JP2002049390A - Speech recognition method and server and speech recognition system - Google Patents

Abstract

(57) [Summary] [PROBLEMS] To recognize voice information that could not be recognized by a client with low processing capability by a server with high processing capability and use the recognition result. An audio input signal (00) input to a client (1) is converted into a digital audio signal and transmitted to a server (2). The server 2 performs voice recognition processing on the voice information from the client, determines whether or not a predetermined recognition result is obtained by the voice recognition processing. If the recognition result is transmitted and the predetermined recognition result is not obtained, the voice information from the client 1 is transferred to another server having more information for voice recognition than the server with respect to the voice information, and another In the server, the voice information is subjected to voice recognition processing, and the result is transmitted to the client 1.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a speech recognition method, a server, and a speech recognition system. More particularly, in a distributed processing environment of a client / server system, speech information that cannot be recognized by a client having a low processing capacity is processed. The present invention relates to a speech recognition method, a server, and a speech recognition system characterized in that recognition is performed by a high server.

[0002]

2. Description of the Related Art Generally, a speech recognition system includes an acoustic analysis unit for converting a speech sample uttered by a speaker into a sequence of characteristic parameters, and a sequence of characteristic parameters obtained by the acoustic analysis unit in a memory or a hard disk. Of the vocabulary words stored in the storage device of the vocabulary word, and a speech matching unit that recognizes the speech having the highest similarity as a recognition result.

As an acoustic analysis method for converting a speech sample into a sequence of characteristic parameters, cepstrum analysis and linear prediction analysis are known, and "digital signal processing of speech / sound information" (Kiyohiro Kano, Satoshi Nakamura, It is also described in detail in Shiro Ise, co-authored by Shokodo. In speech recognition, a technique for recognizing the voice of an unspecified speaker is generally referred to as unspecified speaker speech recognition. In the speaker-independent speaker recognition, since the information on the characteristic parameters of the vocabulary words is stored in the storage device in advance, there is no need to register a word that the user wants to perform voice recognition as in specific speaker speech recognition. .

[0004] Further, as a method of generating information on characteristic parameters of vocabulary words and a method of voice collation between the information and a sequence of characteristic parameters converted from input voice,
A method based on a Hidden Markov Model (HMM) is generally used. In the HMM method, speech units such as syllables, syllables, phonemes, and acoustic events are modeled by the HMM. The modeling of the speech unit by the HMM is described in detail in the above-mentioned "digital signal processing of speech / sound information". Also, by using the Viterbi algorithm described in Chapter 4 of "Digital Signal Processing of Voice / Sound Information", those skilled in the art can easily configure an unspecified speaker voice recognition device.

On the other hand, some digital mobile phones include:
Devices that have a function as an unspecified speaker voice recognition device for realizing a voice dialing function have already been commercialized.

[0006] For example, NTT Mobile Communication Network Co., Ltd. im
As represented by the "mode", the demand for mobile phones not only for voice calls but also as access terminals for the Internet and e-mail is rapidly expanding. Also, IMT
Commercialization of next-generation mobile phones conforming to the -2000 and cdma2000 standards is also planned for next year or later, and the demand for access to the Internet and e-mail by mobile phones is expected to increase more and more.

[0007]

However, as a man-machine interface of a mobile phone, button input is still mainstream due to physical restrictions such as size and shape, and it is not easy to use as an Internet or e-mail access means. . In order to solve this problem, it is desired to provide a voice interface function using a voice recognition technology.

[0008] From the viewpoint of battery life, mobile phones are
The processing power of the installed CPU is low,
Operating clock is also kept low. Therefore, the functions supported by the voice recognition in the mobile phone are also limited. Specifically, if the vocabulary to be recognized is not sufficiently large, and if the user speaks a speech that includes a word that is not included in the vocabulary, the word is not recognized in principle, and the word is not recognized. It often happens that the word is misrecognized.

