JPH09114493A - Interaction controller - Google Patents

Abstract

PROBLEM TO BE SOLVED: To reduce the number of utterings by a user at an erroneous recognition and to make the conversation smooth, in an interaction controller which has an user uttering content recognition means. SOLUTION: The controller has a recognition word estimating section 12. The section 12 is provided with a similar word table 122 which stores separating information to judge the similarity and the non-similarity of the words stored in a recognition word dictionary 121, word narrowing-down table 123 which stores the information to express the presence or the absence of erroneous recognition histories of each word in the dictionary 121, a similar word retrieving section 124 which retrieves similar words and a narrowing-down word retrieving section 125 which performs word narrowing-down retrieval. At the time of an erroneous recognition, the controller does not ask the user to re-input his voice, but asks the user the correctness or the incorrectness of the similar word that is retrieved by referring to the table 122 as the next estimating word. If more than two words having no erroneous recognition history are narrowed down, they are successively used as recognized words and the user is asked to make an appropriate selection. The user answers the question with only 'yes' or 'no'.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a voice service system which utilizes voice recognition and voice synthesis techniques to interact with a user and realizes a predetermined voice service in accordance with the content of the conversation, and more particularly to the user and the system. The present invention relates to a dialogue control device that controls a dialogue between a user and a user.

[0002]

2. Description of the Related Art A voice service system is known which realizes a predetermined voice service while talking with a user by voice. FIG. 13 is a general block configuration diagram of this voice service system. The voice uttered by the user is
It is input to the voice recognition unit 31 of the voice interface 3, recognized by the voice recognition unit 31, and sent to the dialogue control device 4. The dialogue control device 4 generates a voice output sentence (character code) for carrying out the next dialogue after the topic is decided or when the user inputs a voice, and sends it to the voice synthesizer 32. When the dialog control device 4 acquires necessary and sufficient information from the dialog with the user, the dialog control device 4 sends the acquired information to the application processing unit 2 to execute the service desired by the user.

The dialogue between the voice service system and the user is carried out by repeating that the system side outputs voice messages such as instructions and questions for each topic and the user responds to them (interaction pair). The dialogue control device 4 controls the dialogue pair. A schematic configuration of a conventional dialogue control device 4 is as shown in FIG. 14, and includes a topic determination unit 41 that determines a topic, a recognition word dictionary 422 that stores a plurality of recognition words for each topic, and a word search unit 421. A recognition word estimation unit 42 that estimates a recognition word using the speech recognition unit 3, a speech output sentence generation unit 43 that generates a speech output sentence that is a target of speech synthesis from the estimation result and sends the speech output sentence to the speech synthesis unit 32, and a speech recognition unit 3.
1 and the application processing unit 2, and a recognition processing unit 44, and a correctness determination unit 45 that determines whether the estimated recognition word is correct based on the recognition result sent from the voice recognition unit 31. .

The operation of the dialogue control device 4 will be described by taking a service of reservation of a conference room as an example. When making a reservation for a conference room, five topics are required: the name of the person who made the reservation, the date, the use start time, the use end time, and the name of the conference room to be reserved. The dialogue control device 4 controls a dialogue pair performed between the application processing unit 2 and the user for each of these topics.

FIG. 15 is an explanatory diagram of a control flow performed by the dialogue control device 4 regarding the topic of the conference room name, and FIG.
It is a figure which shows the mode of the dialog actually performed between the whole voice service system and a user in this case. Referring to FIGS. 15 and 12, first, the topic determination unit 41 determines the “meeting room name” that is the topic of this time (S301). As a result, the voice output sentence generation unit 43 generates a desired voice output sentence (“Please give the name of the conference room”) necessary to inform the user what information the application processing unit 2 wants to input by voice. Yes (S302). When the user who hears the synthesized voice corresponding to this intended voice output sentence inputs the voice as "Corner A" and this is recognized by the voice recognition unit 31 (S303: Yes), the recognized word estimation unit 42 displays " The recognized word is estimated by referring to the recognized word dictionary 422 that stores the recognized word related to "meeting room name" (S304). Assuming that the recognition word “corner B” has been estimated, the voice output sentence generator 43 asks the user if the estimation result is correct (“is it corner B?”).
Is generated (S305). The user who hears the synthesized voice corresponding to this answers "No" because the recognition result is incorrect.

