JP2002132288A - Enroll sentence voice input method, enroll sentence voice input device, and recording medium storing program for realizing the same - Google Patents

Abstract

(57) [Summary] [Problem] To register a feature of a speaker's voice and customize it for an individual in a voice recognition device, a large amount of sentences registered in advance in the voice recognition device is used by the speaker. It is necessary to perform the work of enrolling to input voice to the user. Even if the speaker inputs a large amount of sentences by voice, it is very troublesome work. However, if the speaker cannot collate well with the acoustic model at the time of enrollment, it is necessary to re-input a part that cannot be collated. There is no problem if you can input the voice again and collate immediately.
There was a problem that the collation could not be easily performed even if it was entered many times, and the enrollment work did not proceed and had to be interrupted sometimes. In order to solve the above-mentioned problem, if the input speech cannot be collated even after repeating the voice input more than a predetermined number of times, the speech recognition apparatus proceeds to input a sentence next.
Prevent re-input processing from being performed any more, or re-input sentences that cannot be collated at the end. In the case of re-input, only the portion that cannot be verified is re-input, or the user is re-input a predetermined range including the portion that cannot be verified.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to an enroll sentence voice inputting method and an enrolled sentence voice input device for inputting an enroll sentence for registering a personal voice for personal customization in a voice collating device. And a recording medium on which a computer program for realizing the same is recorded.

[0002]

2. Description of the Related Art An ordinary speech recognition apparatus holds speech data for collating an unspecified number of speeches. Therefore, in order to be able to accurately collate the voice of a specific individual, it is necessary to register the voice data of the individual using an enroll sentence voice input device attached to the voice recognition device. However, in order to register personal voice data and customize the personal data to improve the accuracy of voice recognition, it is necessary to read a large amount of sentences and input the voice.

Usually, an enroll speech input device is a part of a speech recognition device, and software for enroll speech input includes data of an enroll sentence to be input, and is used together with the speech recognition software and data. It has been incorporated.

[0004] A voice recognition device used by an individual usually has an enroll function in order to customize it for an individual. When an enroll sentence is input, an enroll sentence to be input as a voice on a screen is displayed. The sentence is displayed one by one, and the user looks at the displayed enrollment sentence and performs voice input by reading it out. However, the reading operation of the enroll sentence usually takes 20 to 30 minutes, and is very troublesome for the user. In many cases, speech input is often given up on the way, and the accuracy of speech recognition often does not increase much depending on the amount and content of the acquired personal speech data. For this reason, the software for inputting enroll sentences is equipped not only with the function of inputting detailed enroll sentences, but also with a simplified version of the enroll sentence voice input function, which is particularly useful for improving the speech recognition of enroll sentences to be input. In many cases, the number of enroll sentences to be input is suppressed.

As described above, the work of inputting the enroll sentence is very troublesome for the user. In addition, the enroll sentence is input, and there is a part where the input sound cannot be matched with the acoustic model. However, there is a problem that the user may be prompted to re-enter a part that cannot be collated, but may enter a state where the collation is not successful even if input is repeated many times and the process cannot proceed.

[0006] When the input speech cannot be collated, it often occurs at the end of the enroll sentence.
For example, if you enter an enroll sentence that says "I'm going to school," enter the last part of "Ru" or "Iru" and input only a part of the word, Therefore, the pronunciation was unnatural, and it was often difficult to collate.

Whether the input enroll statement can be compared and where the collation cannot be performed can be displayed by changing the color of the collated part of the enroll statement or the color of the uncollated part,
In general, this is done by pointing the part to be re-input with a pointer or the like.

When the input of the detailed enroll sentence is stopped halfway, or when a simplified version of the enroll sentence voice input function is used, sufficient personal voice data cannot be collected. In comparison, the speech recognition accuracy is lower.

[0009] The tendency is more remarkable for a person whose pronunciation is inaccurate, a person who has an accent, or a person who has a habit of pronunciation. Therefore, in a personal voice recognition device or the like, it is preferable to input all the enroll sentences as much as possible in order to increase the voice recognition rate as much as possible.

[0010]

As described above, if the input amount of the enrolled sentence is reduced, the speech recognition rate does not increase so much. Therefore, it is necessary to keep the input amount of the enrolled sentence as small as possible.

