JP2009282835A - Method and device for voice search - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To provide a voice search device allowing a user to easily search for a wanted data. <P>SOLUTION: The voice search device comprises a voice recognition part B11, a candidate management part B21, a result outputting part B31, a relevancy calculating part B41, a recognition dictionary grammar D101, a search database D201, and a candidate database D301. A search voice for searching which is vocalized by a user and a narrowing-down voice for narrowing down the search are voice-recognized to acquire a search candidate and a narrowing-down candidate. Based on the search candidate, the search database D201 is searched for acquiring a plurality of data and attribute information which the data have as search information. Based on the narrowing-down recognition candidate, the search database D201 is searched for acquiring a plurality of data and the attribute information which the data have as narrowing-down information. The relevancy whose value becomes larger as relevancy between a plurality of respective data contained in the search information and the plurality of respective data contained in the narrowing-down information rises is calculated from the attribute information of respective data contained in the search information and the attribute information of respective data contained in the narrowing-down information. <P>COPYRIGHT: (C)2010,JPO&INPIT

Description

  The present invention relates to a speech search apparatus and method for selecting data desired by a user from a large database having a plurality of attributes.

With recent advances in audio signal processing technology and semiconductor technology, car navigation systems that allow users to search and set destinations from a large facility database by repeatedly exchanging voices between the system and users, even for small devices such as in-vehicle devices The system is realized.

  In car navigation systems that are currently on the market, there are several destinations such as address search that specifies a destination by address, zip code search that specifies a destination by zip code, and facility search that directly specifies a target facility. Means are provided.

  The facility search is designed so that facility names can be entered after narrowing the vocabulary of the facility names registered in the database to hundreds of thousands of facilities by utterances that limit the region in order from the prefecture name. Many. This is due to the fact that limited computational resources such as in-vehicle devices cannot correctly recognize large-scale vocabularies in real time, forcing users to take time and effort to operate at practical performance and speed. Designed to be

  For this reason, particularly when the user knows the facility name clearly, the utterance for designating attributes other than the facility name becomes redundant, and it involves a problem that the interface is troublesome and difficult to use.

  As a method for solving this problem, in Patent Document 1, when speech recognition is operated to extract a plurality of recognition candidates for a large-scale recognition vocabulary and the candidates cannot be sufficiently narrowed down, , Prompting the input of related questions to narrow down candidates, recognizing utterances to related questions, obtaining recognition candidates for narrowing down, and narrowing down from the relationship between the candidates to be narrowed down and candidates as the narrowing conditions Has been proposed.

  Further, this method has a feature that the narrowed candidates are rearranged based on the recognition likelihood of candidates to be narrowed down and the recognition likelihood of candidates as narrowing conditions.

However, this method has a problem in that the dialogue is broken when the user cannot answer the related question. Considering the facility name search as an example, there is a situation in which the name of the target facility is known, but the address and the genre to which the facility belongs are not known, so the question cannot be answered. In other words, this method is a solution based on the premise that many attribute values related to the target data to be searched are remembered, and cannot be said to be a method that can reliably narrow down.
Japanese Patent No. 3420965

  As described above, the conventional technique has a problem in that various attribute values must be remembered with respect to the target data in order to answer related questions that narrow down the target data from the large vocabulary.

  Accordingly, in view of the above problems, the present invention provides a voice search device and method for narrowing down using only relevant information known by the user when the initial utterance cannot be sufficiently narrowed down. It is to provide.

  The present invention recognizes a search voice uttered by a user and a search voice for narrowing down the search, and obtains a search candidate and a search candidate, a plurality of data, A first database storing attribute information corresponding to each data; and (1) searching the first database based on the search candidates to search a plurality of the data and the attribute information corresponding to the data. (2) a candidate management unit that searches the first database based on the refinement recognition candidates and obtains a plurality of the data and the attribute information as refinement information, and a plurality of the search information included in the search information The degree of relevance that increases as the relevance between each of the data and each of the plurality of data included in the refinement information increases. Relevance calculation unit for calculating from the attribute information of each data included in the search information and the attribute information of each data included in the narrowing down information, and the relevance degree having a high value And an output unit that outputs one or a plurality of the data included in the search information.

  According to the present invention, even when it is not possible to narrow down by the search voice, it is possible to easily narrow down from information related to the target data known by the user.

  Hereinafter, a voice search device according to an embodiment of the present invention will be described with reference to the drawings.

(First embodiment)
A voice search device according to a first embodiment of the present invention will be described with reference to FIGS.

  The voice search apparatus according to the present embodiment is for selecting data desired by a user by voice from a large-scale database having a plurality of attributes.

  More specifically, 1) When the target data cannot be sufficiently narrowed down by the first utterance of the user, 2) The related question is presented to the user and the attribute input by the next utterance is designated. 3) After receiving an utterance related to the target data, 4) calculating a relevance level indicating the semantic relationship between the candidate included in the first utterance and the candidate included in the next utterance; ) Narrow down candidates in view of the recognition likelihood and relevance of each of the first utterance and the next utterance.

(1) Configuration of Voice Search Device FIG. 1 is a configuration diagram of a voice search device according to the present embodiment.

  As shown in FIG. 1, the voice search apparatus includes a voice recognition unit B11, a candidate management unit B21, a result output unit B31, a relevance calculation unit B41, a recognition dictionary grammar D101, a search database D201, and a candidate database D301.