When this "no" voice is recognized (S306: Yes), the correctness determination section 45 determines that the recognized word is incorrect (S307). In response to this, the voice output sentence generation unit prompts the user to input voice information again (“Tell me again”)
Is generated (S308). When the user answers "Yes" by repeating the procedure of S303 to S308 (S30
7: Yes) regards the recognized word as correct and continues the dialogue on the next topic (S309). If there is no next topic, the dialogue control ends. It is known that "yes" and "no" answered by the user can be recognized with accuracy close to 100%.

[0007]

As described above, in the conventional dialogue control device 4, when the recognition word is estimated from the content of the user's utterance, the correctness is asked to the user, and it is determined that the recognition is incorrect. Requests the user again for voice input ("Please say again"). Therefore, there has been a problem that the burden on the user increases as the number of erroneous recognitions increases. In particular, if erroneous recognition of the same information is repeated, there is a concern that the user may feel uncomfortable with the system.

As a means for solving such a problem, a plurality of recognition word candidates are previously obtained by a recognition process or the like without prompting the user to input a voice again at the time of erroneous recognition, and the most probable recognition word is obtained. It is conceivable to prompt the user to confirm "yes" and "no" in order from the candidate. However, in order to narrow down the most probable one from a plurality of recognized word candidates, not only a considerable amount of calculation and a large memory space are required, but when a truly correct recognized word candidate is in the lower rank, it is necessary to reach it. There is a problem that the number of confirmations to the user increases.

Therefore, an object of the present invention is to speed up the identification of a recognition word in a dialogue control device having a means for recognizing the utterance content of a user, reduce the number of times the user utters at the time of erroneous recognition, and utter the content To facilitate the dialogue with the user.

[0010]

In order to solve the above problems, the present invention estimates a recognition word dictionary storing a plurality of recognition words and a recognition word corresponding to an input voice by referring to the recognition word dictionary. A recognition word estimation unit, a voice output sentence generation unit that generates a voice output sentence for inquiring whether the estimated recognition word is correct, and a correctness determination unit that determines whether the estimated recognition word is correct or incorrect based on the input voice. In the dialogue control device having, the recognition word estimating unit includes a similar word table in which identification information of the recognition words stored in the recognition word dictionary, which is similar or dissimilar between the recognition words, is stored in the identification area of each recognition word, and the estimation. And a first word search unit for identifying an identification area of another recognition word similar to the recognition word when the recognized word is wrong, And the recognized word corresponding to the identification areas and estimating a recognized word for the next candidate.

For similarity or dissimilarity, for example, the similarity between the recognition models of the recognition words is calculated and recorded in advance, or a pair of recognition words which is apt to be mistaken is statistically obtained from the actual use history. We will respond by recording it.

When creating the similar word table,
For example, the identification areas of the recognition words included in the recognition word dictionary are arranged in a matrix, and binary information defining one as similar and the other as dissimilar is stored in corresponding areas of the matrix. Alternatively, multi-valued information of three or more values indicating the similarity between each recognized word is stored in the corresponding area of the matrix. In the latter case, the first word search unit is configured to determine that recognized words whose multi-valued information exceeds a predetermined threshold are similar to each other.

According to this construction of the dialogue control device, when the estimation of one recognized word is incorrect, the corresponding area of the similar word table is referred to immediately identify the identified areas of other recognized words similar to the recognized word. Can be specified. Therefore, if this identification area is made to correspond to the identification information of the recognition dictionary in the recognition word dictionary in a one-to-one correspondence, the identification and search of the next recognition candidate recognition word will be speeded up. In addition, since the next recognized word can be immediately identified, it is possible to ask the user whether the recognized word is correct or not and to answer only “Yes” or “No”. You can also facilitate the dialogue compared to.

