JP2646539B2 - Standard pattern storage section management method - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION [Summary] In a speech recognition apparatus for recognizing a specific speaker's voice word, a voice pattern is used for correcting and registering a standard pattern of a voice word stored in a standard pattern storage unit. In the word recognition operation, the standard pattern of the voice word is automatically corrected and registered at the same time, so that the burden on the user can be reduced and the recognition rate is increased.

[Industrial applications]

The present invention relates to an improvement in a standard pattern storage unit management method for managing correction and registration of a standard pattern of a voice word stored in a standard pattern storage unit in a voice recognition device that recognizes a specific speaker's voice word.

In the standard pattern storage unit management method, it is desirable that the burden on the user is small and the recognition rate is high.

[Conventional technology]

 Hereinafter, a conventional example will be described with reference to the drawings.

 FIG. 6 is a block diagram of a conventional speech recognition apparatus.

In FIG. 6, when a voice word is registered, a registration mode is set from the keyboard 11 and a voice word is input from the microphone 5. The voice word is converted into a digital signal by the A / D converter 6, and the characteristic parameter is extracted by the feature parameter extractor 7. The characteristic parameters are extracted by a method represented by a filter bank or the like, and the identification code for the speech word from the keyboard 11 (for example, a character code of Fuji or the like indicating the speech word, etc.) and the features of the standard pattern storage unit 12 '. It is written and registered at a new location indicated by the pointer in the parameter field 13 and the identification code field 14.

Since the characteristic parameters of the human voice change with aging, it must be corrected.

If you want to make corrections, enter the correction mode using the keyboard 11,
When an identification code is input from the keyboard 11 and a speech word is input, the feature parameters of the speech word are extracted in the same manner as described above, and the feature parameters of the standard pattern storage unit 12 'corresponding to the identification code are extracted. It is modified to this feature parameter.

When performing voice recognition, the recognition mode is set from the keyboard 11, and when a voice word is input from the microphone 5, the A / D converter 6
Is converted into a digital signal, and the characteristic parameter is extracted by the characteristic parameter extraction unit 7. The characteristic parameter comparison unit 8 stores the extracted characteristic parameter and the standard pattern by a method represented by DP matching or the like. Part 1
The distance from the characteristic parameter in the characteristic parameter column 13 of 2 ′ is calculated, and the identification code of the shortest distance (the one with the highest similarity) is output from the identification code 14 of the standard pattern storage unit 12 ′, This is displayed on the display unit 10 via the result display processing unit 17 and transmitted to a necessary place.

[Problems to be solved by the invention]

Conventionally, however, the correction and registration of the standard pattern in the standard pattern storage unit are performed in the correction mode and the registration mode, respectively, in which the user inputs a speech word. There is a problem that the burden on the user is large to perform, and a high recognition rate is required.

[Means for solving the problem]

As shown in the principle flowchart of the present invention in FIG. 1, the above problem is compared in step 1 with the feature parameters of the input speech word and the feature parameters of the speech word stored in the standard pattern storage unit, In step 2, it is determined whether or not there is a corresponding word. When a voice word is recognized, if there is a corresponding feature parameter, in step 3, it is recognized as a voice word stored in the standard pattern storage unit, and the corresponding feature is recognized. The parameters are corrected to the average value of the feature parameter of the input speech word and the corresponding feature parameter, and if there is no corresponding feature parameter, the feature parameter of the newly input speech word is stored in step 4 in the standard pattern temporary storage unit. And the same speech word is re-input in step 5, and the feature parameter of the speech word is stored in the standard pattern temporary storage unit. When it is determined that the parameter is close to the parameter, the problem is solved by the standard pattern storage management method of the present invention in which the average value of the input characteristic parameter and the characteristic parameter of the standard pattern temporary storage is additionally registered as a new characteristic parameter. .

[Action]

According to the present invention, the correction and registration of the standard pattern in the standard pattern storage section are performed in parallel with the recognition of the speech word, so that the burden on the user for managing the standard pattern storage section can be reduced and the recognition rate can be reduced. Can be increased.

〔Example〕

 Hereinafter, embodiments of the present invention will be described with reference to the drawings.

FIG. 2 is a block diagram of the speech recognition apparatus according to the embodiment of the present invention, and FIG. 3 is a flowchart of a program according to the embodiment of the present invention.