In the conventional method, when the input voice is not collated at the time of inputting the voice of the enroll sentence, a keyboard or a mouse or the like is used to indicate a re-input location which is easy for the user to input or a location which cannot be verified by himself. , It is necessary to instruct the user to move on to input of the next enroll sentence or to instruct to interrupt the input of the enroll sentence. In addition, if the part of the enroll sentence that cannot be collated is skipped and the processing of the next enroll sentence is skipped, it is not possible to re-enter the skipped enroll sentence after the voice input of the enroll sentence has been completed. was there.

[0012]

SUMMARY OF THE INVENTION Briefly stated, the present invention displays an appropriate start position for automatically resuming input when re-inputting an unverifiable portion which occurs during enroll sentence voice input. Or to reduce the number of re-entry for the same part to a predetermined number, or to re-enter the enroll sentence to be re-entered at the end so that the input of the enroll sentence does not feel too troublesome for the user. It is.

[0013] That is, according to the enroll sentence voice input method according to the first aspect of the present invention, a plurality of enroll sentences registered in advance are displayed in a speech recognition system.
A method of inputting a voice of a user who has read out the enrolled sentence, collating the input voice with a pre-registered acoustic model, and voice-inputting an enrolled sentence for registering voice data of the verified user. If there is a part that cannot be collated in the input speech, re-input of the enroll sentence is repeated a predetermined number of times, and if there is still a part that cannot be collated, the process proceeds to the next enroll sentence input processing. It is characterized by the following.

With this configuration, when the enroll sentence is read out and the user's voice is registered, if the enroll sentence cannot be collated, the re-input process can be automatically repeated a predetermined number of times.

According to a second aspect of the present invention, there is provided an enroll sentence speech inputting method for inputting a plurality of enrolled sentences registered in advance in a speech recognition system. When enroll statements that cannot be collated remain after the enroll statement has been input, only the enroll statements that could not be collated
The re-input instruction is performed a predetermined number of times.

With this configuration, when the enroll sentence is read aloud and the voice of the user is registered, when the enroll sentence has been completely input, the enrolled sentence that has been skipped because it cannot be collated is automatically re-input. Then, the process returns to the enroll sentence input process, and the re-input can be automatically performed a predetermined number of times again.

Further, in the enroll sentence voice input method according to the third aspect of the present invention, further, according to the position in the enroll sentence of the unverifiable portion, a second predetermined position from the first predetermined position including the portion is determined. A re-input instruction up to a predetermined position is presented.

With such a configuration, when a part of the enroll sentence that cannot be collated is a part of a word, even if that part is re-entered, the pronunciation becomes unnatural and speech recognition cannot be performed normally, In order to prevent unnatural voice data from being stored, it is possible to re-input a part of a predetermined range including a part that cannot be verified, so that a natural easy-to-verify voice input can be performed.

A computer-readable recording medium on which an enroll sentence voice input device and a program for inputting enroll sentence voice according to the present invention are recorded.
A computer-readable recording medium which records an enroll sentence voice input device using the methods of (1) to (3) and a program for inputting the enroll sentence voice, and which operates the enroll sentence voice input device or on the recording medium. By reading the program into a computer and executing the program, the same effect as when the method of claims 1 to 3 is executed can be obtained.

[0020]

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS (Embodiment 1) This embodiment is an example of an enroll sentence voice input device realized by using the enroll sentence voice input method according to claim 1 and claim 2. As long as the enroll sentence voice input device is realized by using the enroll sentence voice input method of item 2, it is not limited to the enroll sentence voice input device described here.

With respect to the basic processing of the present invention, the configuration and the flow of the processing will be described with reference to the configuration diagram of FIG. 1 and the flowchart of FIG.

Here, when the enrolled sentence is input by voice, if there is a part that cannot be collated in the input enrolled sentence, the voice input of the enrolled sentence including the part is repeated a predetermined number of times n times. If it is not possible, the process automatically proceeds to the input processing of the next enroll sentence.
If there is an enroll sentence that leaves a part that cannot be collated, re-input the enroll sentence again, and if the collation cannot be performed even after repeating the voice input a predetermined number of times m, processing of the enroll sentence will no longer be performed explain. Here, in order to explain the basic skeleton part of the present invention, n = m =
Set to 0. In other words, even if there is a part that cannot be collated when the enroll sentence is input by voice, the re-entry is not performed at that point, and the process proceeds to the next enroll sentence. An example will be described in which only the enroll sentence remaining without being re-entered is input.