Another structure of the present invention for solving the above-mentioned problems is a recognition word dictionary storing a plurality of recognition words and a recognition word for estimating a recognition word corresponding to an input voice by referring to this recognition word dictionary. Dialog having an estimation unit, a voice output sentence generation unit that generates a voice output sentence for inquiring whether the estimated recognition word is correct, and a correctness determination unit that determines whether the estimated recognition word is correct or incorrect based on the input voice. In the control device, the recognition word estimation unit, a word narrowing table in which information indicating the presence or absence of misrecognition history of the recognition word is stored in the identification areas of all the recognition words stored in the recognition word dictionary, A second word search unit for detecting an identification area of a recognition word having no erroneous recognition history by referring to this word narrowing-down table, and selecting one of the recognition words corresponding to the detected identification area. Recognizing the next candidate Characterized in that it estimated word.

By this, by storing the history of the recognition word which has been erroneously recognized once in the word narrowing table,
Repeated erroneous recognition due to the same recognition word can be avoided, and the dialogue can be facilitated.

When there are two identification areas detected by the second search unit, that is, when there are two recognition words having no recognition error history, either one of the recognition words is the correct answer. In this case, the recognition word estimation unit estimates a word corresponding to one of the identification areas as a recognition word. Then, the voice output sentence generation unit generates a voice output sentence for asking whether the recognized word is correct or incorrect. As a result, the correct recognition word can be recognized simply by urging the user to input only “Yes” or “No”, and the dialogue can be facilitated.

[0017]

Embodiments of the present invention will be described below in detail with reference to the drawings. FIG. 1 is a block configuration diagram when the dialogue control device of the present invention is applied to a voice service system, and basically has the same configuration as the conventional system shown in FIG. It should be noted that, in FIG. 1, the same components are denoted by the same reference numerals.

The voice uttered by the user is the voice recognition unit 31.
To the dialogue control device 1 of the present invention. The dialogue control device 1 generates a voice output sentence for carrying out the next dialogue when the topic is decided,
This is sent to the voice synthesizer 32. When the dialogue control device 1 obtains necessary and sufficient information from the dialogue with the user, the dialogue control device 1 sends the obtained information to the application processing unit 2,
Execute the service desired by the user.

FIG. 2 is a schematic configuration diagram of the dialogue control device 1 of the present invention. This dialogue control device 1 includes a recognition word dictionary 12
1, similar word table 122, and word narrowing down table 1
The recognition word estimation unit 12 to which 23 is connected, the topic determination unit 11 having the same function as the conventional device shown in FIG. 14, the voice output sentence generation unit 13, and the correctness determination unit 15 are included. The recognized word estimation unit 12 further refers to the similar word table 122 to search for similar words in the recognized word dictionary 121, and the recognized word dictionary 121 by referring to the word narrowing table 123. And a narrowed-down word search unit 125 that searches the recognition words for erroneous recognition and searches for those having no erroneous recognition history. Both search units 124 and 125 may be used independently or in combination.

FIG. 3 is an explanatory diagram of the structure of the recognition word dictionary 121 used in this embodiment. The recognition dictionary number for determining which topic dictionary, the word name as a recognition candidate, and the search process. For convenience, a word number assigned to each word is stored in a predetermined format. In the illustrated example, the topic “meeting room name” corresponds to the recognition dictionary number “3”, and “corner A” ...
And the like, and their word numbers (1) to (5) are stored.

FIG. 4 is an example of the similar word table 122, and "○", "" that indicate whether the words in the recognized word dictionary (recognition dictionary number "3") shown in FIG. 3 are similar or dissimilar. Binary information of “×” is stored in a matrix form in an area corresponding to each word number, that is, an identification area. In the example shown, "Corner A" and "Corner B"
And are similar to each other, and other words are dissimilar.