Therefore, the problem to be solved by the present invention is as follows:
When a voice uttered by a client having only a low processing capacity CPU such as a mobile phone includes a vocabulary that is not supported by the voice recognition function of the client, the client transmits the voice information to a server having a high processing capacity CPU. Transfer and receive voice recognition services.

[0010] Further, if it is determined that the vocabulary not supported by the voice recognition function of the server is included in the voice information, the server transfers the voice information to a higher-level server supporting a larger vocabulary. It is another object of the present invention to be able to entrust a voice recognition service.

[0011]

SUMMARY OF THE INVENTION According to the present invention, a client received via a communication means using at least one of a plurality of servers having information for speech recognition in a stepwise manner. A voice recognition method for recognizing voice information of the voice information, wherein until a predetermined recognition result is obtained for voice recognition of voice information from the client, more voice recognition is performed from the first server with respect to the voice information. At least one higher-level server having information for performing the voice recognition for the voice information from the client in order, and transmitting the recognition result to the client via the communication unit from the server having obtained a predetermined recognition result. It is characterized by doing.

[0012] According to a second aspect of the present invention, in the first aspect, the first server from which the predetermined recognition result is not obtained,
The information to be transferred to the higher-level server having information for voice recognition next to the first server with respect to the voice information is at least the voice information from the client and the recognition result in the first server. It is characterized by including one.

According to a third aspect of the present invention, in the first or second aspect, the predetermined recognition result is obtained based on a result of comparing a similarity of a word most similar to the voice information with a predetermined threshold value. It is characterized by the following.

According to a fourth aspect of the present invention, in any one of the first to third aspects, the audio information transmitted from the client is a sequence of a digital audio signal converted from an analog audio signal.

According to a fifth aspect of the present invention, in any one of the first to third aspects, the audio information transmitted from the client is a sequence of audio parameters extracted from the digital audio signal.

According to a sixth aspect of the present invention, in any one of the first to third aspects, the audio information transmitted from the client is a series of unique numbers of an audio unit model converted from the audio parameters. And

According to a seventh aspect of the present invention, in any one of the first to sixth aspects, the client and the server are devices having the same configuration.

According to an eighth aspect of the present invention, there is provided a server to which the voice recognition method according to any one of the first to seventh aspects is applied, wherein the time required for the voice recognition processing is measured, It is characterized by a usage fee calculation method in a speech recognition server that calculates a usage fee based on the amount of information for recognition.

According to a ninth aspect of the present invention, there is provided a server for receiving voice information from a client transmitted via a communication means and recognizing the voice, wherein the voice recognition performs voice recognition processing on the received voice information from the client. Processing means, determining means for determining whether a predetermined recognition result has been obtained by the voice recognition processing means, and when the predetermined recognition result has been obtained by the determination, the client to the client via the communication means When the predetermined recognition result is transmitted and the predetermined recognition result is not obtained by the determination, the higher-level server having more information for voice recognition than the server with respect to the voice information is transmitted through the communication unit. Output means for transferring voice information from the client.

According to a ninth aspect of the present invention, in the ninth aspect,
The determination unit is configured to make the determination based on a result of comparing a similarity of a word most similar to the voice information with a predetermined threshold.

According to a twelfth aspect of the present invention, in the ninth aspect,
The receiving means receives client transmission information composed of voice information and a client address transmitted from a client, the voice recognition processing means performs voice recognition by inputting the voice information received by the receiving means, Client address storage means for outputting recognition result information comprising a unique number of a word most similar to voice information and the degree of similarity thereof, and the output means for storing a client address extracted from client transmission information received by the reception means A recognition result returning unit for returning a recognition result to the client by referring to the client address stored in the client address storage unit, an upper server address storage unit for storing the address of the upper server, The audio information received by the receiving means and the client Voice information transfer means for transferring server transfer information composed of the client address stored in the address storage means to the upper server by referring to the address of the upper server stored in the upper server address storage means; The determination unit compares the similarity output from the speech recognition processing unit with a predetermined threshold, and when the similarity is equal to or greater than the threshold, and when the similarity is less than the threshold. If the address of the upper server is not stored in the upper server address storage means, the recognition result returning means is controlled to return the recognition result from the voice recognition processing means to the client, If the degree is less than the threshold value and the address of the upper server is stored in the upper server address storage means, the server transfer information And controlling the voice information transfer means to transfer to the server of the upper.

According to a twelfth aspect of the present invention, there is provided a voice recognition system comprising a client for transmitting voice information, and a voice recognition server according to any one of claims 9 to 11, which receives voice information from the client via communication means. Features the system.

[0023]

DESCRIPTION OF THE PREFERRED EMBODIMENTS In general, in speech recognition, input speech is converted into acoustic parameters such as cepstrum by acoustic analysis, and then collated with words included in the vocabulary by a known Viterbi algorithm. In the matching, the similarity between the acoustic parameters and the models of all the words included in the vocabulary is calculated, and the word having the highest similarity is selected as the recognition result.

The size of the vocabulary supported by the speech recognition device is determined by the processing capability of the CPU incorporated in the speech recognition device and its operation clock.

The cellular phone has a low processing capacity of a CPU mounted thereon and a low operation clock of the CPU from the viewpoint of battery life, and thus the vocabulary size that can be supported is as small as about 100 words.

On the other hand, for example, the CPU mounted on a computer system such as a server does not have the restriction of a cellular phone, so a CPU having the highest processing capability is employed, and the operation clock is set extremely high. . Accordingly, the vocabulary size that can be supported can be extremely large, on the order of tens of thousands of words.

Generally, when a word included in the input speech is not supported by the speech recognition device, the above-mentioned similarity in the collation has a small value.

The similarity of the vocabulary word having the highest similarity is compared with a predetermined threshold value. If the similarity is greater than the threshold value, the word included in the input speech is recognized as the speech. It can be determined that it is included in the vocabulary supported by the recognizer. In this case, the word giving the similarity may be used as the recognition result.

On the other hand, if the similarity is smaller than the threshold value, it is determined that there is a high possibility that the word included in the input speech is not included in the vocabulary supported by the speech recognition device. In order to perform accurate speech recognition on this input speech, it is necessary to entrust processing to a speech recognition device that supports a larger vocabulary.

If the location (address) of the speech recognition device that supports a larger vocabulary is known, the speech information extracted from the input speech and its own address are transferred to the speech recognition device, and the speech recognition process is performed. You can request

Hereinafter, the side requesting the voice recognition processing is called a client, and the side requesting the voice recognition processing is called a server.

The server calculates the similarity in the same manner as described above,
If the similarity of the vocabulary word having the highest similarity is greater than the threshold, it can be determined that the word included in the input speech is included in the vocabulary supported by the server, and the similarity is given. The word is returned to the client as a recognition result.

On the other hand, if the similarity is smaller than the threshold, the process is divided into two depending on whether the location (address) of the server supporting a larger vocabulary is unknown or known.

If unknown, the currently obtained recognition result is returned to the client.

If it is known, it is only necessary to transfer the speech information extracted from the input speech and its own address to the server that supports a larger vocabulary, and request speech recognition processing.

According to the present invention, when the number of words included in the input speech uttered by the client is small, the speech recognition processing is completed by the server supporting a small vocabulary.
If the number of words included in the input speech uttered by the client is large, the speech recognition processing can be entrusted to a server that supports a larger vocabulary. Therefore, it is possible to distribute the speech recognition processing to a plurality of servers according to the number of words included in the input speech. As a result, it is possible to prevent traffic of the request for the voice recognition processing from being concentrated on a specific server.

(Embodiment) An embodiment of the present invention will be described with reference to FIGS.

First, the voice input signal 00 is input to the client. The inputted input signal 00 is converted into a digital audio signal by an AD converter built in the client.