First, the enroll sentence display unit 1 reads one enroll sentence from the enroll sentence storage unit 1-1 in which a plurality of enroll sentences are stored in advance (step S).
1). The example of the enroll sentence storage unit in FIG. 3 is an example of the enroll sentence stored in the enroll sentence storage unit 1-1. Then, it is determined whether reading of all enroll statements has been completed (step S2). If all the processing of the enroll statement has not been completed (step S2: NO),
The enroll statement is displayed on the display device 7 (step S
3).

A user who wants to register his voice,
The enroll sentence displayed on the display device 7 is read aloud, and voice is input from the voice input device 8 such as a microphone through the voice input unit 2 (step S4). The input speech is subjected to A / D conversion (analog / digital conversion), and the enrolled sentence matching unit 3 checks the input speech with the acoustic model in the acoustic model storage unit 3-1 (step S5). The A / D conversion may be performed by the voice input unit 2 or the enroll sentence matching unit 3.

Since the enrolled sentence is a sentence, voice recognition is performed by using a continuous voice recognition technique such as a hidden Markov model. The method of continuous speech recognition is not particularly limited, and any method may be used.

The enroll sentence collation unit 3 stores the collated speech data in the personal speech data storage unit 3-2 (step S6). When performing speech recognition by continuous speech recognition by inputting an enroll sentence by voice, it is not always possible to collate the entire sentence input at once. In some cases, the collation cannot be performed in the middle of a sentence. In such a case, the collation pointer management unit 4 stores the collation pointer, which is information of the enroll statement that could not be collated, based on the collation result in the enroll storage unit 1-.
1 and passed to the enroll sentence redisplay unit 5 at the same time.

The enroll sentence redisplay unit 5 highlights the collated portion of the enroll sentence input by voice (step S8). Normally, the part of the enroll sentence displayed for reading out is highlighted by changing the color of the collated part of the enrolled sentence or displaying it in bold characters. In this embodiment, even if there is a part that cannot be collated, the process automatically proceeds to the input of the next enroll sentence, so that the process returns to step S1.

The method of highlighting is not to change the color of the collated portion but to display the uncollated portion in a conspicuous color such as red, to change the color, or to display a high brightness. , Blinking display, bold characters, and the start and end positions of the points that cannot be collated and need to be re-entered may be displayed with a pointer.
The display method is not particularly limited. If there is no part that cannot be collated, the entire enroll sentence may be highlighted or the processing of the next enroll sentence may be performed, and the display of the pointer is unnecessary.

Further, the redisplay of the enroll sentence may be performed by overwriting the enroll sentence displayed by the enroll sentence display unit 1, or may be displayed on the next line. The method is not limited.

In the display example of the re-input instruction shown in FIG. 4, when an enroll sentence “Today is fine weather” is input by speech, and when the end of the sentence “Ne” cannot be recognized by speech, the collation can be performed. This is an example in which a pointer is displayed on the position of “ne” that did not exist, and an example in which the character of “ne” that could not be collated was made bold.

Thus, the user is able to understand at a glance the part that could not be collated, and when reading out the enroll sentence thereafter, if there is a part similar to the part that could not be collated, it is read carefully. It is possible to improve the input efficiency of the enroll sentence.

When the enroll statement ends (step S
2: YES), the enroll sentence re-input determination unit 6 determines whether there is an enroll sentence that could not be compared (step S11).

Whether or not there is an enroll statement that could not be collated can be determined by setting an error flag when an enroll statement that could not be collated occurs. The collation pointer management unit 4 manages the collation pointer indicating which enroll statement actually caused the collation error. In this embodiment,
The information is stored in the enroll sentence storage unit 1-1 by the collation pointer management unit 4.

As the format in which the data is stored, for example, a flag may be set at the head of each enroll statement as in the example of the enroll statement storage section in FIG. In addition, for which part of the enroll statement the collation error occurred is determined, for example, by setting a flag at the head of each enroll statement as in the example of the enroll statement storage unit in FIG. And the end position may be stored. Here, the start position and the end position are indicated not by the character of the enroll statement but by the mora of the reading of the enroll statement. The mora refers to each of the 50 sounds of "A", "I",... "N", and "kya", "kyu", "kyo", "tsu", etc. are counted as one mora.