In the following, at the time of speech word recognition, in parallel with the automatic correction and registration of the standard pattern in the standard pattern storage,
The description will be given with reference to FIG. 2 and the steps in FIG.

In step 20, when a speech word is input from the microphone 5, it is converted into a digital signal by the A / D converter 6 and input to the feature parameter extractor 7. The feature parameters are extracted in step 21. The feature parameter comparison unit 8 calculates the distance between the extracted feature parameter and the feature parameter in the feature parameter column 13 of the standard pattern storage unit 12, and in step 23, there is a feature parameter within a predetermined value. If not, the standard pattern management unit 9 stores the characteristic parameter in the standard pattern storage temporary data area 15 in the standard pattern storage unit 12 in step 24, and indicates that there is no feature parameter in the display unit 10. Display it and ask them to re-enter the spoken word.

In step 23, if there is a feature parameter within a predetermined value, the process proceeds to step 25. If this is the standard pattern storage temporary data area 15, it is displayed on the display unit 10. Is input from the keyboard 11.

Then, the standard pattern management section 9 compares the identification code in the standard pattern storage section 12 with the identification code in step 27 and determines whether or not the same code exists in step 28. In the storage temporary data area 15, an average of the input characteristic parameter and the characteristic parameter of the currently input voice word is used as a characteristic parameter as a standard pattern together with the identification code of the currently input voice word as a standard pattern storage unit. 12 feature parameter columns 1
3. The information is added and registered in the identification code column 14, and the display on the display unit 10 causes the process to return to step 20, where the speech word is re-input.

In step 28, if there is the same code, in step 30, the characteristic parameters of the identification code are stored in the standard pattern storage temporary data area 15 with the characteristic parameters of the currently input speech word. Is replaced with an averaged value, and the display on the display unit 10 returns to step 20 to re-input the speech word.

In step 25, the feature parameters within a predetermined distance are
If it is not in the standard pattern storage temporary data area 15 but in the characteristic parameter column 13 of the standard pattern storage section 12, the standard pattern management section 9 displays the identification code on the display section 10 in step 31. indicate.

In step 32, it is determined whether or not the displayed identification code is a desired identification code. If the identification code is not the desired identification code, the process proceeds to step 24, where the characteristic parameter is stored in the standard pattern storage temporary storage in the standard pattern storage unit 12. The data is stored in the data area 15 and the display unit 10 indicates that there is no data.
Have them enter the spoken word again.

For the input speech word or the re-input speech word, proceed to Steps 20 to 25, and proceed to Steps 31 and 32, and discriminate in Step 32. If desired, store the inputted speech word in the standard pattern. Recognize the voice word of the standard pattern stored in the unit 12 and send the identification code to a necessary place.In step 33, the feature parameter of the voice word just input and the registered feature parameter are The registered feature parameters are corrected to the averaged value, and the operation ends.

By doing so, the correction and registration of the standard pattern in the standard pattern storage unit 12 are automatically performed at the time of speech word recognition, and there is no need to particularly correct and register, so that the burden on the user can be reduced. And the recognition rate can be increased.

 Next, application examples of the present invention will be described below.

FIG. 4 is a block diagram of a main part of a voice recognition telephone device according to an application example of the present invention, and FIG. 5 is a flowchart showing a method of managing a standard pattern storage unit in the case of FIG.

FIGS. 4 and 5 show that Mr. Fuji's telephone number is 111, for example.
In this case, when a speech word Fuji is input, a dial signal of 111 is output and the speech recognition telephone device is used.

The ROM 50 in FIG. 4 stores a program corresponding to the program described in FIG. 3, and as shown in FIG. 5A, the standard pattern storage unit 12 (in FIG. ), The characteristic parameter column 13 shows the characteristic parameter A
(Fuji's feature parameters of voice words) to E are registered, and corresponding to these feature parameters, telephone numbers 111 to 115 are assumed to have been already written from the keyboard 47 in the identification code column 14. Enter the voice word Fuji 111
The case where the number dial signal is transmitted will be described below.

When the handset 40 shown in FIG. 4 is off-hook, the hook switch 42 is turned off, and this fact is transmitted to the processor 49 via the decoder 46.

Then, under the control of the switch control circuit 44, the connection is established as shown in FIG. Then, when the speech word Fuji is input from the handset 40, this is converted into a digital signal by the A / D converter 43 and input, and the characteristic parameter A 'is obtained as described above.

