JPH0229231B2 - - Google Patents

Description

[Detailed description of the invention]

<Technical Field of the Invention> The present invention relates to the improvement of a Japanese speech input device that recognizes input speech syllable by syllable, and more specifically, to improve recognition errors based on the recognition results of speech uttered during input. This is done to erase the feature standard pattern involved. <Technical background of the invention and its problems> In conventional speech recognition devices, a registration mode in which a standard pattern of speech characteristics is registered and a recognition mode (input mode) in which input speech is recognized are separated, and a form of re-registration is used. Although it was possible to replace the syllable feature patterns in , it was not possible to erase the feature standard patterns that were involved in erroneous recognition in the recognition mode. This does not pose much of a problem when recognizing words as a unit, but when recognizing syllables as a unit, the difference in feature patterns between syllables is small, so it may not be possible to use it as a standard feature pattern due to the influence of the extraction position, etc. There was a problem that if a suitable one was found, the recognition performance would deteriorate. <Object of the Invention> The present invention has been made in view of the above points, and is intended to solve the problems associated with misrecognition in a Japanese speech input device that is equipped with a plurality of feature standard patterns for one syllable. The purpose of the Japanese speech input device of the present invention is to improve recognition performance by erasing appropriate characteristic standard patterns. Originally, it is configured to include erasing means for erasing feature standard patterns that are involved in erroneous recognition. Moreover, according to the present invention, by storing the input voice, when an instruction is given to delete a feature standard pattern that is involved in erroneous recognition, the corresponding part of the input voice can be played back and output to accurately cut out the syllable. This allows the person who inputs the information to check whether or not the feature standard pattern has been deleted by mistake. <Embodiments of the Invention> Hereinafter, with reference to the drawings, the present invention will be described in which continuously uttered speech is recognized in units of syllables, the recognition results are corrected using an input device such as a keyboard, and then in units of words etc. A Japanese voice input device having a function of transferring data to an external device will be explained as an example. FIG. 1 is a block diagram showing the configuration of an apparatus according to an embodiment of the present invention. In FIG. 1, the uttered and input voice is input to an analog input section 2 via a microphone 1 etc., and after being amplified by an amplifier 3 in the analog input section 2, it is input to an analog/digital conversion section 4. Therefore, it is converted into a digital signal, and the digital signal is sent to a speech analysis section 5 and a syllable segmentation section 6.
is input. The speech analysis section 5 divides the input speech into frames of about 16 ms, performs spectrum analysis, and sends the syllable segmentation section 6 at intervals of about 8 ms to collect characteristic patterns and information necessary for syllable segmentation (power, number of zero crossings). etc.). The syllable segmentation section 6 uses the type information from the speech analysis section 5 to cut out syllables from the input speech. The feature pattern of the cut out portion and the audio waveform of that section are stored in a waveform/feature pattern temporary memory 7. Then, the CPU 8 is informed that the audio has been cut out, and the address in the waveform/characteristic pattern temporary memory 7 is also transmitted at the same time. The waveform/characteristic pattern temporary memory 7 is configured to be able to store a plurality of syllables. The processing of the syllable segmentation unit 6 is carried out by the CPU 8.
It is configured so that the start and stop are controlled by commands from. Reference numeral 9 denotes a monosyllable recognition unit, and the monosyllable recognition unit 9 reads the pattern memory 1 according to instructions from the CPU 8.
Distance calculation etc. are performed between the characteristic pattern memory 10a in 0 and the standard pattern memory 10b, and the result is returned to the CPU 8. Then, the CPU 8 stores the results in the recognition result storage memory 11 and displays them on the display device 12. The recognition result storage memory 11 is configured to be able to store recognition results for a plurality of syllables. Reference numeral 13 denotes an audio output control section, and the audio output control section 13 sends information stored in an arbitrary part of the waveform/characteristic pattern temporary memory 7 to the analog output section 14 in response to instructions from the CPU 8. The analog output section 14 is configured to reproduce the audio digital signal into an audio waveform using a digital/analog conversion section 14a and an amplifier 14b, and output it. The pattern memory 10 is divided into two parts, and 10a is a feature pattern memory that can store only one feature pattern corresponding to the input syllable. The latter one 10b is a feature standard pattern memory in which feature patterns of a plurality of syllables are stored. The syllable characteristic pattern memory 10b includes an area for storing each syllable name in code, as will be described later.
It is composed of a flag area for storing registration status and a feature standard pattern area for storing feature standard pattern data. Reference numeral 15 denotes an input unit composed of a keyboard, etc. For example, as shown in FIG.
5a, registration mode key 15b, recognition mode key 1
5c, a syllable deletion key 15d, and the like. Reference numeral 16 denotes an I/F unit that controls data transmission and reception when transferring recognition results to an external device. Next, the operation of the apparatus configured as described above will be explained in terms of registration mode and recognition mode. Description of Registration Mode FIG. 3 shows the processing flow of the CPU 8 in the registration mode. In FIG. 3, when the main body of the apparatus is set to the registration mode by operating the registration mode key 15b, the pattern memory 10 is first initialized in step n1, and all standard patterns are erased. Table 1 shows the configuration of the standard pattern memory 10b.