Here, the reason why the position is not represented by the character of the enroll sentence is that there is a case where a part that cannot be collated is in the middle of reading the kanji of the displayed enroll sentence. . For example, as shown in Fig. 4, it is possible to match up to "school" of "going to school"
If "c" cannot be collated, a pointer may be displayed at the "school" character in "school", but the "ko" sound in "school" can be collated, so re-enter. No need. In order to clarify the part that can not be collated, display "Gacuni Unitail" under "Going to school" to indicate with a pointer that the part of "U" can not be collated, or use bold characters Good to do. Of course, two enroll sentences and their readings may be displayed simultaneously. in this case,
In the enroll sentence storage unit 1-1, it is necessary to store the enroll sentence containing kanji and the reading of the enroll sentence as shown in FIG. 3 so that the mora can be clearly recognized.

Further, these information may be separately stored as a table on a memory without being stored in the enroll sentence storage unit 1-1, and the way of managing the information is not particularly limited. By doing so, when the enroll sentence is displayed to re-enter the enroll sentence, the enroll sentence to be re-input can be determined by the flag, and at the same time, the part that could not be collated can be displayed from the beginning.

If there is no enroll sentence that could not be compared (step S11: NO), the process ends. If there is an enroll statement that could not be compared (step S1
1: YES) instructs the enroll sentence display unit 1 to re-input the enroll sentence that could not be collated, and the enroll sentence display unit 1 skips the enroll sentence that could be collated, and reads only the enroll sentence that could not be collated. (Step S1
2).

The operation of the enroll sentence collation unit 3 at the time of re-input is that if the position information of the uncollated part is not stored in the enroll sentence storage unit 1-1, the entire enroll sentence is re-entered. In other words, the entire enrollment sentence will be collated again, but if the location information of the part that could not be collated is stored, the part of the enrolled sentence including the part that could not be collated can be re-entered. , Only that part needs to be checked again.

Next, it is determined whether reading of all enroll statements that could not be collated has been completed (step S1).
3). When all the enroll statements that could not be collated have not been processed (step S13: NO), the processes from step S14 to step S19 are performed in step S19 except that the return destination of step S19 is step S12.
Since the processing is the same as the processing from step 3 to step S8, the description is omitted.

In this example, the re-input of the enroll sentence that could not be collated is performed after the enroll sentence has been completely input, and the enroll sentence that could not be collated is re-entered. Although the processing of the sentence is not performed, the number of times that the enroll sentence that could not be collated is read in order from the beginning is set to a predetermined number k, and the processing from step S11 to step S19 is performed. It may be repeated a predetermined number of times.

In this embodiment, the maximum number of re-inputs of an enroll statement is determined by the values of n, m, and k. No further processing is performed.

With this configuration, when the enrolled sentence is read out and the user's voice is registered, if the enrolled sentence cannot be collated, the processing has been manually shifted to the next enrolled sentence by a key operation or the like. It was possible to move automatically, and for enroll statements that skipped processing without being able to collate, until now the enroll statement was completely input, collation could be performed again. Although it was not possible to re-input only the enrolled sentence that did not exist, it is possible to return to the input processing of the enrolled sentence that could not be automatically collated.

(Embodiment 2) This embodiment is an example different from Embodiment 1 of the enroll sentence voice input device realized by using the enroll sentence voice input method of the first and second aspects. Here, when there is a part that cannot be matched in the enroll sentence that is input when the enroll sentence is input by voice, the voice input of the part that cannot be verified is performed a predetermined number of times n (n> = 1)
Repeatedly, if collation is still not possible, the process automatically proceeds to the input processing of the next enroll sentence. Then, the voice of the enroll sentence is input again, and the voice input is performed a predetermined number of times m (m> =
1) An example will be described in which if the collation cannot be performed even after repeated times, the processing of the enroll sentence is no longer performed.

The difference from the first embodiment is that when the enroll sentence cannot be collated in the first embodiment, the process immediately proceeds to the input of the next enroll sentence. At this point, the input of the part that cannot be collated in the enroll sentence is repeated a predetermined number of times, and the collation pointer information is stored not in the enroll sentence storage unit 1-1 but in the collation pointer storage unit 4-1.

The configuration diagram and processing flow of the present invention will be described with reference to the configuration diagram of FIG. 5 and the flowchart of FIG.

