JP2001100786A - Method and device for speech recognition, and storage medium - Google Patents

Abstract

PROBLEM TO BE SOLVED: To prevent the repeat of erroneous recognition of the same vocabulary in speech recognition processing and to improve the recognition ratio of a voice inputted from a terminal connected through a network. SOLUTION: A penalty value is subtracted by a penalty value subtraction part 204 after the degree of similarity of an input speech is calculated by a calculation part 203, and a vocabulary to be outputted as the recognition result is selected in accordance with the result by parts 205 and 206.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a technique for recognizing input speech.

[0002] The present invention relates to processing when speech is erroneously recognized.

[0003] The present invention relates to speech recognition of unspecified speakers.

[0004]

2. Description of the Related Art There is a technique in which a voice of an unspecified speaker is taken in, a similarity between a vocabulary to be recognized and an input voice is calculated, and a recognition result is output based on the similarity. At this time,
When a plurality of candidates are output as a recognition result, there is a method in which the candidates are arranged in order of similarity and presented to the user, and the user selects one from the candidate group by voice or other input means.

Further, in the case of a system for recognizing voice data input from a plurality of clients, a method for processing voice input in the same environment by learning microphone sensitivity, noise level, and speaker characteristics. Can improve the recognition rate.

[0006]

The speech recognition technology does not always select the vocabulary intended by the user as the first recognition candidate without error. It is not guaranteed that the vocabulary intended and uttered by the user obtains the highest similarity in the vocabulary to be recognized (referred to herein as the first recognition result). This is due to the user's utterance habits, ambient noise, microphone characteristics, and characteristics of the speech recognition system.

For this reason, in a dialogue system using voice input, it is necessary for the user himself to check whether or not the recognition result of the user's utterance is correct and correct it. In existing systems, the following means are generally used. 1. The recognition results up to the Nth position are displayed on the screen, and the user designates the vocabulary intended by the user with a button, a touch panel, or the like, or utters the number displayed with the vocabulary, The system recognizes the speech and obtains the vocabulary intended by the user. 2. The system synthesizes and reads the first to Nth recognition results, and asks the user for each vocabulary.
The user answers with voice “yes”, “no”, buttons, and the like. 3. The recognition results up to the Nth place are all read out first by speech synthesis, and the user designates the vocabulary intended by the user with a button, a touch panel, or the like, or utters a number assigned to the vocabulary, thereby producing a speech. input.

In the first method, since the selection can be made using the screen, the user can quickly select the intended vocabulary. However, since a terminal using a screen is required, the system becomes expensive.

The above two and three methods are generally used in a system using a telephone line, and can be operated by a terminal such as a cellular phone which outputs sound. Furthermore, the second method requires less time loss if the vocabulary intended by the user is higher in the recognition result, but has a longer time until the vocabulary intended by the user is uttered in the lower case. Required and time loss is increased. In addition, when the recognition result does not fall in the N-th place, the vocabulary is spoken again to perform re-input and re-recognition after answering "No" to all N words, so that the burden on the user is large. In addition, there are many systems in which N = 1 in particular in Method 2. This is a system that repeats vocabulary input and confirmation. A simple recognition task will reduce the burden on the user, but a difficult recognition task may not be able to be entered properly for the reasons described above. Comes out.

The third method is effective for difficult recognition tasks, but the user must listen to a long system announcement for confirmation. When there is no vocabulary intended by the user in the recognition results up to the N-th place, the burden on the user is as large as in Method 2.

Also, the speech recognition processing is generally very heavy processing, and therefore requires a computer with high specifications. Further, information equipment terminals carried by users are required to be more portable than how high the specifications. Portable information device terminals, such as Internet browsers, having only a minimum number of applications that can communicate with servers have emerged. For this reason, it is realistic to leave the voice recognition processing to the server and the client to work on the data transmission and reception.