The characteristic parameter A 'is compared with the characteristic parameters A to E in the characteristic parameter column 13 of the standard pattern storage unit 12. If A has a high degree of similarity, for example, the identification code is displayed on the display unit 48 in FIG. The telephone number 111 in the column 14 is displayed.

If this is correct, when transmission is instructed from the keyboard 47, the dial pulse generator 45 transmits the 111th dial pulse to the transmission line.

At this time, as shown in FIG.
Then, a characteristic parameter A ″ is obtained and A ″ is stored in place of the characteristic parameter A as shown in FIG.

If the feature parameter of the input speech word becomes F and the content of the identification code displayed on the display unit 48 is incorrect,
The user goes on-hook the handset 40.

When on-hook, the feature parameter is F at this time,
If it is not in the characteristic parameter column 13, it is stored in the standard pattern storage temporary data area 15 as shown in FIG.

When the voice word is input again from the handset 40, the feature parameter is F 'and the standard pattern storage temporary data area 15 is set.
If the similarity degree is the closest among the characteristic parameters F in this case, it is determined that this voice word has not been registered, and this is displayed on the display unit 48, and the telephone number 116 is input from the keyboard 47. .

If the telephone number 116 is not found in the identification code column 14, if not, the average of the characteristic parameter F in the standard pattern storage temporary data area 15 and the characteristic parameter F 'of the voice word just input is obtained. F "and the telephone number 116 are newly registered as shown in FIG. 5 (E).

If the telephone number input from the keyboard 47 is, for example, 111, and is the same as the telephone number 111 in the identification code column 14,
The averaged feature parameter F ″ is replaced with the feature parameter A as shown in FIG. 5D as the latest feature parameter.

That is, as described above, the standard pattern in the standard pattern storage unit 12 is automatically corrected or registered at the time of voice recognition.

〔The invention's effect〕

As described above in detail, according to the present invention, at the time of speech word recognition, correction of the standard pattern in the standard pattern storage unit,
Since the registration is automatically performed, it is not necessary to perform the correction and the registration, so that the burden on the user can be reduced and the recognition rate can be increased.

[Brief description of the drawings]

1 is a flowchart of the principle of the present invention, FIG. 2 is a block diagram of a speech recognition apparatus according to an embodiment of the present invention, FIG. 3 is a flowchart of a program of the embodiment of the present invention, and FIG. 5 is a flowchart showing a standard pattern storage unit management method in the case of FIG. 4, and FIG. 6 is a block diagram of a conventional speech recognition apparatus. In the figure, 1-4, 20-33 are program steps, 5 is a microphone, 6 is an A / D converter, 7 is a feature parameter extractor, 8 is a feature parameter comparator, 9 is a standard pattern manager, 10, 48 is a display unit, 11 is a keyboard, 12 and 12 'are standard pattern storage units, 13 is a feature parameter column, 14 is an identification code column, 15 is a temporary data storage area for standard patterns, and 17 is a result display processing unit.

 ──────────────────────────────────────────────────続 き Continuation of the front page (72) Inventor Toshihiro Asami 1015 Uedanaka, Nakahara-ku, Kawasaki City Inside Fujitsu Limited (56) References JP-A-60-60697 (JP, A) JP-A-60-207197 (JP) , A) JP-A-60-73592 (JP, A) JP-A-59-170899 (JP, A)

Claims (1)

(57) [Claims] When a speech word of a specific speaker is inputted, the feature parameters of the speech word are compared with the feature parameters of the speech word stored in the standard pattern storage unit (1).
If there is a corresponding feature parameter in the speech recognition device that recognizes the presence or absence of a corresponding word (2) and recognizes a voice word, it is recognized as a voice word stored in the standard pattern storage unit, and the corresponding feature parameter is determined. The average value of the feature parameter of the input speech word and the corresponding feature parameter is corrected (3). If there is no corresponding feature parameter, the feature parameter of the newly input speech word is stored in the standard pattern temporary storage unit. (4) When the same voice word is re-input, and it is determined that the feature parameter of the voice word is close to the feature parameter stored in the standard pattern temporary storage unit, the input feature parameter and the standard pattern temporary storage unit The average value of the feature parameter is additionally registered as a new feature parameter (5).
A standard pattern storage management method, characterized in that: