JP2002162989A - System and method for sound model distribution - Google Patents

Abstract

PROBLEM TO BE SOLVED: To allow an acquirer to obtain a sound model suitable to his or her voice from a host computer by an acquirer's terminal without taking labor for speaker adaptation. SOLUTION: The host computer 1 has a receiving means 11 which receives a sound model provided by a provider-side terminal 2, a storage means 12 which stores the sound model received by the receiving means 11, a retrieving means 13 which retrieves a sound model suitable to the acquirer's voice from sound models stored in the storage means 12 when the acquirer's sound is inputted from the acquirer's side terminal 3 (more in concrete, recognizes the acquirer's voice by using the sound models stored in the storage means 12 and retrieves a sound model by which an excellent recognition result (recognition score) can be obtained), and a sending-out means 14 which sends the sound model retrieved by the retrieving means 13 to the acquirer's side terminal 3.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

The present invention relates to an acoustic model distribution system and an acoustic model distribution method.

[0002]

2. Description of the Related Art In recent years, information input using speech recognition has been put into practical use in mobile phones and personal computers. Devices such as mobile phones and personal computers usually have a so-called “unspecified speaker acoustic model” in an initial state. However, in the case of an acoustic model for unspecified speakers, a high recognition rate is not shown for all voices, and some speakers are easy to recognize and some speakers are difficult to recognize. Also, for a speaker that is easy to recognize, the performance of the acoustic model for unspecified speakers is generally inferior to that of the acoustic model tuned for the speaker.

Therefore, a function of tuning the acoustic model (speaker adaptation function) by having the user utter a learning voice of about several tens to several hundreds of words based on the acoustic model for an unspecified speaker. ) Are provided in many devices. Generally, the greater the number of adaptation voices, the higher the recognition performance can be expected. However, uttering hundreds of words or sentences requires a considerable amount of labor.

[0004]

SUMMARY OF THE INVENTION According to the present invention, it is possible to obtain an acoustic model suitable for the user's own voice from a host computer without spending labor for speaker adaptation at the terminal of the user. It is an object of the present invention to provide an acoustic model distribution system and an acoustic model distribution method.

[0005]

In order to achieve the above object, the invention according to claim 1 comprises a host computer, a terminal of a provider that provides an acoustic model, and a terminal of a provider that acquires an acoustic model. A receiving unit for receiving the acoustic model provided from the terminal on the provider side, a storage unit for storing the acoustic model received by the receiving unit, and a terminal for acquiring from the terminal on the acquiring side. Searching means for retrieving an acoustic model suitable for the acquirer's voice from the acoustic models stored in the storage means when the user's voice is input; And sending means for sending to the terminal.

According to a second aspect of the present invention, in the acoustic model distribution system according to the first aspect, the terminal on the provider side generates a sound model by generating a sound model by performing speaker adaptation by inputting the sound of the provider. It is characterized by having a person adaptation function.

According to a third aspect of the present invention, in the acoustic model distribution system according to the first aspect, the provider side terminal further has a function of providing provider information to a host computer, When the computer receives the provider information from the terminal on the provider side, the computer stores the provider information in the storage means.

According to a fourth aspect of the present invention, in the acoustic model distribution system according to the first aspect, the provider side terminal further includes a provider attribute information and / or a learning score at the time of speaker adaptation. And / or a function of providing a voice sample used at the time of speaker adaptation to a host computer as provider ID information, wherein the host computer receives the provider ID information from a terminal on the provider side. At this time, it is characterized in that the ID information of the provider is stored in a storage means, and is used as additional information for searching for an acoustic model most suitable for the acquirer.

According to a fifth aspect of the present invention, in the acoustic model distribution system according to the first aspect, when the attribute information of the provider is collectively managed by a predetermined management system, the host computer executes the predetermined processing. It is characterized in that the attribute information of the provider is received from the management system and stored in the storage means, and is used as additional information for searching for the acoustic model most suitable for the user.

According to a sixth aspect of the present invention, in the acoustic model distribution system according to the fourth or fifth aspect,
The terminal on the side of the acquirer, together with the voice for acoustic model search,
Attribute information of the acquirer and / or audio samples,
The host computer has a function of giving the ID information of the acquirer to the host computer. When the ID information of the acquirer is given from the terminal of the acquirer, the host computer obtains the acquirer given from the terminal of the acquirer. ID information of the provider
The computer further includes a preliminary search unit that preliminarily selects a likely acoustic model candidate from the storage unit by collating the information with the information. Is searched for an acoustic model suitable for the voice of the acquirer.

According to a seventh aspect of the present invention, in the acoustic model distribution system according to any one of the first to sixth aspects, the terminal on the side of the acquirer is predetermined in the host computer. The speech for the user's acoustic model search for the vocabulary information and its correct information are given to the host computer, and the search means of the host computer uses the predetermined vocabulary information to perform a certain recognition on the acquirer's speech. An acoustic model indicating a rate is automatically retrieved from the storage means.

According to an eighth aspect of the present invention, in the acoustic model distribution system according to any one of the first to sixth aspects, the terminal on the acquirer side provides vocabulary information to the host computer, The voice for searching the acoustic model of the acquirer with respect to the vocabulary information from the acquirer's terminal and its correct information are given to the host computer, and the search means of the host computer uses the vocabulary information given from the acquirer's terminal Then, an acoustic model showing a certain recognition rate with respect to the voice of the acquirer is automatically searched from the storage means.

According to a ninth aspect of the present invention, in the acoustic model distribution system according to the eighth aspect, when the terminal on the acquirer side is a mobile phone and a telephone number list is used as vocabulary information, the telephone number list is used. This is characterized in that only information such as the name and nickname of the user is given to the host computer as vocabulary information, and a filter function is provided in the terminal of the acquirer so that the telephone number itself is not given to the host computer.

The invention according to claim 10 is the first invention.
10. The acoustic model distribution system according to claim 9, wherein the host computer further provides the retrieved acoustic model to a terminal on the side of the acquirer when the acoustic model is retrieved by the retrieval means. Means for confirming the performance of the acoustic model, and the confirming means actually recognizes the voice sent from the terminal on the side of the acquirer using the searched acoustic model, and It is characterized in that the score is given to the terminal of the acquirer as a confirmation result.

The invention according to claim 11 is the first invention.
0, the searching means has a function of searching for not only an optimal acoustic model but also a candidate for an acoustic model up to a predetermined order, and the confirming means has a function of searching for a candidate of an acoustic model up to a predetermined rank. If the user does not like one acoustic model candidate, another acoustic model candidate is presented.

The invention according to claim 12 is the first invention.
12. The acoustic model distribution system according to claim 11, wherein the host computer has a charging unit for charging the acquirer when the terminal on the acquirer side purchases the acoustic model from the host computer. It is characterized by doing.

The invention according to claim 13 is the first invention.
13. The acoustic model distribution system according to claim 12, wherein the host computer has a payment unit for paying a fee to the provider when the acoustic model is provided from a terminal on the provider side. It is characterized by doing.

According to a fourteenth aspect of the present invention, there is provided a host computer, a provider-side terminal for providing an acoustic model, and a provider-side terminal for acquiring an acoustic model. When receiving the acoustic model provided from the terminal of the acquirer, the received acoustic model is accumulated, and when the speech of the acquirer is input from the terminal of the acquirer, the acoustic model suitable for the acquirer's voice is accumulated. It is characterized in that a search is performed from among the acoustic models that have been performed, and the retrieved acoustic model is transmitted to the terminal of the acquirer.

The invention according to claim 15 is the first invention.
5. The acoustic model distribution method according to 4, wherein the host computer charges the acquirer when the acquirer terminal purchases the acoustic model from the host computer.

[0020] Further, the invention according to claim 16 is based on claim 1.
The acoustic model distribution method according to claim 4 or claim 15, wherein the host computer pays a fee to the provider when the acoustic model is provided from a terminal on the provider side.

[0021]

Embodiments of the present invention will be described below with reference to the drawings. According to the present invention, a certain user (acquirer of an acoustic model) acquires an acoustic model that has been learned (speaker adaptation) by another user (acoustic model provider) having voice characteristics and speech characteristics similar to those of the user. It is intended to be available. That is, a mobile phone (i-mode
(Registered trademark)) or a personal computer or other device (terminal) that has a voice recognition function, and the user (provider) of the provider's terminal learns the acoustic model (speaker adaptation) using his / her own voice. , To other user's equipment (the terminal on the side of the acquirer).

FIG. 1 is a diagram showing a configuration example of an acoustic model distribution system according to the present invention. Referring to FIG. 1, the acoustic model distribution system includes a host computer 1, a provider-side terminal 2 that provides an acoustic model, and a acquirer-side terminal 3 that acquires an acoustic model.

Here, the host computer 1 is owned and managed by, for example, an acoustic model intermediary, and includes a receiving unit 11 for receiving the acoustic model provided from the provider terminal 2, and a receiving unit 11. A storage unit 12 for storing the received acoustic model, and an acoustic model suitable for the acquirer's voice when the acquirer's voice is input from the terminal 3 on the acquirer side. Search from inside (more specifically, perform voice recognition processing on the voice of the acquirer using the acoustic model stored in the storage unit 12 and search for an acoustic model that can obtain a good recognition result (recognition score). ) It has a search means 13 and a sending means 14 for sending the acoustic model searched by the search means 13 to the terminal 3 on the side of the acquirer.

In the above acoustic model distribution system, the terminal 2 on the provider side has a speaker adaptation function of generating an acoustic model by learning (speaker adaptation) by inputting the speech of the provider. Specifically, in the terminal 2 on the provider side, the speaker adaptation is performed by the provider's voice input of the contents (about several tens to several hundreds of utterances) of the determined utterance list for speaker adaptation, An acoustic model is constructed, and this acoustic model can be sent to the host computer 1.

In the above acoustic model distribution system, the terminal 2 on the provider side further transmits the provider information (name, address, telephone number, etc. of the provider) to the host computer 1.
May be provided. In this case, when the host computer 1 receives the provider information from the provider-side terminal 2, the host computer 1 provides the provider information to the storage unit 12, for example, from the provider. The stored acoustic model is stored in association with the acoustic model. The provider information is used for, for example, paying a provision fee from the host computer 1 to the provider of the acoustic model, as described later.

In the above acoustic model distribution system, the terminal 2 on the provider side is, for example, the attributes (sex and age of the provider) of the provider stored in the terminal 2 on the provider side.
And / or the learning score at the time of speaker adaptation, and / or the speech sample used at the time of speaker adaptation (a part of speech uttered for speaker adaptation (for example, a speech sample such as the first word) or a speech sample). May be provided to the host computer 1 as ID information. In this case, when the host computer 1 receives the ID information from the provider-side terminal 2, the host computer 1 transmits the ID information to the host computer 1. For example, the information is stored in the storage unit 12 in association with the acoustic model provided by the provider. This ID information can be used as additional information for searching for an acoustic model that is optimal for the user, as described later.

In the above example, the host computer 1
Said that attribute information such as sex and age of the provider is received from the terminal 2 of the provider, but attribute information such as sex and age of the provider is collectively managed by a predetermined management system like i-mode. In this case, as shown in FIG. 2, the host computer 1 receives the attribute information of the provider not from the terminal 2 on the provider side but from the predetermined management system 20 and stores the attribute information in the storage unit 12. It can also be used as additional information for searching for the most appropriate acoustic model for the user.

Further, the terminal 3 on the side of the acquirer gives the host computer 1 the sound for searching for the acoustic model (for example, the contents (about one to several utterances) of the specified utterance list for searching for the acoustic model). Has become. In addition, the terminal 3 on the side of the acquirer has a function of providing the host computer 1 with the attribute information of the acquirer and / or the audio sample for preliminary selection together with the audio for acoustic model search as the acquirer ID information. May be. In this case, as shown in FIG. 3, the host computer 1 compares the ID information of the acquirer given from the terminal 3 of the acquirer with the ID information of the provider (from the terminal 3 of the acquirer). By comparing the given ID information of the acquirer with the ID information of the provider, or by comparing the audio sample for preliminary selection given from the terminal 3 of the acquirer with the audio sample from the provider. ), A preliminary search means 15 for preliminarily selecting a sound model candidate which is likely to some extent may be further provided. At this time, the search means 13 of the host computer 1 searches for the sound model preliminarily selected by the preliminary search means 15. An acoustic model suitable for the voice of the acquirer is searched from the candidates.

That is, in the acoustic model distribution system of FIG. 3, the number of acoustic models stored in the storage means 12 of the host computer 1 reaches a certain number, and among all the acoustic models, the optimal for the voice of the acquirer. If it becomes difficult to search for a sound model in terms of the amount of calculation, the attribute information of the user, such as age and gender, submitted at the same time as the sound for the sound model search from the user is provided by the provider. The effort of calculating the recognition score by applying an acoustic model by comparing with attribute information, or by directly comparing the audio sample submitted by the provider with the audio sample submitted by the provider. Without performing the processing that takes a certain amount of time, a candidate for the acoustic model that is likely to some extent (for example, the attribute information of the provider almost close to the attribute information of the acquirer is The acoustic model provided by the provider at the time of the acquisition or the acoustic model provided by the provider when there is an audio sample of the provider which is almost similar to the audio sample of the acquirer) is preliminarily selected by the preselecting means 15. Preliminary selection makes it possible to realize a quick search for the acoustic model in the search means 13.

Specifically, the preliminary selection means 15 compares information such as the sex and age of the acquirer and the audio sample for preliminary selection submitted by the acquirer with the information submitted by the provider. The simple processing described above narrows down the candidates of the acoustic model considered to be suitable for the acquirer from among the huge acoustic models registered in the storage unit 12.

The search means 13 of the host computer 1 applies a small number of acoustic models narrowed down by preliminary selection to the input speech from the acquirer and performs speech recognition, as described later, and performs a speech recognition on each acoustic model. By calculating scores (likelihoods) and sorting the values in descending order, the most suitable model for the voice of the acquirer is obtained up to the N-th candidate.

Also, in the acoustic model distribution system shown in FIG. 1, FIG. 2 or FIG. 3, vocabulary information predetermined in the host computer 1 exists. It is necessary to provide the host computer 1 with the acoustic model search voice for the vocabulary information predetermined in the host computer 1 and its correct answer information from the terminal 3 on the side of the acquirer in order to obtain the information. As a result, the search means 13 of the host computer 1 uses the vocabulary information that matches the correct answer information to perform a speech recognition process of the acquirer, and generates an acoustic model that shows a constant recognition rate for the acquirer's speech. Automatic search can be performed from the storage unit 12. Here, when the reading information for the predetermined vocabulary is not set in the host computer 1, the reading information can be given to the host computer 1 as the correct answer information from the terminal 3 on the obtaining side. When reading information for a predetermined vocabulary is set in the host computer 1, a number for designating the number of reading information as correct answer information is sent from the terminal 3 on the side of the acquirer. To the host computer 1.

FIG. 4 is a diagram for explaining the above processing in detail. Referring to FIG. 4, information 101 about vocabulary information determined in advance is stored in the host computer 1.
Exists. In the example of FIG. 4, the reading information on the predetermined vocabulary is set in the information 101 on the predetermined vocabulary information. Specifically, in the example of FIG. 4, “Tokyo”,
"Osaka",... Are registered in advance, and "Tokyo", "Osaka",. In this case, the acquirer utters an acoustic model search voice “Osaka” for vocabulary information (for example, “Osaka”) determined in advance in the host computer 1 as an acoustic model search voice, and obtains a terminal on the side of the obtainer. 3 and, as the correct answer information, a number "2" designating the "second reading information" is input to the terminal 3 on the side of the acquirer. As a result, the host computer 1 can detect that the sound for searching the acoustic model from the terminal 3 on the side of the acquirer is that of the reading information “Osaka”, and the searching unit 13 determines the acoustic model of each phoneme of “Osaka”. Among them, the acoustic model most suitable for the acoustic model retrieval speech “Osaka” can be retrieved using the speech recognition technology.

In the above example, the host computer 1 searches for the most suitable acoustic model for the vocabulary predetermined for acoustic model search. In order not to search for an optimal acoustic model for the vocabulary but to allow the host computer 1 to select an optimal acoustic model for the vocabulary that the acquirer actually wants to use, the terminal 3 on the acquirer side uses the speech recognition system used. Grammar information (vocabulary information to be recognized) for host computer 1
And the search means 13 of the host computer 1 can be configured to automatically search for an acoustic model showing the highest recognition rate when the grammatical information is used.

FIG. 5 is a diagram for explaining the processing of the acoustic model distribution system having such a configuration. Referring to FIG. 5, after obtaining the vocabulary information to the host computer 1, the terminal 3 on the acquirer side obtains the acoustic model search voice (voice sample) of the acquirer for the vocabulary information from the terminal 3 on the acquirer side
And the correct information thereof to the host computer 1, and the search means 13 of the host computer 1 uses the vocabulary information given from the terminal 3 on the side of the acquirer to set the acoustic model showing a certain recognition rate for the speech of the acquirer. Is automatically retrieved from the storage means 12. That is, the search means 13 of the host computer 1 generates an acoustic model capable of correctly recognizing the speech (speech sample) of the acquirer under the constraint of the vocabulary information (grammar) given from the terminal 3 on the acquirer side. It is designed to search.

In the example of FIG. 5, the terminal 3 on the acquirer side is a mobile phone, and the terminal 3 on the acquirer side gives a telephone number list as vocabulary information. That is, in the example shown in FIG. 5, the terminal 3 on the side of the acquirer gives the name reading information “Suzuki-san”, “Sato-san”,... It has become. In this case, the terminal 3 on the side of the acquirer causes the acquirer to utter words present in the vocabulary information given from the terminal 3 on the acquirer side to the host computer 1, for example, "Suzuki-san", "Sato-san" or the like. Thus, the host computer 1 can be provided with the voice data and the correct answer information (for example, a number indicating the position of the vocabulary information corresponding to the voice data). Then, the search information 13 of the host computer 1
Then, instead of searching for an optimal acoustic model for a predetermined vocabulary, using a vocabulary information provided from the terminal 3 on the side of the acquirer, an acoustic model showing a certain recognition rate for the voice of the acquirer is used. The model is automatically retrieved from the storage unit 12,
It can be provided to the terminal 3 on the side of the acquirer.

As described above, the vocabulary information is provided to the host computer 1 from the terminal 3 on the side of the acquirer, and the voice for searching the acoustic model of the acquirer and the correct information thereof for the vocabulary information from the terminal 3 on the side of the acquirer are With the configuration provided to the host computer 1, it is possible to search for and provide a more suitable acoustic model according to the purpose of use of the user.

In the above example, a telephone number list is given to the host computer 1 as vocabulary information.
A filter function in which only information to be directly recognized, such as names and nicknames in the telephone number list, is given to the host computer 1 as vocabulary information, and the telephone numbers assigned to individual names are not given to the host computer 1. Can be provided in the terminal 3 on the side of the acquirer, and in this case, protection of privacy can be realized.

In the acoustic model distribution system shown in FIG. 1, FIG. 2 or FIG. 3, as shown in FIG.
3 is based on FIG. 3), when the acoustic model is retrieved by the retrieval means 13, the host computer 1 sends the retrieved acoustic model to the terminal 3 on the side of the acquirer before the acoustic model is retrieved. Confirmation means 1 for confirming model performance
6 may be further provided. In this case, the confirmation means 1
6 can actually recognize the voice sent from the terminal 3 on the acquirer side by using the searched acoustic model, and give the score to the terminal 3 on the acquirer side as a confirmation result.

That is, in the acoustic model distribution system of FIG. 6, before the acoustic model automatically retrieved (selected) by the host computer 1 is actually downloaded to the terminal 3 of the acquirer, the terminal 3 of the acquirer is required to download the acoustic model. Can check the performance of the acoustic model automatically searched (selected) in the host computer 1 by the checking means 16 of the host computer 1 online. Specifically,
The confirmation means 16 of the host computer 1 actually recognizes the voice (voice sample for evaluation test submitted by the obtainer) transmitted from the terminal 3 on the obtainer side by using the selected acoustic model, and The score (confirmation result) is transmitted to the terminal 3 on the side of the acquirer. Thus, the acquirer can determine in advance whether or not the acoustic model automatically searched (selected) by the host computer 1 may be actually acquired.

In the acoustic model distribution system of FIG. 6, the search means 13 of the host computer 1 may have a function of searching for not only an optimal acoustic model but also acoustic model candidates up to a predetermined order. In this case, if the acquirer does not like one acoustic model candidate retrieved by the retrieval unit 13, the confirmation unit 16 can present another acoustic model candidate. In other words, not only the optimal acoustic model, but also the second-
If the search means 13 of the host computer 1 is provided with a function of searching for a candidate of the rank, and the user does not like the performance of the automatically selected acoustic model, the candidate of the second rank and the candidate of the third rank are determined. It can also be presented to the acquirer one after another until a certain number is reached.

In the acoustic model distribution system shown in FIG. 1, FIG. 2, FIG. 3 or FIG. 6, as shown in FIG. 7 (the example of FIG. 7 is based on FIG. 6), When the terminal 3 has purchased an acoustic model from the host computer 1, the host computer 1 may be provided with a charging means 17 for charging the user.

FIG. 1, FIG. 2, FIG. 3, FIG. 6 or FIG.
When the acoustic model is provided from the provider-side terminal 2 to the host computer 1 as shown in FIG. 8 (the example of FIG. 8 is based on FIG. 7), Payment means 18 for paying a fee to the provider may be provided in the host computer 1.

Next, a specific example of the processing operation in the above-described acoustic model distribution system of the present invention (for example, the acoustic model distribution system of FIG. 8) will be described. Now, a user using a terminal having a voice recognition function (a user of the terminal 2 on the provider side) actually speaks a designated word or sentence for speaker adaptation in his / her own voice, and Suppose that an acoustic model tuned to the voice of the user has been constructed.

Thereafter, the user (provider) connects the provider-side terminal 2 to the host computer 1 for distributing the acoustic model via a telephone line (for example, a public telephone line) or a computer network such as the Internet. Performs key operations for providing an acoustic model. As a result, the acoustic model parameter information is transmitted from the provider terminal 2 to the host computer 1. From the terminal 2 on the provider side, in addition to the parameters of the acoustic model, provider information, information for specifying the sex and age of the provider, and speaker adaptation stored in the terminal 2 ID information of the provider such as a part (about several words) of the voice sample spoken for the user can be simultaneously transmitted to the host computer 1 as additional information.

On the other hand, a user who wants to obtain an acoustic model suitable for his / her voice (user of the terminal 3 on the side of the user)
After connecting the terminal 3 on the side of the acquirer to the host computer 1 for distributing the acoustic model via a telephone line (for example, a public telephone line) or a computer network such as the Internet, a voice (voice of several words) and its correct answer information are obtained. Further, it transmits to the host computer 1 the acquirer information or the acquirer ID information for specifying the sex, age, etc. of the acquirer.

Host computer 1 for distribution of acoustic model
Above, preliminary selection of some acoustic models provided by providers close to the sex and age of the acquirer, and from among them,
Using the voice sample submitted by the acquirer and its correct answer information, the top N searchers showing the highest recognition rate (or average recognition score) for the acquirer's voice are searched.

For the acoustic model selected in this way, the acquirer can submit the specified confirmation word or sentence to the acoustic model distribution host computer 1. If this is submitted, host computer 1
Performs a recognition process on the voice for confirmation using a grammar of a certain scale including a specified word or sentence for confirmation, and transmits information of the recognition result or the recognition rate to the terminal 3 on the side of the acquirer.

If the acquirer likes the acoustic model, the acoustic model is automatically downloaded to the acquirer's terminal 3 by performing a key operation or the like on the acquirer's terminal 3 to determine the acquisition. .

If the acquirer does not like the acoustic model, the user operates the terminal 3 on the acquirer side to perform a key operation or the like to notify the user of the acoustic model. The same processing is repeated for the acoustic model of the order (the value of N is increased by one until the number reaches a certain number, such as 2, 3,... According to the number of repetitions).

When the acquirer finally decides to acquire (purchase) the acoustic model, the host computer 1 charges the acquirer and transfers a part (or all) of the charge to the host computer 1 (intermediary agent). ), And pay the rest to the provider by remittance or discount on call charges.

More specifically, when the acquirer finally decides to purchase a specific acoustic model, and the acoustic model is distributed to the terminal 3 of the acquirer, the host computer 1 The information of the user and the information of the provider of the acoustic model provided to the acquirer are stored as a pair, and at a later date, the acquirer is charged or the reward is paid to the provider.

[0053]

As described above, according to the first to sixteenth aspects of the present invention, a host computer,
A terminal of a provider that provides an acoustic model, and a terminal of a provider that acquires an acoustic model, wherein the host computer receives the acoustic model provided from the terminal of the provider, and receives the received acoustic model. When the model is stored and the acquirer's voice is input from the acquirer's terminal, an acoustic model suitable for the acquirer's voice is searched from the stored acoustic models, and the retrieved acoustic model is obtained. Since the data is transmitted to the terminal on the speaker side, the terminal on the acquirer side can obtain an acoustic model suitable for the user's own voice from the host computer without spending labor for speaker adaptation. Further, even if the terminal of the acquirer does not have the speaker adaptation function, the terminal of the acquirer can acquire an acoustic model suitable for the voice of the acquirer.
That is, in order to realize speaker adaptation processing on a terminal,
According to the present invention, even a user of an inexpensive terminal (a terminal on the side of the acquirer) not equipped with such a CPU needs to be equipped with a CPU having a high processing capability to some extent. , And a high recognition rate is realized, so that the cost of the terminal (the terminal on the side of the acquirer) can be reduced as a result.

In particular, according to the fourth aspect of the present invention, in the acoustic model distribution system according to the first aspect, the terminal on the provider side further includes attribute information of the provider and / or a speaker adaptation time. It has a function of providing a learning score and / or a voice sample used at the time of speaker adaptation to a host computer as ID information of a provider, and the host computer transmits ID information of the provider from a terminal on the provider side. At the time of reception, the ID information of the provider is stored in the storage means and is used as additional information for searching for an acoustic model that is optimal for the acquirer. It is possible to search for an acoustic model accurately.

According to the fifth aspect of the present invention, in the acoustic model distribution system according to the first aspect, when the attribute information of the provider is collectively managed by a predetermined management system, the host computer is configured to: The attribute information of the provider is received from a predetermined management system and stored in the storage means, and is used as additional information for searching for an acoustic model that is optimal for the acquirer. The model can be searched accurately.

According to the sixth aspect of the present invention, in the acoustic model distribution system according to the fourth or fifth aspect, the terminal on the acquirer side acquires the attribute information of the acquirer together with the sound for acoustic model search. And / or a function of giving the audio sample to the host computer as the ID information of the acquirer, and the host computer, when the ID information of the acquirer is given from the terminal of the acquirer, A preliminary search means for preliminarily selecting a likely acoustic model candidate from the storage means by comparing the ID information of the acquirer provided from the terminal on the side with the ID information of the provider, and The means retrieves an acoustic model suitable for the voice of the acquirer from among the acoustic model candidates preliminarily selected by the preliminary retrieval means. In, it is possible to find the optimal acoustic model to obtain user even more suitable probability.

According to an eighth aspect of the present invention, in the acoustic model distribution system according to any one of the first to sixth aspects, the terminal on the acquirer side provides the vocabulary information to the host computer. Also, a speech for searching the acoustic model of the acquirer with respect to the vocabulary information from the acquirer's terminal and its correct answer information are given to the host computer, and the search means of the host computer uses the lexical information given from the acquirer's terminal. Is automatically searched from the storage means for an acoustic model showing a certain recognition rate for the voice of the acquirer, so that a more appropriate acoustic model suitable for the purpose of use of the acquirer is searched and provided. can do.

According to the ninth aspect of the present invention, in the acoustic model distribution system according to the eighth aspect, when the terminal on the acquirer side is a mobile phone and a telephone number list is used as vocabulary information, the telephone number is used. Only the information such as names and nicknames in the list are given to the host computer as vocabulary information, and a filter function that does not give the telephone number itself to the host computer is provided to the terminal of the acquirer, so that privacy protection can be realized. it can.

According to the tenth and eleventh aspects of the present invention, in the acoustic model distribution system according to any one of the first to ninth aspects, the host computer further includes: When a model is searched, prior to providing the searched acoustic model to the terminal of the acquirer, the model has a confirmation unit for confirming the performance of the acoustic model, and the confirmation unit includes a terminal of the acquirer. Is actually recognized using the searched acoustic model, and the score is given to the terminal on the side of the acquirer as a confirmation result, so that the acquirer automatically searches (selects) in the host computer 1. ) It is possible to determine in advance whether the acquired acoustic model is actually acceptable.

[Brief description of the drawings]

FIG. 1 is a diagram showing a configuration example of an acoustic model distribution system according to the present invention.

FIG. 2 is a diagram showing another configuration example of the acoustic model distribution system according to the present invention.

FIG. 3 is a diagram showing another configuration example of the acoustic model distribution system according to the present invention.

FIG. 4 is a diagram for explaining processing between the terminal on the side of the acquirer and the host computer when reading information for a predetermined vocabulary is set in the host computer.

FIG. 5 is a process between the terminal on the acquirer side and the host computer when the grammar information (recognition target vocabulary information) for the speech recognition system being used is provided from the terminal on the acquirer side to the host computer. FIG.

FIG. 6 is a diagram showing another configuration example of the acoustic model distribution system according to the present invention.

FIG. 7 is a diagram showing another configuration example of the acoustic model distribution system according to the present invention.

FIG. 8 is a diagram showing another configuration example of the acoustic model distribution system according to the present invention.

[Explanation of symbols]

 DESCRIPTION OF SYMBOLS 1 Host computer 2 Provider's terminal 3 Acquirer's terminal 11 Receiving means 12 Storage means 13 Search means 14 Sending means 15 Preliminary search means 16 Confirmation means 17 Billing means 18 Payment means 20 Predetermined management system

──────────────────────────────────────────────────続 き Continued on the front page (51) Int.Cl. ⁷ Identification symbol FI Theme coat ゛ (Reference) G06F 17/60 302 G06F 17/60 506 506 G10L 3/00 521S G10L 15/00 551A

Claims (16)

[Claims] 1. A host computer, a provider terminal for providing an acoustic model, and a provider terminal for acquiring an acoustic model, wherein the host computer is provided from the provider terminal. Receiving means for receiving the acoustic model, storage means for storing the acoustic model received by the receiving means, and when the voice of the acquirer is input from the terminal on the acquirer side, the acoustic model suitable for the voice of the acquirer, An acoustic model comprising: retrieval means for retrieving from acoustic models stored in a storage means; and transmission means for transmitting the acoustic model retrieved by the retrieval means to a terminal on the side of the acquirer. Delivery system. 2. The acoustic model distribution system according to claim 1, wherein the terminal on the provider side has a speaker adaptation function of generating an acoustic model by adapting a speaker by inputting a speech of the provider. An acoustic model distribution system, characterized in that: 3. The acoustic model distribution system according to claim 1, wherein the provider terminal further has a function of providing provider information to a host computer, and wherein the host computer has a provider terminal. Wherein the provider information is stored in a storage unit when the provider information is received from the computer. 4. The acoustic model distribution system according to claim 1, wherein the terminal on the provider side further includes attribute information of the provider and / or a learning score at the time of speaker adaptation and / or speaker adaptation. The host computer has a function of providing the used audio sample to the host computer as provider ID information. When the host computer receives the provider ID information from the provider side terminal, Is stored in a storage means, and is used as additional information for searching for an acoustic model most suitable for the user. 5. In the acoustic model distribution system according to claim 1, when the attribute information of the provider is collectively managed by a predetermined management system, the host computer transmits the attribute information of the provider from the predetermined management system. And receiving the stored information in a storage means, and using the received information as additional information for searching for an acoustic model most suitable for the user. 6. The acoustic model distribution system according to claim 4 or 5, wherein the terminal on the acquirer side acquires the attribute information of the acquirer together with the sound for acoustic model search.
Alternatively, the host computer has a function of providing the audio sample to the host computer as ID information of the acquirer, and when the ID information of the acquirer is given from the terminal of the acquirer, the host computer And a preliminary search unit that preliminarily selects, from the storage unit, candidates of the acoustic model that are likely to some extent by comparing the ID information of the acquirer given from with the ID information of the provider, and the search unit of the host computer includes: An acoustic model distribution system wherein an acoustic model suitable for the voice of the acquirer is retrieved from acoustic model candidates preliminarily selected by preliminary retrieval means. 7. The acoustic model distribution system according to claim 1, wherein the terminal on the side of the acquirer is an acoustic model of the acquirer with respect to vocabulary information predetermined in the host computer. A search voice and its correct answer information are provided to the host computer, and the search means of the host computer stores the acoustic model indicating a certain recognition rate with respect to the voice of the acquirer using predetermined vocabulary information. An acoustic model distribution system characterized in that the acoustic model is automatically searched for from a means. 8. The acoustic model distribution system according to claim 1, wherein the terminal on the side of the acquirer provides the vocabulary information to the host computer, and receives the vocabulary information from the terminal on the side of the acquirer. The voice for searching the acoustic model of the user with respect to the vocabulary information and the correct information thereof are provided to the host computer, and the search means of the host computer uses the vocabulary information provided from the terminal on the side of the obtainer to perform the search for the voice of the obtainer An acoustic model distribution system wherein an acoustic model having a constant recognition rate is automatically retrieved from the storage means. 9. The acoustic model distribution system according to claim 8, wherein the terminal of the acquirer is a mobile phone, and when a telephone number list is used as vocabulary information, only information such as names and nicknames in the telephone number list is provided. , As a vocabulary information, and a filter function is provided in the terminal of the acquirer so that the telephone number itself is not given to the host computer. 10. The acoustic model distribution system according to any one of claims 1 to 9, wherein the host computer further obtains the retrieved acoustic model when the acoustic model is retrieved by the retrieval means. Prior to giving to the terminal on the side of the user, the apparatus has a confirmation unit for confirming the performance of the acoustic model, and the confirmation unit uses the retrieved acoustic model to convert the voice sent from the terminal on the side of the acquirer. An acoustic model distribution system, which performs actual speech recognition and gives the score as a confirmation result to a terminal on the side of the acquirer. 11. The acoustic model distribution system according to claim 10, wherein said search means has a function of searching for not only an optimal acoustic model but also acoustic model candidates up to a predetermined order. An acoustic model distribution system characterized by presenting another acoustic model candidate when the acquirer does not like one acoustic model candidate retrieved by the retrieval means. 12. The acoustic model distribution system according to any one of claims 1 to 11, wherein the host computer transmits to the acquirer when the terminal on the acquirer side purchases the acoustic model from the host computer. An acoustic model distribution system characterized by having a charging means for performing charging of (1). 13. The acoustic model distribution system according to claim 1, wherein the host computer charges a fee to the provider when the acoustic model is provided from a terminal on the provider side. An acoustic model distribution system comprising payment means for making a payment. 14. A host computer, a provider terminal for providing an acoustic model, and a provider terminal for obtaining an acoustic model, wherein the host computer is provided from the provider terminal. When the acoustic model is received, the received acoustic model is stored, and when the acquirer's voice is input from the acquirer's terminal, the acoustic model suitable for the acquirer's voice is selected from the stored acoustic models. A method for distributing acoustic models, characterized by searching and transmitting the searched acoustic models to a terminal on the side of the acquirer. 15. The acoustic model distribution method according to claim 14, wherein the host computer, when the terminal on the side of the purchaser has purchased the acoustic model from the host computer,
An acoustic model distribution method characterized by charging a user. 16. The acoustic model distribution method according to claim 14 or 15, wherein the host computer receives the acoustic model from a terminal on a provider side.
A method of delivering an acoustic model, comprising paying a fee to a provider.