p. 83-84 (58) Fields surveyed (Int.Cl. ⁷ , DB name) G10L 15/12 G10L 15/18 G10L 15/28 JISST file (JOIS)

Claims (12)

(57) [Claims] 1. A continuous-sentence speech recognition apparatus for recognizing continuous-sentence speech by phrase matching, wherein a case-dominant grammar table in which clauses are classified and stored according to their case and meaning, and a clause to be selected next A clause group flag table that stores the clause group flag to be instructed, a collating means that collates the input voice and outputs a recognition score, and refers to the clause flag to select the clause to be collated next in the collating means. The phrase selecting means refers to the case governing grammar table, flags the phrase group appearing in the phrase of the input speech of the phrase group flag table, and sets the flag. The bunsetsu and the bunsetsu of the input speech are collated by the collating means, and it is checked whether or not the bunsetsu included in the bunsetsu group flagged actually appears in the bunsetsu of the input speech, and the bunsetsu is selected based on the result. The continuous sentence voice recognition device is characterized in that the continuous sentence voice is recognized based on the recognition score output by the matching means. 2. As a result of matching the flagged phrase and the phrase of the input voice by the matching means, if the phrase included in the flagged phrase group can actually be matched in the input voice, it is matched. 2. The continuous sentence speech recognition apparatus according to claim 1, wherein a flag of the phrase group flag table corresponding to the phrase is reset. 3. The continuous sentence speech recognition device according to claim 1, wherein the collating means narrows down the number of phrase candidates to be collated with the phrase of the input voice to a predetermined number by the beam search method. 4. A relevance score calculation means for calculating a relevance score from the relevance between phrases and adding the calculated relevance score to a matching result of input speech. The continuous sentence voice recognition device according to claim 2 or claim 3. 5. The continuous sentence speech recognition apparatus according to claim 4, wherein the relevance score between the phrases is calculated using the co-occurrence relation data between the phrases. 6. The continuous sentence speech recognition device according to claim 4, wherein the relevance score between the phrases is calculated by using the adjacency relation data between the phrases. As 7. relevance data, using the relationship <br/> communication of data between two clauses, each matching a phrase in the input speech, and calculates the relevance score of clause collated before that Remember, the relevance score of that time is the relevance score calculated last time,
The continuous sentence voice recognition device according to claim 4, wherein the continuous sentence voice recognition device is calculated from the relevance score calculated this time. 8. The continuous sentence speech recognition device according to claim 7, wherein the relevance score for the current time is calculated from the sum of the relevance score calculated last time and the relevance score calculated this time. 9. The continuous sentence speech recognition apparatus according to claim 7, wherein the relevance score for the current time is calculated from the product of the relevance score calculated last time and the relevance score calculated this time. 10. The continuous sentence speech recognition apparatus according to claim 7, wherein the larger of the previously calculated relevance score and the relevance score calculated this time is used as the relevance score for the current time. 11. The continuous sentence speech recognition apparatus according to claim 7, wherein the smaller one of the previously calculated relevance score and the relevance score calculated this time is used as the relevance score for the current time. 12. The continuous sentence speech recognition apparatus according to claim 7, wherein the relevance score for the current time is calculated from the average value of the relevance score calculated last time and the relevance score calculated this time.