The similarity relation between words is not necessarily limited to binary information and may be represented by multivalued information. FIG. 5 shows the degree of similarity between words as continuous low-valued multivalues (either discrete values or analog values) L1 to L5. In this case, a predetermined threshold is set and compared with each similarity, and if the similarity is higher than the threshold, the similarity is determined, and if the similarity is lower than the threshold, the similarity is determined. The advantage of expressing the similarity relation with binary information is that the configuration of the similar word search unit 121 is simplified, and the advantage of expressing it with multivalued information is that the degree of similarity is expressed step by step, so that the recognition word dictionary It is to be able to define the similarity between the words in 16 more finely. When a plurality of recognized words exceed the threshold value when represented by multivalued information, the recognized word candidates may be specified in descending order of similarity.

FIG. 6 shows an example of the word narrowing table 123. The recognition word is stored in an area corresponding to the word numbers of all the words stored in the recognition word dictionary 121 shown in FIG. Binary information indicating the presence / absence of the misrecognition history of (No misrecognition history: ○, No misrecognition history: ×) is stored. As the initial value, “◯” is stored in all areas, and is updated to “×” at each misrecognition to be removed from the next recognition word candidate. It should be noted that this word narrowing-down table 123 is created each time the recognition word dictionary 121 to be used is changed, and has the same number of identification areas as the number of words in the new recognition word dictionary.

Next, the operation of the voice service system and the dialogue control apparatus 1 of this embodiment will be described with reference to FIGS. Here, for the sake of convenience, an example of the topic of reservation of a conference room will be described as in the conventional example.

FIG. 7 is an explanatory view of a control flow of the dialogue control device 1 when the similar word search unit 124 and the narrowed word search unit 125 are used together, and FIG. 9 shows the entire voice service system and the user in this case. It is a figure which shows the mode of the dialog actually performed between. As a premise, the word narrowing table 12
It is assumed that all the values in the identification area 3 are in the desired state (all are “◯”).

In this example, first, the topic determining section 11 determines the "meeting room name" (S101). As a result, the voice output sentence generation unit 13 generates a desired voice output sentence (“Please call the meeting room name”) (S102). The user who heard the synthesized voice corresponding to this expected voice output sentence is "Corner A".
When the voice recognition unit 11 recognizes this (S103: Yes), the recognition word estimation unit 12 recognizes the recognition word dictionary 12 with the recognition dictionary number “3” corresponding to the “meeting room name”.
The recognized word is estimated by referring to 1 (S104). Assuming that the recognition word "corner B" has been estimated, the voice output sentence generation unit 13 generates a voice output sentence ("Corner B?") For asking the user if the estimation result is correct (S105). ). The user who hears the synthesized voice corresponding to this answers "No" because the recognition result is incorrect.

When this "no" voice is recognized (S106: Yes), the correctness determination unit 15 determines that the recognized word is incorrect (S107: No). In response to this, the recognized word estimation unit 12 specifies the word number of the erroneously recognized word from the recognized word dictionary 121 and determines the word narrowing table 12
The relevant identification area on 3 is updated to "x" (S108).
The recognized word estimation unit 12 also checks the number of candidate words, that is, the number of identification areas in which “◯” is stored in the word narrowing table 123 (S109), and when 3 or more (S10).
9: Yes), a similar word search process is performed (S110).
Specifically, first, referring to the similar word table 122,
The word number of a word similar to the misrecognized word (corner B: word number (2)) is specified. In the example of FIG. 4 or 5, “corner A” of word number (1) is specified. Next, the word corresponding to this word number (1) is searched from the recognition word dictionary 16. Then, the searched word (“corner A”) is estimated as the next recognized word (S11).
1), the process returns to S105.

If the number of candidate words is 2 or less in S109 (S109: No) and it is confirmed that there are candidate words (S112: Yes), the similar word table 122 is not referred to. The corresponding word in the recognized word dictionary 121 is searched from the ascending order of word numbers, and this is sequentially estimated as a recognized word (S113), and the process returns to S105.

