JPS647680B2 - - Google Patents

Description

[Detailed description of the invention]

[Industrial Field of Application] The present invention relates to a correction method for a speech recognition device, and in particular, to a method that can efficiently correct misrecognition by utilizing the regularity of misrecognition for monosyllabic words to be recognized. be. [Prior Art] Recently, as the speech recognition rate of speech recognition devices has improved, voice input typewriters using speech recognition devices have been put into practical use. In this type of voice input typewriter, a sentence to be typed is input into a voice input device while being divided into monosyllables, and the results of voice recognition in the voice input device are printed by a printing device. Conventionally, it is sufficient for the speech recognition device used in such speech input devices to be capable of recognizing monosyllabic speech; for example, in the case of Japanese, if it can recognize 68 monosyllables, it is almost satisfactory. It is known that it has a voice recognition function. First, FIG. 1 shows a schematic configuration of an example of a conventional speech recognition device. In FIG. 1, 1 is a microphone that inputs voice, 2 is a voice recognition unit main body that processes the voice signal from microphone 1 and converts it into a bit pattern of a digital voice signal, and 3 is a voice recognition unit main body that processes the voice signal from microphone 1 and converts it into a bit pattern of a digital voice signal. A storage device that stores bit patterns. Reference numeral 4 denotes a display section that displays the voice as a result of recognition in the form of characters, and 5 denotes a printing section that prints characters representing the voice that is the result of final recognition. The conventional speech recognition device having the configuration shown in Figure 1 starts with A (A) and ends with N (N).
48 sounds, as well as voiced sounds such as BA (BA) and Pa (PA),
Semi-voiced sounds are input from a microphone 1, and bit patterns of digital audio signals representing those sounds are registered in advance in a storage device 3. In the subsequent voice recognition, when the voice to be recognized is similarly input from the microphone 1, the analog voice signals representing the voice are converted into corresponding digital voice signals in the voice recognition section body 2, and the digital voice signals are converted into digital voice signals. The bit pattern is compared with the bit pattern previously registered in the storage device 3.
Through the comparison, when the registered bit pattern that most closely approximates the bit pattern corresponding to the input voice is recognized as the input bit pattern, the voice represented by that bit pattern is displayed on the display unit 4.
to be displayed. In this case, if the sound to be recognized is input in exactly the same way as when the sound registered in advance in the storage device 3 was input, there will be no error in recognizing the bit pattern by the comparison described above, but if the bit pattern is even slightly different, If the voice is input in a slow tone, the probability of erroneous recognition will significantly increase. Moreover, the misrecognition of the audio bit pattern
As mentioned above, the probability of erroneously selecting other words belonging to the same vowel row was extremely high. In conventional speech recognition devices, in response to such misrecognition, the user presses the erase key on the display unit 4 to erase the display of the misrecognized word, and then inputs the same desired word into the microphone 1 again. The above recognition was repeated. Therefore, correcting misrecognition requires considerable effort and time. [Problems to be Solved by the Invention] As described above, although the speech recognition rate of conventional speech recognition devices has been improved as described above, there is still a high probability of erroneous recognition.
Moreover, in order to correct such misrecognition, conventional speech recognition devices use a correction method that deletes the recognized word and re-enters the speech to re-recognize it. The drawback was that it took time. Generally, however, there are regularities in the mode of misrecognition of speech, and most of these misrecognitions occur in the consonant part of a monosyllable. for example,
When MA is input, the word that is incorrectly recognized and output is the word that belongs to the group of the same vowel (A) row.
(A), Ka (KA), Sa (SA), Ta (TA), Na (NA),
HA (HA), ya (YA), RA (RA) and WA (WA), and the other vowel rows (E, I,
Similar erroneous recognition of only the consonant part generally occurs for each group (O, U). An object of the present invention is to provide a new correction method for a speech recognition device that can efficiently correct misrecognitions by focusing on the regularity of speech misrecognition mentioned above, with the aim of eliminating the conventional drawbacks. It's about doing. [Means for solving the problems] In order to achieve such an object, the present invention has the following features:
When recognizing multiple monosyllabic words for a speech recognition device, multiple monosyllabic words are classified and registered for each vowel group, and from the address registered in the recognized monosyllabic word other words in the same vowel group are registered. Using a deriving means for deriving other addresses in which monosyllabic words are registered, the address of the misrecognized monosyllabic word is converted to another address by the deriving means, and other monosyllabic words in the same vowel group are It is characterized by correcting the misrecognized word by replacing it with the misrecognized monosyllabic word. [Operation] The basic point of the correction method according to the present invention that quickly corrects the above-mentioned misrecognition is that, when registering words to be recognized in the storage device, consecutive words in the storage device are registered in advance. In a series of addresses, monosyllabic words are classified and registered for each word group in each vowel line, and furthermore, the address for the recognized word can be modified with an arbitrary number, for example, an arbitrary number can be added to such address. The object of the present invention is to add a means for adding and subtracting , and to make it possible to efficiently correct erroneously recognized words without having to re-input the required words. According to the present invention, by using a deriving means for deriving other addresses in which other monosyllabic words in the same vowel group are registered from the registered address of the recognized monosyllabic word, the erroneously recognized monosyllabic word is Since the address of is converted to another address by the derivation means,
For example, by automatically correcting the address,
Erroneously recognized monosyllabic words can be corrected simply and easily, and therefore, erroneously recognized words can be corrected much more quickly and reliably than in the past. [Example] Hereinafter, the present invention will be described in detail with reference to the drawings. FIG. 2 shows an example of the configuration of a speech recognition device that corrects misrecognized words according to the present invention. Table 1 shows an example of the assignment of addresses in the storage device to the monosyllabic words in row A when registering the monosyllabic words to be recognized in the storage device.

[Table] Here, as a result of inputting the sound "na" into a speech recognition device and performing speech recognition, it is assumed that the sound is mistakenly recognized as "sa".The method for correcting misrecognition according to the present invention is shown in the configuration shown in the second figure. I will explain about it. The address of the storage device for the input sound "na" is "04" as shown in Table 1, and the address of the storage device for the sound "sa" resulting from erroneous recognition is "02". Therefore, in that case, adding "02" to the address "02" of the misrecognized word "sa" will give the address "04" of the required monosyllabic word "na".
becomes. If this relationship is reversed and the desired monosyllabic word "na" is input and the monosyllabic word that is misrecognized is "ra", the address of the misrecognized word "ra" is "08". Therefore, in this case, the address “08” to “04” of the misrecognized word “ra” is
By subtracting , the address "04" of the required monosyllabic word "na" is obtained. In order to perform addition and subtraction of the number of addresses as described above with the configuration shown in the second figure, the recognition output signal from the voice recognition section main body 6 is separated into two systems, one of which is sent via the OR circuit 7 to the display. The input word is supplied to section 8 to display the input word, and the other is supplied to shift register 9 in the other system for temporary storage. Now, suppose that a word with an address different from the address "04" of the monosyllabic word "na" displayed on the display section 8,
For example, if it is mistakenly recognized as “sa”,
Address "02" of "SA" is temporarily stored in the shift register 9. Therefore, when the correction switch 18 is immediately connected to the U side, a positive voltage is applied from the voltage source 19 to the one-shot multivibrator 16 via the switch 18 and the diode 15, causing the one-shot multivibrator 16 to perform monostable operation. Generate a pulse signal. The stored contents of shift register 9 are input to adder 11 and subtracter 12 by the pulse signal. On the other hand, the AND circuit 13 applied with the above-mentioned pulse signal from the one-shot multivibrator 16 and the positive voltage from the voltage source 19 via the U side of the switch 18 operates, and the signal is input from the shift register 9 to the adder 11. “1” is added to the address number “02” of the stored content. The addition output "03" is taken out from the adder 11 and input to the shift register 9 via the OR circuit 10, and further input to the display section 8 via the OR circuit 7. Therefore, on the display section 8, the previously displayed incorrectly recognized word "sa" is erased, and the word "ta" at the address "03" is newly displayed. However, this word "ta" is still incorrect, so if you operate the correction switch 18 again and connect to the U side, the word "ta" will be erased from the display section 8 in the same way as described above. Then, the required word "NA" corresponding to the next address "04" is displayed, and the correction is completed. Furthermore, as another example, if the word "na" is input but the word "la" is mistakenly displayed, the shift register 9 will display the erroneously recognized "la".
address "08" is temporarily stored, so when the correction switch 18 is operated to connect to the D side, a positive voltage from the voltage source 19 is applied to the one-shot multivibrator 16 via the switch 18 and the diode 17. , a monostable operation is performed to generate a pulse signal. The stored contents of shift register 9 are input to adder 11 and subtracter 12 by the pulse signal. On the other hand, the AND circuit 14 to which both the pulse signal from the one-shot multivibrator 16 and the positive voltage from the voltage source 19 via the D side of the switch 18 are applied operates. Therefore, "1" is subtracted from the memory content "08" of the shift register 9 that is input to the subtracter 12, and the subtraction result "07" is taken out and transmitted through the OR circuit 10 to the shift register 9 and the OR circuit. The signal is supplied to the display section 8 via the circuit 7. Therefore, the display section 8
