JP2002229584A - Speech recognition method, speech information search method, program, recording medium, speech recognition system, speech recognition server computer, and speech information search server computer - Google Patents

Abstract

(57) [Summary] [PROBLEMS] To improve the accuracy of speech recognition by considering not only the speech but also other factors related to the speech in a method of recognizing speech emitted from a speaker by a computer. Let it. In a recognition dictionary memory used by a voice recognition unit, a plurality of types of voice recognition dictionaries including a language model and an acoustic model, a plurality of types of personal attributes assumed in advance for a speaker, and It is stored in advance in association with attributes of a plurality of types of environments in which it is assumed that a speaker is placed when uttering a voice. The voice recognition unit 82 is based on the actual attributes of the speaker who utters the voice to be recognized and the attributes of the environment where the speaker is actually placed when uttering the voice.
In the recognition dictionary memory 80, a language model and an acoustic model suitable for the speaker are selected, and the voice generated by the speaker is recognized by using the selected model.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a technique for recognizing a voice uttered by a specific or unspecified speaker by a computer.

[0002]

2. Description of the Related Art There is already a technique for recognizing a voice uttered from a speaker by a computer. This speech recognition technology includes a specific speaker system that recognizes only a voice uttered from a speaker registered in advance, and an unspecified speaker system that does not limit the speaker.

[0003] In any of the systems, the speech recognition technology generally recognizes actual speech by using a speech recognition dictionary based on actual speech data representing actual speech actually uttered by a speaker.

[0004] As one application of such speech recognition technology,
There is already a speech information search for retrieving information using speech.

[0005] One conventional example related to this voice information search is described in Japanese Patent Application Laid-Open No. 2000-216888. In this conventional example, when a user makes an inquiry to a client computer by voice, the inquiry information is transmitted to a server computer connected to the client computer via a network. The server computer having received the inquiry information performs a speech recognition process on the inquiry information, and searches a specific database for information that matches the speech recognition result. The retrieved information is transmitted from the server computer to the client computer, so that the user obtains necessary information at the client computer.

[0006]

SUMMARY OF THE INVENTION In speech recognition,
In general, its accuracy depends on the performance of the speech recognition dictionary.
Therefore, in order to improve the accuracy of speech recognition, it is necessary to improve the performance of the speech recognition dictionary.

The present inventor has conducted research on a speech recognition dictionary, and has obtained the following knowledge.

When a speech recognition dictionary is used, the most similar to the actual speech data is searched for from a plurality of types of speech data (assumed speech data) stored in advance, whereby the actual speech is retrieved. Be recognized. However, when the speech recognition process is performed only depending on the actual speech data, it is difficult to improve the accuracy of the speech recognition.

[0009] On the other hand, the speech recognition technology may be used without any limitation, but may be used in a limited manner. In the latter case, a certain tendency may be easily found in the voice emitted from the speaker.

For example, when speech recognition technology is used for speech information retrieval, and the speech information retrieval is designed to provide information on a geographical environment in response to an information retrieval request from a user (for example, , To provide information on the location of each facility or provide traffic congestion information for each road)
The content of inquiry information issued by a speaker by voice for information retrieval is naturally limited. For example, the content of the inquiry information depends on the environment in which the speaker is placed when the speaker issues the inquiry information. Therefore, for example, if the environment in which the speaker is placed is different, the linguistic content (meaning) of the voice uttered from the speaker is also different, and if the environment in which the speaker is placed is known, the speaker It is possible to predict the linguistic content (meaning) of the voice uttered from the language to some extent accurately.

[0011] Further, the acoustic characteristics of speech emitted from a speaker for information retrieval also depend on the environment in which the speaker is placed. For example, when the speaker is driving on a highway in a car, the running sound of the car is louder than in the case where the car is stopped in a parking lot, and as a result, the sound is mixed in the voice emitted from the speaker. The running sound as background sound also becomes loud.

Further, based on these findings, the present inventor has found that the linguistic content and acoustic characteristics of speech emitted from a speaker depend on the environment (speech environment) in which the speaker is placed. In the case that it is anticipated, it has been found that it is optimal to construct a speech recognition dictionary so that the final speech recognition result is obtained taking into account not only the actual speech from the speaker but also its environment. Was.

Some of the findings described above are effective irrespective of whether the speech recognition method is the specific speaker method or the unspecified speaker method. Focuses on the fact that the acoustic features of the speech to be recognized depend on the attributes (eg, gender and age) of the target speaker who is the speaker that emits the speech. The inventor has noticed that it is also effective to construct a speech recognition dictionary so that a final speech recognition result is obtained in consideration of not only the speaker's own attribute but also the speaker's own attribute.

[0014]

SUMMARY OF THE INVENTION Against this background, the present invention considers not only voices uttered by a speaker but also other factors related to the utterances, thereby making it possible to obtain voices. The object of the present invention is to improve recognition accuracy, and the following aspects are obtained by the present invention. Each mode is described in the same manner as in the claims, divided into sections, each section is numbered, and described in the form of citing the numbers of other sections as necessary. This is to facilitate understanding of some of the technical features described in the present specification and some of the combinations thereof, and the technical features and the combinations thereof described in the present specification have the following aspects. It should not be construed as limited.

