JP3815110B2 - Voice input device and voice input method - Google Patents

Description

[0001]
BACKGROUND OF THE INVENTION
The present invention relates to a man-machine interface, and more particularly to a voice recognition that is a means for inputting data into an apparatus and converts voice into character information.
[0002]
[Prior art]
In a device using voice recognition, it is inevitable that the device makes a recognition error. Therefore, in order to implement as a device, it is essential to provide a means for correcting the recognition result. Although there is a method of designating a correction location with a pointing device (mouse or touch panel) as a method for designating a correction location, the present invention assumes a situation where only the speech is used or only the speech can be used. In such a case, conventionally, since a correction location cannot be specified for a recognition error, correction is performed by re-phrasing the same utterance content as the previous time.
[0003]
[Problems to be solved by the invention]
When trying to correct a recognition error using only speech, it is troublesome to restate the entire utterance. In particular, it is very inconvenient and inefficient to rephrase the whole word for a part of a misrecognized part of a long utterance such as a continuous number with a large number of digits or an address.
[0004]
An object of the present invention is to provide a man-machine interface that enables a user to make corrections efficiently with only a small amount of speech even when an inevitable recognition error occurs in a speech input device using speech recognition. An object of the present invention is to provide a voice input device provided, or various information processing devices including the voice input device.
[0005]
[Means for Solving the Problems]
In the present invention, a recognition target generation unit that generates a recognition target dictionary that enables input intended to be corrected, and a previous recognition result are held. When a correction intention is detected, the recognition result that is held is retained. The problem is solved by providing a result holding means for correcting and outputting the above.
[0006]
In the speech input device of the present invention, speech input means for inputting speech, speech recognition means for recognizing the input speech within the range of the designated recognition target dictionary, and result holding for holding the recognition result of the speech recognition means A recognition result output means for outputting a recognition result, and a recognition target generation means for generating a recognition target dictionary used by the voice recognition means based on a basic recognition target dictionary specified in advance. When the generation unit receives the recognition result of the previous utterance that can be corrected from the result holding unit, the generation unit generates the correction target dictionary from the basic recognition target dictionary and the recognition result, and specifies the correction target dictionary A new recognition target dictionary is created by combining the corrected correction term dictionary and the basic recognition target dictionary, and the result holding means determines whether the recognition result includes the correction term. If the recognition result includes a correction term, a correction corresponding to the intention of the correction term is applied to the stored content, and the corrected result is output as a recognition result output unit and a recognition target generation unit. When the recognition result does not include a correction term, the result is output as it is to the recognition result output means and the recognition target generation means.
[0007]
Further, in the voice input device of another configuration of the present invention, voice input means for inputting voice, voice recognition means for recognizing the inputted voice within the range of the designated recognition target dictionary, and recognition by the voice recognition means A result holding means for holding the result; a recognition result output means for outputting the recognition result; a recognition target generating means for generating a recognition target dictionary used by the voice recognition means based on a basic recognition target dictionary specified in advance; The recognition target generation means combines the basic recognition target dictionary with a correction term dictionary specified in advance to create a new recognition target dictionary, and the result holding means recognizes the speech recognition means. When the correction term is included in the result, a result obtained by performing correction corresponding to the intention of the correction term on the retained content is output.
[0008]
According to another aspect of the present invention, there is provided a voice input device comprising: voice input means; recognition means for recognizing input voice; and output means for outputting the recognition result of the input voice. When a part of the recognition result is a misrecognized word, the misrecognized word, a correct word, and a phrase indicating which of those words is correct are input as corrections from the input means, Recognizing the correction, the result of correcting the misrecognized word in a part of the recognition result to a correct word is output as the final result.
[0009]
According to another aspect of the present invention, there is provided a speech input device comprising: means for inputting numeric speech; recognition means for recognizing input speech; and output means for outputting a recognition result of the input speech. When the number of the recognition result is a misrecognized word, the means inputs, as a correction, a phrase indicating a mathematical arithmetic difference between the misrecognized number and the correct number from the input unit. Then, a number obtained by recognizing the correction and performing an operation instructed by the correction on the erroneously recognized number is output as a final result.
[0010]
DETAILED DESCRIPTION OF THE INVENTION
A first embodiment of the present invention will be described.
[0011]
FIG. 1 shows a block diagram of the voice input device. The spoken voice is converted into a digital signal by voice input means (corresponding to from the microphone to A / D conversion), and the voice recognition means gives a range of a recognition target dictionary that is given in advance or is designated from outside each time. Then, the digital signal is recognized and the recognition result is output. The result holding means holds the recognition result and presents it to the user with a screen or voice as appropriate via the recognition result output means. The recognition target generation unit generates a recognition target dictionary from the basic recognition target dictionary, and reconfigures the recognition target dictionary so that the user's utterance including the correction intention can be recognized using the correction term dictionary at the same time. In addition, when the recognition target generation unit receives the recognition result from the result holding unit, the recognition target generation unit sets the recognition target dictionary so that the user's utterance including the correction intention can be recognized using the recognition result at the same time. Reconfigure.
[0012]
An example will be described in which consecutive digits with unlimited digits are set as recognition targets. For the sake of simplicity, the basic recognition target dictionary includes only arbitrary continuous numbers. However, when an actual application system is assumed, various control commands may be added.
[0013]
If the user utterance is U and the device output is S, the user input and device output are
U “010” (spoken by user)
S “080” (The device mistakenly replaces “1” with “8”)
U “8 to 1” (User speaks with intention to correct)
S “010” (Device corrects “8” in previous recognition result to “1”)
It becomes.
[0014]
In order to recognize the input “8 to 1” of the second user, the recognition target generation means reconstructs the recognition target dictionary. In addition to the “basic recognition target dictionary”, the recognition target dictionary is a dictionary in which a combination of “correction target dictionary” + “correction term” + “basic recognition target dictionary corresponding to the correction target dictionary” is arranged in parallel. In this case, the correction term is “no”, but it is not limited to this and may be “from” or “more”. The correction target dictionary is all combinations of one or more digits for the recognition result. In this example, “0”, “8”, “08”, “80”, and “080” are set for “080”. However, since “0” alone cannot identify the first digit “0” or the third digit “0”, it is better not to use “0” as a correction target dictionary.
[0015]
In addition, the “basic recognition target dictionary corresponding to the correction target dictionary” can be corrected more reliably by removing the correction target from an arbitrary continuous number. For example, “8 is 1”, “8 is 88”, “8 is 11”, and the like may be included in the recognition target, but “8 is 8” may not be the recognition target.
[0016]
The previous example is an example of a substitution error (a number is misrecognized as another number) when consecutive numbers are to be recognized. However, an error (a number that is not actually spoken is included in the recognition result) ) And omission errors (errors in which the actually uttered numbers are missing from the recognition result) can be corrected in the same way. For example, as an example of an error
U “4100”
S “42100” (The device springed up “2” and made an error)
U “421 to 41” (Speaking with intention of correction)
S “4100” (Device corrects previous recognition result)
As an example of omission error,
U “0123”
S “023” (Equipment is missing “1”)
U “02 to 012” (Voice including intention of correction)
S “0123” (Device corrects previous recognition result)
It becomes.
[0017]
By the way, the example here is 3-4 digits and a small number of digits for the sake of simplicity, but when the number of digits increases, the burden on the user is less compared to rephrasing everything. It is clear. It is also possible to make corrections repeatedly using the same method.
As described above, the voice input device of the present application holds the voice input means for inputting voice, the voice recognition means for recognizing the input voice within the designated recognition target dictionary, and the recognition result of the voice recognition means. And a recognition result output means for outputting a recognition result, and a recognition target generation means for generating a recognition target dictionary used by the speech recognition means based on a basic recognition target dictionary specified in advance. Further, the recognition target generation unit creates a new recognition target dictionary by combining the basic recognition target dictionary and a correction term dictionary specified in advance, and the result holding unit corrects the recognition result of the voice recognition unit. If a term is included, a result obtained by performing correction corresponding to the intention of the correction term on the retained content is output. The basic recognition target dictionary includes a dictionary for recognizing a bar reading that does not include a digit position, and a dictionary for recognizing a digit reading that includes a digit position, and the result holding means includes the voice recognition means If the recognition result includes a correction term and a digit reading, the correction position is specified by the digit information of the digit reading and the correction corresponding to the intention of the correction term is retained. The corrected result can also be output.
[0018]
Furthermore, the present invention can also be used when the recognition target is a place name such as an address. For example,
U “Higashi Koigakubo, Kokubunji, Tokyo”
S “Nishikoigakubo, Kokubunji, Tokyo” (The equipment mistakenly replaces “Higashikoigakubo” with “Nishikoigakubo”)
U “East Koigakubo to Nishikoigakubo” (Speaking with intention of correction)
S “Tokyo Higashi Koigakubo, Kokubunji, Tokyo” (The device corrects the previous recognition result)
It can be corrected as follows. In this case, the basic recognition target dictionary is, for example, addresses all over Japan, and the correction term is “O”. The correction target dictionary is obtained from the basic recognition target dictionary and the recognition result, and may be, for example, three words of “Tokyo”, “Kokubunji City”, and “Nishikoigakubo”. In addition, it is possible to obtain misrecognized units in consideration of restrictions as real addresses (Kokubunji City is only in Tokyo, etc.). In such a case, the dictionary to be corrected is “Nishikoigakubo, Kokubunji City, Tokyo” It becomes "Kokobunji City Nishikoigakubo", "Nishikoigakubo", "West".
[0019]
A second embodiment of the present invention will be described. FIG. 2 shows a block diagram of the voice input device. The difference from the block diagram of FIG. 1 is that the recognition target generation means does not use the recognition result.
[0020]
“OK” or the like is used as a correction term.
[0021]
For example,
U “010”
S "080" (Device replaces 1 with 8)
U "OK, 1" (The first digit is not corrected, the second digit is 1)
S “010” (Device corrects previous recognition result)
It becomes. The correction part is specified by the number of “OK”.
[0022]
The same applies when the recognition target is a place name. For example,
U “Higashi Koigakubo, Kokubunji, Tokyo”
S “Nishikoigakubo, Kokubunji, Tokyo” (The equipment mistakenly replaces “Higashikoigakubo” with “Nishikoigakubo”)
U “OK, OK, Nishi-Kogakubo” (Speaking with intention of correction)
S “Tokyo Higashi Koigakubo, Kokubunji, Tokyo” (The device corrects the previous recognition result)
It becomes.
[0023]
Next, a third embodiment of the present invention will be described. The third embodiment is an example in which recognition targets are limited to numerals, but the block diagram of the voice input device is the same as that in FIG. As correction terms, “add”, “subtract”, “plus”, “minus”, etc. are used.
[0024]
For example, using "Add"
U "67890"
S "67190" (Device replaces 8 with 1)
U “Add 700 (Nanahyaku)” (Voice including intention of correction)
S “67890” (Device corrects previous recognition result)
With “Draw”,
U “010”
S "080" (Device replaces 1 with 8)
U “Take 70” (Voice including intention of correction)
S “010” (Device corrects previous recognition result)
It becomes. What is important here is that the utterance including the intention of correction by the user is an utterance including the positions such as “Hyaku” and “Ju”, and the correction position can be designated by the position information.
[0025]
From the above embodiments, in the present invention, the user who inputs correction does not re-input only the word to be corrected, but always adds an auxiliary word (correction term). In other words, the correction method is indicated by the correction term. Speak the correct contents and correct contents at the same time, and use a word indicating direction such as “to”, “from”, “more” to indicate which is correct or correct, or “OK” for each digit. The correction term is made to correspond, and the correction term is used as a signal for shifting the digit. Of course, the correction term is not limited to “OK”. In the case of number recognition, recognition and calculation can be combined. An operator is used as the correction term, and a number is calculated for the misrecognized word according to the operator attached to the number input to the number of erroneously recognized numbers to be corrected, and the result is set as a correct recognition result.
[0026]
The present invention is not an invention that specifically limits a speech recognition algorithm. However, for example, an HMM (Hidden Marko Model) can be used to implement the invention. Details of speech recognition by the HMM are described in “Seiichi Nakagawa: Speech recognition by probability model, IEICE, 1988” and the like is omitted here.
[0027]
【The invention's effect】
According to the present invention, it is possible to easily correct a recognition result of a particularly long utterance by using only a voice and efficiently input a voice.
[0028]
In addition, since correction is a partial input or an input method that is different from the previous input using correction terms is used, it is possible to provide a voice input device that realizes various corrections and is easy for a user to input. In addition, it is possible to provide a voice input device that not only simplifies input but also improves the final recognition result.
[Brief description of the drawings]
FIG. 1 is a block diagram showing an embodiment of a voice input device according to the present invention.
FIG. 2 is a block diagram showing another embodiment of the voice input device according to the present invention.
[Explanation of symbols]
DESCRIPTION OF SYMBOLS 101 ... Voice input means, 102 ... Voice recognition means, 103 ... Result output means, 104 ... Recognition result output means, 105 ... Recognition object generation means, 106 ... Basic recognition object dictionary, 107 ... Correction term dictionary

Claims (2)

Input means for inputting a voice,
A recording means for recording a basic recognition target dictionary for storing at least numbers and a correction dictionary for recording operators as correction terms;
Recognition means for recognizing the input speech using the basic recognition target dictionary and the correction dictionary ;
Output means for outputting the recognition result of the input speech;
Have
The input means, when the numbers of the recognition result is incorrect recognition, receives the correct input words indicated mathematical operations specific differences of the correct numbers as the mis-recognized number in combination with the operators and numbers,
The recognition means recognizes the number and operator of the correction input, and the final result is a number obtained by performing an operation using the operator on the number of the recognition result and the recognized correction input number. A voice input device for outputting. An input means for inputting voice;
A recording means for recording a basic recognition target dictionary for storing at least numbers and a correction dictionary for recording operators as correction terms;
Recognition means for recognizing the input speech using the basic recognition target dictionary and the correction dictionary;
A voice input method executed in a voice input device having output means for outputting a recognition result of the input voice,
When the number of the recognition result output is incorrect, correction of a phrase indicating a mathematical arithmetic difference between the number of the recognition result and the correct number by a combination of the operator and the number via the input means Receiving input,
Recognize the correction input number and operator using the recognition means,
A speech input method, wherein a number obtained by performing an operation using the operator on the number of the erroneous recognition result and the recognized number of corrected input is output as a final result.

Images