[Table] The initialization process in step n1 is realized by putting "0" in the flag area of the standard pattern memory 10b indicating the presence or absence of registration. Next, proceeding to step n2, the monosyllable to be uttered is displayed on the display device 12 as follows. `` <sub>A1</sub> '' Here, the subscript ``1'' indicates that it is the first character in the ``A'' pattern. The operator looks at the display on the display device 12 and inputs a predetermined monosyllabic voice by uttering it. In response to this voice input, the process moves to step n3, where the syllable segmentation unit 6 is instructed to start cutting out the voice, and the syllable segmentation unit 6 cuts out a single syllable, and the intermediate waveform and voice analysis unit 5 The characteristic pattern obtained is stored in the waveform/characteristic pattern temporary memory 7. In step n4, the syllable segmentation section 6 checks whether a single syllable has been segmented, and if it has been segmented, the process moves to the next step 5. At step n5, the syllable segmentation section 6 is commanded to stop the segmentation process, and the registration process is continued. In step n6, the audio portion corresponding to the syllable just cut out is read out from the waveform/characteristic pattern temporary memory 7 and is reproduced and output from the analog output section 14 via the audio output control section 13. In step n7, it is determined whether the operator has correctly extracted the audio based on the reproduced audio, and in accordance with the resulting instruction from the keyboard 15, it is determined whether to perform re-extracting or registration. In this step n7, if the operator listens to the playback output and determines that the clip has been cut out accurately, he or she operates the execution key 15i, and as a result, the process moves to step n8, and if the operator instructs re-cutting, , the process returns to step n3 in response to the operation of the release key 15h. In step n8, the characteristic standard pattern memory 1 corresponding to the syllable displayed on the display device 12 is
The feature standard pattern is stored in the 0b position, and "1" is set in the flag indicating whether or not the corresponding registration exists.
Set. At step n9, it is determined whether or not all standard patterns have been registered. If not, the process returns to step n2, displays the next monosyllable, for example, " _A2 ", and performs the same process. In this way, when the registration is completed, several characteristic standard patterns of all monosyllabic characters are registered in the standard pattern memory 11b. Next, the operation in recognition mode will be explained. Description of Recognition Mode FIG. 4 shows the processing flow of the CPU 8 in the recognition mode. First, the apparatus is set to recognition mode by operating the recognition mode key 15c, and when the operator utters a voice to be recognized, in step n11, the syllable segmentation section 6 is instructed to start cutting out syllables according to the input voice. give orders. Then, the syllable segmentation unit 6 initializes the waveform/feature pattern temporary memory 7, and thereafter stores the feature pattern and waveform corresponding to the cut out syllable from the first address, starting from the beginning and end of the waveform and feature pattern of each syllable. street address information
Give to CPU8. In step n12, a check is made to see if a syllable has been cut out, and if it has been cut out, the process moves to step n13. In step n13, the characteristic pattern of the waveform/characteristic pattern temporary memory 7 is transferred to the characteristic pattern memory 1 of the pattern memory 10.
It is transferred to the 0a area and recognized. That is, by giving a recognition command to the monosyllable recognition unit 9, recognition is performed by comparing the contents of the characteristic pattern memory 10a with the contents of the standard pattern memory 10b, and the result is stored in the recognition result storage memory 7 and displayed on the display device 12. (Step n14). For example, if the first recognition result when uttering "kaimono" as an input voice is "mogimono", the display device 12 will display "mogimono- _" , and the recognition result storage memory 11 will display each recognition result. A plurality of recognition result candidates for syllables are stored, for example, as shown in Table 2.