  The function of each part B11-B41 is realizable also by the program transmitted or stored in the computer.

(2) Explanation of each part Hereinafter, each composition of a voice search device is explained.

(2-1) Search database D201
The search database D201 is a database that stores one or more data to be searched in the voice search device and attribute information corresponding to the data.

  An example of the search database is a facility name database stored in attribute information such as a prefecture name, city name, city name, latitude / longitude, and genre stored in the car navigation system.

  Here, “attribute information” refers to the prefecture, city name, latitude / longitude, and genre of the data of the location unique to the data in order to specify the data such as the facility name.

(2-2) Recognition dictionary grammar D101
The recognition dictionary grammar D101 describes a vocabulary list, grammar, etc. for accepting words, multiple word sequences, phrases, or sentences that a user inputs as search keys to select target data from the search database D201. It is a thing.

(2-3) Speech recognition unit B11
The speech recognition unit B11 recognizes speech input from the user based on the recognition dictionary grammar D101 and generates recognition candidates.

  Except for the case where no candidate close to the input speech is output because the input speech has noise or the like, one or more recognition candidates are generated by the speech recognition unit B11.

  Moreover, you may operate | move so that the recognition score showing the probability of a recognition candidate may be acquired with respect to each candidate.

  Note that the voice recognition unit B11 has a configuration in which a microphone, an amplifier that amplifies voice, an AD converter, and the like are incorporated in the previous stage, and operates so as to perform voice recognition after converting the input voice to a digital voice signal. Alternatively, the digital audio signal may be stored in a recording medium, and the audio signal may be read from the recording medium as necessary.

(2-4) Candidate manager B21
When the candidate management unit B21 receives the recognition candidate and the recognition score generated by the speech recognition unit B11, the candidate management unit B21 generates a search key from the recognition candidate, searches the search database D201, and acquires data that matches the search condition.

  Further, at this time, it may operate so as to acquire a plurality of attribute information possessed by the data.

  The retrieved data and recognition candidates are stored in the candidate database D301 as history information.

  Further, the candidate management unit B21 refers to the candidate database D301, and when a plurality of utterances are managed as history information, the candidate management unit B21 determines the utterances to be narrowed down and the utterances to be narrowed down, and In order to narrow down the candidates based on the relationship, the relevance calculation unit B41 is inquired about the semantic similarity between candidates included in each utterance, and the relevance is acquired.

  It operates to narrow down candidates to be narrowed down using the acquired degree of association and recognition score.

  When candidates are narrowed down, the candidate database D301 is updated and the result is notified to the result output unit B31.

(2-5) Relevance calculation unit B41
The relevance calculation unit B41 operates to calculate the semantic similarity between candidates inquired from the candidate management unit B21 as the relevance and return the calculation result to the candidate management unit B21.

  Detailed operations of the relevance calculation unit B41 and the candidate management unit B21 will be described later in detail using a specific example.

(2-6) Result output unit B31
The result output unit B31 presents the search result notified from the candidate management unit B21 to the user.

  As a configuration including a display device in the result output unit B31, the search result may be displayed on the screen, or as a configuration including a speech synthesizer and a speaker, it may operate to read out candidates in order. Any configuration may be used as long as the result is presented.

(3) Specific Example Next, a specific operation of the voice search device will be described.

  As a specific example, a description will be given by taking as an example a situation in which a facility name is retrieved by voice from a large-scale facility database in order to achieve a task of setting a destination using a car navigation system as a motif.

  FIG. 2 shows an example of a facility name database to be searched. As shown in FIG. 2, in the facility name database, names of 10,000 facility data, names of prefectures and cities where the facilities are located, genre names of facilities expressed hierarchically, and attribute information of latitude and longitude are stored. ing.

  FIG. 3 shows an example of a recognition dictionary grammar for recognizing a voice input from a user. As shown in FIG. 3, a facility name of 10,000 words in the facility data is registered, and is described so as to accept a word utterance of the facility name.

  This is a detailed operation of the facility search task, taking as an example a situation where the user wants to set “Mt. Takao” in Tokyo as a destination.

(4) Operation of Voice Recognition Unit B11 First, the operation of the voice recognition unit B11 will be described.

  When the user who is prompted to input the facility name utters “Takao”, which is the facility name to be searched, the speech recognition unit B11 is input according to the recognition dictionary grammar described in FIG. Output multiple words that match the voice.

  There are various methods for setting the number of recognition word candidates to be output. For example, only candidates whose recognition score exceeds a certain threshold can be selected. In this embodiment, words whose recognition score exceeds 50 are output as candidates.

  FIG. 4 illustrates a recognition result in which 100 candidates are output by the speech recognition unit B11. Here, in FIG. 4, words having the same facility name are displayed as recognition candidates. However, when entries having the same reading that are treated equivalently in speech recognition are collectively output in the recognition dictionary grammar as a recognition result, It is efficient to generate candidates with the same reading.

(5) Operations of Candidate Management Unit B21 and Candidate Database D301 Next, operations of the candidate management unit B21 will be described with reference to the flowchart of FIG. FIG. 5 shows a flowchart of the candidate management unit B21 that searches the search database D201 using the received recognition result and narrows down candidates from a plurality of utterances.

  First, the candidate management unit B21 acquires the speech recognition result illustrated in FIG. 4 (step S2101).