Regarding the configuration diagram of FIG. 5, the same processing blocks as those of the configuration diagram of FIG. 1 are denoted by the same reference numerals, and description thereof will be omitted. 6, the same processing steps as those in the flowchart of FIG. 2 are assigned the same step names, and description thereof will be omitted.

Steps S1 to S5 are the same as those of the first embodiment except that enroll sentence collation is performed on the re-input speech based on the information of the collation pointer storage unit 4-1, so that the description is omitted.

The enroll sentence matching unit 3 compares the voice data with the acoustic model stored in the acoustic model storage unit 3-1, and converts the user's voice data of the matched portion into the personal voice data storage unit 3.
-2 (step S6). When one enroll sentence can be compared only halfway, only the part that could be compared may be stored in the personal voice data storage unit 3-2, or stored only when the entire enroll sentence is compared. However, when an enroll sentence having a part that cannot be collated occurs, the former is more preferable because more personal voice data can be collected.

The collation pointer management unit 4 receives, from the enroll sentence collation unit 3, information as to where and how much collation was possible and what part was unsuccessful in the form of a character string or a character position / number of mora. The position information or the position information of the part that could not be collated is obtained, and the collation pointer information is stored in the collation pointer storage unit 4-1 (step S7). In the present embodiment, the collation pointer information is provided and stored in the collation pointer storage unit 4-1.
1 may be stored together with the enroll statement.

FIG. 7 shows an example of the collation pointer storage section 4-1. There is an enroll statement as shown on the left of FIG.
The following describes what happens to the contents of the collation pointer storage unit 4-1 when there is a part that cannot be collated in the second and nth enroll statements. FIG. 7 shows a sentence number indicating the number of the enroll sentence stored in the enroll sentence storage unit 1-1, the start position and the end position of the part that could not be compared. is there.

The enroll sentence redisplay unit 5 obtains the position information of the collated part from the collation pointer storage unit 4-1.
The collated part or the collated part is highlighted (step S8).

When the enroll sentence re-display unit 5 highlights the display, the control is transferred to the enroll sentence re-input determination unit 5. The enroll sentence re-input determination unit 5 determines whether there is a part that cannot be collated in the enroll sentence (step S9), and if there is no part that cannot be collated (step S9: N).
O) returns to step S1 to move to the processing of the next enroll statement. If there is a part that cannot be collated (step S9: YES), it is determined whether or not the re-entry of the enroll sentence has been repeated a predetermined number of times n or more (step S10). If it has not been repeated more than once (step S10: YES), step S4
Then, the enroll sentence is read out from the position of the collation pointer indicating the re-entry start position of the enrolled sentence, and voice input is performed again. If the re-input is repeated a predetermined number of times n or more (step S10: YES), step S1
Returns to the processing of the next enroll statement.

At the time of re-inputting the enroll sentence, the enroll sentence collation unit 3 preferably performs collation from the collation pointer displayed on the enroll sentence redisplay unit 5. Again, it is possible to re-check the entire enrolled sentence, but this is not very desirable because the already-checked part may not be checked again this time.

When the input of the enroll statement has been completed (step S2: YES), it is determined whether or not the enroll statement which could not be collated remains, and whether or not the collation pointer information remains in the collation pointer storage unit 4-1 (step S2). Step S1
1), if enroll statements that could not be collated remain (step S11: YES), enroll statements that could not be collated are read from the enroll statement storage unit 1-1 (step S12). The collation pointer storage unit 4-1 stores information indicating the position of the enroll sentence to which collation sentence was reached, and when there is no uncollated part of the enrolled sentence, the collation pointer management unit 4 returns to the enrolled sentence. The collation pointer information of the sentence is stored in the collation pointer storage unit 4-1.
Remove from. If there are no enroll statements that could not be collated (step S11: NO), the process ends.

Steps S14 to S2 when the enroll statement can be read normally (step S13: NO)
The process 1 is the same as the processes of steps S3 to S11 except that the process ends when the reading of the enroll statement ends, and the return destination is the position of the corresponding step, and thus the description is omitted.

Although not provided in the present embodiment, a function for manually interrupting the processing by a key operation or the like or moving to the processing of the next enroll statement may be provided as necessary.

With this configuration, when the enroll sentence is read out and the voice of the user is registered, the part that cannot be compared with the enrolled sentence can be automatically re-entered up to a predetermined number of times. If you enter the number of times again, even if there is a part that cannot be collated, you can automatically proceed to the processing of the next enroll statement. The skipped enroll sentence can be automatically re-entered up to a predetermined number of times again, and the number of repetitions can be set so as not to dislike repetition of voice input for the same enroll sentence.

(Embodiment 3) In the first and second embodiments, the position of a part that cannot be collated is indicated as a re-input start position by a pointer, and the enroll sentence is read out again from the position of the pointer and speech input is performed. Was. But,
Even if you re-enter only a part of the enroll sentence by voice,
In many cases, the collation could not be easily performed, so that the user was unable to proceed even after inputting many times, and the input of the enroll sentence ended without being able to perform the collation. Therefore, in the present embodiment, depending on the position in the enroll sentence where a part that cannot be collated using the method of claim 3 is located, a voice is reproduced from a predetermined start position to a predetermined end position, including a part that cannot be collated. An example of an instruction to input will be described.

The configuration diagram and the processing flow of the present embodiment are as follows. In the highlighting of the enroll sentence in steps S8 and S18, the collated part and the uncollable part are displayed in different colors, highlighted, bold characters, and pointers. In this case, the process is the same as that of the second embodiment except that the system specifies not only a part that cannot be collated but also a range in which it is preferable to re-input the data so that the system can easily collate. This processing is performed in steps S7 / S8 and S18 / step S1 of the first and second embodiments.
9 is included in the process of calculating the collation pointer and highlighting the enroll sentence.

The collation pointer management unit 4 first collates the audio data with the acoustic model from the enroll sentence collation unit 3,
In the present embodiment, the position information of the collated part and the uncollated part is received in the form of the number of moras. Next, the collation pointer management unit 4 refers to the reading of the enroll sentence in the enroll sentence storage unit 1-1 that has already been read in the enroll sentence display unit 1, and refers to the phrase one phrase before the start position of the part that cannot be collated. The position of one position after the position and the end position is obtained and stored in the collation pointer storage unit 4-1. The enroll sentence redisplay unit 5 refers to the collation pointer in the collation pointer storage unit 4-1 and displays an input range with a pointer or the like so as to input from one sentence before to one sentence after.

If the unmatchable part is within a predetermined range from the beginning of the sentence (for example, within 4 moras from the beginning of the sentence), the input is made up to one phrase after the unmatchable part from the beginning of the sentence. If the unmatchable part is within a predetermined range from the end of the sentence (for example, within 4 moras from the end of the sentence), the input is made from one phrase before the end of the unmatchable part to the end of the sentence.

FIG. 4 shows the case where the character “Ru” of “Going to school” cannot be collated, the characters that cannot be collated are displayed in bold characters, and the input range is indicated by a pointer so that the character can be entered one sentence before. This is an example. FIG. 7 shows the collation pointer stored in the collation pointer storage unit 4-1 when there is a part that cannot be collated in the second and nth enroll statements when there is an enroll statement as shown on the left side of FIG. Is an example,
And the start position and the end position of the re-input position are additionally stored. Here, the start position and the end position of the part that could not be compared and the start position and the end position of the re-input position are represented by the number of mora from the head. When the collation pointer is not stored in the collation pointer storage unit 4-1 but is stored in the enroll sentence storage unit 1-1, the result is as shown in FIG.

Here, the range is specified by a phrase such as one phrase before or after one phrase that cannot be collated, but the number of phrases is not limited to one. In addition, the sentence may be delimited by punctuation marks instead of phrases (when there is no punctuation in the enroll sentence stored in the enroll sentence storage unit 1-1, it may be regarded that there is a punctuation at the end of the sentence). However, when punctuation is used, if the range delimited by the punctuation of the sentence of the enroll sentence is wide, the range of re-input may be long. The re-entry range may be determined not only by phrases and punctuation marks but also by other units.

The display method may be, for example, that the collated part is blue, the uncollated part is red, the part to be re-input is yellow, and the yellow and red parts are re-input. The part is blue, the part that cannot be matched is red, the start position and the end position of the part to be re-entered may be specified with a pointer, the part that can be matched is displayed in thin characters, the part that cannot be matched is displayed in bold, The start position and the end position of the part to be re-input may be designated by a pointer. Besides this,
You can change the combination of color, font and pointer,
The part to be re-entered may be displayed and specified together with a message instructing re-entry, and the method is not particularly limited. If there is no part that cannot be collated, the fact that the entire enrolled sentence has been collated has been highlighted to indicate to the user, and the display of the pointer is unnecessary.