[Table] In Table 2 above, the underlined items are the top recognition results for each syllable name.
By operating the "Next Syllable Candidate" key 15e,
This corresponds to a characteristic standard pattern that is displayed on the display device 12 and that can be erased according to the present invention. When the operator has finished inputting the word "kaimono" mentioned above, the operator should
5. Operate the "end" key 15g. As a result, the syllable segmentation unit 6 is commanded to stop segmentation (steps n15, n16).
Then, if all the character strings are correct, by inputting the "transfer" key 15j, the I/F unit 16
Kana characters can be output to an external device via (steps n18, n19). If you look at the display of the recognition results and find that most of the characters are incorrect or you have made a mistake, you can return to the initial state by inputting the "cancel" key 15f as determined in step n17. If some of the recognition results are incorrect, the operator will make corrections using keys, as shown in step n20. There are two ways to modify using keys. First, move the cursor to the position you want to correct using the 1 key.
5k, 15l Use "→" and "←" to move the cursor to the character you want to correct. For example, if you want to correct the second character "gi", the display changes as follows by operating the cursor movement key "←" 15l. One method is to enter characters at this cursor position using the Japanese keypad 15a of the keyboard 15, and make the following corrections. <Key input>"I" Momono Another method is to input the "Next syllable candidate" key 15e on the keyboard 15. <Key input>"Next syllable candidate" MOMONO " Next syllable You can modify it like "Candidate" Moimo . Note that at this time, the configuration is such that items with the same syllable name are displayed only once. That is, the candidates for the recognition result of the second syllable are " _gi3 , _ni1 , _gi2 , _i2 ,...", but " _gi2 " is not displayed. Although the recognition result is corrected by the above-described procedure, the feature standard pattern is deleted by the following procedure according to the present invention. In the example of "kaimono" mentioned above, the appearance of "mo" as a recognition result for the input speech "ka" is considered abnormal in terms of syllable similarity. Therefore, use cursor movement 15k, 15l "←" and "→" to move the cursor on the display device 12 to the position of the first "mo" and change the display screen to the following (n20) Momo Here , To delete this characteristic standard pattern of "mo", operate the "syllable deletion" key 15d (step n21). If you want to delete the characteristic standard pattern other than "mo", operate the "Next Syllable Candidate" key 15e to select and display syllable candidates one by one, and when they appear on the display, press the "Delete Syllable" key 15d. You will be instructed to operate on the syllable you want to erase. When the above display is displayed, "Syllable deletion" key 15d
By inputting , the process moves to step n22 based on the judgment in step n21, and the audio output control section 13
The addresses of the start and end of the waveform of the syllable to be reproduced are specified, and the waveform of the corresponding part is read out from the waveform/characteristic pattern temporary memory 7 and given to the analog output section 14 via the audio output control section 13. , the audio corresponding to "ka" is reproduced and output from the analog output section 14. By listening to this reproduced audio, the operator judges whether the cutting position is accurate or not, and then inputs the "execute" key 15i or the "cancel" key 15h. When the key 15i is pressed, the process proceeds from step n23 to step n24, and when the "cancel" key 15h is pressed, the process proceeds to step n17. When moving from step n23 to n24
The CPU 8 stores the syllable feature standard pattern corresponding to the instructed recognition result (in this case, " _Mo1 ") in the standard pattern memory 10b of the pattern memory 10.
At the same time, the registration presence/absence flag of the corresponding syllable name part is set to "0". In the above example, the standard feature pattern " _Mo1 " is deleted, so
This is achieved by setting "0" to the flag indicating whether or not the syllable name " _Mo1 " is registered. If all the registration flags for that syllable name become "0" (step n
25), the following display is made to encourage re-registration (step n26). All "mo" patterns have been deleted. The operator sees this display, sets the device to registration mode, and registers the standard pattern of "Mo". In the above embodiment, the standard pattern memory 10b
Although a part of the memory content of the input user is erased based on the input person's own judgment, the present invention is not limited to this, and for example, contributes to recognition by determining the quality of the standard pattern. It is also possible to make it impossible to erase the feature standard pattern. As a method for determining the quality of a standard pattern, for example, the method previously proposed by the applicant in Japanese Patent Application No. 57-217296 ``Speech Recognition Device'' is used, in which a counter means is provided for each feature standard pattern, and a recognition judgment of input speech is performed. There is a method in which the counter value is increased or decreased depending on the result, and a feature standard pattern contributing to recognition is determined according to the counter value. As described above, according to the embodiment described above,
In a speech input device that has a plurality of feature standard patterns for one syllable, recognition performance can be improved by erasing inappropriate feature standard patterns that are involved in erroneous recognition. Therefore,
For example, when using a voice input device, an inappropriate standard pattern may be registered as a feature standard pattern due to an operational error during registration, and syllables that are far from the input voice may appear as recognition results. Even if you say "I", "Niya" always appears in the first place as a recognition result, and you may have to make corrections each time. By erasing "niya" from the feature standard pattern, "niya" will no longer appear in first place, improving recognition performance. Also. For example, if you say "kaimono" and misrecognize "ka" as "mo", by storing the input voice, when you instruct to delete the characteristic standard pattern of "mo", By reproducing and outputting the part of the audio corresponding to the input speech "ka", the inputter can check whether the syllable has been correctly extracted, and the recognition result for the input speech in which the feature standard pattern that was instructed to be deleted is normal. Therefore, it is possible to avoid mistakenly erasing feature standard patterns that contribute to recognition. <Effects of the Invention> As described above, according to the present invention, in a Japanese speech input device that recognizes input speech in units of syllables by comparing input speech with standard patterns of characteristics of multiple types of syllables registered in advance, temporary storage means for temporarily storing uttered voice data; display means for displaying the recognition result of the voice; and erasure for erasing feature standard patterns involved in misrecognition based on the recognition result displayed on the display means. The erasing means includes a selection means for selecting characters to be erased from the recognition results displayed on the display means, an erasure key, and a selection means for selecting characters based on operations of the selection means and the erasure key. means for reading audio data corresponding to the selected character from the temporary storage means and outputting the audio data; an execution key for deleting the feature standard pattern of the character selected by the selecting means after confirmation by the audio output; It is characterized by including a key to cancel the erasure,
Therefore, it is possible to judge whether or not the feature standard pattern that was instructed to be deleted is the recognition result for the target input voice, and it is possible to avoid the inconvenience caused by mistakenly deleting the feature standard pattern that contributes to recognition. , recognition performance can be improved by eliminating inappropriate feature standard patterns.