(1) A method for recognizing a voice uttered by a specific or unspecified speaker by a computer, wherein the speaker utters a voice by using an attribute of an environment where the speaker is actually placed. Based on real environment attribute data representing a certain real environment attribute, a plurality of types of speech recognition dictionaries prepared in advance in association with a plurality of types of environment attributes which are assumed to be placed by the speaker when uttering a voice. A first step of selecting a speech recognition dictionary corresponding to the actual environment attribute from among the following, and using the selected speech recognition dictionary based on actual speech data representing actual speech actually emitted by the speaker. A second step of recognizing the actual voice and generating voice recognition result data representing the recognition result. In this method, a plurality of types of speech recognition dictionaries are prepared in advance in association with attributes of a plurality of types of environments in which speakers are assumed to be placed when uttering voice. Then, in this method, when a speaker utters a voice, based on actual environment attribute data representing an actual environment attribute which is an attribute of an environment where the speaker is actually placed, among a plurality of types of speech recognition dictionaries, The speech recognition dictionary corresponding to the actual environment attribute is selected. Further, in this method, the actual speech is recognized by using the selected speech recognition dictionary based on the actual speech data representing the actual speech actually uttered by the speaker, and the speech recognition result representing the recognition result is obtained. Data is created. Therefore, according to this method, the actual voice is recognized by considering not only the actual voice of the speaker but also the attributes of the environment in which the speaker is actually placed at the time of uttering. In comparison with speech recognition taking into account
The accuracy of voice recognition can be easily improved. In this section, “environment” can be interpreted to include, for example, the geographical environment where the speaker is located. The attributes of the geographical environment include, for example, a sightseeing spot, a downtown area, an amusement park, and an expressway. "Environment" can also be interpreted to include the acoustic environment in which the speaker is located. The attributes of the acoustic environment include, for example, a place where a background sound such as a natural sound and a noise exists, and a vehicle on which a speaker is placed and moved. When the speaker is in the vehicle, the running sound of the vehicle is a factor for specifying the attribute of the acoustic environment. The traveling sound of the vehicle depends on factors such as the type of vehicle, traveling speed, road surface conditions, and the like, and these factors are also factors that specify the attributes of the acoustic environment. (2) The method further includes a third step of causing the plurality of types of speech recognition dictionaries to learn so that the contents reflect the relationship between the corresponding actual environment attribute data, the actual speech data, and the speech recognition result data ( A speech recognition method according to the first aspect [Claim 2]. According to this method, the learning of the voice recognition dictionary is performed so as to reflect not only the actual voice data and the voice recognition result data corresponding to each other but also the actual environment attribute data corresponding to them. It is easy to reliably adapt the content to a new speech recognition situation. (3) A method of recognizing a voice uttered from an unspecified speaker by a computer, the method representing an actual attribute which is an actual attribute of a target speaker who is a speaker emitting a voice to be recognized. A plurality of types of speech recognition dictionaries based on personal attribute data and actual environment attribute data representing actual environment attributes which are attributes of an environment where the target speaker is actually placed when emitting the voice. Are prepared in advance in association with the attributes of a plurality of types of environment in which the speaker is assumed to be placed when emitting the speaker and the attributes of the plurality of types of person assumed in advance for the speaker. A first step of selecting a voice recognition dictionary corresponding to the real personal attribute and the real environment attribute from among the actual voice data and real voice data representing a real voice actually emitted by the target speaker; A second step of recognizing the actual voice data by using the obtained voice recognition dictionary and generating voice recognition result data representing the recognition result. [Claim 3]. In this method, paying attention to adopting an unspecified speaker method, a plurality of types of speech recognition dictionaries are used. It is prepared in advance in association with the assumed attributes of a plurality of types of environments. Then, in this method, the actual speaker attribute data representing the actual speaker attribute, which is the actual attribute of the target speaker who is the speaker emitting the voice to be recognized, and the target speaker is used when the voice is emitted. Based on the actual environment attribute data representing the actual environment attribute, which is the attribute of the environment actually placed, the actual speaker attribute data and the actual environment attribute data of the target speaker are selected from a plurality of types of speech recognition dictionaries. The voice recognition dictionary is selected. Further, in this method, the actual speech data is recognized by using the selected speech recognition dictionary based on the actual speech data representing the actual speech actually uttered by the target speaker, and the speech representing the recognition result is obtained. Recognition result data is created. Therefore, according to this method,
Considering not only the actual voice of the speaker but also the attributes of the speaker itself and the attributes of the environment where the speaker was actually placed when speaking, the actual voice is recognized. In comparison with speech recognition taking into account
The accuracy of voice recognition can be easily improved. In this section, “person's attribute” can be interpreted to include at least one of, for example, profile, gender, age, and the like. In this section, “environment” can be interpreted in the same manner as in the above section (1). (4) Further, the plurality of types of speech recognition dictionaries are trained so that the contents reflect the relationship among the actual speech data, the speech recognition result data, the actual personal attribute data, and the actual environment attribute data corresponding to each other. The speech recognition method according to claim 3, including a third step [Claim 4]. According to this method, the learning of the speech recognition dictionary is performed so as to reflect not only the actual speech data and the speech recognition result data corresponding to each other, but also the corresponding actual personal attribute data and actual environment attribute data. In addition, it is easy to reliably adapt the contents of the speech recognition dictionary to a new speech recognition situation. (5) The speech recognition method according to any one of (1) to (4), wherein each type of the speech recognition dictionary includes an acoustic model and a language model. According to this method, by considering both the acoustic characteristics of the actual speech uttered by the speaker and the linguistic constraints between words in a character string corresponding to the actual speech, the actual Voice is recognized. (6) Each of the server computers that can communicate with a plurality of client computers used by a plurality of users who desire to search for necessary information by voice, the information from voices from unspecified users. Search for information that matches the search request
A method of transmitting data to a computer, wherein actual personal attribute data representing an actual personal attribute which is an actual attribute of a target speaker who is a speaker that emits a voice to be recognized, and the target talk when the voice is generated A plurality of types of speech recognition dictionaries based on actual environment attribute data representing actual environment attributes, which are attributes of the environment where the speaker is actually placed, and it is assumed that the speaker will be placed when uttering a voice. The attributes of the plurality of types of environment and the attributes of the plurality of types of individuals presumed for the speaker are prepared in advance in association with the respective attributes of the individual. A first step of selecting a corresponding speech recognition dictionary, and using the selected speech recognition dictionary based on actual speech data representing actual speech actually uttered by the target speaker. A second step of recognizing voice data and generating voice recognition result data representing the recognition result; and at least a plurality of types of provided information that can be provided to each user based on the generated voice recognition result data. ,
In the provided information database stored in advance in association with a plurality of types of speech recognition result data assumed in advance, the provided information corresponding to the created speech recognition result data is searched, and the searched provided information is A third step of transmitting to the client computer used by the target user. According to this method, not only the actual voice of the user, but also the attribute of the user and the user's actual Since the actual voice is recognized by also considering the attribute of the environment placed in the environment, the accuracy of voice recognition can be easily improved as compared with the case where the voice recognition is performed in consideration of only the actual voice. .