This re-input method does not need to be performed every time re-input is performed. If only a part that cannot be collated is re-input and collation can be performed immediately, there is no need to use this method. Therefore, for example, if there is a part that cannot be collated by inputting the enroll sentence, the first re-entry re-enters only the part that cannot be collated, and if it is still not collatable, the collation is performed using the method described above. You may make it re-input the part of the predetermined range including the part which cannot be performed. In addition, the pointer indicating the re-input range may not be displayed, and the enroll sentence re-display unit 5 may output a predetermined message such as a message to reread from the previous phrase.

With such a configuration, when a part of the enroll sentence that cannot be collated is a part of a word, even if that part is re-entered, the pronunciation becomes unnatural and speech recognition cannot be performed normally, In order to prevent unnatural voice data from being stored, it is possible to re-input a part of a predetermined range including a part that cannot be verified, so that a natural easy voice input can be performed.

(Embodiment 4) Next, a recording medium in which a program for realizing an enrolled voice input device according to an embodiment of the present invention is described is a CD-ROM 9-like a recording medium shown in FIG. In addition to the portable recording medium 7 such as the floppy disk 1 or the floppy disk 9-2, any other storage device 10 stored at the end of a communication line or the recording medium 11 such as a hard disk or a RAM of the computer 12 may be used. When using the data display device 7 according to the invention, the data is read by the computer 12 and executed on the main memory.

The enroll sentence is read out by loading the program from the recording medium into the computer 12, and when the enroll sentence cannot be collated when registering the voice of the user, the enroll sentence has not been manually operated by a key operation or the like until now. In the process of processing the next enroll statement, it is now possible to move automatically, and for enroll statements skipped because they could not be collated, input of the enroll statement has been completed until now It was difficult to re-enter only the enroll statement that could not be collated again.However, it is possible to automatically return to the input processing of the enroll statement that could not be collated. It is possible to realize an enrollment voice input device that does not become uncomfortable due to repetition of voice input.

[0069]

According to the enroll sentence voice input device using the enroll sentence voice input method of the present invention, when the enroll sentence is read out and the user's voice is registered, when the enroll sentence cannot be collated, In the past, it was possible to automatically move to the processing of the next enroll sentence manually by key operation, etc., and for enroll statements that skipped processing without collation, Is
Once the enroll statement has been entered,
Although it was not possible to re-input only the enroll statement that could not be collated, it is possible to automatically return to the input processing of the enroll statement that could not be collated.

Further, according to the enroll sentence voice input device using the enroll sentence voice input method of the present invention, when the enroll sentence is read aloud and the voice of the user is registered, the portion where the enroll sentence cannot be collated is automatically processed. Re-entry is possible up to a predetermined number of times, and if the collation is completed or re-entered a predetermined number of times, even if there is a part that can not be collated, it is possible to automatically proceed to the next enroll sentence processing, When the enroll sentence has been completely entered, if the enroll sentence was skipped because it could not be matched, it can be automatically re-entered up to the specified number of times again, and voice input repeated for the same enroll sentence It is possible to set the number of repetitions to such an extent that the user does not dislike.

Further, according to the enroll sentence speech input device using the enroll sentence speech input method of the present invention, when an unrollable part of the enroll sentence is a part of a word,
Even if you re-input only that part, the pronunciation will be unnatural, and it will not be possible to recognize speech normally or to store unnatural speech data. It is possible to re-input a part in a predetermined range including a part that cannot be compared.

Further, according to the recording medium in which the program for realizing the enroll sentence voice input device of the present invention is described, when reading the enroll sentence and registering the voice of the user,
If the collation and collation of the enroll statement cannot be performed, the process has been manually shifted to the next enroll statement by key operation, etc. Until now, for enroll statements,
Once the enroll statement has been entered,
Although it was difficult to re-enter only the enroll sentence that could not be collated, it was possible to automatically return to the input processing of the enroll sentence that could not be collated, and voice input was repeated many times in a row with the same enroll sentence Thus, an enrolled voice input device that does not become unpleasant because it cannot be collated can be realized.