The client 1 transmits the client transmission information 11 to the server 2. Client transmission information 11 includes audio information 22 obtained from a digital audio signal.
And the client address 33. here,
The client address indicates an ID number or the like that can identify a client such as a mobile phone number. In addition, transmission
This is performed according to the communication protocol of the mobile phone network.

The server 2 which has received the client transmission information 11 via the mobile telephone network decomposes the client transmission information 11 into voice information 22 and a client address 33 in the voice information receiving means a, and converts the voice information 22 into voice recognition means. c and the voice information transfer means e, and the client address 33 is stored in the client address storage means b.

Voice recognition means c receiving voice information 22
Performs a matching process using a known Viterbi algorithm or the like, and outputs recognition result information 44 composed of the unique number of the word having the highest similarity and the similarity to the similarity comparison means d and the recognition result reply means g. I do.

The similarity comparison means d compares the similarity contained in the recognition result information 44 received from the speech recognition means c with a preset threshold value.

If the similarity is larger than the threshold value, reply instruction information 66 is output to the recognition result reply means g, and the recognition result reply means g outputs the reply instruction information via the mobile phone network.
Instructs the client to return the recognition result.

If the similarity is smaller than the threshold value, the processing is divided into two types depending on whether the upper server address 77 is stored in the upper server address storage means f.

If the high-order server address 77 is not stored in the high-order server address storage means f (the high-order server address 77 is unknown), reply instruction information 66 is output to the recognition result reply means g. The recognition result returning means g instructs the client to return the recognition result via the mobile phone network.

If the upper server address 77 is stored in the upper server address storage means f (the upper server address 77 is known), the transfer instruction information 55 is output to the voice information transfer means e, Voice information transfer means e
Through the communication means such as a cellular phone network to instruct the server to transfer the server transfer information 88 to a higher-level server.

The voice information transfer means e receiving the transfer instruction signal 55 from the similarity comparison means d sends the voice information reception means a to the upper server having the higher server address 77 stored in the higher server address storage means f. It sends out server transfer information 88 composed of the sent voice information 22 and the client address 33 stored in the client address storage means b, and requests voice recognition processing.

Upon receiving the reply instruction information 66 from the similarity comparison means d, the recognition result reply means g sends the recognition result received from the voice recognition means c to the client having the client address 33 stored in the client address storage means b. The information 44 is returned as server return information 99 via the mobile phone network.

The voice information receiving means a, voice recognition means b, similarity comparison means d, voice information transfer means e, and recognition result reply means g are realized by the CPU of the client. The client address storage means b and the upper server address storage means f are realized by the memory of the client.

The server and the host server described above can be regarded as a client and a server, respectively. The server transfer information 88 can be regarded as the client transmission information 11 from the viewpoint of the host server.

The voice information 22 can be regarded as a sequence of general information obtained from the input signal 00 input to the client.

The audio information 22 can be specified as a sequence of digital audio signals converted by an AD converter built in the client. In this case, the client only needs to have an AD converter.

Further, the audio information 22 can be specified as a sequence of acoustic parameters such as cepstrum extracted from a digital audio signal converted by an AD converter built in the client. In this case, the client only needs to have an AD converter and a CPU for calculating acoustic parameters.

Further, the audio information 22 is specified as a sequence of a unique number of an audio unit model converted from audio parameters such as cepstrum extracted from a digital audio signal converted by an AD converter built in the client. can do. In this case, the client only needs to be provided with an AD converter and a CPU for calculating a sequence of acoustic parameters and a unique number of a speech unit model. It is easy for those skilled in the art to extract a sequence of acoustic parameters having the highest similarity to the sequence of acoustic parameters using a known Viterbi algorithm. Alternatively, a method of calculating the probability of each speech unit model represented by the HMM for the acoustic parameter of each frame and determining the speech unit model giving the highest probability may be used.

In addition, the client and server in FIG. 1 may be devices having the same configuration.

Next, another embodiment of the present invention will be described with reference to FIGS.
This will be described with reference to FIG.

FIG. 3 is a diagram for explaining a voice recognition mode according to another embodiment of the present invention. A billing system for a voice recognition service to which the present invention is applied will be described with reference to FIG.
In FIG. 3, in addition to the client 1, three servers, a first server 12, a second server 13, and a third server 14 (all of which are as shown in FIG. 2) are installed in the mobile phone network. Upper server address storage means f of the first server 12
Stores the address of the second server 13, and stores the address of the third server 14 in the upper server address storage unit f of the second server 13, and stores the address of the upper server of the third server 14. It is assumed that the means f does not store the address of a higher-order server.

The voice recognition service provided by the first server 12, the second server 13, and the third server 14 can be used once (voice recognition processing) for ¥ 1, ¥ 5, ¥ 10 (each per unit time). Is charged to the client user. The service fee differs depending on the server. The higher the server, the larger the vocabulary is supported, so an expensive computer is required, and the fee is set higher in order to recover the capital investment.

Hereinafter, the first server 12, the second server 13,
The billing system will be described by taking as an example the case where the service time at each of the third servers 14 is one unit time.

When the user of the client 1 receives the voice recognition service in the first server 12 in response to the input signal 00, a service charge of 1 yen is charged.

As a result of the speech recognition service in the first server 12, all the words included in the input signal 00 are the first words.
If it is determined that the word is included in the vocabulary handled by the server 12, the first server 12 returns the recognition result to the client 1 as first server return information 99A. Therefore, the user of the client must pay a fee of 1 yen to the provider of the first server 12.

On the other hand, when it is determined that words not included in the vocabulary handled by the first server 12 are included in the input signal 00 as a result of the voice recognition service in the first server 12, The first server 12 transfers the first server transfer information 88A to the second server 13 based on the client transmission information 11. If, as a result of the speech recognition service in the second server 13, it is determined that all the words included in the input signal 00 are included in the vocabulary handled by the second server 13, 13 returns the recognition result to the client 1 as the second server return information 99B.
Therefore, the user of the client must pay a fee of 5 yen to the provider of the second server 13. At this point, the user of the client 1 has
A total of 6 yen must be paid for the service charge of 1 yen for the first server and the service charge for the second server 13 of 5 yen.

On the other hand, when it is determined that words not included in the vocabulary handled by the second server 13 are included in the input signal 00 as a result of the voice recognition service in the second server 13, The second server 13 transmits the first server transfer information 8
The second server transfer information 88B is transferred to the third server 14 based on 8A.

The third server 14 returns the result of the voice recognition service to the client as third server return information 99C. Therefore, the user of the client 1 must pay a fee of 10 yen to the provider of the third server 14. At this point, the user of the client 1 is charged 16 yen including the service charge of the first server 12 of 1 yen, the service charge of the second server 13 of 5 yen, and the service charge of the third server 14 of 10 yen. .

Of course, the user of the client 1 can directly send the client transmission information 11 to the third server 14. In this case, the burden on the user of the client 1 is only 10 yen. In this case, the user of the client 1 bears a burden when receiving all services of the first server 12, the second server 13, and the third server 14 in order.
You can save 6 yen compared to 6 yen service fee.
Compared to the case where the service of the server 12 is sufficient, the loss is 9 yen. First, the client user can freely select which server receives the service in consideration of the content of the words included in the input signal 00 and the service fee of the server.

Next, a method of reflecting the result of the voice recognition service in the first server 12 on the voice recognition service in the second server 13 to reduce the load of the voice recognition processing in the second server 13 will be described.

For example, when the user of the client 1 wants to check the information of the restaurant of the Mediterranean cuisine in the Shinjuku-ku area of Tokyo using a mobile phone, a "Mediterranean restaurant in Shinjuku"
To the mobile phone. At this time, the voice of “Mediterranean restaurant in Shinjuku” is input as signal 00,
It is input to the mobile phone as the client 1. The client 1 transmits a client transmission signal 11 to the first server 12 based on the input signal 00.

It is assumed that the speech recognition service of the first server 12 happens to support abundant vocabulary for Japanese place names, but does not support foreign place names as vocabulary. In this case, the recognition result information 44 output from the voice recognition unit c of the first server 12 is output in a format as shown in FIG.

Here, each line is the client transmission information 11
This is information about words extracted from. The first column from the left represents the label of the word, the second column from the left represents the score representing the certainty of the word, and the third column from the left represents the time at the end of the utterance of the word. I have. Sile
nce is a label representing silence. ? ? ? Indicates that the first server 12 has determined that words not included in the vocabulary supported by the first server 12 are uttered between 450 ms and 700 ms, and the likelihood of the words is 0% Is displayed.

The recognition result information 44 of FIG. 4 is attached to the transfer instruction signal 55 of FIG. 2 and sent to the voice information transfer means e of the first server 12, and further, the voice information transfer means e of the first server 12 Is transferred to the second server 13 as a part of the first server transfer information 88A.

The second server 13 comprises a voice information receiving means a
Then, the first server transfer information 88A is analyzed, and the voice information 22 is analyzed.
Then, the recognition result information 44 of FIG. The voice recognition means c of the second server 13 refers to the recognition result information of FIG. 4 and regards a word having certainty greater than a certain value as a correct answer. Then, a voice recognition process is performed.

If the word “Mediterranean” is included in the vocabulary supported by the second server 13, 450 m
Since the probability that the section from s to 700 ms corresponds to the word "Mediterranean" indicates a value of, for example, 95%, it can be recognized that the word "Mediterranean" is uttered.

As a result, the second server 13 receives the input signal 0
Since all words included in 0 can be recognized with high certainty, the second server reply information 99B can be returned to the client 1 via the recognition result transmitting means g.

If the word “Mediterranean” is not included in the vocabulary supported by the second server 13, the probability is low, and the second server 13 entrusts the third server 14 with the recognition service. Will be. Also in this case, it is possible to realize the same method as the above method.

In the above description, the client has been described as a mobile phone, and the server has been described as a voice recognition device installed in a server in the mobile phone network. However, the client is a computer such as a personal computer, and the server is a server in the Internet network. May be a voice recognition device installed in the device. Further, the combination may be such that the client is a mobile phone, and the server is a voice recognition device installed on a server in an internet network connected to the mobile phone network, or the client is a computer such as a personal computer, and the server is an internet network. Needless to say, the combination may be a speech recognition device installed in a server in a mobile telephone network connected to the mobile phone.

[0076]

As described above, according to the present invention,
Voice information that could not be recognized by a client with low processing capability can be recognized by a server with high processing capability, and the recognition result can be used.

[Brief description of the drawings]

FIG. 1 is a diagram illustrating a speech recognition mode by a client / server.

FIG. 2 is a diagram illustrating a configuration of a server.

FIG. 3 is a diagram illustrating a speech recognition mode by three servers.

FIG. 4 is a diagram showing a format of recognition result information.

[Explanation of symbols]

 00 Input signal 1 Client 2 Server 11 Client transmission information 12 First server 13 Second server 14 Third server 22 Voice information 33 Client address 44 Recognition result information 55 Transfer instruction signal 66 Reply instruction information 77 Upper server address 88 Server transfer information 88A First server transfer information 88B Second server transfer information 99 Server reply information 99A First server reply information 99B Second server reply information 99C Third server reply information a Voice information receiving means b Client address storage means c Voice recognition means d Similarity Comparison means e Voice information transfer means f Upper server address storage means g Recognition result return means

──────────────────────────────────────────────────続 き Continued on the front page (51) Int.Cl. ⁷ Identification symbol FI Theme coat ゛ (Reference) G10L 3/00 561A 561H 571C 571C 571J