Further, according to this method, information desired by a user can be retrieved with high accuracy based on a highly accurate speech recognition result and provided to the user. In this section, the communication between the "client computer and the server computer" may be a communication method using radio waves or a communication method using a dedicated or public telephone line. Further, the “communication” is performed via an open network such as the Internet, LAN, WA, or the like, regardless of the communication method.
N, etc., over a closed network. Such an interpretation can also be applied in the following sections. (7) In the providing information database, the third step is based on the created speech recognition result data and refers to at least one of the actual personal attribute data and the actual environment attribute data. The voice information search method according to claim 6, including a step of searching for the provided information corresponding to the generated voice recognition result data [Claim 6]. In this method, the information desired by the user is searched in the provided information database based on not only the searched voice recognition result data but also at least one of the actual personal attribute data and the actual environment attribute data. Therefore, according to this method, it is possible to easily improve the accuracy of the search as compared with a case where the information desired by the user is searched based only on the searched voice recognition result data. Further, according to this method, at least one of the actual personal attribute data and the actual environment attribute data can be effectively used for both voice recognition and information retrieval. (8) The voice information search method according to (6) or (7), wherein the plurality of types of provided information include a plurality of types of geographic environment-related information on a geographic environment. In this section, `` geographical environment '' is interpreted to include, for example, the environment related to facilities such as amusement parks, accommodation facilities, shops, factories, event venues, parking lots, parks, museums, museums, public facilities, etc. It can be interpreted to include environments related to roads such as general roads, highways, gravel roads, and snowy roads. (9) A program executed by a computer to execute the method according to any one of (1) to (8) [Claim 8]. When this program is executed by a computer, the same functions and effects as those of the method according to any one of the above items (1) to (8) can be realized. The program according to this section can be interpreted to include not only a combination of commands executed by a computer to perform its function but also a file or data processed by each command. (10) A recording medium on which the program according to (9) is recorded in a computer-readable manner. If the program recorded on the recording medium is executed by a computer, the same operation and effect as the method according to any one of the above modes (1) to (8) can be realized. The “recording medium” in this section can adopt various formats, for example, a magnetic recording medium such as a floppy (registered trademark) disk, a CD,
At least one of an optical recording medium such as a CD-ROM, a magneto-optical recording medium such as an MO, and an unremovable storage such as a ROM can be employed. (11) A system for recognizing a voice uttered from an unspecified speaker by a computer, wherein a plurality of types of environment attributes in which the speaker is supposed to be placed when uttering the voice, A plurality of types of speech recognition dictionaries prepared in advance in association with a plurality of types of attributes of a person assumed in advance for a speaker, and actual attributes of a target speaker who is a speaker emitting a voice to be recognized Based on the actual personal attribute data representing the actual personal attribute, and the actual environmental attribute data representing the actual environmental attribute which is the attribute of the environment where the target speaker is actually placed when emitting the voice. And selecting a speech recognition dictionary corresponding to the actual personal attribute and the actual environment attribute from among the speech recognition dictionaries based on actual speech data representing the actual speech actually emitted by the target speaker. A voice recognition system for recognizing the actual voice data by using the selected voice recognition dictionary, and generating voice recognition result data representing the recognition result [Claim 10]. According to this system, not only the actual voice of the speaker, but also the attributes of the speaker itself and the environment in which the speaker was actually placed at the time of uttering, according to basically the same principle as the method according to the above item (3). In addition, since the actual speech is recognized by taking into account the attribute, the accuracy of speech recognition can be easily improved as compared with the case where speech recognition is performed in consideration of only the actual speech. (12) A server computer for recognizing a voice emitted from an unspecified user by communicating with a plurality of client computers used by a plurality of users who emit voices to be recognized, It is prepared in advance in association with the attributes of a plurality of types of environment in which the speaker is assumed to be placed when emitting a voice and the attributes of a plurality of types of individuals presumed for the speaker. Plural types of speech recognition dictionaries, actual personal attribute data representing the actual personal attributes that are the actual attributes of the target speaker who is the speaker that utters the voice to be recognized, and the target speaker when the voice is uttered Is based on actual environment attribute data representing an actual environment attribute which is an attribute of an environment in which the actual personal attribute and the actual environment attribute are selected from the plurality of types of speech recognition dictionaries. And a speech recognition dictionary corresponding to the selected speaker, and based on the actual speech data representing the actual speech actually uttered by the target speaker, using the selected speech recognition dictionary to recognize the actual speech data. ,
A speech-recognition server computer including a recognition unit that creates speech recognition result data representing the recognition result. According to the server computer, not only the actual voice of the user but also the attribute of the user and the user are actually placed at the time of uttering according to basically the same principle as the method according to the above item (3). Considering the attributes of the environment, the actual speech is recognized.
The accuracy of voice recognition can be easily improved. (13) Communication with a plurality of client computers used by a plurality of users who wish to search for necessary information by voice, respectively, thereby adapting to an information search request by voice from an unspecified user. A server computer that searches for and sends the information to each client computer, and includes attributes of a plurality of types of environments in which the speaker is assumed to be placed when emitting a voice, and A plurality of types of speech recognition dictionaries prepared in advance in association with a plurality of types of personal attributes assumed in advance, and a plurality of types of provision information that can be provided to the respective users, A provided information database pre-stored in association with the speech recognition result data, and a user who emits, by voice, information necessary for searching for the necessary information. The actual personal attribute data representing the actual personal attribute which is the actual attribute of a certain target user, and the real environment representing the actual environmental attribute which is the attribute of the environment where the target user is actually placed when uttering the voice Based on the attribute data, from among the plurality of types of speech recognition dictionaries, a speech recognition dictionary corresponding to the actual personal attribute and the actual environment attribute is selected, and the actual speech actually emitted by the target user is Recognition means for recognizing the actual speech data by using the selected speech recognition dictionary based on the actual speech data to be represented and creating speech recognition result data representing the recognition result; Based on the recognition result data, the provided information corresponding to the voice recognition result data is searched in the provided information database, and the searched provided information is searched for. , The server voice information retrieval including the information retrieval means for transmitting to the client computer used by the target user,
Computer [Claim 12]. According to the server computer, not only the actual voice of the user but also the attribute of the user and the user are actually placed at the time of uttering according to basically the same principle as the method according to the above item (3). Since the actual voice is recognized by also considering the attribute of the environment, the accuracy of the voice recognition can be easily improved as compared with the case where the voice recognition is performed in consideration of only the actual voice. Further, according to the server computer, the information desired by the user can be retrieved with high accuracy based on the highly accurate speech recognition result and provided to the user. (14) The information retrieval means, based on the created speech recognition result data and referring to at least one of the actual personal attribute data and the actual environment attribute data, The voice information search server computer according to mode (13), further comprising means for searching for the provided information corresponding to voice recognition result data. According to this server computer,
As in the method according to the above mode (7), the information desired by the user is compared with a case where the information is searched based only on the searched voice recognition result data, and the accuracy of the search can be easily improved. Further, according to the server computer, at least one of the actual personal attribute data and the actual environment attribute data can be effectively used for both voice recognition and information retrieval, as in the method according to the above mode (7). Available.

[0016]

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS Hereinafter, a more specific embodiment of the present invention will be described in detail with reference to the drawings.

FIG. 1 is a functional block diagram showing a voice information search system for implementing a voice information search method according to the present embodiment. The voice information search method utilizes a voice recognition method according to an embodiment of another aspect of the present invention. In addition, the voice information search system uses a voice recognition system according to still another aspect of the present invention, and uses a server computer for voice recognition according to yet another aspect of the present invention. And a server computer for voice information search according to the above aspect.

The voice information search method according to the present embodiment (hereinafter, simply referred to as “search method”) is as shown in FIG.
A plurality of users (only one user is shown in the figure as a representative) and a server computer 10 installed in a voice information search center (this is an example of the above-described voice recognition server computer) This is an example of the voice information search server computer). Each user obtains desired information from the server computer 10 by communicating with the server computer 10 via each client computer 12.

As shown in FIG. 1, each client computer 12 is configured to include a voice input unit 20 and an information presenting unit 22.

The voice input unit 20 generates voice data (actual voice data) by taking in the voice when the user utters inquiry information which is necessary information for the server computer 10 for information search. This is the part to do.

On the other hand, when information matching the inquiry information uttered from the user by voice is retrieved and transmitted by the server computer 10, the information presenting unit 22 visually or indirectly retrieves the information. And / or a part that is presented audibly to the user.

Each client computer 12 is further configured to include a position acquisition unit 24 and an ID storage unit 26, as shown in FIG.

The position obtaining section 24 is a part for obtaining the current position of the user on the earth. On the other hand, ID
The storage unit 26 stores a user ID which is identification information of the user.
Is the part that memorizes.

The client computer 12 is configured as shown in FIG. 2 to perform these four functions, namely, voice input, information presentation, position acquisition, and ID storage. As is well known, the client computer 12 includes a processing unit 30 (hereinafter, referred to as a processing unit 30).
"PU") and a memory 32 are connected to each other by a bus 34 to include a computer unit 36. The memory 32 is configured to include a recording medium such as a ROM, a RAM, a magnetic disk, and an optical disk. The memory 32 stores a voice input program and a user ID in advance. The part of the memory 32 that stores the user ID is an example of the ID storage unit 26.

As shown in FIG. 2, the client computer 12 further includes a transmission / reception unit 40, and the transmission / reception unit 40 can communicate with the server computer 10 by radio waves.

The client computer 12 further includes a global positioning system, as shown in FIG.
A system 42 (hereinafter abbreviated as “GPS”) is provided. As is well known, the GPS 42 is a device that measures the current position of the GPS 42 using an artificial satellite. The GPS 42 is an example of the position acquisition unit 24.

The client computer 12 further includes a switch 46 operated by a user, a microphone 48 for inputting a voice uttered by the user, a display 50 for displaying information on a screen, and outputting information by sound. And a speaker 52 to be used. The microphone 48 is an example of the voice input unit 20, and the display 50 and the speaker 52 are each an example of the information presenting unit 22.

In recent years, as shown in FIG. 3, mounting a navigation device 60 on an automobile as a vehicle has become widespread. The navigation device 60 generally includes a computer and a GPS (both not shown), a switch 62, a display 64, and a speaker 6.
6 are included. Further, connecting a mobile phone 70 such as a mobile phone or a PHS to such a navigation device 60 to enable transmission / reception to / from the outside by radio waves has become widespread. The mobile telephone 70
A receiver 72 is provided for receiving a voice from a speaker.

By using the navigation device 60 and the mobile telephone 70, the client computer 12 can be realized. in this case,
Computer section 36 of client computer 12
Is the computer of the navigation device 60, GPS
42 is the GPS of the navigation device 60, switch 46 is the switch 62, microphone 48 is the receiver 72 of the mobile telephone 70, display 50 is the display 64, speaker 52 is the speaker 66, and transmitter / receiver 4
0 can be constituted by the receiving unit of the navigation device 60 and the transmitting / receiving unit of the mobile telephone 70, respectively. When the client computer 12 is realized by using the navigation device 60 and the mobile telephone 70 in this way, a special hardware resource must be added to the vehicle so that the search method can be implemented in the vehicle. Becomes indispensable, so that this search method can be implemented at low cost.

The hardware resources of the client computer 12 have been described above. Next, the hardware resources of the server computer 10 will be described.

As shown in FIG. 1, the server computer 10 has a recognition dictionary memory 80, a voice recognition unit 82, and a reference information database 8 in order to perform a function of voice recognition.
4 (hereinafter simply referred to as “reference information DB”; the same applies to other databases), a data storage unit 86, a storage data memory 88, and a dictionary learning unit 90.

The recognition dictionary memory 80 is a memory for storing a voice recognition dictionary referred to for identifying and recognizing a plurality of types of voices. In the present embodiment, the recognition dictionary memory 80 is configured to store the acoustic model and the language model as a speech recognition dictionary.

The acoustic model is a model for performing speech recognition in consideration of acoustic feature information of the speech. In this embodiment, by using this acoustic model,
The actual voice (inquiry information) from the speaker is converted into an inquiry sentence (character string).

As shown in FIG. 4, this acoustic model has attributes of a plurality of supposed types of speakers (for example, sex,
By associating with each assumed personal attribute data representing the age (eg, age) and each assumed environment attribute data representing attributes of a plurality of assumed environments (speech environments) (for example, downtown areas, expressways, etc.). A plurality of types are provided in the recognition dictionary memory 80.

On the other hand, the language model includes a plurality of types of linguistic constraint information (for example, constraint information when a plurality of words are connected to each other) between a plurality of words in a character string represented by a series of sounds. This is a model for performing speech recognition in consideration of grammatical constraint information). In the present embodiment, the above query sentence is converted into a query command by using this language model.

Here, the inquiry command can be considered to mean a search condition that the server computer 10 needs to refer to in an information search service desired by the user (described in detail later). For example, if the query sentence is expected to be composed of two nouns “I” and “you” connected to each other by the case particle “to”, the search condition Leads to the condition that the word "I" and the word "you" are included, together with the association that the two words exist simultaneously.

This language model, like the acoustic model,
As shown in FIG. 5, each assumed personal attribute data representing the attributes (for example, gender, age, etc.) of the assumed plurality of types of speakers, and the attributes of the assumed plurality of types of environment (speech environment) For example, a plurality of types are provided in the same recognition dictionary memory 80 by being associated with each assumed environment attribute data representing sightseeing spots, expressways, and the like.

The speech recognition unit 82 performs so-called speech recognition processing using the acoustic model and the language model, thereby converting inquiry information actually issued as speech from a speaker into an inquiry sentence. The converted query is converted into a query command.

Prior to using the acoustic model and the language model, the speech recognizing unit 82 is a user who is a user who has issued a speech to be recognized from a plurality of types of language models and a plurality of types of acoustic models. The acoustic model and the language model corresponding to the actual personal attribute data representing the actual attribute of the user and the actual environment attribute data representing the attribute of the actual environment of the user when the user utters the voice are selected. . By using the selected acoustic model and language model, the speech recognition unit 82 recognizes speech emitted from a user as a speaker.

Here, the actual attribute of the user (actual personal attribute) is specified by the user ID of the user.
Similarly, attributes of the user's actual environment (actual environment attributes)
Is specified by information (current position information) indicating the current position of the user received by the server computer 10 from the client computer 12 of the user.

In the present embodiment, the relationship between the actual personal attribute and the user ID and the relationship between the actual environment attribute and the current location information are basically based on the fact that the user is a server computer. Before using the information search service in 10, for example, the information search service is stored in the reference information DB 84 by a business entity that operates the information search service. The reference information DB 84 includes a position information DB 92 and a user information DB 94 as shown in FIG.

Here, the relationship between the personal attribute and the user ID will be described. The personal attribute includes, for example, gender, age, height, weight, etc., whereas the user ID is, for example, the server computer 10. Means an identification number as a member number assigned from the server computer 10 when the user is registered as a member permitted to use the information search service. At the time of the registration, if the user registers the personal attribute, the relationship between the personal attribute and the user ID can be stored in the user information DB 94 prior to using the information search service.

Next, the relationship between the environmental attribute and the location information will be described. The environmental attribute includes, for example, a downtown area, an expressway, and the like, whereas the location information includes, for example, information provided from the GPS 42, For example, it includes latitude and longitude. The relationship between these environmental attributes and location information can be found by prior research. Therefore, in the present embodiment, the relationship between the environment attributes and the position information is stored in the position information DB 92 in advance.

As shown in FIG. 1, the server computer 10 further provides a provided information DB 10 for information search.
0 and an information retrieval / providing unit 102.

In the provided information DB 100, provided are a plurality of types of provided information which can be provided to the user, which are stored in advance in association with the respective voice recognition result data. The plurality of types of provided information include a plurality of types of geographic environment-related information on the geographic environment (for example, facility guidance information, road congestion information, store information, travel information, and the like).

On the other hand, the information retrieval / providing unit 102
Is based on the speech recognition result by the speech recognition unit 82,
The provided information DB 100 is searched for information expected to match the information desired by the user. At the time of the search, the information search / providing unit 102 stores the actual personal attribute data corresponding to the user ID of the user in the user information DB 94.
, And the actual environment attribute data corresponding to the current position of the user is searched in the position information DB 92. Then, the information search / provider 102 first
Based on the speech recognition result by the speech recognition unit 82, information that is expected to match the information desired by the user is provided to the provided information D.
In B100, the provisional provision information is searched as provisional provision information, and among the retrieved provisional provision information, information matching the retrieved actual personal attribute data and actual environment attribute data is used as final provision information by the user. To provide.

The data storage section 86, storage data memory 88, and dictionary learning section 90 are provided for learning the recognition dictionary.

The data storage unit 86 includes the audio data storage unit 8
6a and an inquiry data storage unit 86b. The voice data storage unit 86a stores the voice data actually received by the server computer 10 from the client computer 12 in the storage data memory 88. The inquiry data storage unit 86b is inquiry data, (a) an inquiry command as speech recognition result data acquired by the speech recognition unit 82 based on the speech data, and (b) a reference for acquisition thereof. The data representing both the actual personal attribute and the actual environment attribute are stored in the storage data memory 88. The voice data and the inquiry data are stored in the storage data memory 88 in association with each other.

The dictionary learning section 90 has a storage data memory 88
Read out each voice data and each inquiry data from
Further, the recognition dictionary memory 80 is updated so that the contents are reflected by the read voice data and inquiry data. The learning of the recognition dictionary memory 80 is performed. In the present embodiment, the learning is performed when the amount of data stored in the storage data memory 88 is larger than a set value.

FIG. 6 conceptually shows a hardware configuration of server computer 10. The server computer 10 includes a PU 120 and a memory 122, like the computer unit 36 of the client computer 12.
Are connected to each other by a bus 124. The memory 122 is configured to include a recording medium such as a ROM, a RAM, a magnetic disk, and an optical disk, like the memory 32 of the computer unit 36. The memory 122 has several areas, and these areas constitute a program memory 130, the recognition dictionary memory 80, the reference information DB 84, the provided information DB 100, and the stored data memory 88. Have been. The program memory 130 includes a voice information search program for performing the above-described voice recognition and information search, and a recognition dictionary memory 8.
A dictionary learning program for learning 0 is stored in advance.

The server computer 10 further includes a transmission / reception unit 132 as shown in FIG.
2 makes it possible to communicate with an arbitrary client computer 12 by radio waves.

FIG. 7 shows the client computer 1
The content of the voice input program executed by the server 2 and the content of the voice information search program executed by the server computer 10 are represented in a flowchart conceptually and associated with each action of the user in a timely manner. I have.

A user who desires to receive the information search service is referred to as step S1 (hereinafter simply referred to as "S1").
In the other steps, the switch 46 is turned ON. Then, the determination in S21 of the voice input program becomes YES, and in S22, a call is made to the voice information search center by the transmitting / receiving unit 40 (for example, the mobile telephone 70).

As a result, the voice information search program S4
In 1, the server computer 10 used by the voice information retrieval center is connected to the client computer 12 used by the current user by a telephone line.

Thereafter, in S23 of the voice input program, a call from the server computer 10 is confirmed, and
The determination is YES.

As described above, the server computer 10
And the client computer 12 shift to a state where they can communicate with each other by radio waves.

Thereafter, in S24 of the voice input program, the user ID of the current user is read from the memory 32 and transmitted to the server computer 10. Server computer 1 that has received the user ID
0 stores the received user ID in the memory 122 in S42.

Thereafter, the client computer 12
Indicates that the GPS 42
To detect the current position of the user this time. The detected current position is digitized by, for example, latitude and longitude, and transmitted to the server computer 10 as current position data. Subsequently, the server computer 10 stores the received current position in the memory 122 in S43. Thereafter, the server computer 10 reads the question data from the memory 122 in S44, and
Send to

Subsequently, the client computer 12
Receives the question data in S26, and then outputs the content of the question represented by the question data to the user by voice via the speaker 52 in S27. In response to this question, the user, at S2,
Inquiry information to the server computer 10
The voice is input through the microphone 48.

Thereafter, the client computer 12
Transmits the voice uttered by the user to the server computer 10 as voice data in S28.

Subsequently, the server computer 10
At 45, an audio information search is performed based on the received audio data.

The details of step S45 are shown in the flowchart of FIG. 8 as a voice information search routine. In this routine, first, in S61, the memory 122
The actual personal attribute data corresponding to the user ID stored in the user information DB 94 is searched in the user information DB 94. next,
In S62, the actual environment attribute data corresponding to the current position stored in the memory 122 is searched in the position information DB 92.

Thereafter, in S63, an acoustic model corresponding to the retrieved actual personal attribute data and actual environmental attribute data is retrieved in the recognition dictionary memory 80. Subsequently, in S64, a language model corresponding to the retrieved actual personal attribute data and the actual environment attribute data is retrieved in the recognition dictionary memory 80.

Thereafter, in S65, the client
By using the selected acoustic model and language model based on the voice data received from the computer 12, an inquiry command corresponding to the inquiry information issued by the user by voice is created as described above. You.

Subsequently, in S66, based on the created inquiry command, provided information matching the user's inquiry information is searched in the provided information DB 100. Specifically, as described above, the actual personal attribute data corresponding to the user ID of the user is searched in the user information DB 94, and the actual environmental attribute data corresponding to the current position of the user is retrieved from the position information DB 92. The provided information is searched based on the searched actual personal attribute data and the actual environment attribute data, and the created inquiry command.

After that, in S67, the retrieved provided information is transmitted to the client computer 12. Subsequently, in S68, the voice data used in the current voice recognition is stored in the storage data memory 88. Further, in S69, the inquiry command, the actual personal attribute, and the real environment attribute acquired in the current voice recognition are stored. The inquiry data to be represented is stored in the storage data memory 88. As described above, one execution of the voice information search routine, that is, one execution of S45 in FIG. 7 ends, and as a result, one execution of the voice information search program ends.

The client computer 12 which has received the retrieved provided information, in S29 of FIG.
The provided information is presented to the user via at least one of the display 50 and the speaker 52. This completes one execution of the voice input program. As a result, in S3, the user acquires the retrieved provided information, that is, the information desired by the user. Thus, a series of procedures by the user for voice information search is completed.

FIG. 9 conceptually shows the contents of a dictionary learning program executed by the server computer 10. This program is executed repeatedly. At the time of each execution, first, in S81, it is determined whether or not the data storage amount, which is the amount of data stored in the storage data memory 88, is larger than a set value. If it is not more than the set value, the determination is NO, and one cycle of the program is immediately terminated.

On the other hand, if the data storage amount is larger than the set value, the determination in S81 becomes YES, and S8
Move to 2. In this S82, the acoustic model is
The learning is performed so that the voice data and the inquiry data stored in the storage data memory 88 are reflected. Thereafter, the language model is trained so that the inquiry data stored in the storage data memory 88 is reflected.
This completes one execution of the dictionary learning program.

As is clear from the above description, in the present embodiment, the part of the server computer 10 that executes S61 to S65 in FIG. 8 constitutes the voice recognition unit 82, and the part that executes S66 and S67. Constitute the information retrieval / providing unit 102, the part executing S68 constitutes the voice data storage unit 86a, and the part executing S69 constitutes the inquiry data storage unit 86b. Furthermore, the part of the server computer 10 that executes the dictionary learning program constitutes the dictionary learning unit 90.

As is clear from the above description, in the present embodiment, the acoustic model and the language model cooperate with each other to constitute an example of the "speech recognition dictionary" in claim 1 or 3, and FIG. S61 to S64 cooperate with each other to constitute an example of the “first step” in the claim, and S65 in the same figure corresponds to “second step” in the claim.
This constitutes an example.

Further, in the present embodiment, S in FIG.
5. The method according to claim 2, wherein 81 to S83 cooperate with each other.
Constitutes an example of the “third step” in FIG.

Further, in this embodiment, the acoustic model and the language model cooperate with each other to constitute an example of the "speech recognition dictionary" in claim 5, and S61 to S64 in FIG. An example of a “first step” in the claim is constituted, S65 of the same figure constitutes an example of a “second step” in the same claim, and S66 and S67 of the same figure are cooperated with each other. Or an example of the “third step” in 6 and the provision information DB 100
Constitutes an example of the “provided information database” in the claim, and the provided information constitutes an example of the “geographical environment-related information” in claim 7.

Further, in this embodiment, among all the steps of the voice information search program, S61 to S65 of FIG. 8 and all the steps of the voice information search program are each an example of the "program" according to claim 8. It is composed.

Further, in the present embodiment, the memory 1
22 constitutes an example of the “recording medium” according to claim 9.

Further, in the present embodiment, the acoustic model and the language model cooperate with each other to constitute an example of the “speech recognition dictionary” in claim 10 or 11, and the speech recognition unit 82 uses the “voice recognition dictionary” in the claim. It constitutes an example of "recognition means".

Further, in the present embodiment, the acoustic model and the language model cooperate with each other to constitute an example of the “speech recognition dictionary” according to the twelfth aspect.
0 constitutes an example of the “provided information database” in the claim, the voice recognition unit 82 constitutes an example of the “recognition unit” in the claim, and the information search / provide unit 102 reads the “information search” in the claim. It constitutes an example of "means".

The embodiment of the present invention has been described in detail with reference to the drawings. However, this is merely an example, and the embodiments described in the above-mentioned "Means for Solving the Problems and Effects of the Invention" will be described. As various modifications based on the knowledge of those skilled in the art,
The invention can be implemented in other modified forms.

[Brief description of the drawings]

FIG. 1 is a functional block diagram conceptually showing a voice information search system used for implementing a voice information search method according to an embodiment of the present invention.

FIG. 2 is a client computer 12 in FIG.
FIG. 2 is a block diagram conceptually showing the hardware configuration of FIG.

FIG. 3 shows a client computer 12 in FIG.
Vehicle-mounted navigation device 60 and mobile telephone 7
It is a front view which illustrates 0.

FIG. 4 is a diagram conceptually showing a configuration of an acoustic model in FIG. 1 in a tabular form.

FIG. 5 is a diagram conceptually showing a configuration of a language model in FIG. 1 in a tabular form.

FIG. 6 is a block diagram conceptually showing a hardware configuration of server computer 10 in FIG.

7 is a flowchart conceptually showing the content of a voice input program in FIG. 2 and the content of a voice information search program in FIG.

8 is a flowchart conceptually showing the contents of S45 in FIG. 7 as a voice information search routine.

9 is a flowchart conceptually showing the contents of a dictionary learning program in FIG.

[Explanation of symbols]

 Reference Signs List 10 server computer 12 client computer 80 recognition dictionary memory 82 voice recognition unit 90 dictionary learning unit 100 provided information database 102 information search / providing unit

──────────────────────────────────────────────────続 き Continued on the front page (51) Int.Cl. ⁷ Identification symbol FI Theme coat ゛ (Reference) G10L 15/28 G10L 3/00 521S 551A 551P

Claims (12)

[Claims] 1. A method for recognizing a voice uttered by a specific or unspecified speaker by a computer, wherein the attribute is an attribute of an environment where the speaker is actually placed when the speaker utters the voice. Based on actual environment attribute data representing actual environment attributes, a plurality of types of speech recognition dictionaries prepared in advance are respectively associated with attributes of a plurality of types of environments in which the speaker is supposed to be placed when uttering a voice. A first step of selecting a speech recognition dictionary corresponding to the actual environment attribute from among the following, and using the selected speech recognition dictionary based on actual speech data representing actual speech actually emitted by the speaker. A second step of recognizing the actual voice and creating voice recognition result data representing the recognition result. 2. A third step of learning the plurality of types of speech recognition dictionaries so that the contents reflect the relationship between the actual environment attribute data, the actual speech data, and the speech recognition result data corresponding to each other. The speech recognition method according to claim 1, comprising: 3. A method for recognizing a voice uttered from an unspecified speaker by a computer, wherein a real personal attribute which is a real attribute of a target speaker who is a speaker emitting a voice to be recognized is determined. A plurality of types of speech recognition dictionaries based on the actual personal attribute data representing the actual speaker attribute data and the actual environment attribute data representing the actual environment attribute which is the attribute of the environment where the target speaker is actually placed when uttering the voice. ,
It is prepared in advance in association with the attributes of a plurality of types of environment in which the speaker is assumed to be placed when emitting a voice and the attributes of a plurality of types of individuals presumed for the speaker. First selecting a speech recognition dictionary corresponding to the actual personal attribute and the actual environment attribute from among the objects;
Based on actual speech data representing actual speech actually emitted by the target speaker, using the selected speech recognition dictionary to recognize the actual speech data, and a speech recognition result representing the recognition result. And a second step of creating data. 4. The method according to claim 1, wherein the contents of the plurality of types of voice recognition dictionaries reflect the relationship among the corresponding real voice data, voice recognition result data, real personal attribute data, and real environment attribute data. The voice recognition method according to claim 3, further comprising a third step of learning. 5. A server computer which can communicate with a plurality of client computers used by a plurality of users who desire to search for necessary information by voice, respectively. Is a method of searching for information conforming to the information search request by the user and transmitting the information to each client computer, which represents the actual attribute of the target speaker who is the speaker emitting the voice to be recognized. A plurality of types of speech recognition dictionaries based on actual personal attribute data and actual environment attribute data representing an actual environment attribute which is an attribute of an environment where the target speaker is actually placed when emitting the voice,
It is prepared in advance in association with the attributes of a plurality of types of environment in which the speaker is assumed to be placed when emitting a voice and the attributes of a plurality of types of individuals presumed for the speaker. First selecting a speech recognition dictionary corresponding to the actual personal attribute and the actual environment attribute from among the objects;
Based on actual speech data representing actual speech actually emitted by the target speaker, using the selected speech recognition dictionary to recognize the actual speech data, and a speech recognition result representing the recognition result. A second step of creating data; and at least a plurality of types of provided information that can be provided to each user based on the created voice recognition result data,
In the provided information database stored in advance in association with a plurality of types of speech recognition result data assumed in advance, the provided information corresponding to the created speech recognition result data is searched, and the searched provided information is Transmitting to the client computer used by the target user. 6. The provision information database according to claim 3, wherein the third step is based on the created speech recognition result data and refers to at least one of the actual personal attribute data and the actual environment attribute data. 6. The voice information search method according to claim 5, further comprising the step of searching for the provided information corresponding to the generated voice recognition result data. 7. The plurality of types of provided information include a plurality of types of geographical environment-related information on a geographical environment.
Or the voice information search method according to 6. 8. A computer-executable program for performing the method according to claim 1. 9. A recording medium recording the program according to claim 8 in a computer-readable manner. 10. A system for recognizing a voice uttered from an unspecified speaker by a computer, wherein a plurality of types of environment attributes are assumed in advance that the speaker is placed when uttering a voice; A plurality of types of speech recognition dictionaries prepared in advance in association with a plurality of types of attributes of the speaker presumed for the speaker, and the actual status of the target speaker who is the speaker emitting the voice to be recognized Based on the actual personal attribute data representing the actual personal attribute which is the attribute of the actual environmental attribute data representing the actual environmental attribute which is the attribute of the environment where the target speaker is actually placed when emitting the voice, A speech recognition dictionary corresponding to the actual personal attribute and the actual environment attribute is selected from a plurality of types of speech recognition dictionaries, and an actual speech data representing an actual speech actually produced by the target speaker is provided. The basis, by using the selected speech recognition dictionary, the speech recognition system comprising a recognition unit for the actual recognizing the speech data, to create a speech recognition result data representing the recognition result. 11. A server computer for recognizing a voice emitted from an unspecified user by communicating with a plurality of client computers used by a plurality of users who emit voices to be recognized. The attributes of a plurality of types of environment in which the speaker is assumed to be placed when uttering a voice, and the attributes of a plurality of types of person assumed in advance for the speaker are prepared in advance. And the actual attribute data representing the actual attribute of the target speaker who is the speaker who emits the voice to be recognized, and the target Based on the actual environment attribute data representing the actual environment attribute which is the attribute of the environment where the speaker is actually placed, the actual personal attribute and the actual A voice recognition dictionary corresponding to the boundary attribute is selected, and the actual voice data is obtained by using the selected voice recognition dictionary based on actual voice data representing actual voice actually uttered by the target speaker. A server for voice recognition, comprising: recognition means for recognizing and generating voice recognition result data representing the recognition result.
Computer. 12. An information search request by voice from an unspecified user by communicating with a plurality of client computers used by a plurality of users who desire to search for necessary information by voice. A server computer for searching for information that conforms to and transmitting the information to each client computer, wherein a plurality of attributes of the environment in which the speaker is supposed to be placed when uttering a voice, and the speaker A plurality of types of speech recognition dictionaries prepared in advance in association with a plurality of types of personal attributes assumed in advance, and a plurality of types of provided information that can be provided to each user, Provided information database stored in advance in association with the type of speech recognition result data, and information necessary for searching for the necessary information is issued by voice The actual personal attribute data representing the actual personal attribute, which is the actual attribute of the target user who is the user, and the real environment attribute, which is the attribute of the environment where the target user is actually placed when uttering the sound, Based on the actual environment attribute data to be represented, the speech recognition dictionary corresponding to the actual personal attribute and the actual environment attribute is selected from the plurality of types of speech recognition dictionaries, and the target user actually utters the speech recognition dictionary. Recognition means for recognizing the actual speech data by using the selected speech recognition dictionary based on actual speech data representing actual speech and creating speech recognition result data representing the result of the recognition; Based on the obtained speech recognition result data, the provided information corresponding to the speech recognition result data is searched in the provided information database, and the searched An information search means for transmitting provided information to the client computer used by the target user;