At , the word "ra" that had been displayed up to that point is erased and the word "ya" corresponding to address "07" is displayed. However, this word “ya” also has the required word “na”.
is a word with three different addresses, so if you repeatedly connect the correction switch 18 to the D side three times, the display on the display section 8 will sequentially change from the word "ya" to "ma",
“Ha” changes to “Na” until the desired word “Na” is reached.
The correction is now complete. In the present invention, the address of the erroneously recognized monosyllabic word is obtained using a derivation means that derives other addresses in which other monosyllabic words in the same vowel group are registered from the registered address of the recognized monosyllabic word. is converted into another address by the derivation means. In the above example, the correction switch 18 is used as the derivation means for that purpose.
However, in carrying out the present invention, it is of course possible to automatically correct the address using the address counter without using the switch 18. [Effects of the Invention] As is clear from the above description, the present invention merely intends to recognize monosyllabic words,
Multiple monosyllabic words are registered by vowel group, and other addresses where other monosyllabic words in the same vowel group are registered are derived from the registered address of the recognized monosyllabic word. Since the address of the erroneously recognized monosyllabic word is converted to another address by the deriving means, the erroneously recognized monosyllabic word can be corrected simply and easily. Unlike the case of recognizing entire speech patterns, where several patterns are prepared according to their certainty and speech recognition is performed in the order of their certainty, it is necessary to memorize several such patterns. In addition, a dictionary is provided in which monosyllabic words are registered so that they can be classified into the same vowel group, and monosyllabic words can be corrected by simply controlling the addresses of the dictionary. Syllable words can be recognized and corrected, and misrecognition of monosyllabic words can be corrected much more quickly and reliably than in the past.

[Brief explanation of drawings]

FIG. 1 is a block diagram showing a schematic structure of a conventional speech recognition device, and FIG. 2 is a block diagram showing an example of the structure of a speech recognition device implementing the correction method of the present invention. DESCRIPTION OF SYMBOLS 1... Microphone, 2, 6... Recognition part, 3... Storage device, 4, 8... Display part, 5... Printing part, 7, 10...
OR circuit, 9...shift register, 11...adder,
12...Subtractor, 13, 14...AND circuit, 15,
17...Diode, 16...One-shot multivibrator, 18...Correction switch, 19...Voltage source.

Claims (1)

[Claims] 1. When recognizing a plurality of monosyllabic words for a speech recognition device, the plurality of monosyllabic words are classified and registered for each vowel group, and from the address where the recognized monosyllabic word is registered, the words within the same vowel group are By using a deriving means for deriving other addresses in which other monosyllabic words are registered, and converting the address of the misrecognized monosyllabic word into another address by the deriving means, other monosyllabic words in the same vowel group are registered. A method for correcting a speech recognition device, characterized in that the misrecognized word is corrected by replacing the monosyllabic word with the misrecognized monosyllabic word.