  Next, a search key for searching the search database D201 is extracted from the acquired speech recognition result (step S2102). In this embodiment, since the word received from the user is the facility name, the recognition result itself becomes a search key for subtracting the facility name that is an attribute of the search database D201, and 100 search keys including duplicates are described. Generated.

  Next, the search database D201 is searched using a plurality of generated search keys, data that matches the search condition and its attributes are acquired (step S2103), and the recognition result and the search result are linked to obtain the candidate database D301. Is recorded as history information (step S2104). FIG. 6 shows an example of history information written in the candidate database D301, and the order in which results are acquired, recognition results (scores and ranks), attribute information of the corresponding search database D201, and the like are associated with each other. Although the candidate database D301 shows an example in which all attribute information of data is stored, only the pointer to the search database D201 may be held.

  Next, it is determined whether the newly acquired recognition result is an utterance for narrowing down or a newly input utterance (S2105). Here, since the recognition result is acquired first, it is determined that the utterance is not a narrowed utterance but a newly input utterance, and the operation of the candidate management unit B21 is finished.

(6) Result output and refined utterance When the candidate management unit B21 that searches the search database D201 using the key acquired from the user's utterance finishes, the result output unit B31 receives the updated candidate list, The result is notified to the user.

  For example, when notifying users by reading the candidate results in order from the most likely, the prompt “100 candidates found. The first is in Kumamoto Prefecture. Nakaoyama, the second is Mt. Takefu in Ibaraki Prefecture. When notification is made to the user using the screen, the top candidates can be displayed as a list on the screen.

  When the target data is included in the upper rank of a plurality of candidates, it is possible to achieve the purpose by input specifying it, but in the present embodiment, it is assumed that the target data is not found immediately.

  By confirming the output result, the user grasps that the target data could not be narrowed down by the first utterance, and performs further input for narrowing down.

  For example, consider a case in which the user remembers Lake Sagami near Mt. Takao and inputs “Lake Sagami” as the utterance for narrowing down as related information.

  The speech recognition unit B11 outputs a plurality of words that match the input speech according to the recognition dictionary grammar written in FIG. FIG. 7 illustrates a recognition result in which 30 candidates are output by the speech recognition unit B11.

(7) Operations of Candidate Manager B21 and Candidate Database D301 Next, operations of the candidate manager B21 when a recognition result for the narrowed utterance is generated will be described with reference to the flowchart of FIG.

  First, the candidate management unit B21 acquires the speech recognition result illustrated in FIG. 7 (step S2101).

  Next, a search key for searching the search database D201 is extracted from the acquired speech recognition result (step S2102). In this embodiment, since the word received from the user is the facility name, the recognition result itself becomes a search key for the facility name that is an attribute of the search database D201, and 30 search keys including duplicates are generated. .

  Next, the search database D201 is searched using a plurality of generated search keys, data that matches the search condition and its attributes are acquired (step S2103), and the recognition result and the search result are linked to obtain the candidate database D301. Is recorded as history information (step S2104). FIG. 8 shows an example of history information written in the candidate database D301, in which the order in which results are acquired, recognition results (scores and ranks), attribute information of the corresponding search database D201, and the like are associated with each other. Although an example in which all the attribute information of data is stored in the candidate database D301 is shown, only a pointer to the search database D201 may be held.

  Next, it is determined whether the newly acquired recognition result is an utterance for narrowing down or a newly input utterance (S2105).

  Since there are two recognition result groups input here, it is determined that the utterance input for the second time is a narrowed utterance, and the degree of association for each candidate between utterances is inquired to the association degree calculation unit B41. (S2106). In the present embodiment, 100 candidates from the first utterance to be narrowed down and 30 candidates from the next utterance to be narrowed down are recorded in the candidate database D301, so the degree of association between 100 × 30 = 3000 candidates is calculated. A specific method of calculating the degree of association will be described in detail subsequently.

(8) Operation of Relevance Calculation Unit B41 Next, the operation of the relevance calculation unit B41 will be described.

  The degree-of-association calculation unit B41 calculates the degree of association representing the semantic similarity by comparing data attribute information for each candidate pair inquired from the candidate management unit B21.

  The attribute information necessary for calculating the relevance may be received from the candidate management unit B21 as a set with the candidate, or the relevance calculation unit B41 may directly refer to it from the pointer to the search database D201.

  The specific calculation method of the degree of association is based on the assumption that, for example, a target existing at one point, for example, a facility with a close latitude and longitude, that is, a nearby facility is uttered as related information when searching for a facility name. The degree of relevance is calculated by paying attention to the latitude and longitude attribute information among various attributes. By comparing the latitude and longitude of the candidates, for example, the Euclidean distance can be obtained. Based on the obtained Euclidean distance, the relevance level between candidates is calculated by normalizing the relevance level so that the relevance level decreases as the distance increases and the relevance level increases as the distance decreases.

  An example of the result of performing this calculation for all candidates 100 × 30 is shown in FIG.

(9) Narrowing Operation by Candidate Manager B21 Returning to the operation of candidate manager B21 again with reference to FIG.

(9-1) Step S2107
As shown in FIG. 9, the degree of association between candidates is calculated by the degree-of-association calculation unit B41, and when the result is notified to the candidate management unit B21, the candidate to be narrowed down is narrowed down using the calculated degree of association. The narrowing-down score is calculated (S2107).