If the recognition result is determined to be correct in S107 (S107: Yes), that is, if the user answers "yes", or if it is confirmed in S112 that no candidate word exists (S112: No). Determines whether there is a next topic (S114), and if there is a topic, S101
Return to the process of (3), and if there is no topic, the dialogue control ends.

FIG. 8 shows an example in the control flow of FIG. 7, in which word narrowing is performed immediately after the topic is decided instead of generating the desired voice output sentence. In this case, when the topic is determined and the recognized word dictionary 121 corresponding thereto is specified (S201), the recognized word estimation unit 12 refers to the narrowing-down table 123 to check the number of candidate words (S202,
S203). Then, if the number of candidate words is 3 or more, an arbitrary word (“corner B”) is searched from the recognized word dictionary 121, and this is estimated as a recognized word (S204). Subsequent processes S205 to S214 are the same as S105 to S11 in FIG.
Same as 4.

On the other hand, if the number of candidate words is 2 or less in S203 (S203: No), the process jumps to S212. In other words, if there are two candidate word numbers, one is selected from the ascending order of word numbers, and if there is one candidate word, the word is estimated as a recognized word and the process returns to S205 (S2
13) If there is no candidate word, move on to the next topic.
FIG. 10 is a diagram showing a state of dialogue actually performed between the entire voice service system and the user when there are two candidate words and when there is one candidate word.

As described above, by using the similar word table 122, the next recognition word candidate at the time of erroneous recognition can be quickly identified, and comparison with FIG. 12 showing an example of a conventional dialogue on the same topic is made. Obviously, the number of times the user answers the system by voice after uttering the desired conference room name is reduced, and the voice is only "Yes" or "No" with a high recognition rate, so the dialogue is smooth. Be converted.

Further, since the history of the once erroneously recognized words is stored in the word narrowing table 123 so as to be excluded from the candidates of the next recognized word, repetition of erroneous recognition is prevented. Furthermore, this word narrowing table 123
By performing the word narrowing process using the word prior to the voice input to the user, the user can select "Yes" from the beginning, such as "Corner A?" Or "Corner B?"
The correct word can be recognized simply by answering "no", and the number of times the user speaks is further reduced. As a result, it is possible to simultaneously reduce the burden on the user and facilitate the dialogue.

In this embodiment, ".smallcircle." And "x" are used as the binary information stored in the similar word table 122 shown in FIG. 4 and the word narrowing table 123 shown in FIG. However, of course, if it is information that can discriminate either one of the two values, it may be another type of information such as logic 1 and logic 0.

[0035]

As is apparent from the above description, according to the present invention, the similar relation between words in the recognized word dictionary is referred to by using the similar word table at the time of erroneous recognition. This has the effect of significantly shortening the time required to specify a word, as compared with the case of calculating similarity. In addition, since the next candidate word can be quickly identified, it becomes easy to immediately give the user a synthetic voice asking whether the word is estimated to be correct or incorrect, and the user can only give a "Yes" or "No" voice. It can be input only.

Further, regarding a word for which misrecognition has occurred, the misrecognition history is stored in the identification area in the word narrowing-down table, so that there is an effect of avoiding misrecognition again. Further, when the number of words having no misrecognition history is narrowed down to two or less, one of the words is estimated to be a recognized word and the user is inquired of whether the word is correct, thereby reducing the user's voice input load. Can be made.

In this way, the dialogue control can facilitate the dialogue by reducing the number of times the user utters as much as possible and by making each utterance content a short voice such as "Yes" or "No". The device is realized.

[Brief description of the drawings]

FIG. 1 is a block configuration diagram of a voice service system to which a dialogue control device of the present invention is applied.

FIG. 2 is a block configuration diagram of an embodiment of a dialogue control device according to the present invention.

FIG. 3 is an explanatory diagram showing an example of contents of a recognized word dictionary.

FIG. 4 is an explanatory diagram showing an example of a case where binary information is stored in a similar word table.

FIG. 5 is an explanatory diagram showing an example of a case where multi-valued information is stored in a similar word table.

FIG. 6 is an explanatory diagram showing an example of a word narrowing table.

FIG. 7 is a diagram showing a control flow of the dialogue control device according to the present embodiment.

FIG. 8 is a diagram showing another control flow of the dialogue control device according to the present embodiment.

9 is a diagram showing a state of actual dialogue between the entire voice service system and a user according to the control flow of FIG. 7. FIG.

FIG. 10 is a diagram showing a state of dialogue actually performed between the entire voice service system and a user according to the control flow of FIG. 8;

FIG. 11 is a diagram showing a state of dialogue when there is one candidate word in the control flow of FIG. 8.

FIG. 12 is a diagram showing a state of dialogue actually performed between the entire voice service system and a user by a conventional dialogue control device.

FIG. 13 is a block configuration diagram of a voice service system to which a conventional dialogue control device is applied.

FIG. 14 is a block configuration diagram of a conventional dialogue control device.

FIG. 15 is a diagram showing a control flow of a conventional dialogue control device.

[Explanation of symbols]

 1 Dialogue Control Device 11 Topic Determining Unit 12 Recognized Word Estimation Unit 121 Recognized Word Dictionary 122 Similar Word Table 123 Word Narrowing Table 124 Similar Word Searching Unit (First Word Searching Unit) 125 Narrowing Word Searching Unit (Second Word) Search unit) 13 voice output sentence generation unit 14 recognition control unit 15 correctness determination unit 2 application processing unit 3 voice interface 31 voice recognition unit 32 voice synthesis unit

Claims (5)

[Claims] 1. A recognition word estimating unit for estimating a recognition word corresponding to an input voice by referring to a recognition word dictionary storing a plurality of words, and a voice output sentence for asking whether the estimated recognition word is correct or incorrect. In the dialogue control device having a voice output sentence generation unit that performs, and a correctness determination unit that determines the correctness of the estimated recognition word based on the input voice, the recognition word estimation unit is stored in the recognition word dictionary. A similar word table storing similar or dissimilarity information between words in the identification area of each word, and referring to the similar word table when the estimated recognized word is incorrect, and similar to the recognized word And a first word search unit for identifying the identification region of the word, and estimating the word corresponding to the identified identification region as the next candidate recognition word. 2. The similar word table arranges the identification regions of the words included in the recognized word dictionary in a matrix form, and associates the matrix with binary information that defines one as similar and the other as dissimilar. The dialogue control device according to claim 1, wherein the dialogue control device is stored in an area. 3. The similar word table arranges the identification regions of words included in the recognized word dictionary in a matrix form, and associates the matrix with multivalued information of three or more values representing the similarity between the words. The dialogue according to claim 1, wherein the dialogue is stored in an area, and the first word search unit is configured to determine that words whose multi-valued information exceeds a predetermined threshold value are similar to each other. Control device. 4. A recognition word estimating unit for estimating a recognition word corresponding to an input voice by referring to a recognition word dictionary storing a plurality of words, and a voice output sentence for asking whether the estimated recognition word is correct or incorrect. In the dialogue control device having a voice output sentence generation unit that performs, and a correctness determination unit that determines the correctness of the estimated recognition word based on the input voice, the recognition word estimation unit is stored in the recognition word dictionary. A word narrowing-down table in which information indicating the presence or absence of a misrecognition history of the word is stored in the identification area of each word, and a word identification area having no misrecognition history is detected by referring to the word narrowing-down table. And a word search unit, and estimates any one of the words corresponding to the detected identification region as the next candidate recognition word. 5. The recognition word estimating unit estimates a word corresponding to one of the identification regions as a recognition word when the number of identification regions detected by the second word searching unit is two. The dialogue control device according to claim 4.