JP2002268684A - Sound model distributing method for voice recognition - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide a sound model distributing method for voice recognition which can increase variations of sound models by limiting the double registration of a sound model by the same provider. SOLUTION: This method for gathering a sound model from a provider A and registering it together with identification information on the provider A, and then distributing the sound model at a request made by a purchaser B has a step for inputting parameter information on the sound model to a server 8 together with the identification information on the provider A and a step for limiting the double registration of the sound model by checking the identification information on the provider A when the server 8 registers the sound model.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to an acoustic model distribution method for speech recognition.

[0002]

2. Description of the Related Art In recent years, information input using voice recognition of mobile phones, personal computers, and the like has begun to be put into practical use. Normally, these devices have a so-called “unspecified speaker acoustic model” in an initial state.

[0003] However, although this is an acoustic model for unspecified speakers, it does not show a high recognition rate for truly every voice, and some speakers are easy to recognize and some speakers are difficult to recognize. . Also, for a speaker who is easy to recognize, when the speaker is compared with an acoustic model tuned for himself,
In general, acoustic models for unspecified speakers have poor recognition performance. Therefore, a function (speaker adaptation function) for tuning the acoustic model is provided by having the user utter a learning voice of about several tens to several hundred words based on the acoustic model for unspecified speakers. There are not a few devices.

In general, the higher the number of learning voices, the higher the recognition performance can be expected. However, in order to actually utter words or sentences of several hundred words, a considerable amount of labor is required for the user. . Therefore, in order to save the user's effort for speaker adaptation required in advance, the distribution of an acoustic model using an acoustic model provided by another user who has similar voice quality and speech characteristics to the user. A method is conceivable.

[0005]

However, in order for this method to work well, it is necessary to prepare a large number of acoustic models having various variations. Even if many are registered, the variation of the acoustic model does not increase, and it is impossible to respond to purchasers of various acoustic models.

SUMMARY OF THE INVENTION The present invention has been made to solve the above-mentioned problems of the prior art, and it is possible to increase the number of variations of acoustic models by restricting duplicate registration of acoustic models from the same provider. It is another object of the present invention to provide an acoustic model distribution method for speech recognition.

[0007]

According to the first aspect of the present invention,
A method of collecting an acoustic model from a provider, registering the acoustic model together with the identification information of the provider, and distributing the acoustic model in response to a request from the purchaser. Inputting information to a server; and restricting duplicate registration of the acoustic model by examining the provider's identification information when registering the acoustic model.

According to a second aspect of the present invention, in the first aspect of the present invention, the learning information is input to the server together with the acoustic model parameter information.

The invention described in claim 3 is the first or second invention.
In the invention described above, overlapping registration of acoustic models is characterized in that learning scores of acoustic models are compared and limited.

[0010] The invention according to claim 4 is the invention according to claim 1 or 2.
In the described invention, overlapping registration of acoustic models is limited by selection by a provider of the acoustic model.

According to a fifth aspect of the present invention, in the fourth aspect, an index is presented to the provider when the provider selects an acoustic model.

[0012]

BRIEF DESCRIPTION OF THE DRAWINGS FIG. 1 is a block diagram showing a configuration of an audio model distribution method for speech recognition according to the present invention. FIG. 1 is a system configuration diagram showing an embodiment of a speech recognition acoustic model distribution method according to the present invention. Reference numeral 1 denotes an acoustic model provider terminal (hereinafter, referred to as a terminal).
"Terminal 1"), 2 is a terminal of a purchaser of the acoustic model (hereinafter, "Terminal 2"), 7 is a communication network,
Reference numeral 8 denotes an acoustic model distribution host computer (hereinafter, referred to as “server 8”).

The terminal 1 is an information processing device capable of communicating with the server 8 via the communication network 7 and has a voice recognition function. The provider A of the acoustic model is the terminal 1
By inputting the contents (about several tens to several hundreds of utterances) of the determined utterance list for speaker adaptation using, the speaker adaptation is performed, an acoustic model for the provider is constructed, and the server 8 To enter. Not only the acoustic model but also a part of the speech uttered for speaker adaptation (for example, the first word, “Hachinohe” in this example) is input to the server 8.

The terminal 2 is an information processing device capable of communicating with the server 8 via the communication network 7, and has a voice recognition function. The purchaser B of the acoustic model uses the terminal 2
Is input to the server 8 with the contents of the determined utterance list for model search (one to several utterances). The sound for preliminary selection of the acoustic model (for example, “Hachinohe”) is also input to the server 8.

Examples of the communication network 7 include a computer network represented by the Internet and a LAN. The terminals 1 and 2 and the server 8 are respectively connected to a public switched telephone network (PSTN), a wireless telephone network, a CATV network, and a satellite. It is connected to a communication network 7 via a communication line such as a communication network.

The server 8 distributes the acoustic model collected from the provider A to the purchaser B, that is, information processing managed and operated by an intermediary for mediating between the acoustic model provider A and the purchaser B. Device. The server 8 includes therein a survey / correspondence unit 3 for duplicate registration, a processing unit 4 for a purchaser, an acoustic model database 5, and an utterance list 6 for acoustic model search.

The duplicate registration investigation / correspondence unit 3 searches the acoustic model database 5 to determine whether or not the acoustic model of the same provider has already been registered.
Perform processing to eliminate duplication.

The processing unit 4 for the purchaser provides each function of "preliminary acoustic model selection function", "optimal model search function", "recognition accuracy calculation function", and "billing processing function". Hereinafter, each function will be described.

The preselection function of the acoustic model is performed by uploading (submitting) the sex and age of the buyer B and the buyer B.
By a simple process of comparing information such as voice for preliminary selection in the obtained voice sample for model search with information already uploaded from the provider A, from among a huge number of registered acoustic models, A narrowing down of acoustic model candidates considered to be suitable for the purchaser B is performed.

The search function of the optimal model is to apply a small number of acoustic models narrowed down by the preliminary selection to the voice of Buyer B, calculate a score (likelihood, etc.) for each model,
By sorting the values in descending order, the purchaser B
The most suitable model for the voice is obtained up to the n-th candidate.

The recognition accuracy calculation function performs an online speech recognition experiment using a speech sample for an evaluation test uploaded from the purchaser B in order to finally confirm the performance of the automatically searched acoustic model. . Although the grammar is prepared in advance, if grammar information is uploaded from the purchaser B, it can be used.

The billing function is such that when the purchaser B finally decides to purchase a specific acoustic model and the acoustic model is distributed to the terminal 2 of the purchaser B, information on the purchaser B, And the provider A of the acoustic model provided to the buyer B
Is stored as a pair, and the purchaser B is paid or the provider A is paid later.

The acoustic model database 5 stores the acoustic model provided by the provider A together with its additional information.

The utterance list 6 for acoustic model search is information shared by the provider A, the purchaser B, and the server 8.
This enables automatic search (including preliminary selection) of an acoustic model suitable for purchaser B.

Hereinafter, a method for restricting the duplicate registration of the acoustic model to the server 8 will be described. Note that, here, the provider A using the terminal 1 actually utters a specified word or sentence for speaker adaptation in his / her own voice, and has already created an acoustic model tuned to his / her own voice. Assume that you are building.

The acoustic model provider A uses the terminal 1 to connect to the server 8 via the communication network 7 and uploads the acoustic model parameter information. At this time, in addition to the acoustic model parameter information, terminal identification information, information for specifying the sex and age of the provider, and information on acoustic model learning stored in the terminal (number of learning words and learning score) Upload at the same time.

The server 8 stores the provider A received from the terminal 1
And unique provider identification information (provider ID) is provided based on the personal information described above, and the provider ID, parameter information of the uploaded acoustic model, and information at the time of learning are centrally managed on the acoustic model database 5. .

Each time the server 8 receives an acoustic model,
The provider ID is checked, and if the acoustic model uploaded from the same provider has already been registered, duplicate registration is restricted by one of the following two methods.

One method is to automatically select one of an already registered acoustic model and a newly uploaded acoustic model and formally adopt it. At that time, the learning score of the already registered acoustic model is compared with the learning score of the newly uploaded acoustic model,
The one with the larger value is officially adopted. That is, when the already registered acoustic model is adopted, the uploading process itself at that time is invalidated. On the other hand, when the newly uploaded acoustic model is adopted, the already registered acoustic model is used. The parameter information and the learning information of are replaced with newly uploaded ones.

Another method is a method in which the provider A selects one of an already registered acoustic model and a newly uploaded acoustic model, and formally adopts the selected acoustic model. From the terminal 1, the provider A designates which one of the already registered acoustic model and the newly uploaded acoustic model is to be officially registered, so that only one acoustic model is finally left.

At this time, the server 8 receives the upload date of the acoustic model already registered and the information at the time of learning, the past performance of the acoustic model (the frequency suitable for the voice of the purchaser, or the N-best candidate May be transmitted to the terminal 1, and the provider A may present an index for prompting a more accurate determination.

According to the embodiment described above, an acoustic model is collected from the provider A, registered with the identification information of the provider A, and distributed in response to a request from the purchaser B. Inputting the parameter information of the acoustic model together with the identification information of the provider A to the server 8, wherein the server 8 checks the identification information of the provider A when registering the acoustic model. Restricting the duplicate registration of acoustic models, it is possible to prevent the acoustic models to be collected from being registered in duplicate, thereby increasing variations of various acoustic models. As a result, it is possible to respond to purchasers of various acoustic models.

When the provider A selects the acoustic model, the index is presented to the provider A, so that the provider A can make a more accurate determination. It is possible to realize improvement of the quality of the acoustic model obtained.

[0034]

According to the first aspect of the present invention, it is possible to prevent the acoustic models to be collected from being registered repeatedly, so that it is possible to increase the variations of various acoustic models and to purchase various acoustic models. Can respond to people.

According to the fifth aspect of the present invention, when the provider selects an acoustic model to restrict duplicate registration, the provider can make a more accurate determination, and the registered It is possible to improve the quality of the acoustic model.

[Brief description of the drawings]

FIG. 1 is a system configuration diagram showing an embodiment of a speech recognition acoustic model according to the present invention.

[Explanation of symbols]

 Reference Signs List 1 acoustic model provider terminal (terminal 1) 2 acoustic model purchaser terminal (terminal 2) 7 communication network 8 acoustic model distribution host computer (server)

Claims (5)

[Claims] 1. A method of collecting an acoustic model from a provider, registering the acoustic model together with the identification information of the provider, and distributing the acoustic model in response to a request from a purchaser, comprising the steps of: Together with inputting parameter information of the acoustic model to the server, and the server restricting duplicate registration of the acoustic model by examining the identification information of the provider when registering the acoustic model. An acoustic model distribution method for speech recognition characterized by the following. 2. The server according to claim 1, wherein the learning information is input to the server together with the acoustic model parameter information.
The described acoustic model distribution method for speech recognition. 3. The acoustic model distribution method for speech recognition according to claim 1, wherein the duplicate registration of acoustic models is limited by comparing learning scores of acoustic models. 4. The acoustic model distribution method for speech recognition according to claim 1, wherein the duplicate registration of the acoustic model is restricted by selecting the acoustic model provider. 5. The acoustic model distribution method for speech recognition according to claim 4, wherein when the provider selects an acoustic model, an index is presented to the provider.