  As a method of calculating the i th narrowing score Di (0 <i <= 100) from the relevance degree Wij (relationship between the i th narrowing target candidate and the j th narrowing condition candidate) and the j th recognition score Rj. There are the following methods.

Di = Σj (Wij) x (Rj)

However, 0 <j <= 30. FIG. 10 shows the result of calculating the narrowing score according to the above formula. As shown in the example above, the score for determining whether or not to narrow down is not limited to the semantic relevance, but the reliability of recognition in the utterance for narrowing down is used to narrow down the reliability. The purpose is to do.

(9-2) Step S2108
When the narrowing score is calculated in this way, the first 100 facility candidates to be narrowed down are narrowed down in view of the narrowing score and the recognition score (S2108).

  For example, if a candidate for which the refinement score is 0 is removed as a refinement and an operation is performed so as to perform a resort with the total value of the recognition score and the refinement score, the three facilities illustrated in FIG. 11 remain as candidates, and these are stored in the candidate database D301. Is written as history information.

(10) Result output When the search database D201 is searched using the key acquired from the user's utterance and the process of narrowing down the first utterance is completed, the result output unit B31 receives the updated candidate list, Notify the user of the result.

  For example, when notifying users by reading the candidate results in order from the most likely one, the following prompt “Three candidates were found. The first one is in Tokyo. Mt. Takao, the second is ... "

  Here, the user can confirm that the target facility has been read out as the first candidate and confirm that the target candidate has been searched, and can input the target target by utterance or designating the screen to specify the first candidate. Data can be specified.

(11) Effects According to the present embodiment, 1) When selecting data desired by a user from a large-scale database having a plurality of attributes, 2) sufficiently narrowing down target data by the first utterance If not, 3) present the related question to the user and not specify the attribute to be input in the next utterance; 4) accept the utterance related to the target data; The degree of association representing the semantic relation between the candidate included in the utterance and the candidate included in the next utterance is calculated, and 6) narrowing down candidates in consideration of the recognition likelihood and the degree of association of the first utterance and the next utterance, respectively. To do.

  Therefore, it is possible to avoid the situation where the user fails to answer the related question from the system and the dialogue is broken, and the target data can be selected from a large-scale database.

(Second Embodiment)
A voice search device according to a second embodiment of the present invention will be described with reference to FIGS.

(1) Configuration of Voice Search Device FIG. 12 is a schematic configuration diagram of a voice search device according to the present embodiment.

  In the speech search apparatus according to the present embodiment, the relevance calculation unit B42 operates differently from the relevance calculation unit B41 in the first embodiment, and the relevance calculation method determination unit B51 and the relevance calculation method database D401 are new. It has been added.

  In addition, since the thing to which the same number as the block diagram in 1st Embodiment was attached | subjected performs the same operation | movement, description is abbreviate | omitted.

  This embodiment is different from the previous embodiment in that the relevance calculation method is dynamically switched with reference to the relevance calculation database D401.

  The degree of association calculation method database D401 stores a method for calculating the degree of association between data and one or more conditions for adopting the calculation method.

  The relevance calculation method determination unit B51 determines a method for the relevance calculation unit B42 to calculate the relevance with reference to candidate attribute information to be calculated and the relevance calculation method database D401.

(2) Relevance calculation method determination unit B51
FIG. 13 shows an example of the relevance calculation method database D401. In the facility search task in the car navigation handled in this embodiment, the genre of the facility exists in the attribute in the search database D201, and the condition is such that the calculation method to be applied is determined by the attribute information of this genre.

  A specific calculation method is as follows.

  First, a facility such as a station that serves as a user's action base point is calculated so as to have a relevance value only for a facility with a narrow range such as a walking distance.

  Second, for facilities that exist over a certain area rather than points such as roads and rivers, a plurality of points registered as facility data with the same facility name are complemented with straight lines, and the straight lines supplemented with candidates for narrowing down Only the facilities where the shortest distance from is within a certain range are calculated so as to have a relevance value.

  In this way, by switching the calculation method in correspondence with the facility that is input as the narrowing-down condition and intuitive proximity, it is possible to avoid a situation in which the degree of association of unintended candidates becomes high and cannot be narrowed down sufficiently.

  Then, the relevance calculation method database D401 stores the calculation information as described above using a calculation function with the attribute information as input information and the relevance as output information, and selects the calculation function according to the genre attribute information.

(3) Effect According to the present embodiment, it is possible to calculate the relevance by a method close to the user's intuition by holding a plurality of relevance calculation methods and switching the calculation method according to the candidates to be compared. As a result, the user can narrow down candidates efficiently.

(Third embodiment)
A voice search device according to a third embodiment of the present invention will be described with reference to FIGS.

(1) Configuration of Voice Search Device FIG. 14 is a schematic configuration diagram of a voice search device according to the present embodiment.

  The speech search apparatus according to this embodiment is different in the operations of the recognition dictionary grammar D102, the candidate management unit B22, the relevance calculation unit B43, and the relevance calculation method determination unit B52 in the second embodiment, and additionally includes a phrase analysis unit B61. Is a newly added configuration.

  Components having the same reference numerals as those in the previous embodiment have the same operations, and thus description thereof is omitted.

  In the present embodiment, the point that the utterance including the phrase suggesting the relationship between utterances is received, the phrase is analyzed, and the calculation method of the degree of association is switched according to the phrase obtained from the analysis result is different from the previous embodiment. Is different.