[Brief explanation of drawings]

Fig. 1 is a block diagram showing the configuration of a device according to an embodiment of the present invention, Fig. 2 is a plan view showing an example of an input section such as a keyboard, and Fig. 3 is a processing flow diagram for explaining the operation in registration mode. , FIG. 4 is a processing flow diagram for explaining the operation in the recognition mode. 5... Speech analysis unit, 7... Waveform/feature pattern temporary memory, 8... CPU, 9... Monosyllable recognition unit, 10...
Pattern memory, 10a... Feature pattern memory,
10b... Standard pattern memory, 11... Recognition result storage memory, 15... Keyboard, 15b... Registration mode key, 15c... Recognition mode key, 15d... Syllable deletion key.

Claims (1)

[Scope of Claims] 1. A Japanese speech input device that recognizes input speech in units of syllables by comparing the input speech with standard patterns of characteristics of multiple types of syllables registered in advance, which temporarily stores speech data uttered at the time of input. a temporary storage means for displaying the voice recognition result, a display means for displaying the recognition result of the voice, and an erasing means for erasing the feature standard pattern involved in the misrecognition based on the recognition result displayed on the display means, the erasing means , a selection means for selecting a character to be erased from among the recognition results displayed on the display means, an erase key, and audio data corresponding to the character selected by the selection means based on the operation of the selection means and the erase key. means for reading out from the temporary storage means and outputting it audibly, an execution key for deleting the character feature standard pattern selected by the selection means after confirmation by the audible output, and a key for canceling the deletion. A Japanese voice input device characterized by: