JP2007004052A - Voice interactive device and voice understood result generation method - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To reduce omission of selection of recognition result by voice recognition. <P>SOLUTION: Voice, input via a microphone 20, is recognized, on the basis of a recognition object words. On the basis of the words included in the recognition result candidates that are the recognition results, similar words are detected from a disk 40, in which the recognition object words of the voice are stored in association with the similar words that are likely to be incorrectly recognized. The detected similar words are added to the recognition result candidates, and the understanding result to be the response to the speech uttered is generated from the recognition result candidates, to which the similar words have been added. <P>COPYRIGHT: (C)2007,JPO&INPIT

Description

  The present invention relates to a voice dialogue apparatus that performs a dialogue according to spoken voice, and more particularly, to a voice dialogue apparatus and a voice understanding result generation method with improved voice recognition rate.

  2. Description of the Related Art A voice dialogue apparatus has been devised in which a voice uttered by a user is input and a system response is made according to a voice recognition result of the inputted voice, thereby performing a dialogue with the user. In such a speech dialogue apparatus, various speech recognition methods are used in order to improve the recognition rate of input speech.

  For example, in order to improve the speech recognition rate, a plurality of recognition result candidates obtained by recognizing input speech are output as speech recognition results, the reliability for the speech recognition results is obtained, and this newly obtained A method of generating an understanding result of the utterance content up to now is disclosed by adding the obtained reliability to the reliability of the speech recognition result of the speech uttered so far (Patent Document 1). .

As a result, it is possible not only to recognize the input speech as a word, but also to select a more likely word in consideration of the relationship with the context of the speech-recognized word. Recognition accuracy can be improved.
JP 2004-251998 A

  However, in the method disclosed in Patent Document 1, it is assumed that any of the recognition result candidates output as the speech recognition result always includes a word spoken by the user. The speech recognition result may have been leaked.

  Thus, if the correct speech recognition result is excluded from the correct recognition result candidates, naturally, the generated understanding result is not a reliable result.

  Therefore, the present invention has been proposed in view of the above-described circumstances, and correct words and phrases uttered by the user are excluded from the final recognition result candidates, thereby reducing the possibility of omission of the voice. An object of the present invention is to provide a speech dialogue apparatus and a speech understanding result generation method capable of improving the recognition rate of speech recognition.

  In the spoken dialogue apparatus of the present invention, the input means for inputting the spoken voice, the storage means for storing the speech recognition target word and the similar word that is easily misrecognized by the recognition target word, and the storage means, The speech recognition means for recognizing the speech input by the input means based on the recognition target word, and the similar word is detected from the storage means based on a word included in a recognition result candidate as a recognition result by the speech recognition means. An understanding result for generating an understanding result as a response to the spoken voice from a detection means and the similar result detected by the detecting means to the recognition result candidate, and from the recognition result candidate to which the similar word is added The above-described problem is solved by providing the generation unit.

  In the speech understanding result generation method of the present invention, the input step of inputting the spoken speech, the speech recognition step of recognizing the speech input by the input step based on the recognition target word, and the speech recognition step Detection based on a word included in a recognition result candidate that is a recognition result, and detecting the similar word from a storage unit that stores the speech recognition target word and a similar word that is easily misrecognized by the recognition target word. And an understanding result generation for generating an understanding result as a response to the spoken speech from the recognition result candidate obtained by adding the similar word to the recognition result candidate. The above-described problems are solved by providing the process.

  The speech dialogue apparatus according to the present invention adds a similar word that is easily misrecognized to the recognition target word detected from the storage unit based on the word included in the recognition result candidate that is the recognition result by the speech recognition unit, An understanding result that is a response to the spoken voice is generated from the recognition result candidate to which the word is added.

  As a result, not only the recognition result candidate obtained as a recognition result by the speech recognition means, but also leaked in the speech recognition processing, the word included in the recognition result candidate is a similar word that is acoustically close and easily misrecognized. It can be handled as a final recognition result candidate.

  Therefore, even if it is a legitimate utterance by the user, it can be selected as an understanding result even if it is excluded from the recognition result candidates by voice recognition processing, so the recognition rate is improved. It is possible to make it.

  Further, the speech understanding result generation method of the present invention uses, as a recognition result candidate, a similar word that is easily misrecognized by the recognition target word detected from the storage unit based on the word included in the recognition result candidate that is the recognition result by speech recognition. In addition, an understanding result that is a response to the spoken voice is generated from the recognition result candidates to which similar words are added.

  As a result, not only the recognition result candidate obtained as a recognition result by speech recognition but also leaked in the speech recognition processing, but the word included in the recognition result candidate is a similar word that is acoustically close and easily misrecognized. Can be treated as a typical recognition result candidate.

  Therefore, even if it is a legitimate utterance by the user, it can be selected as an understanding result even if it is excluded from the recognition result candidates by voice recognition processing, so the recognition rate is improved. It is possible to make it.

  Hereinafter, embodiments of the present invention will be described with reference to the drawings.

  First, the configuration of a voice interactive apparatus shown as an embodiment of the present invention will be described with reference to FIG. The voice interaction apparatus shown in FIG. 1 as an embodiment of the present invention is configured when applied to a navigation apparatus mounted on a moving body such as a vehicle. For example, when the navigation device is mounted on a vehicle that is a moving body, the navigation device detects the current position of the vehicle, displays a map corresponding to the current position of the vehicle displayed from the map data, and displays a route to a desired destination. You can give guidance.

  When this voice interactive apparatus is applied to a navigation apparatus, various functions required by the navigation apparatus can be operated interactively by the interaction between the user and the system.

  As shown in FIG. 1, the voice interaction device is stored in a switch 10, a microphone 20, a memory 30, a disk 40 for storing map data used for route guidance, voice data for guidance voice, and the like. A disk reader 41 that reads various data, a controller 50 that recognizes voice input through the microphone 20, understands the contents of the voice recognition result, and generates a system response; and a map showing a route search result , A menu screen, a voice recognition result by the control device 50, etc., for example, a monitor 60 such as a liquid crystal display, and a speaker for outputting a guidance voice or a system response voice (hereinafter simply referred to as a system response) in a dialogue with the user. 70.

  The switch 10 instructs the input control unit 51 of the control device 50, which will be described later, to start a speech recognition process on the speech uttered by the user and input through the microphone 20 when pressed by the user.

  The microphone 20 inputs the voice uttered by the user to a voice recognition unit 52 of the control device 50 described later. For example, the user utters words and sentences used for the operation of the navigation device, that is, operation commands and proper nouns such as place names, facility names, road names, and sentences including these phrases, and the voice is output from the microphone 20. input.

  The memory 30 is a random accessible RAM (Random Access Memory) or the like, and a storage area 31 for storing a voice recognition dictionary / grammar read from the disk 40 by the disk reading device 41 when a voice recognition process is executed. And a storage area 32 for storing an understanding result of utterances up to the present time output from a language understanding unit 54 of the control device 50 described later.

  The understanding result of the utterance up to the present time stored in the storage area 32 of the memory 30 is that when the utterance is input next to the current utterance, the understanding of the newly input utterance is combined with the past utterance understanding result. Used to understand.

  The disk 40 is a storage medium that stores a dictionary / grammar for speech recognition used for speech recognition, a map database, a proper noun database, a general database, a reading database, and the like.

  Generally, in a speech recognition system using a speech recognition dictionary / grammar, only input sentences using the recognition target words and grammar described in the speech recognition dictionary / grammar are accepted as speech recognition results. Can do. Information about the recognition target word is also managed in a proper noun database, a general database, and a reading database.

  For example, assuming that the main task of the navigation device is a destination setting for route search, the user can input only words related to facilities such as “Kanagawa Prefecture” and “Yokohama Station” as input sentences input from the microphone 20 by the user. It is possible to envisage both input by sentences combining a plurality of keywords such as “Yokohama Station in Kanagawa Prefecture” and “Yokohama Station on the Tokaido Line”.

  Therefore, the speech recognition dictionary / grammar stored in the disk 40 is configured to be able to handle both such a word-only input and a document including a plurality of keywords.

  The map database contains map data used for map display and route search.

  In the proper noun database, proper nouns indicating facilities such as “Oppama Station” and “Tokyo ○ × Park” and proper nouns indicating places such as “Saitama Prefecture” and “Tama City” are associated with detailed information on each proper noun. Stored.

  The general database includes general nouns indicating the type of facility such as “station” and “gas station”, general nouns indicating administrative divisions such as “prefecture”, “city” and “town”, affirmative words such as “yes” and “no” Negative words, verbs such as “go”, “find”, and “return”, numbers such as “1” ... “1000”, particles such as “so”, auxiliary verbs such as “is”, etc. are associated with each detailed information. Stored.

  Proper noun database, detailed information of general database, for example, if the word is a facility, Kana kanji notation for display and the location of the facility, attribute information such as restaurants and gas stations, and in the case of other words, This is attribute information indicating its meaning.

  FIG. 2 shows an example of a proper noun database. As shown in FIG. 2, the proper noun database includes each proper noun such as “Oppama Station”, “Tokyo ○ × Park”, “Saitama Prefecture”, and “Tama City” for each word ID that uniquely identifies these proper nouns. “Notation” indicating the form of the proper name, “Type” indicating the type of each proper noun, “Address” specifying the location of each proper noun by prefecture, city, and village, and how to read each proper noun prepared separately It is managed by information such as a “reading list” indicating reading IDs linked to the stored reading database.

  Further, as shown in FIG. 2, proper nouns whose “type” is a station among proper nouns are also managed in route information as “routes”. Although not shown, road information is also given as “route” to proper nouns whose “type” is interchange.

  FIG. 3 shows an example of a general database. As shown in FIG. 3, the general database includes “station”, “prefecture”, “city”, “town”, “yes”, “no”, “so”, “is”, “go”, “search”, “return”, “1”, “1”, etc. For example, a general noun other than a proper noun such as “1000”, a verb, a particle, an auxiliary verb, etc., for each word ID that uniquely identifies them, “notation” indicating the notation form of each word, “category” roughly classifying each word The information is managed by information such as “type” which is a subordinate concept of “category” and “reading list” indicating reading IDs linked to a reading database storing how to read each word separately prepared.

  The reading database is linked to each word classified into the proper noun database and the general database, and stores how to read each word.

  The words stored in the proper noun database and the general database may be different in notation and reading even if they have the same meaning. For example, “Tokyo ○ × Park” having a word ID of 5201 registered in the proper noun database of FIG. 2 is shortened from “Tokyo ○ × Park”, which is the official notation, as “○ × Park”. It is often written and simply called “Marubatsu Park”. Although “Tokyo ○ × Park” and “○ × Park” are recognized as the same facility by the user, the notation and reading are different.

  In addition, the word written as “town” with the word ID 0704 registered in the general database in FIG. 3 is read as either “town” or “cho” depending on the case.

  Therefore, in order to cope with such paraphrasing of words, a reading database is provided separately for the proper noun database and the general database.

  FIG. 4 shows a reading database corresponding to the proper noun database shown in FIG. 2 and the general database shown in FIG.

  As shown in FIG. 4, the reading database reads “reading” indicating the reading for each reading ID that uniquely specifies all readings of words stored in the proper noun database and the general database. The notation of word proper noun database and general database and word ID are held as a “semantic list”. Also, as shown in FIG. 4, the reading database is a “similar word reading list”, a reading ID of a word that is easily misrecognized as itself, a word whose reading is uniquely specified by this reading ID, The degree of similarity in which the degree of misrecognition is quantified is stored in a list format.

  Next, the control device 50 will be described. The control device 50 includes an input control unit 51, a speech recognition unit 52, a word reliability calculation unit 53, a language understanding unit 54, a response generation unit 55, a GUI display control unit 56, and a speech synthesis unit 57. The voice recognition process is performed on the voice input through the microphone 20 and a system response is made according to the voice recognition result.

  The input control unit 51 instructs the voice recognition unit 52 to start the voice recognition process in response to the switch 10 being pressed by the user.

  In response to an instruction from the input control unit 51, the speech recognition unit 52 executes speech recognition processing by capturing a speech signal uttered by a user input from the microphone 20 and digitized by an A / D converter (not shown).

  The voice recognition unit 52 performs voice recognition by matching the captured digitized voice signal with a standby sentence made up of recognition target words held in the voice recognition dictionary / grammar constructed in the storage area 31 of the memory 30. The speech recognition result is output to the language understanding unit 54.

  During the matching process, the speech recognition unit 52 calculates a likelihood that is the acoustic proximity between the speech feature data and each standby sentence, and recognizes a speech recognition result that has a likelihood equal to or greater than a certain value. The result is a candidate.

  The speech recognition unit 52 uses the top N recognition result candidates (hereinafter also referred to as N-best candidates) of the speech recognition results with the highest likelihood as the recognition result candidates and their likelihoods to the word reliability calculation unit 53. Output. In this recognition result candidate, reading IDs, which are identification codes in the reading database shown in FIG. 4 stored in the disk 40, which are given for each reading of words included in the recognition result candidate, are arranged in descending order of likelihood. Expressed as a word string.

  The word reliability calculation unit 53 calculates word reliability for each word reading for all words included in the recognition result candidates output as the speech recognition result from the speech recognition unit 52. That is, the word reliability calculation unit 53 treats words having the same meaning as different words if they are read differently, and calculates the word reliability.

Therefore, the word reliability is calculated in units of reading IDs of the reading database shown in FIG. The word reliability is a value indicating the possibility that a word is uttered in a single utterance, and the word reliability with respect to a reading IDW of a certain word W is set to Conf (W) and N-best candidates, respectively. If the log likelihood is Li, it can be obtained by the following equation (1).

  The word reliability for each reading ID calculated by the word reliability calculation unit 53 is all stored in the memory 30 as “a reading ID list of words included in recognition result candidates”. The calculation of word reliability by the word reliability calculation unit 53 is disclosed in Japanese Patent Application Laid-Open No. 2004-251998.

  The language understanding unit 54 first calculates a word reliability for a similar word whose reading is similar to the recognition result candidate that is the voice recognition result of the voice recognition unit 52 and is likely to be misrecognized. As a result of the speech recognition processing by the unit 52, it can be handled as a recognition result candidate in the same way as a word that is a recognition result candidate. Note that the word reliability setting procedure for similar words will be described in detail later.

  Further, the language understanding unit 54 classifies the words included in the recognition result candidates including similar words into categories that are semantically grouped, and adds the word reliability of words belonging to the same category, thereby adding a category score. Is calculated.

  If the main task of the navigation device is a destination setting when performing route search, the categories are, for example, “prefecture category”, “city category”, “route name category”, “ "Facility name category" can be considered. For example, prefecture names such as “Tokyo” and “Kanagawa” are categorized in “prefecture category”, and final purpose in destination setting such as “Yokohama Station” and “Yokohama Aoba Inter” in “facility name category” The destination name is classified.

  The language understanding unit 54 selects words to be output as an understanding result one by one from categories whose category score exceeds a predetermined threshold, and verifies all combinations. The language understanding unit 54 uses only combinations that are semantically consistent as understanding result candidates, adds the reliability of words constituting the understanding result candidates, and adds the results to the number of words constituting the understanding result candidate. The comprehension correction is made to calculate the understanding result score.

  The language understanding unit 54 outputs the understanding result candidate having the maximum understanding result score to the response generation unit 55 as the understanding result.

  The response generation unit 55 generates a response sentence based on the understanding result output from the language understanding unit 54 and outputs the response sentence to the GUI display control unit 56 and the speech synthesis unit 57.

  The GUI display control unit 56 controls the disk reading device 41 to read the map data stored in the disk 40 and display a map via the monitor 60 as needed, or the response generation unit 55 generates the map data. The response content corresponding to the response sentence is displayed via the monitor 60.

  The voice synthesizer 57 synthesizes a digital voice signal in accordance with the response sentence according to the response sentence generated by the response generator 55, and passes through a D / A converter and an output amplifier (not shown) included in the voice synthesizer 57. To the speaker 70.

  Next, processing operations from when the speech recognition process by the control device 50 is started until a response sentence is output will be described using the flowchart shown in FIG. In this flowchart, the main task of the navigation device will be described as a destination setting for route search.

  First, when the navigation device is activated in step S <b> 1, the control device 50 controls the disk reading device 41 to read the speech recognition dictionary / grammar from the disk 40 and store it in the storage area 31 of the memory 30.

  Then, when the user presses the switch 10, the input control unit 51 instructs the start of voice recognition, and the voice recognition unit 52 enters a voice recognition enabled state. In response to this, the voice recognizing unit 52 starts taking in a voice signal that is spoken by the user and input through the microphone 20 and digitized by an A / D converter (not shown).

  The voice recognition unit 52 continues to calculate the average power of the digitized voice signal (hereinafter also simply referred to as a digital signal) until the switch 10 is pressed. After the switch 10 is pressed, when the instantaneous power of the digital signal becomes greater than a predetermined value compared to the average power, it is determined that the user has spoken and the capturing of the digitized audio signal is started. .

  In step S <b> 2, the speech recognition unit 52 compares the captured digitized speech signal with the standby sentence held in the speech recognition dictionary / grammar constructed in the storage area 31 of the memory 30, thereby The speech recognition process is executed by calculating the likelihood.

  The speech recognition unit 52 outputs the top N recognition result candidates with high acoustic likelihoods and the likelihoods to the word reliability calculation unit 53 as speech recognition results. In this recognition result candidate, reading IDs, which are identification codes in the reading database shown in FIG. 4 stored in the disk 40, which are given for each reading of words included in the recognition result candidate, are arranged in descending order of likelihood. Expressed as a word string.

  In step S3, the word reliability calculation unit 53 calculates all words included in the recognition result candidates, that is, word reliability for each reading ID, from the recognition result candidates and the likelihood of each recognition result candidate.

  In step S4, the language understanding unit 54 stores in the disk 40 detailed information of all the words included in the recognition result candidates whose word reliability is calculated by the word reliability calculation unit 53 using the reading ID. It is taken out from the proper noun database and general database.

  In addition, the language understanding unit 54 reads a word whose word reliability is calculated from a similar word reading list set for each reading ID of the reading database stored in the disk 40 and a similar word that is easily misrecognized. And the similarity indicating the degree of misrecognition.

  In step S5, the language understanding unit 54 calculates the word reliability for the similar word using the reading ID and the similarity of the similar word extracted from the reading database of the disk 40, and sets each similar word. As a result, the language understanding unit 54 can handle similar words as recognition result candidates in the same manner as words that are recognized as recognition result candidates as a result of the speech recognition processing by the speech recognition unit 52.

  The procedure for setting the word reliability for similar words by the language understanding unit 54 will be described in detail later.

  In step S6, the language understanding unit 54 corrects the word reliability of all the words that may have been uttered in the dialogue up to this point.

  The correction of the word reliability is executed by increasing or decreasing the word reliability calculated in step S3 according to the presence or absence of consistency with the other words in the recognition result candidate or consistency.

  For example, if there is the word “Tokyo Station” in the recognition result candidate of the first utterance and “Tokyo” is in the recognition result candidate of the second utterance, ”And“ Tokyo Station ”have a vertical relationship, so the word reliability of each other is strengthened, and the word reliability is corrected in the direction of increasing. Also, if there is “Kyoto Station” instead of “Tokyo Station” in the recognition result candidate of the first utterance, it is between “Tokyo” and “Kyoto Station”. Since the hierarchical relationship does not hold, the mutual word reliability is weakened, and the word reliability is corrected in a decreasing direction.

  In step S7, the language understanding unit 54 classifies words included in recognition result candidates including similar words into categories that are semantically grouped, and adds the word reliability of words belonging to the same category. A category score is calculated.

  In step S8, the language understanding unit 54 selects a category for which the calculated category score exceeds a predetermined threshold.

  In step S9, the language comprehension unit 54 selects one word each as an understanding result from each category selected in step S8, verifies all combinations, and outputs them as understanding result candidates. Search for a combination that has sex.

  In step S <b> 10, the language understanding unit 54 determines whether one or more combinations that are semantically consistent as candidate understanding results have been found. If one or more combinations that can be semantically consistent are found, the language understanding unit 54 proceeds to step S11. If no combination is found, the language understanding unit 54 returns to step S8 and performs category selection again.

  In step S11, the language understanding unit 54 adds the word reliability of the words constituting each understanding result candidate, which is a combination that can be semantically consistent with the words selected from each category, and understands the result of the addition. An understanding result score is calculated by performing correction according to the number of words constituting the result candidate. Then, the language understanding unit 54 selects an understanding result candidate having the maximum understanding result score as the best understanding result, and outputs it to the response generation unit 55.

  In addition, the processing content from step S6 to step S11 by the language understanding part 54 is disclosed by Unexamined-Japanese-Patent No. 2004-251998.

  In step S <b> 12, the response generation unit 55 generates a response display content and a response sentence based on the understanding result generated by the language understanding unit 54, and outputs the response display content and the response sentence to the GUI display control unit 56 and the speech synthesis unit 57, respectively.

  In response to the understanding result, the response generation unit 55 generates a response sentence that prompts the user to input insufficient information if the information necessary for destination setting is insufficient, and configures the selected understanding result. If the word reliability of the word is low and it is determined that confirmation is necessary, a response sentence for confirming the understanding content is generated. In addition, when the destination is determined, the response generation unit 55 searches for a map up to the destination and generates a response sentence informing that the map is to be displayed.

  In step S13, the speech synthesizer 57 synthesizes a digital speech signal corresponding to the response sentence according to the response sentence generated by the response generator 55, and a D / A converter (not shown) provided in the speech synthesizer 57, It outputs to the speaker 70 via an output amplifier.

  In step S14, the GUI display control unit 56 displays the response display content generated by the response generation unit 55 on the monitor 60 and, if map display is necessary, the map data from the disc 40 using the disc reader 41. , The map is displayed on the monitor 60, and the series of input processes is completed.

(Word reliability setting procedure for similar word Wn: When obtaining from word reliability of word W)
Next, the procedure for setting the word reliability for the similar word in step S5 of the flowchart shown in FIG. 5 will be described using the flowchart shown in FIG.

  Before explaining the procedure for setting the word reliability of similar words, an example of a recognition result candidate list Lr1 that is a result of the speech recognition unit 52 performing speech recognition processing in step S2 of the flowchart shown in FIG. 7 shows. As shown in FIG. 7, the recognition result candidate list Lr1 describes a recognition result and its likelihood for each recognition result candidate rank. The recognition result indicates the reading ID and how to read the word.

  In the example illustrated in FIG. 7, the recognition result “11041 (Okutama)” with a likelihood of 60 is selected as the first recognition result candidate, and the likelihood is 40 as the second recognition result candidate. The recognition result “11041 (Okutama) +17011 (Eki)” is selected.

  On the other hand, FIG. 8 shows an example of the word reliability list Lc, which is a result of the word reliability calculation unit 53 calculating the word reliability in step S3 of the flowchart shown in FIG. As shown in FIG. 8, in the word reliability list Lc, the reading ID of the word that is the recognition result candidate described in the recognition result candidate list Lr1, how to read the word, and the calculated word reliability are described. Yes.

  In the example shown in FIG. 8, the reliability of the recognition result candidate “11041 (Okutama)” is 1.0, and the reliability of “17011 (Eki)” is 0.4.

  In the flowchart shown in FIG. 6, assuming that the recognition result candidate list Lr1 as shown in FIG. 7 and the word reliability list Lc as shown in FIG. 8 are obtained in step S2 of the flowchart shown in FIG. The procedure for setting the word reliability of the similar word in step S5 will be described.

  First, in step S <b> 21, the language understanding unit 54 extracts one word that is a recognition result candidate from the word reliability list Lc output from the word reliability calculation unit 53.

  For example, assume that the word extracted in this step is “11041 (Okutama)” in the unit reliability list Lc shown in FIG.

  In step S22, the language understanding unit 54 refers to the reading database stored in the disk 40, and obtains the reading ID and similarity of the similar word of the word (word W) that is the recognition result candidate extracted in step S21. A similar word reading list Lw is extracted.

  In step S23, the language understanding unit 54 extracts one similar word Wn from the extracted similar word reading list Lw.

  In the reading database shown in FIG. 4, as a similar word list of the word W “11041 (Okutama)”, “11061 (Oppama)” with a similarity of 0.5 and “ 11031 (Tama) "is registered, and here," 11061 (Oppama) "is extracted as the similar word Wn.

  In step S24, the language understanding unit 54 determines whether the similar word Wn extracted in step S23 is registered in the word reliability list Lc. In step S23, when “11061 (Oppama)” is extracted as the similar word Wn, the language understanding unit 54 refers to the word reliability list Lc to determine whether or not the word having the reading ID 11061 is already registered. to decide.

  If the similar word Wn is already registered in the word reliability list Lc, the language understanding unit 54 returns to step S22 and executes the process related to the next similar word Wn. If the similar word Wn is not registered in the similar word list Lc, the language understanding unit 54 proceeds to step S25.

  The purpose of checking whether or not the similar word Wn is present in the word reliability list Lc in step S24 is not appearing in the recognition result candidate list Lr1, but is easily misrecognized by a word appearing in the recognition result. Since the possibility of the word is to be checked, it is not necessary to check the similar word Wn if it is already registered in the word reliability list Lc obtained from the recognition result candidate list Lr1. Therefore, if a word with the same reading ID is registered in the word reliability list Lc, the next similar word Wn is examined.

  In step S25, in response to the fact that the similar word Wn taken out in step S23 is not registered in the word list Lc, the language understanding unit 54 now selects the similar word Wn as a similar word Wn of another word. It is determined whether or not it is already registered in the similar word Wn list Ln.

  The language understanding unit 54 proceeds to step S26 when the similar word Wn is not registered in the similar word Wn list Ln, and proceeds to step S29 when it is registered.

  In step S26, the language understanding unit 54 refers to the proper noun database and the general database on the disc 40 in response to the similarity word Wn extracted in step S23 being not registered in the similarity word list Ln, and this similarity word Detailed information of Wn is extracted.

  In step S27, the language understanding unit 54 retrieves the detailed information of the similar word Wn from the database of the disk 40, and then sets the word reliability for the similar word Wn. When “Oppama”, which is a word having a reading ID of 11061, is selected as the similar word Wn, the language understanding unit 54 sets the word reliability for this “Oppama”.

  The word reliability of the similar word Wn is expressed as follows when the word reliability of the word W obtained in step S3 of the flowchart shown in FIG. 5 is used and the similarity between the word W and the similar word Wn is β (2) ) Can be expressed as:

    Word reliability of similar word Wn = word reliability of word W × β (2)

  “Okutama”, which is a word having a reading ID of 11041, is selected as the word W, and “Oppama” having a reading ID of 11061 is selected as a similar word Wn whose reading is similar to the word W, and a word reliability calculation is performed. It is assumed that the word reliability of “Okutama” is calculated as 1.0 by the unit 53. At this time, the word reliability of “Oppama”, which is the similar word Wn, is 0.5 because the similarity between “Okutama” and “Oppama” is 0.5 from the reading database shown in FIG. 2) Using the equation, it can be calculated as follows.

  Word reliability of “Oppama” = Word reliability of “Okutama” (1.0) × Similarity between “Okutama” and “Oppama” (0.5) = 0.5

  In step S28, the language understanding unit 54 adds the reading ID of the similar word Wn for which the word reliability has been calculated to the similar word list Ln together with the detailed information acquired from the disk 40 and the word reliability.

  In step S29, the language understanding unit 54 determines whether the similar word Wn in the similar word list Ln is already registered as a similar word of another word in the similar word list Ln. Update confidence.

  The language understanding unit 54 determines the word reliability before update of the similar word Wn already registered in the similar word list Ln, the word reliability of the word W obtained in step S3 of the flowchart shown in FIG. Using the similarity β of the similar word Wn, it can be updated by the following equation (3).

  Word reliability after update of Wn = Word reliability before update of Wn + Word reliability of W × β (3)

  The pre-update word reliability of Wn on the right side is the word reliability of similar words registered in the similar word list Ln, and β is the similarity between the word W and the similar word Wn.

  As shown in the equation (3), when the similar word Wn is already registered in the similar word list as the similar word Wn of another word, by adding the current degree of similarity to the word reliability, A new word reliability, that is, an updated word reliability can be calculated.

  In step S30, the language understanding unit 54 sets the word reliability for the similar word Wn, then refers to the similar word reading list Lw of the word W acquired in step S22, and all the similar words for the word W are determined. Determine whether word confidence is set.

  The language understanding unit 54 proceeds to step S31 when the word reliability of all the similar words Wn is set, and returns to step S22 when there is a similar word Wn for which the word reliability is not yet set. The word reliability calculation or the word reliability update process for the similar word Wn is executed.

  In step S31, the language understanding unit 54 determines whether or not the word reliability is set for the similar words Wn for all the words W registered in the word reliability list Lc.

  If the word reliability is set for the similar word Wn for all the words W, the language understanding unit 54 proceeds to step S32. If there is a word W that has not yet been processed for the similar word Wn, the language understanding unit 54 returns to step S21 and repeats steps S21 to S31.

  In step S32, the language understanding unit 54 completes the setting of the word reliability for the similar words Wn for all the words W registered in the word reliability list Lc, and the contents of the similar word list Ln are stored in the word reliability list Lc. Add to

  FIG. 9 shows a similar word list Ln of similar words corresponding to all the words W registered in the word reliability list Lc shown as an example in FIG.

  In the word reliability list Lc shown in FIG. 7, “Okutama” whose reading ID is 11041 is “Oppama” whose reading ID is 11061, as can be seen from the reading database shown in FIG. 4. And two similar words “Tama”, which is a word having a reading ID of 11031, exist.

  Further, in the word reliability list Lc shown in FIG. 8, “Eki” whose reading ID is 17011 is “upeki” whose reading ID is 11051, as can be seen from the reading database shown in FIG. 4. One similar word exists.

  Since these are words that are not registered in the word reliability list Lc shown in FIG. 8, they are all added to the similar word list Ln as shown in FIG.

  FIG. 10 shows an example in which the contents of the similar word list Ln shown in FIG. 9 are added to the word reliability list Lc shown in FIG. Thus, in step S6 and subsequent steps after step S5 of the flowchart shown in FIG. 5, the similar word Wn for all the words W in the recognition result candidate list Lr1 is equivalent to the word W in the recognition result candidate list Lr1. Will be able to handle.

  As described above, not only the word W obtained as a result of the speech recognition, but also the similar recognition word Wn that is easily misrecognized as the word W has been leaked in the speech recognition processing by the speech recognition unit 52, but is finally recognized. Can be treated as a result candidate. Therefore, even if the utterance is a legitimate utterance by the user, the possibility of being selected as an understanding result can be left even if it is excluded in the process of selecting a recognition result candidate.

  In addition, the similar word is uttered by the user by indicating the closeness of the similarity between the similar word and the recognition target word by using the degree of similarity of how much the recognition target word is easily misrecognized. Since the possibility can be determined based on the similarity, an accurate understanding result can be generated.

  Furthermore, since the word reliability of a word included in the recognition result candidate and the word reliability of a similar word are obtained, respectively, it is possible to determine the utterance possibility of each word based on the same determination criterion. When a similar word having a high is included in the recognition result candidate, the possibility of being uttered according to the level of word reliability can be verified, so that a more accurate understanding result can be generated.

(Word reliability setting procedure for similar word Wn: When obtaining from likelihood)
Next, another procedure for setting the word reliability of the similar word Wn will be described. In the example described above, the word reliability of the similar word Wn having a high similarity with the word W in the recognition result candidate list Lr1 is calculated from the word reliability of the word W in the recognition result candidate list Lr1 using the formula (2) or ( 3) Calculated using the equation.

  On the other hand, the speech recognition process is performed again on the speech uttered by the user and input via the microphone 20, the acoustic likelihood is obtained, and the word reliability of the similar word Wn is calculated from the likelihood. The method to do is explained.

  With reference to the flowchart shown in FIG. 11, another procedure for setting the word reliability of the similar word Wn in step S5 of the flowchart shown in FIG. 5 will be described.

  In FIG. 11, steps S41 to S45 for checking whether or not the similar word Wn is registered in the similar word list Ln are exactly the same as steps S21 to S25 in the flowchart shown in FIG. Therefore, the description is omitted.

  Further, in step S1 of the flowchart shown in FIG. 5 which is the preceding stage of steps S41 to S52 shown in FIG. 11, the voice signal uttered by the user and inputted through the microphone 20 is digitized, for example, It is assumed that the memory 30 is temporarily buffered.

  In step S46, the language understanding unit 54 refers to the proper noun database and the general database on the disk 40 in response to the similarity word Wn extracted in step S23 being not registered in the similarity word list Ln, and this similarity word Detailed information of Wn is extracted.

  In step S47, the language understanding unit 54 adds the similar word Wn to the similar word list Ln.

  In step S48, the language understanding unit 54 reads the grammar that can recognize the similar word Wn added to the similar word list Ln from the disk 40 by controlling the disk reading device 41, and the storage area of the memory 30. 31 is stored.

  In response to this, the voice recognition unit 52 reads the digitized voice signal temporarily buffered in the memory 30, and further recognizes the similar word Wn stored in the storage area 31 of the memory 30. The speech signal recognition process is executed again by calculating the acoustic likelihood in comparison with the possible grammar.

  In step S49, the language understanding unit 54 recognizes the recognition result together with the likelihood in the recognition result list Lr2 in response to the recognition result obtained by the second speech recognition process for the speech signal digitized by the speech recognition unit 52. Add to

  In step S50, the language understanding unit 54 refers to the similar reading list Lw of the word W acquired in step S42, and determines whether or not the word reliability is set for Wn for all similar words for the word W.

  The language understanding unit 54 proceeds to step S51 when the word reliability of all the similar words Wn is set, and returns to step S42 when there is a similar word Wn for which the word reliability is not yet set. .

  In step S45, if the similar word Wn exists in the similar word list Ln, the language understanding unit 54 has already performed speech recognition processing by the speech recognition unit 52 using a grammar that always accepts the similar word Wn. Therefore, it will progress to this step S50.

  In step S51, the language understanding unit 54 determines whether or not the word reliability is set for the similar words Wn for all the words W registered in the word reliability list Lc.

  If the word reliability is set for the similar word Wn for all the words W, the language understanding unit 54 proceeds to step S52. If there is a word W that has not yet been processed for the similar word Wn, the language understanding unit 54 returns to step S41 and repeats steps S41 to S51.

  In step S <b> 52, the word reliability calculation unit 53 is the result of performing the speech recognition process on the same speech signal for the second time as the recognition result candidate list Lr <b> 1 that is the result of speech recognition processing of the speech signal first by the speech recognition unit 52. The word reliability is calculated from the recognition result candidate string combined with the recognition result candidate list Lr2.

  In this way, the voice interaction apparatus performs the voice recognition process again on the voice uttered by the user and input through the microphone 20, and obtains the acoustic likelihood, and the similar word Wn is obtained from this likelihood. The word reliability of is calculated.

  Thereby, the voice interactive apparatus can select an optimum understanding result while comparing the word reliability of the word in the recognition result candidate included in the first recognition result with the word reliability of the similar word Wn.

(Expansion of recognition result candidates)
As described above, the spoken dialogue apparatus shown as the embodiment of the present invention can treat the similar word Wn that is easily misrecognized as the word W as a final recognition result candidate. By using not only the similar word Wn corresponding to 1 but also a similar phrase that is easily misrecognized as the word W by a combination of readings, it is possible to reduce words that are leaked from the final recognition result candidates. .

  FIG. 12 shows a state where the reading database shown in FIG. 4 is expanded and a similar phrase reading list for words for each reading ID is added. A similar phrase is a word string that is easily misrecognized as a way of reading a target word when two or more words are combined and has a similarity.

  For example, as shown in FIG. 12, the similar phrases for “Ueki” whose reading ID is 11051 are “Shingu” whose reading ID is 11071 and “Eki” whose reading ID is 17011. The word string “Shingu Ueki” is a combination of As shown in FIG. 12, the similarity of this similar phrase is 0.8.

  As a result, when the recognition result “Ueki” is obtained, “Shingu” and “Eki” may be the final recognition result candidates, and the word reliability can be calculated from the similarity and the utterance can be made The sex can be verified.

  Therefore, even if the utterance is a legitimate utterance by the user, the possibility of being selected as an understanding result can be left even if it is excluded in the process of selecting a recognition result candidate.

  On the other hand, by using similar words having a high degree of similarity with a word string composed of two or more words included in the recognition result candidates, it is possible to reduce words that are leaked from the final recognition result candidates. it can.

  For example, as shown in FIG. 13, a similar word having a high degree of similarity with respect to a word string in which “shingu” having a reading ID of 11071 and “eki” having a reading ID of 17011 are combined. Then, “Ueki”, which is a word with a reading ID of 11051, is registered. As shown in FIG. 13, the similarity of this similar word is 0.8.

  As a result, when the recognition results “Shingu” and “Eki” are obtained, “Ueki”, which is part of the word string, may be the final recognition result candidate. It is possible to calculate and verify the utterance possibility.

  Therefore, in this case as well, there is a possibility of being selected as an understanding result even if it is excluded in the process of selecting a recognition result candidate in spite of a legitimate utterance by the user.

  Note that the similarity between word strings composed of two or more words as shown in FIG. 13 is calculated in advance only for combinations of words that can be actually recognized.

  The above-described embodiment is an example of the present invention. For this reason, the present invention is not limited to the above-described embodiment, and various modifications can be made depending on the design and the like as long as the technical idea according to the present invention is not deviated from this embodiment. Of course, it is possible to change.

It is a figure for demonstrating the structure of the voice interactive apparatus shown as embodiment of this invention. It is the figure which showed an example of the proper noun database. It is the figure which showed an example of the general database. It is the figure which showed an example of the reading database. It is a flowchart for demonstrating the processing operation | movement after starting a speech recognition process until it outputs a response sentence. It is a flowchart for demonstrating the procedure which sets the word reliability of a similar word. It is the figure which showed an example of the recognition result candidate list. It is the figure which showed an example of the word reliability list. It is the figure which showed an example of the similar word list. It is the figure which showed a mode that the similar word list was added to the word reliability list. It is a flowchart for demonstrating the procedure which sets the word reliability of a similar word. It is the figure which showed the similar punctuation list. It is the figure which showed the similar word and similarity of the word sequence comprised by two or more words.

Explanation of symbols

DESCRIPTION OF SYMBOLS 10 Switch 20 Microphone 30 Memory 40 Disk 50 Control apparatus 51 Input control part 52 Speech recognition part 53 Word reliability calculation part 54 Language understanding part 55 Response generation part 56 GUI display control part 57 Speech synthesis part 60 Monitor 70 Speaker

Claims (7)

Input means for inputting spoken speech;
Storage means for storing the speech recognition target word and a similar word that is easily misrecognized by the recognition target word;
Speech recognition means for recognizing the speech input by the input means based on the recognition target word;
Detecting means for detecting the similar word from the storage means based on a word included in a recognition result candidate that is a recognition result by the voice recognition means;
Understanding result generating means for adding the similar word detected by the detecting means to the recognition result candidate and generating an understanding result as a response to the spoken voice from the recognition result candidate obtained by adding the similar word; A spoken dialogue apparatus characterized by comprising: The degree of misrecognition of the similar word and the recognition target word is indicated by a degree of similarity obtained by quantifying how much the similar word is erroneously recognized by the recognition target word. The spoken dialogue apparatus according to 1. Word reliability calculation means for calculating a word reliability indicating the possibility that each of the words included in the recognition result candidate and the similar word detected by the detection means is spoken,
The understanding result generation means generates an understanding result as a response to the spoken speech from the recognition result candidate to which the similar word is added based on the word reliability calculated by the word reliability calculation means. The voice interactive apparatus according to claim 1. The speech recognition means recognizes the speech using a first grammar that recognizes the recognition target word, recognizes the speech using a second grammar that recognizes the similar word,
The understanding result generation means generates an understanding result as a response to the spoken speech from recognition result candidates that are recognition results using the first grammar and the second grammar by the speech recognition means. The voice interactive apparatus according to claim 1. The understanding result generation means determines how much a word included in the recognition result candidate is erroneously recognized in a word string formed by connecting two or more recognition target words or a part of the word string. In accordance with the quantified similarity, the two or more recognition target words connected as the word string are added to the recognition result candidates,
The spoken dialogue apparatus according to claim 1, wherein an understanding result as a response to the spoken voice is generated from a recognition result candidate to which the two or more recognition target words are added. The understanding result generation means is a numerical value indicating how much a word string in which two or more words included in the recognition result candidate are connected or a part of the word string is erroneously recognized by the recognition target word. Depending on the similarity, the recognition target word is added to the recognition result candidate,
The spoken dialogue apparatus according to claim 1, wherein an understanding result that is a response to the spoken voice is generated from a recognition result candidate to which the recognition target word is added. An input process for inputting spoken voice;
A speech recognition step for recognizing the speech input by the input step based on the recognition target word;
Based on a word included in a recognition result candidate that is a recognition result by the speech recognition step, the similarity is stored from a storage unit that stores the speech recognition target word and a similar word that is easily misrecognized by the recognition target word in association with each other. A detection step of detecting a word;
An understanding result generation step of adding the similar word detected by the detection step to the recognition result candidate, and generating an understanding result as a response to the spoken speech from the recognition result candidate to which the similar word is added. A speech understanding result generation method characterized by comprising:

Images