(2) Recognition dictionary grammar D102
FIG. 15 shows an example of a recognition dictionary grammar D102 for accepting an utterance including a phrase optimal for the present embodiment.

  As shown in FIG. 15, the facility data is described so as to accept the utterance in which the facility name is 10,000 words alone or the facility name is followed by a phrase suggesting the relationship between utterances.

(3) Phrase analysis unit B61
The phrase analysis unit B61 receives a plurality of recognition results obtained from the speech recognition, separates the facility name included in the recognition result and the “phrase suggesting the relationship between utterances”, and then the phrase suggests The attribute to be identified is identified.

  Here, the phrase “suggesting a relationship between utterances” refers to a phrase indicating a relationship between both sounds, that is, a relationship between both utterances, for example, a search sound and a sound for narrowing down. It is a word like that.

  The separated facility name candidates are transferred to the candidate management unit B22 to search the search database D201, and the classified phrases are transferred to the relevance calculation method determination unit B52.

(4) Candidate manager B22
The candidate management unit B22 first receives the recognition candidate and the recognition score generated by the phrase analysis unit B61. Next, the search database D201D201 is searched after generating a search key from the recognition candidates. Next, data matching the search condition is acquired. Next, the retrieved data and recognition candidates are stored in the candidate database D301D301 as history information.

  Further, the candidate management unit B22 refers to the candidate database D301, and when a plurality of utterances are managed as history information, the candidate management unit B22 determines utterances to be narrowed down and utterances to be narrowed down, and In order to narrow down the candidates based on the relationship, the relevance calculation unit B43 is inquired about the semantic similarity between the candidates included in each utterance, and the relevance is acquired.

  Next, an operation is performed to narrow down candidates to be narrowed down using the acquired degree of association and recognition score. Next, when candidates are narrowed down, the candidate database D301 is updated, and the result is notified to the result output unit B31B31.

(5) Relevance calculation unit B43
The degree-of-association calculating unit B43 calculates the degree of relevance representing the semantic similarity between the candidates inquired from the candidate managing unit B22 after inquiring the degree-of-calculation method to the degree-of-association calculating method determining unit B52. A result is returned to candidate management part B22.

(6) Relevance calculation method determination unit B52
The relevance calculation method determination unit B52 uses a relevance calculation method database based on candidate attribute information to be calculated and a phrase output by the phrase analysis unit 61, as a method for the relevance calculation unit B43 to calculate relevance. This is determined with reference to D401.

(7) Specific Example In order to explain more specifically, a user who is prompted to enter a facility name under the same task as in the first embodiment is the name of the facility to be searched for, “Takao ) ”, The facility name candidates are narrowed down to 100, and“ near Sagamiko ”is said to narrow down the candidates.

(8) Operation of Speech Recognition Unit B11 The speech recognition unit B11 outputs a plurality of recognition results that match the input speech according to the recognition dictionary grammar written in FIG.

  FIG. 16 illustrates a recognition result in which 30 candidates are output by the speech recognition unit B11. In order to simplify the explanation, it is assumed that a result of “Nearby” continuously acquired from the voice recognition unit B11 with the same facility name as that of the narrowed utterance in the first embodiment.

(9) Operation | movement of phrase analysis part B61 The phrase analysis part B61 will isolate | separate the phrase which suggests the relationship between a facility name and utterance from each recognition candidate, if 30 recognition results acquired are received.

  In the present embodiment, the accepted word sequence described in the recognition dictionary grammar D102 is limited to {facility name} + {phrase suggesting the relationship between utterances}. Therefore, the former can be assigned as the facility name and the latter as the phrase from the boundary information “/” of the recognition result.

  Here, the recognition result related to the facility name is updated, and as a result, the same recognition result as in FIG. 7 is passed to the candidate management unit B21. Subsequently, the phrase “near” is analyzed to determine which attribute of the data is focused on to determine the relationship between utterances.

  As a method for analyzing a phrase, a method in which a table in which a phrase described in the grammar is associated with an attribute to be noticed is held and an attribute to be noticed is determined with reference to the table can be considered. For example, from the “near / latitude / longitude” entry in the table, the attribute to be noted is determined as “latitude / longitude”, and the result is notified to the relevance calculation method determination unit B51.

(10) Operation of Candidate Management Unit B22 Next, the operation of the candidate management unit B22 when a recognition result for the narrowed utterance is generated will be described with reference to the flowchart of FIG.

  First, the candidate management unit B22 acquires the speech recognition result illustrated in FIG. 7 (step S2101).

  Next, a search key for searching the search database D201 in the acquired speech recognition result is extracted (step S2102). In the present embodiment, thirty search keys relating to facility names described including duplicates are generated.

  Next, the search database D201 is searched using a plurality of generated search keys, data that matches the search condition and its attributes are acquired (step S2103), and the recognition result and the search result are linked to obtain the candidate database D301. Is recorded as history information (step S2104). FIG. 8 shows an example of history information written in the candidate database D301, in which the order in which results are acquired, recognition results (scores and ranks), attribute information of the corresponding search database D201, and the like are associated with each other.

  Next, it is determined whether the newly acquired recognition result is an utterance for narrowing down or a newly input utterance (S2105).