Claims (12)

[Claims] 1. Using at least one of a plurality of servers having information for progressively more speech recognition,
A voice recognition method for recognizing voice information from a client received via a communication unit, wherein a first recognition is performed on the voice information until a predetermined recognition result is obtained for voice recognition of the voice information from the client. At least one upper server having more information for speech recognition from the server,
A voice recognition method comprising sequentially performing voice recognition on voice information from the client, and transmitting the recognition result to the client via a communication unit from a server that has obtained a predetermined recognition result. 2. The system according to claim 1, wherein, from the first server for which the predetermined recognition result has not been obtained, the first server has more information for voice recognition next to the first server with respect to the voice information. A speech recognition method, wherein the information to be transferred to the server includes at least one of speech information from the client and a recognition result in the first server. 3. The method according to claim 1, wherein the predetermined recognition result is obtained based on a result obtained by comparing a similarity of a word most similar to the voice information with a predetermined threshold value. Voice recognition method. 4. The voice recognition method according to claim 1, wherein the voice information transmitted from the client is a series of digital voice signals converted from an analog voice signal. 5. The speech recognition method according to claim 1, wherein the speech information transmitted from the client is a sequence of acoustic parameters extracted from the digital speech signal. 6. The speech recognition method according to claim 1, wherein the speech information transmitted from the client is a series of unique numbers of a speech unit model converted from the acoustic parameters. . 7. The speech recognition method according to claim 1, wherein the client and the server have the same configuration. 8. A server to which the voice recognition method according to claim 1 is applied, wherein a time required for voice recognition processing is measured, and the measured time and information for voice recognition related to the voice information are measured. A usage fee calculation method in a speech recognition server, wherein the usage fee is calculated based on the amount of the usage fee. 9. A server which receives voice information from a client transmitted via a communication means and recognizes the voice, wherein the voice recognition processing means performs voice recognition processing on the received voice information from the client; Determining means for determining whether or not a predetermined recognition result has been obtained by the voice recognition processing means; and if the predetermined recognition result has been obtained by the determination, the predetermined recognition result is transmitted to the client via communication means. Transmitting, if the predetermined recognition result is not obtained by the determination, the voice information from the client via the communication means to a higher-level server having more information for voice recognition than the server with respect to the voice information And an output means for transferring the speech. 10. The speech recognition apparatus according to claim 9, wherein said judgment means makes said judgment based on a result of comparing a similarity of a word most similar to said speech information with a predetermined threshold value. server. 11. The method according to claim 9, wherein the receiving unit receives client transmission information including voice information transmitted from the client and a client address, and the voice recognition processing unit receives the voice received by the receiving unit. Inputting information and performing voice recognition, outputting recognition result information including a unique number of a word most similar to the voice information and its similarity, the output unit is extracted from the client transmission information received by the reception unit. A client address storage unit for storing the client address stored in the client address storage unit, a recognition result return unit for returning a recognition result to the client by referring to the client address stored in the client address storage unit, and an address of the upper server. An upper server address storing means for storing, and the receiving means To transfer server transfer information composed of the voice information and the client address stored in the client address storage unit to the upper server by referring to the address of the upper server stored in the upper server address storage unit. The judgment means compares the similarity output from the speech recognition processing means with a predetermined threshold value, and when the similarity is equal to or greater than the threshold value, If the degree is less than the threshold value and the address of the upper server is not stored in the upper server address storage unit, the recognition result is returned so that the recognition result from the voice recognition processing unit is returned to the client. Means for controlling the means, and when the similarity is less than the threshold value, the address of the upper server is stored in said upper server address storing means. If so, the voice information transfer means controls the voice information transfer means to transfer the server transfer information to the host server. 12. A client for transmitting voice information,
A speech recognition system comprising: the speech recognition server according to claim 9, wherein the speech information from the client is received via communication means.