Since the speech recognition process is greatly affected by the environment in which the speaker is present and the voice of the speaker, even a speech recognition device corresponding to an unspecified speaker learns the voice of the speaker and noise in which the speaker is present. This greatly improves the performance. But,
If the speech recognition system is a system that is divided into a server and a number of clients connected by a certain line, even if learning about a certain client, the result cannot be used by other clients, so the client connects to the server. You have to relearn every time, or save and manage the default learning results for all clients. For this reason, it is important that the speech recognition server authenticates the client and performs a speech recognition process suitable for the client.

[0013]

SUMMARY OF THE INVENTION In order to solve the above-mentioned problems of the prior art, the present invention provides a method for inputting voice, obtaining a similarity between the input voice and dictionary data, and obtaining a penalty corresponding to the dictionary data. A speech recognition method, apparatus, and storage medium for obtaining an evaluation value in consideration of a penalty value by subtracting a value from the obtained similarity and selecting a vocabulary to be output as a recognition result of the input voice based on the evaluation value are provided. .

[0014] In order to solve the above-mentioned problems of the prior art, the present invention preferably selects a word having a higher evaluation value as a vocabulary to be output as the recognition result.

In order to solve the above-mentioned problems of the prior art, the present invention preferably determines whether or not the selected vocabulary is incorrect, and updates the penalty value if it is determined that the selected vocabulary is incorrect.

In order to solve the above-mentioned problems of the prior art, according to the present invention, preferably, the penalty value is a value held for each vocabulary.

In order to solve the above-mentioned problem of the prior art, the present invention preferably updates the penalty value by increasing the value.

In order to solve the above-mentioned problems of the prior art, the present invention preferably updates the penalty value to a smaller value each time the recognition process is repeated.

In order to solve the above-mentioned problem of the prior art, the present invention preferably inputs the voice via a network.

In order to solve the above-mentioned problems of the prior art, the present invention preferably outputs the selected vocabulary.

In order to solve the above-mentioned problems of the prior art, the present invention preferably outputs the selected vocabulary via a network.

In order to solve the above-mentioned problem of the prior art, the present invention is preferably configured such that, when an instruction to select a vocabulary other than the first place is input to the output vocabulary, the error is determined. judge.

In order to solve the above-mentioned problem of the prior art, the present invention preferably inputs the instruction via a network.

[0024] In order to solve the above-mentioned problems of the prior art, the present invention preferably changes the processing for recognizing the voice according to the identification information of the terminal that has transmitted the voice via the network.

In order to solve the above-mentioned problems of the prior art, the present invention preferably changes parameters used for recognizing the voice according to the identification information of the terminal that has transmitted the voice via the network.

[0026] In order to solve the above-mentioned problems of the prior art, the present invention preferably holds identification information of a terminal connectable via the network and information relating to speech recognition suitable for the terminal.

In order to solve the above-mentioned problems of the prior art, the present invention preferably changes the processing at the time of recognition in accordance with the stored information.

In order to solve the above-mentioned problems of the prior art, according to the present invention, the parameters at the time of recognition are preferably changed according to the stored information.

In order to solve the above-mentioned problems of the prior art, the present invention is preferably arranged such that, when identification information of a terminal which has transmitted a voice via the network is not stored in advance, the identification information and the terminal New information relating to speech recognition suitable for is registered.

[0030]

FIG. 1 is a functional block diagram of a speech recognition apparatus according to the present invention. In FIG. 1, 101, 107
Is a telephone, 102 and 106 are public networks, 103 and 108 are microphones, 104 is a speaker, 105 is a display screen,
Reference numeral 109 denotes a button, and 201 denotes a voice recognition device.

Next, each element constituting the speech recognition apparatus (201) will be described. The voice of the user captured from the microphone (103) and the telephone (101) enters the voice recognition device (201) from the voice capturing unit (202).
The similarity calculation unit (203) uses the data (21
0) is calculated. The similarity is calculated by a penalty value subtracting unit (204).
1) is subtracted. That is, as the penalty value increases, the score of the similarity decreases. Then, the result sorting unit (20
5) and output the vocabulary from the top to the Nth order as a candidate group obtained as a result of recognition on the display screen (105) with character information or by voice to the speaker (104) and the user. Telephone connected to line (10
7) Announce (notify) to the user by outputting.

The user receives the recognition result and, based on the output N recognition results, inputs the vocabulary intended by the user from a voice from the telephone directory (107), a push button, a microphone (10).
8), buttons (109) attached to the display screen,
Select by touch panel. This is the user's recognition processing for the speech recognition result of the system. This confirmation result is sent to the misrecognition detection unit (207). Here, it is determined whether or not the speech recognition result is incorrect recognition. If not, the vocabulary selected by the user is determined as a result. Output from the output unit (209). The determination as to whether or not the recognition is erroneous is made by determining that the first candidate is selected by the user's confirmation processing, that no erroneous recognition is performed, otherwise, erroneous recognition is performed.

The result of the user's confirmation is also sent to the penalty value setting section (208). Here, the penalty value stored in the penalty value database (211) is updated according to the user's confirmation result.

FIG. 2 is a flowchart of the voice recognition system according to the present invention. Here, an example of a system that recognizes a station name spoken by a user and confirms the station name will be described.

When the system starts up, the penalty arrays corresponding to all the recognition target words are initialized to 0 (S101 to S103). This value corresponds to a penalty value. Next, the system outputs an announcement “Please give me the station name” (S104) to prompt the user to say the station name. Thereafter, a voice uttered by the user is captured (S105), and a voice recognition process is performed.

First, the parameter max_ruijido
Is initialized with -∞ (S106). This parameter is
This is a storage area that holds the maximum similarity in consideration of the penalty value, and is a variable that always contains the maximum value up to that point. The degree of similarity between the input speech and the recognition target vocabulary is calculated for all the recognition target vocabularies (S108), and a value obtained by subtracting a penalty value for each vocabulary from the value is compared with the value of this parameter. The vocabulary [recognition target vocabulary [max] "that maximizes the similarity in consideration of the penalty value by selecting and storing the vocabulary is searched for (S109 to S109).
111).

Next, a process for confirming the result with the user is performed. As a system announcement, ""
x] "Is it OK?" (S112), and prompts the user by pressing the "Yes" or "No" button (S113). The vocabulary uttered in S112 is the vocabulary of the number stored in max in S110. If the "No" button is pressed, it is a misrecognition,
Announce "recognition failure" (S115), and the S11
A 1 is given as a penalty value to the recognized vocabulary uttered in 2, that is, the vocabulary specified by the value stored in the parameter max (S116). If it is determined in S114 that the "Yes" button has been pressed, the recognition is successful and the recognition result is announced again (S117).

Finally, the penalty values associated with all the words to be recognized are updated (S118 to S120). Actually, each penalty value is multiplied by 0.8 (S1
20). As a result, the penalty value immediately after erroneous recognition is 0.8, and the penalty value is continuously multiplied by 0.8 as the number of times of recognition increases, so that the value gradually approaches zero.

When these processes are completed, the system asks the user whether or not to end the process (S121). If the process is to be continued, an announcement prompting the utterance of the station name is issued (S121).
Return to 104).

Next, the processing of the extension agent system server using the above-described speech recognition will be described with reference to the flowchart of FIG.

When the system starts up, the penalty value database is initialized with 0 (S301), and waits for a call from the user (S302). When the call is received, an announcement is sent to "Who will you connect?" (S303), and the user is prompted to speak a personal name (S30).
4). The speech taken in here calculates the similarity (ruijido [i]) with the vocabulary to be recognized, and three vocabularies in the vocabulary in the descending order of ruijido [i] −penalty [i], i = max ₁ , i = max ₁ , Search for max ₂ and max ₃ (S305).

The recognition results of the first to third ranks are synthesized and output by voice to indicate to the user, and allow the user to select a desired person name. First, “a vocabulary to be recognized [max i] (i = 1 to
3) to insert ““ recognition target vocabulary [ma
x ₁ ] ”is 1, 1 is“ recognition target vocabulary [max ₂ ] ”, and _{3 is} “ recognition target vocabulary [max ₃ ] ”.
If not, say 4. Is generated and a system announcement is output (S30).
6). Next, the user's voice is captured (S307), and the recognition target vocabulary is set to "1", "2", "3", or "4", and a voice recognition process is performed to determine which recognition target vocabulary is closer to the voice captured in S307 (S308). .

If the recognition result i is 1 to 3,
A call is forwarded to “recognition target vocabulary [max _i ]” (S
309 to S311), the system returns to the penalty value initialization process (S301), and returns to the standby state for the user's telephone (S302). If the recognition result i is 4, the recognition result is determined to be erroneous recognition, and the similarity with the input speech of the recognized vocabulary is substituted for the penalty value of the recognized vocabulary as the recognition result (S312). Then, the penalty values of all the recognition target vocabularies are updated, and the process returns to the process of prompting the user to speak a personal name again (S303).

Next, a system for inputting a station name by using the above-described voice recognition will be described with reference to the flowchart of FIG.

When the system starts up, the penalty array is initialized with 0 (S401 to S403). Next, an announcement urging the utterance of the station name is played (S404), and the utterance of the user is taken in (S405). From here, speech recognition processing using the station name as the vocabulary to be recognized is performed, but the penalty value is 0.
Only the vocabulary for the word is recognized (S408). As a result, the processing for recognition can be reduced somewhat.
The recognition target vocabulary whose penalty value is 0 calculates the similarity with the input voice (S409). One of the vocabulary words having the highest similarity is selected from the recognition target vocabulary having the penalty value of 0, and the selected vocabulary is used as the recognition result (S406 to S412).

The recognition result is confirmed to the user by an announcement (S413). The user answers with voices of “yes” and “no” (S414, S415), and if “yes”, announces the recognition result again (S416), and terminates this processing. If “No”, a penalty value of the recognition target vocabulary as a recognition result is set. This value is an integer part obtained by multiplying the similarity of the recognition result by 10 (S4
17).

When the penalty value is set, an announcement to prompt the user to speak the station name is played again (S418),
The entire penalty value is updated (S419 to S42)
3). This means that for non-zero penalty values,
The penalty value is subtracted by one (S4).
22).

In the present system, the penalty value can be considered to be the number of times that a certain vocabulary to be recognized is not added to the recognition processing after the recognition error.

The system described so far is a system in which a user accesses a device having a voice recognition function using a telephone and inputs a voice, and also returns a recognition result of the input voice to the user's telephone. A system for connecting a user terminal to a speech recognition server via a network (the Internet, a LAN, or the like) connected by some kind of line will be described.

FIG. 5 is a diagram showing the configuration of such a system.

This speech recognition system comprises a speech recognition server (521) and a plurality of speech recognition clients (501) connected thereto via a network.

The voice recognition client (501) captures a user's voice waveform from a voice capture device such as a microphone (502), and sends the user's voice waveform to the voice recognition server (521) by the voice waveform transmission unit (503). The speech recognition client does not perform the speech recognition process directly, but only captures the speech and transforms the speech waveform into a form suitable for transmission over the network. Since the processing is light, the operation is possible even in an environment where the processing performance is not so high.

The voice recognition server (521) is a part for voice-recognizing the voice waveform obtained via the network.
Here, a speech recognition unit (527) for performing actual speech recognition
And a client management unit (522) for acquiring speech recognition parameters necessary for speech recognition from the client information. The client ID acquisition unit (523) acquires a unique ID for each client from the IP address and port number of the client attached to the audio waveform data obtained via the network, and stores the ID and the client database (526). Compare.

If the data corresponding to the client ID exists in the client database, it means that the client has previously been accessed, and the speech recognition parameters learned at that time are input to the speech recognition parameter acquisition unit (525). ) To get. If there is no data corresponding to the client ID in the client data, it is considered that the client has accessed the voice recognition server for the first time, and the voice recognition parameters are learned and stored in the client database together with the client ID.

The voice recognition parameters of the client obtained in this way are stored in the voice recognition unit (52) together with the voice waveform.
7) to perform voice recognition processing.

FIG. 6 is a flowchart showing the processing executed in the system of FIG.

The voice recognition client fetches a voice waveform (S601) and transmits the waveform to the voice recognition server (S602).

When data is transmitted from the client (S611), the voice recognition server acquires a client ID from the transmitted IP address and port number of the client (S612). Next, the client database is searched, and the obtained client ID is obtained.
If there is a voice recognition parameter corresponding to (S613), the parameter is acquired (S61).
4). If not, it is assumed that the client has accessed for the first time, and the speech recognition parameters are learned (S
615), and store the result in the client database (S616). Using these speech recognition parameters,
The voice waveform is recognized (S617), and the recognition result is transmitted to the client (S618).

The speech recognition client displays the result from the speech recognition server (S603), and ends the processing.

FIG. 8 is a flowchart showing a processing example different from that of FIG. 6 in the system of FIG.

The client here can be considered a computer connected to the Internet. The client software is installed on a computer as a Web browser plug-in. At the time of this installation, an ID unique to this client is generated. This I
D needs to be unique among all clients, and is generated from the machine, the serial number of the client software, the IP address, and the like.

When the user accesses the voice recognition server with a Web browser, the voice of the user is captured by a microphone connected to the computer (S601), and the ID created by the plug-in at the time of installation is acquired (S70).
1), is transmitted to the server together with the audio waveform (S703).
Here, the audio waveform transmitted to the server is not limited to the raw audio data captured from the microphone, but may be audio data compressed specifically for voice recognition. Further, if the task is a simple speech recognition task such as "yes" or "no" that requires only a small amount of processing, the speech recognition may not be sent to the speech recognition server but may be performed by the plug-in of this client. At this time, the entire audio waveform may be fetched and then transmitted to the server all at once, or the audio waveform may be transmitted in real time for the fetched amount.

The voice recognition server constantly monitors the data sent from the client (S61).
1) When it is sent, a search is made to see if the client ID is present in the client database (S61).
2). If the client ID is registered in the client database, the voice recognition parameter is obtained (S613). If not registered, a default value is set as a voice recognition parameter (S616).
Voice recognition is performed using these voice recognition parameters and voice waveforms (S614), and the result is transmitted to the client (S615).

After transmitting the data to the server, the client monitors for the recognition result returned from the server (S).
614). When the result is returned, the result is shown to the user (S615), and the processing of the client ends.

After the server sends the result to the client,
The speech recognition parameters unique to the client are calculated again (S617). Specifically, 1. The lowest power in the audio waveform is calculated, and the value is used as the noise level. 2. The lowest power in the speech waveform is calculated, and the power spectrum in the vicinity is set as the noise power spectrum. 3. The highest power in the audio waveform is calculated, and the calculated value is used as the microphone sensitivity level. 4. Speaker clustering is performed, and the result is used as speaker information. Can be considered. These voice recognition parameters are stored in the client database together with the client ID (S618). This makes it possible to perform speech recognition using the speech recognition parameters obtained from the speech waveform before one access, so that it is relatively easy to adapt to the environment of the client.

By using the noise level of 1 and the microphone sensitivity level of 3, the threshold value for voice / non-voice determination can be set according to the client environment. In addition, Spectra using the noise power spectrum of 2
By performing lSubtraction, a clean speech waveform without noise can be estimated from the input power spectrum. By using the speaker information of No. 4, an acoustic model suitable for the speaker can be used.

In the case of a large-scale speech recognition system in which many clients are connected to one server, the client database may be too large. Therefore, if only the latest 10,000 client databases are registered, the client data is more likely to be used for frequently used clients, and has been used only once in the past. The client's client data will quickly disappear from the database.

By adopting the system as described above,
Even in the case where the speech recognition system is divided into a server and a large number of clients, speech recognition parameters corresponding to each client can be used.

[Brief description of the drawings]

FIG. 1 is a functional configuration diagram of a speech recognition device according to the present invention.

FIG. 2 is a flowchart showing a speech recognition process according to the present invention.

FIG. 3 is a flowchart showing processing of an extension intermediary system.

FIG. 4 is a flowchart showing processing of a station name input system.

FIG. 5 is a configuration diagram of a system in which a server and a client are connected via a network.

FIG. 6 is a flowchart showing a first process in the system of FIG. 5;

FIG. 7 is a flowchart showing a second process in the system of FIG. 5;

 ──────────────────────────────────────────────────続 き Continuing on the front page (72) Inventor Tsuyoshi Yagisawa 3-30-2 Shimomaruko, Ota-ku, Tokyo Within Canon Inc. (72) Inventor Katsuhiko Kawasaki 3-30-2 Shimomaruko, Ota-ku, Tokyo (72) Inventor Hiroki Yamamoto 3-30-2 Shimomaruko, Ota-ku, Tokyo Canon Inc. (72) Inventor Masaaki Yamada 3-30-2, Shimomaruko, Ota-ku, Tokyo Canon Inc. F term (reference) 5D015 AA02 HH05 KK02 LL02 LL04 LL05 LL07 LL12

Claims (51)

[Claims] 1. A voice is input, a similarity between the input voice and dictionary data is obtained, and a penalty value corresponding to the dictionary data is subtracted from the obtained similarity to obtain an evaluation value in consideration of a penalty value. A speech recognition method comprising selecting a vocabulary to be output as a recognition result of the input speech based on the evaluation value. 2. The speech recognition method according to claim 1, wherein a word having a higher evaluation value is selected as a vocabulary to be output as the recognition result. 3. The speech recognition method according to claim 1, wherein it is determined whether or not the selected vocabulary is incorrect, and if the selected vocabulary is determined to be incorrect, the penalty value is updated. 4. The speech recognition method according to claim 1, wherein the penalty value is a value held for each vocabulary. 5. The speech recognition method according to claim 3, wherein the updating of the penalty value increases the value. 6. The speech recognition method according to claim 1, wherein the penalty value is updated to a smaller value each time the recognition process is repeated. 7. The speech recognition method according to claim 1, wherein the speech is input via a network. 8. The speech recognition method according to claim 1, wherein the selected vocabulary is output. 9. The speech recognition method according to claim 1, wherein the selected vocabulary is output via a network. 10. The speech recognition according to claim 3, wherein, when an instruction to select a vocabulary other than the first vocabulary is input to the output vocabulary, the vocabulary is determined to be the error. Method. 11. The speech recognition method according to claim 10, wherein the instruction is input via a network. 12. The voice recognition method according to claim 7, wherein a process for recognizing the voice is changed according to the identification information of the terminal that has transmitted the voice via the network. 13. The apparatus according to claim 7, wherein parameters used for recognizing the voice are changed according to the identification information of the terminal that has transmitted the voice via the network.
Voice recognition method described in. 14. The speech recognition method according to claim 1, wherein identification information of a terminal connectable via the network and information relating to speech recognition suitable for the terminal are stored. 15. The speech recognition method according to claim 12, wherein the processing at the time of recognition is changed according to the held information. 16. The speech recognition method according to claim 13, wherein parameters for the recognition are changed according to the stored information. 17. If the identification information of the terminal that has transmitted the voice via the network is not stored in advance,
13. The speech recognition method according to claim 12, wherein the identification information and information relating to speech recognition suitable for the terminal are newly registered. 18. An input unit for inputting a voice, a similarity deriving unit for obtaining a similarity between the input voice and the dictionary data, and a penalty value obtained by subtracting a penalty value corresponding to the dictionary data from the obtained similarity. A speech recognition apparatus comprising: an evaluation value deriving unit that obtains an evaluation value in consideration of a value; and a selection unit that selects a vocabulary to be output as a recognition result of the input speech based on the evaluation value. 19. The speech recognition apparatus according to claim 18, wherein said selection means selects a word having a higher evaluation value as a vocabulary to be output as said recognition result. 20. A determination unit for determining whether or not the selected vocabulary is incorrect, and a penalty value updating unit for updating the penalty value when the determination unit determines that the word is incorrect. The speech recognition device according to claim 18, wherein: 21. The speech recognition apparatus according to claim 18, further comprising a holding unit that holds the penalty value for each vocabulary. 22. The speech recognition apparatus according to claim 20, wherein the penalty value updating unit updates the value so as to increase the value. 23. The method according to claim 2, wherein the penalty value is updated to a smaller value each time the recognition process is repeated.
2. The speech recognition device according to 1. 24. The speech recognition apparatus according to claim 18, wherein said input means inputs speech via a network. 25. The speech recognition apparatus according to claim 18, further comprising output means for outputting the selected vocabulary. 26. The speech recognition apparatus according to claim 18, further comprising output means for outputting the selected vocabulary via a network to a network. 27. The method according to claim 20, wherein the determination unit determines that the error is the error when an instruction to select a vocabulary other than the first vocabulary is input to the output vocabulary. The speech recognition device according to the above. 28. The speech recognition apparatus according to claim 27, wherein the instruction is input via a network. 29. The voice according to claim 24, further comprising control means for controlling a process for recognizing the voice in accordance with the identification information of the terminal which has transmitted the voice via the network. Recognition device. 30. The speech recognition apparatus according to claim 24, further comprising control means for changing a parameter used for recognizing the speech in accordance with the identification information of the terminal that has sent the speech via the network. 31. The voice according to claim 18, further comprising identification information of a terminal connectable via said network, and recognition information holding means for holding information relating to voice recognition suitable for the terminal. Recognition device. 32. The speech recognition apparatus according to claim 29, wherein the control unit changes the processing at the time of the recognition in accordance with the information held in the holding unit. 33. The speech recognition apparatus according to claim 30, wherein the control unit changes the parameter at the time of the recognition according to the information held in the holding unit. 34. If the identification information of the terminal that has transmitted the voice via the network is not stored in advance,
30. The speech recognition apparatus according to claim 29, wherein the identification information and information relating to speech recognition suitable for the terminal are newly registered. 35. A control program for inputting voice, a control program for obtaining a similarity between the input voice and dictionary data, and a penalty value corresponding to the dictionary data is subtracted from the obtained similarity. A control program for obtaining an evaluation value in consideration of a penalty value, and a control program for selecting a vocabulary to be output as a recognition result of the input voice based on the evaluation value. A readable storage medium. 36. The computer-readable storage medium according to claim 35, wherein a word having a higher evaluation value is selected as a vocabulary to be output as the recognition result. 37. A control program for determining whether the selected vocabulary is incorrect, and a control program for updating the penalty value when the selected vocabulary is determined to be incorrect. A storage medium readable by a computer according to claim 35. 38. The computer-readable storage medium according to claim 35, wherein the penalty value is a value held for each vocabulary. 39. The computer-readable storage medium according to claim 37, wherein the updating of the penalty value increases the value. 40. The method according to claim 3, wherein the penalty value is updated to a smaller value each time the recognition process is repeated.
A storage medium readable by a computer according to claim 5. 41. The computer-readable storage medium according to claim 35, wherein the voice is input via a network. 42. The computer-readable storage medium according to claim 35, wherein a control program for outputting the selected vocabulary is stored. 43. The computer-readable storage medium according to claim 35, wherein a control program for outputting the selected vocabulary via a network is stored. 44. A control program for determining an error when an instruction to select a vocabulary other than the first vocabulary is input to the output vocabulary. Item 38. A computer-readable storage medium according to Item 37. 45. The computer-readable storage medium according to claim 44, wherein the instruction is input via a network. 46. The computer according to claim 41, wherein a control program for changing a process of recognizing the voice according to the identification information of the terminal that has transmitted the voice via the network is stored. Storage medium readable by. 47. The computer according to claim 44, wherein a control program for changing a parameter used for recognizing the voice according to the identification information of the terminal that has transmitted the voice via the network is stored. Storage medium readable by. 48. The apparatus according to claim 35, wherein identification information of a terminal connectable via the network and a control program for reading and using information relating to speech recognition suitable for the terminal are stored. Computer readable storage medium. 49. The computer-readable storage medium according to claim 46, wherein the processing at the time of recognition is changed according to the stored information. 50. The computer-readable storage medium according to claim 47, wherein the parameter at the time of recognition is changed according to the stored information. 51. When the identification information of the terminal that has transmitted the voice via the network is not stored in advance,
47. The computer-readable storage medium according to claim 46, wherein a control program for newly registering the identification information and information relating to speech recognition suitable for the terminal is stored.