  Here, since there are two input recognition result groups, it is determined that the utterance input the second time is a narrowed utterance, and the degree of association for each candidate between utterances is inquired to the association degree calculation unit B43. (S2106). In the present embodiment, 100 candidates from the first utterance to be narrowed down and 30 candidates from the next utterance to be narrowed down are recorded in the candidate database D301, so the degree of association between 100 × 30 = 3000 candidates is calculated. become. A specific method of calculating the degree of association will be described in detail subsequently.

(11) Operation of Relevance Level Calculation Unit B43 The relevance level calculation unit B43 calculates the relevance level representing the semantic similarity by comparing the attribute information of the data for each candidate pair inquired from the candidate management unit B22. To do.

  As a specific calculation method of the degree of association, the degree of association is calculated by inquiring to the degree-of-association calculation determination unit B52.

  The relevance calculation method determination unit B52 pays attention to the attribute “latitude / longitude” to be noticed notified from the phrase analysis unit B61, and puts on the premise that the facility where the latitude and longitude are approached, that is, the near facility is uttered, Relevance is calculated by paying attention to the latitude and longitude attribute information among various attributes. By comparing the latitude and longitude of the candidates, for example, the Euclidean distance can be obtained. Based on the obtained Euclidean distance, the degree of association can be calculated by normalizing so that the degree of association is smaller as the distance is larger and the degree of association is larger as the distance is smaller. An example of the result of performing this calculation for all candidates 100 × 30 is shown in FIG.

(12) Narrowing operation by candidate management unit B22 The operation returns to the operation of candidate management unit B22 again.

(12-1) Step S2107
When the degree of association between candidates is calculated by the association degree calculation unit B43 as shown in FIG. 9 and the result is notified to the candidate management unit B22, the candidate to be narrowed down is narrowed down using the calculated degree of association. The narrowing score is calculated (S2107).

  As a method of calculating the i th narrowing score Di (0 <i <= 100) from the relevance degree Wij (relationship between the i th narrowing target candidate and the j th narrowing condition candidate) and the j th recognition score Rj. Is as follows.

Di = Σj (Wij) x (Rj)

However, 0 <j <= 30. FIG. 10 shows the result of calculating the narrowing score according to the above formula.

  As shown in the example above, the score for determining whether or not to narrow down is not limited to the semantic relevance, but the reliability of recognition in the utterance for narrowing down is used to narrow down the reliability. The purpose is to do.

(12-2) Step S2108
When the narrowing score is calculated in this way, the first 100 facility candidates to be narrowed down are narrowed down in view of the narrowing score and the recognition score (S2108).

  For example, if a candidate for which the refinement score is 0 is removed as a refinement and an operation is performed so as to perform a resort with the total value of the recognition score and the refinement score, the three facilities illustrated in FIG. 11 remain as candidates, and these are stored in the candidate database D301. Is written as history information.

(13) Result output When the search database D201 is searched using the key acquired from the user's utterance and the narrowing process for the first utterance is completed, the result output unit B31 receives the updated candidate list, Notify the user of the result.

  For example, when notifying users by reading the candidate results in order from the most likely one, the following prompt “Three candidates were found. The first one is in Tokyo. Mt. Takao, the second is ... " Here, the user can confirm that the target facility has been read out as the first candidate and confirm that the target candidate has been searched, and can input the target target by utterance or designating the screen to specify the first candidate. Data can be specified.

(14) Modification Example In the present embodiment, when a narrowed utterance is received from a user, a phrase suggesting a relationship between utterances is simultaneously received, and the calculation method of the degree of association is dynamically changed.

  In the above example, since the attribute of interest obtained by analyzing “near” is the same latitude and longitude as in the first embodiment, the same result was obtained.

(14-1) Modification 1
The modification example 1 will be described.

  It is assumed that a narrowed utterance such as “same town as Zenkoji” is input to the first utterance “Toshiba Ryokan”. Then, paying attention to the city name from the phrase “same town”, the degree of association is calculated as 1 if the city names between candidates match, and the degree of association is calculated as 0 if they do not match. Thereby, candidates can be narrowed down from the similarity of different attributes.

(14-2) Modification 2
The modification example 2 will be described.

  For the first utterance “Shopping Mall Kawasaki”, focus on the genre attribute from the refined utterance such as “Department Store-like”, and by switching the relevance calculation method according to the phrase, narrow down to the user's intention The method can be realized.

(15) Effects According to the present embodiment, 1) When selecting data desired by a user from a large-scale database having a plurality of attributes, 2) sufficiently narrowing down target data by the first utterance If not, 3) Rather than presenting relevant questions to the user and specifying the attributes to be input in the next utterance 4) Accepting utterances related to the target data with phrases suggesting relevance In the above, 5) the degree of relevance representing the semantic relevance between the candidate included in the first utterance and the candidate included in the next utterance is calculated based on the result of analyzing the phrase while switching the calculation method, and 6) The candidates are narrowed down in consideration of the recognition likelihood and the relevance of each of the first utterance and the next utterance.

  Therefore, it is possible to efficiently select target data from a large-scale database by avoiding a situation where the user fails to answer a related question from the system and the dialogue is broken.

(Fourth embodiment)
A voice search device according to a fourth embodiment of the present invention will be described with reference to FIGS.

(1) Configuration of Voice Search Device FIG. 17 is a schematic configuration diagram of a voice search device according to this embodiment.

  The speech search apparatus according to the present embodiment has a configuration in which the operations of the speech recognition unit B12 and the candidate management unit B23 in the second embodiment are different, and in addition, a similarity determination unit B71 and an input database D501 are newly added. Yes.

  Components having the same reference numerals as those in the previous embodiment have the same operations, and thus description thereof is omitted.

  In the present embodiment, the input voice D or voice recognition result is held in the input database D501, and the newly input voice or voice recognition result is compared by the similarity determination unit B71, so that the previous voice or voice recognition result is compared. It is different from the previous embodiment in that it is determined whether the utterance is overwritten utterance or narrowed utterance.

(2) Voice recognition unit B12
The speech recognition unit B12 recognizes speech input from the user based on the recognition dictionary grammar D101 and generates recognition candidates.

  Except for the case where no candidate close to the input speech is output because the input speech has noise or the like, one or more recognition candidates are generated by the speech recognition unit B12.

  In addition, the input speech for recognizing the similarity between utterances later or the recognized result is stored in the input database D501. As an example of the recognized result, here, a phoneme sequence obtained in the midway of speech recognition is held. An example of the stored input database D501 is shown in FIG.

(3) Candidate manager B23
The candidate management unit B23 performs the input recognition result from the search of the search database D201 to the narrowing down of candidates according to the flowchart shown in FIG.

  However, in the present embodiment, when the second and subsequent recognition results are acquired, the step of determining whether the newly acquired result is a narrowed utterance (S2105) and the step of not being a narrowed utterance (S2109) Different from the flowchart of FIG.

  In the present embodiment, when determining whether or not the voice is narrowed down, the similarity calculation unit B71 is inquired.

  Here, as shown in FIG. 18, the first recognized phoneme sequence is {N, a, k, a, o, s, a}, and the recognized phoneme sequence corresponding to the second utterance is {T, a, k. , A, o, s, a, N}.

  The similarity determination unit B71 measures the distance between two phoneme sequences according to a distance scale such as an edit distance, and determines whether the two utterances are similar depending on whether the result exceeds a threshold.

  If it is determined that the phoneme series is similar to the threshold, the operation corresponding to the first utterance is deleted, and only the second utterance candidate that is overwritten utterance is left To do.

  In this way, by eliminating candidates corresponding to past utterances corresponding to overwrite utterances, even if unnatural utterances contain only candidates that are different from the intended ones, the narrowed dialogue continues naturally it can.

(4) Modified example In the present embodiment, a measure of phoneme sequence editing distance is adopted as a method for measuring the similarity between utterances, but the present invention is not limited to this.

  For example, it is possible to apply a method in which the ease of replacement of phoneme pairs is observed in advance and the distance between phonemes is calculated by DP matching using this as a cost.

(5) Effect According to this embodiment, when narrowing down from a large-scale database with multiple utterances, it is automatically determined whether the newly input voice is an overwrite of a past utterance or a narrowed utterance. By determining, even when the user repeatedly utters the target data facility name, the operation is performed without unnatural narrowing down.

(Fifth embodiment)
A voice search device according to a fifth embodiment of the present invention will be described with reference to FIGS.

(1) Configuration of Voice Search Device FIG. 20 is a schematic configuration diagram of a voice search device according to this embodiment.

  The speech search apparatus according to this embodiment has a configuration in which the operations of the candidate management unit B24 and the recognition dictionary grammar D103 in the previous embodiment are different.

  Components having the same reference numerals as those in the previous embodiment have the same operations, and thus description thereof is omitted.

  This embodiment is different from the previous embodiment in that it operates to reduce the vocabulary to be recognized and improve the recognition performance by removing facilities already listed as candidates in the refined utterance from the recognition candidates. .

(2) Specific Example To explain with a specific example, it is assumed that 100 recognition results shown in FIG. 4 are obtained for the first utterance as recognition candidates for the recognition dictionary grammar 103 shown in FIG. 3 as in the first embodiment.

(3) Candidate manager B24
The candidate management unit B24 changes the calculation so as to remove the 100 candidates shown in FIG. 4 from the recognition dictionary grammar 103 at the same time as calculating the search database D201 and updating the candidate database D301 as in the first embodiment. Thus, a small-scale recognition dictionary grammar 103 is generated as shown in FIG.

  Because it listens for narrowed vocabulary and narrowed utterances, it can be expected that recognition performance will improve.

(4) Effect According to the present embodiment, by removing the facilities already listed as candidates from the recognition candidates, the vocabulary to be recognized can be reduced, and as a result, the recognition performance can be improved. Can narrow down candidates efficiently.

(Example of change)
The present invention is not limited to the above-described embodiments, and various modifications can be made without departing from the gist thereof.

It is a figure showing the schematic structural example of the speech search device which concerns on the 1st Embodiment of this invention. It is a figure showing an example of the search database which concerns on 1st Embodiment. It is a figure showing an example of the recognition dictionary grammar which concerns on 1st Embodiment. It is a figure showing an example of the recognition result output from the speech recognition part which concerns on 1st Embodiment. It is a flowchart which shows operation | movement of the candidate management part which concerns on 1st Embodiment. It is a figure showing an example of the data stored in the candidate database which concerns on 1st Embodiment. It is a figure showing an example of the recognition result output from the speech recognition part which concerns on 1st Embodiment. It is a figure showing an example of the data stored in the candidate database which concerns on 1st Embodiment. It is a figure showing an example of the calculation result of the related degree which concerns on 1st Embodiment. It is a figure which shows an example of the calculation result of the narrowing score which concerns on 1st Embodiment. It is a figure which shows an example of the updated candidate database which concerns on 1st Embodiment. It is a figure showing the example of schematic structure of the speech search device which concerns on 2nd Embodiment. It is a figure showing an example of the data stored in the relevance calculation method database which concerns on 3rd Embodiment. It is a figure showing the schematic structural example of the speech search device which concerns on 3rd Embodiment. It is a figure showing an example of the recognition dictionary grammar which concerns on 3rd Embodiment. It is a figure showing an example of the recognition result output from the speech recognition part which concerns on 3rd Embodiment. It is a figure showing the schematic structural example of the speech search device which concerns on 4th Embodiment. It is a figure showing an example of the input database which concerns on 4th Embodiment. It is a flowchart which shows operation | movement of the candidate management part which concerns on 4th Embodiment. It is a figure showing the example of schematic structure of the speech search device which concerns on 5th embodiment. It is a figure showing an example of the recognition dictionary grammar which concerns on 5th embodiment.

Explanation of symbols

B11 Speech recognition unit B21 Candidate management unit B31 Result output unit B41 Relevance calculation unit D101 Recognition dictionary grammar D201 Search database

Claims (9)

A voice recognition unit for recognizing a search voice and a search voice for narrowing down the search, and obtaining a search candidate and a search candidate,
A first database for storing a plurality of data and attribute information corresponding to each of the data;
(1) Search the first database based on the search candidates to obtain a plurality of the data and the attribute information corresponding to the data as search information, and (2) the first data based on the refinement recognition candidates. A candidate management unit that searches one database and acquires a plurality of the data and the attribute information as narrowing information;
Each of the plurality of pieces of data included in the search information and each of the plurality of pieces of data included in the narrowing information includes a degree of relevance that increases as the relevance increases. A degree-of-association calculation unit that calculates from the attribute information of each data and the attribute information of each data included in the narrowing-down information;
An output unit that outputs one or a plurality of the data included in the search information having the high degree of relevance;
A voice search device. The attribute information is input information, and a calculation function having the relevance as an output result further includes a second database in which at least one is stored;
The relevance calculation unit
Determining the attribute information of interest from among the plurality of attribute information corresponding to the data included in the search information or the refinement information;
One calculation function is selected from the second database based on the determined attribute information,
Calculating the relevance based on the selected calculation function;
The voice search device according to claim 1. A phrase analysis unit that extracts a phrase suggesting a relationship between the search voice and the refined voice from the refinement candidates;
The relevance calculation unit calculates the relevance based on the attribute information of each data included in the search information, the attribute information of each data included in the refinement information, and the phrase.
The voice search device according to claim 1. A third database that stores voice input from the user or the recognition candidates;
A similarity determination unit that compares the speech stored in the third database or the recognition candidates to obtain a similarity between the utterances of the user;
Further comprising
The candidate management unit
When the similarity is lower than a threshold, narrowing down based on the voice of the user,
When the similarity is higher than a threshold value, the recognition candidate related to the voice of the user is overwritten on the search candidate or the refinement candidate.
The voice search device according to claim 1. The voice recognition unit deletes the search candidates and the narrowing candidates from a recognition dictionary used for the voice recognition.
The voice search device according to claim 1. The data to be searched is a target existing at one point,
The attribute information is the latitude and longitude of the point, or an address,
The relevance is a distance centered on the point.
The voice search device according to claim 1. The voice recognition unit
By referring to the recognition dictionary used for the speech recognition, the search speech is recognized to obtain one or more search candidates and a search recognition score indicating the probability of the recognition. The above-mentioned narrowing-down candidates and a narrowing-down recognition score representing the certainty of the recognition are obtained,
The relevance calculation unit
Calculating the relevance based on the attribute information of each data included in the search information, the attribute information of each data included in the narrowing information, the search recognition score, and the narrowing recognition score;
The voice search device according to claim 1. A voice recognition step for recognizing a search voice and a search voice for narrowing down the search, and obtaining a search candidate and a search candidate.
(1) Based on the search candidates, search a first database that stores a plurality of data and attribute information corresponding to each data, and search the plurality of data and the attribute information corresponding to the data. (2) a candidate management step of searching the first database based on the refinement recognition candidates and obtaining a plurality of the data and the attribute information as refinement information;
Each of the plurality of pieces of data included in the search information and each of the plurality of pieces of data included in the narrowing information includes a degree of relevance that increases as the relevance increases. A relevance calculation step of calculating from the attribute information of each data and the attribute information of each data included in the refinement information;
An output step of outputting one or a plurality of the data included in the search information having the high degree of relevance;
A voice search method comprising: A voice recognition function for recognizing a search voice and a search voice for a search uttered by a user and a search voice for narrowing down the search;
(1) Based on the search candidates, search a first database that stores a plurality of data and attribute information corresponding to each data, and search the plurality of data and the attribute information corresponding to the data. (2) a candidate management function for searching the first database based on the refinement recognition candidates and obtaining a plurality of the data and the attribute information as refinement information;
Each of the plurality of pieces of data included in the search information and each of the plurality of pieces of data included in the narrowing information includes a degree of relevance that increases as the relevance increases. A relevance calculation function for calculating from the attribute information of each data and the attribute information of each data included in the narrowing-down information;
An output function for outputting one or a plurality of the data included in the search information having the high degree of association;
Voice search program for realizing the above with a computer.