[Brief description of the drawings]

FIG. 1 is a configuration diagram of a first embodiment of the present invention.

FIG. 2 is a flowchart showing a flow of processing according to the first embodiment of the present invention;

FIG. 3 shows an example of an enroll statement storage unit

Fig. 4 Example of display of recognized and unrecognized parts

FIG. 5 is a configuration diagram of a second embodiment of the present invention.

FIG. 6 is a flowchart showing a processing flow according to the second embodiment of the present invention;

FIG. 7 shows an example of a collation pointer storage unit.

FIG. 8 is an exemplary diagram of a storage medium.

[Explanation of symbols]

1 Enroll sentence display section 1-1 Enroll sentence storage section 2 Voice input section 3 Enroll sentence collation section 3-1 Acoustic model storage section 3-2 Personal voice data storage section 4 Collation pointer management section 4-1 Collation pointer storage section 5 Enroll Sentence re-display unit 6 Enroll sentence re-input determination unit 7 Display device such as display 8 Voice input device such as microphone 9 Portable storage medium such as CD-ROM or floppy (registered trademark) disk 9-1 CD-ROM 9-2 Floppy disk 10 Other storage device stored at the end of communication line 11 Storage medium such as RAM / hard disk on computer 12 Computer

 ────────────────────────────────────────────────── ─── Continuing on the front page (72) Inventor Shoji Harada 4-1-1, Kamiodanaka, Nakahara-ku, Kawasaki-shi, Kanagawa F-term within Fujitsu Limited (reference) 5D015 AA01 AA04 BB01 GG03 GG06 LL04 LL05 LL10 LL10

Claims (5)

[Claims] In a voice recognition system, a plurality of enrolled sentences registered in advance are displayed, a voice of a user who reads out the enrolled sentence is input, and the input voice is input to an acoustic model registered in advance. This is a method of collating and inputting an enroll sentence for registering the collated user's voice data.If there is a part that cannot be collated in the input speech, repeat the re-input of the enroll sentence a predetermined number of times. A method for inputting enroll sentence speech, characterized in that if there is still a part that cannot be verified, the process proceeds to the next enroll sentence input process. 2. A method for inputting a plurality of enroll sentences registered in advance in a speech recognition system, wherein an enroll sentence that cannot be collated at the stage when the input of a plurality of enroll sentences to be inputted has been completed has been completed. An enroll sentence voice input method characterized in that, if there is a remaining, only a predetermined number of re-entry instructions are given for enroll sentences that could not be verified. 3. The enroll sentence voice input method according to claim 1 or 2, further comprising, in accordance with a position in the enroll sentence of a part that cannot be matched, from a first predetermined position including the part. A re-entry instruction up to a predetermined position is presented. 4. An audio collating apparatus, wherein: an enroll sentence storage unit for storing a plurality of enroll sentences, an enroll sentence display unit for displaying an enroll sentence to be input, a voice input unit for inputting the enroll sentence by voice, An enroll sentence voice input device comprising: an acoustic model storage unit that stores a model; an enroll sentence matching unit that matches the voice with the acoustic model; and a personal voice data storage unit that stores matched voice data. A collation pointer management unit that manages information about a part that could not be collated and information about a range to be re-entered based on the collation result of the enroll sentence collation unit; and displays a re-entry part based on the two pieces of information. If there is a part that cannot be collated at the time of voice input of the enroll sentence and the enroll sentence re-display part, Repeat voice input,
When the number of times exceeds a predetermined number, the processing moves to the next enroll statement,
An enroll sentence voice input device, comprising: an enroll sentence re-input determination unit for re-inputting the remaining enroll sentence if there is a remaining enroll sentence that cannot be collated at the end of reading all the enroll sentence. 5. In a voice recognition system, a step of displaying a pre-registered enrolled sentence, a step of inputting a voice of a user who has read out the enrolled text, and a step of converting the input voice into a pre-registered sound A computer-readable recording medium storing a program for causing a computer to execute a step of collating with a model and an enrolling step for registering voice data of a user who has been collated, and cannot collate with input voice If there is a part, a program that causes the computer to execute a step of repeatedly instructing re-entry of the enroll sentence a predetermined number of times and a step of proceeding to input processing of the next enroll sentence if there is still a part that cannot be verified. A computer-readable recording medium that has been recorded.