JP2002140094A - Device and method for voice recognition, and computer- readable recording medium with voice recognizing program recorded thereon - Google Patents

Abstract

PROBLEM TO BE SOLVED: To solve such problem that correcting operation becomes complicated if a word unit is incorrectly sectioned since the correcting operation needs to be instructed for every word unit. SOLUTION: This device is provided with a voice recognition part 2 which generates a plurality of syllable strings by recognizing the voice from a microphone 1, generates multiple word strings from the syllable string by using an n-gram dictionary in an n-gram dictionary storage part 4 and sorts them into an initial candidate and multiple candidates, and stores the initial candidate in a display area of a RAM 3, a display part 5 which displays the word string stored in the display area of the RAM 3 to a user, and a candidate retrieval part 10 which retrieves a word string wherein only all character strings corresponding to the character position of a misrecognized character string have characters changed from the next candidate and stores them in the display area of the RAM 3 when the misrecognized character string is specified through a keyboard 6 or with a mouse 7.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a speech recognition apparatus applied to a speech word processor or the like and having a function of correcting a result of continuous speech recognition, a speech recognition method, and a computer-readable recording medium recording a speech recognition program. It is.

[0002]

2. Description of the Related Art Text input by a speech recognition apparatus is useful as a means for creating a document. Generally, 100% correct accuracy is not always obtained by text input by continuous speech recognition, so a correction function is provided.

For example, in Japanese continuous speech recognition, a word string as a recognition result is divided into units such as words, and the next candidate different from the initial candidate is presented to the user in that unit, and the user can select a correct word string. Like that.

In this case, the break is not always recognized correctly in the first solution.
Japanese Patent No. 180462 discloses a speech recognition device that changes the break.

FIG. 11 is a diagram showing a configuration of a conventional speech recognition apparatus. In FIG. 11, reference numeral 101 denotes a microphone for converting an input voice into an electric signal; 102, a voice recognition unit for performing voice recognition processing on the electric signal from the microphone 101; 103, a random access memory (hereinafter referred to as RAM);
04 is a word dictionary storage unit that stores a word dictionary used when generating a word string from a syllable string, 105 is a display unit that displays word strings and syllable strings, 106 is a keyboard that performs correction operations,
A word correction unit 108 corrects the initial candidates according to a correction operation.

Next, the operation will be described. First, a case where there is no need to change the break will be described. When the user speaks into the microphone 101, the voice recognition process starts. First, consider a user's utterance example A, "Onsei Nishikichi (voice recognition device)".

When the microphone 101 converts the input voice into an electric signal, the voice recognition unit 102 A / D converts the electric signal from the microphone 101, quantizes the electric signal, analyzes the spectrum, and separates it into syllable units. The plausible syllable string is stored in the RAM 103. In order to explain the correction operation of the conventional speech recognition device, it is assumed that a syllable string “Onsei_Ninchi_Soch” is stored for utterance example A. The symbol "_" means a word break.

Subsequently, the speech recognition section 102 refers to the word dictionary in the word dictionary storage section 104 for storing words corresponding to the syllable strings, and generates word strings corresponding to the respective syllable strings stored in the RAM 103. . Then, the maximum likelihood word string “speech_recognition_device” is output as an initial candidate.

[0009] The syllable string “Onsei_Ninchi_Soch” and the word string “Speech_Recognition_Device” of the initial candidates output from the speech recognition unit 102 are transferred to the word correction unit 108. The display unit 105 displays the word string “voice_recognition_device” of the word correction unit 108.

[0010] The user looks at the initial candidate “speech_cognition_device” on the display unit 105, and inputs a misrecognized word on the keyboard 106.
And perform a correction operation by selecting a candidate or changing a break. In this example, the word segment is correctly recognized, and the second word “recognition” is a misrecognized word.

Therefore, when the user specifies the misrecognized word “cognition” as shown in the display unit 105A of FIG. 12, the word correction unit 108 displays the next candidate word for “cognition”. Since “recognition”, “authentication”, “authorization”, and the like are displayed as the next candidates, the user selects the keyboard 10
6, "recognition" is selected as a candidate. Word correction unit 108
Is the word string “voice_
The “recognition_device” is displayed on the display unit 105 (display unit 105B in FIG. 12).

As described above, when there is no need to change the delimiter, when the candidate word “recognition” is selected as a candidate, the word “recognition” of the next candidate is obtained, and the misrecognized word string “speech_recognition_device” Can be easily corrected to a word string “speech_recognition_device” that is correctly recognized.

Next, a case where a break change is required will be described. Here, considering the utterance example B “Let's come before (prepare before coming)”, the following correction operation is performed from the incorrectly recognized word string to the correctly recognized word string shown below. This misrecognized word string incorrectly recognizes word breaks.

[0014] Word string of misrecognition ""

A correction operation is performed using a syllable string “car______________” corresponding to the word string of the misrecognition. First, a character “yo” in a syllable string “car__________” to be corrected is specified and corrected to the character “yo” to obtain a corrected syllable string 1.

Modified syllable sequence 1

Next, the first word "car" of the modified syllable string 1
Is incorrectly separated from the second word “de”, these are changed to “coming” and “until” to change the modified syllable sequence 2
Get.

Modified syllable string 2

Further, the second word "to" of the modified syllable string 2
Is incorrectly separated from the third word "niyoi"
The modified syllable string 3 is obtained by dividing these into “by” and “you”.

Corrected syllable string 3 "by _ by _ by _ do"

When the user confirms that all word divisions have been changed correctly, finally, the corrected syllable string 3 is converted into kanji characters to obtain a correctly recognized word string.

A word sequence of correct recognition "preparing_before_coming_"

[0023]

Since the conventional speech recognition apparatus is configured as described above, it is necessary to instruct a correction operation on a word-by-word basis. There has been a problem that the operation tends to be complicated.

That is, as in the case of the utterance example B, it may be necessary to perform a modification operation by changing the delimiter in units of words. The correction operation must be complicated.

In the modification operation by changing the delimiter, it is difficult to know which delimiter should be changed. In particular, when the word “until” is registered as only one word in the word dictionary, no matter how much the delimiter is changed, A correctly recognized word string cannot be obtained.

Further, the conventional speech recognition apparatus has a problem that when a plurality of erroneously recognized words are present, the correction operation must be performed a plurality of times, which imposes a heavy burden on the user.

The present invention has been made to solve the above-described problem, and does not require a single corrective operation even when there are a plurality of erroneously recognized portions without being aware of the delimiter in word units. It is an object of the present invention to configure a speech recognition device, a speech recognition method, and a computer-readable recording medium that records a speech recognition program, which can easily correct the word string of the above.

[0028]

SUMMARY OF THE INVENTION A speech recognition apparatus according to the present invention recognizes an input speech to generate a syllable string and performs n-
a voice recognition unit for generating a plurality of word strings corresponding to a syllable string using a gram dictionary, selecting initial candidates and a plurality of next candidates, and storing the initial candidates in a display area, and a word string stored in the display area When the misrecognized character string in the initial candidate displayed on the display means is specified, and only the character string corresponding to the character position of the misrecognized character string is changed in character Candidate searching means for searching for a column from the next candidate and storing it in the display area.

In the speech recognition apparatus according to the present invention, when the character string of the next candidate corresponding to the character position of the misrecognized character string is not the same as the misrecognized character string, the candidate retrieving means expands and defines the character change. It is like that.

The speech recognition apparatus according to the present invention includes a display changing means for displaying an erroneously recognized character string in any one of a display mode of a shaded display, an inverted display, a luminance change display, a color change display, and an animation display. It was made.

The speech recognition apparatus according to the present invention extracts the next candidate character string corresponding to the character position of the erroneously recognized character string from the next candidate searched by the candidate searching means, and matches the next candidate character string with the erroneously recognized character string. Next candidate character string display means for displaying the next candidate character string together with the initial candidate, and next candidate character string selecting means for correcting the misrecognized character string of the initial candidate to the next candidate character string by selecting a candidate Things.

In the speech recognition method according to the present invention, a syllable string is generated by speech recognition of an input voice, a plurality of word strings corresponding to the syllable string are generated using an n-gram dictionary, and initial candidates and a plurality of word strings are generated. A voice recognition step of selecting the next candidate and storing the initial candidate in the display area, a display step of displaying a word string stored in the display area to the user, and a misrecognized character string in the initial candidate displayed in the display step And a candidate search step of searching the next candidate for a word string in which only the character string corresponding to the character position of the misrecognized character string has been changed from the next candidate and storing it in the display area. It is provided with.

In the speech recognition method according to the present invention, in the candidate search step, if the character string of the next candidate corresponding to the character position of the erroneously recognized character string is not the same as the erroneously recognized character string, it is defined as a character change. It is like that.

[0034] The voice recognition method according to the present invention includes a display change step of displaying an erroneously recognized character string in any of the display modes of shaded display, inverted display, luminance change display, color change display, and animation display. It was made.

In the voice recognition method according to the present invention, the next candidate character string corresponding to the character position of the erroneously recognized character string is extracted from the next candidate searched in the candidate search step, and the next candidate character string is displayed on the display unit in accordance with the erroneously recognized character string. A next candidate character string displaying step of displaying the next candidate character string together with the initial candidate, and a next candidate character string selecting step of correcting the misrecognized character string of the initial candidate into the next candidate character string by selecting a user candidate are provided. Things.

A computer-readable recording medium storing a speech recognition program according to the present invention recognizes an input speech as a speech, generates a syllable string, and uses an n-gram dictionary to generate a plurality of word strings corresponding to the syllable string. Is generated and selected as an initial candidate and a plurality of next candidates, a voice recognition procedure for storing the initial candidate in a display area, a display procedure for displaying a word string stored in the display area to a user, and a display procedure for displaying A misrecognition character string designation procedure in which the misrecognition character string in the initial candidate is designated, and a word string in which only the character string corresponding to the character position of the misrecognition character string is changed from the next candidate is searched. And a candidate search procedure to be stored in the display area.

In the computer readable recording medium storing the speech recognition program according to the present invention, in the candidate search procedure, the character string of the next candidate corresponding to the character position of the erroneously recognized character string is not the same as the erroneously recognized character string. It is defined as an extended character change.

The computer-readable recording medium on which the speech recognition program according to the present invention is recorded has a hatched display,
A display change procedure for displaying an erroneously recognized character string is provided in any of the display modes of the reverse display, the luminance change display, the color change display, and the animation display.

The computer-readable recording medium storing the speech recognition program according to the present invention extracts a next candidate character string corresponding to the character position of the misrecognized character string from the next candidate searched in the candidate search procedure, A next candidate character string display procedure for displaying the next candidate character string along with the initial candidate on the display unit in accordance with the recognition character string, and a next candidate character for correcting the misrecognized character string of the initial candidate to the next candidate character string by selecting a candidate And a column selection procedure.

[0040]

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS One embodiment of the present invention will be described below. Embodiment 1 FIG. FIG. 1 is a diagram showing a configuration of a speech recognition apparatus according to Embodiment 1 of the present invention. In FIG. 1, reference numeral 1 denotes a microphone (speech recognition unit) for converting an input voice into an electric signal, 2 denotes a speech recognition unit (speech recognition unit) for performing a speech recognition process on an electric signal from the microphone 1, and 3 denotes a readable RAM.
(Speech recognition means) 4 is an n-gram dictionary storage unit (speech recognition means) for storing an n-gram dictionary used when generating a word string from a syllable string, and 5 is a word string in a display area of the RAM 3 It is a display unit (display means).

Reference numerals 6 and 7 respectively denote a keyboard and a mouse for performing a correction operation, 8 a range designating section for storing the character position of the misrecognized character string specified by the correction operation in the RAM 3, and 9 a display form of the misrecognized character string. A display change unit (display change unit) 10 to be changed is a candidate search unit (candidate search unit) that searches a next candidate for a word string that satisfies search condition 1 described below with reference to an initial candidate in which an erroneously recognized character string is specified. It is.

Next, the operation will be described. FIG. 2 is a flowchart showing a speech recognition method according to Embodiment 1 of the present invention. When the user speaks into the microphone 1, voice recognition processing is started (step ST1). Here, the above-mentioned utterance example B “to come (be prepared)” is considered.

Recognition of a word string is based on the probability P (W |
It is obtained by calculating W that maximizes Y). The word string probability P (W | Y) is obtained from equation (1).

P (W | Y) = P (Y | W) · P (W) / P (Y) (1)

Here, W is an uttered word string, Y is a syllable string, P (Y | W) is an appearance probability of the syllable string Y given the word string W, and P (W) is a word string W Is the appearance probability of.
Since the word string W that maximizes the word string probability P (W | Y) may be obtained, the term P (Y) common to the word string W on the right side of the equation (1) can be omitted, and P (Y | W It is sufficient to find a word string W that maximizes P (W).

At time t = 1, 2,..., L, when the syllable string Y corresponding to the word string W is determined by the equation (2), P
(Y | W) is the syllable probability P (Y ₁ ), P (Y ₂ ),.
It can be calculated from P (Y _L ) by equation (3).

Y = Y ₁ , Y ₂ ,..., Y _L (2)

(Equation 1)

The appearance probability P (W) of the word string W is m
When the word string W of a word is determined by Expression (4), it is obtained from Expression (5) (word n-gram information) independently of the syllable probability.

W = w ₁ , w ₂ ,..., W _m (4)

(Equation 2)

According to the above-described calculation, of the syllable strings, if the word string corresponding to the syllable string exists in the dictionary,
A word string W that maximizes the word string probability P (W | Y) is calculated. For the calculation of the combination, for example,
High-speed processing may be performed by using bi or stack decoding, or the probabilities may be calculated by summation using logarithmic probabilities. The appearance probability P (W) of each word is n-gr
It is stored in the am dictionary storage unit 4 in advance.

FIG. 3 shows an example of the n-gram dictionary storage unit 4 when n = 1 (reference numeral 4A) and when n = 2 (reference numeral 4B). As shown in FIG. 3, the n-gram dictionary storage unit 4 stores a syllable string, a word string corresponding to the syllable string, and the appearance probability P (W) of the last word as one record. In the first embodiment, the number of word chains n
Handles up to two chains, but the number n of chains may of course be three or more.

When the microphone 1 converts the input voice into an electric signal (step ST2), the voice recognition unit 2 performs A / D conversion of the electric signal, quantizes the electric signal, and analyzes the spectrum.
The plausible syllable strings separated in syllable units are stored in the RAM 3 in the order of their likelihood (step ST3).

Here, it is assumed that the following two syllable strings are stored in the RAM 3 together with the syllable string likelihood P (Y | W).

Syllable string 1 “Say to come” P (Y | W) = 0.6 Syllable string 2 “Say to come” P (Y | W) = 0.4

Next, the voice recognition unit 2 generates word strings corresponding to the syllable strings 1 and 2, respectively, and stores the word strings together with the likelihood in the RAM 3 (step ST4). here,
Likelihood P (Y | W) · P corresponding to the following word string
(W) is stored.

Word sequence 1 “I want to urinate in a car” P (Y | W) · P (W) = 0.3 Word sequence 2 “Prepare before coming” P (Y | W) · P (W) = 0 .16 Word string 3 "Urious until coming" P (Y | W) · P (W) = 0.06

The candidate retrieval unit 10 stores the most likely word string from the word strings 1, 2, and 3 stored in the RAM 3 as initial candidates in the display area of the RAM 3 (step ST5).
In the above case, the word string 1 “I want to urinate in a car” is stored as an initial candidate in the display area of the RAM 3 as the maximum likelihood word string. The above steps ST1 to ST5 are the speech recognition steps.

Also, the word string 2 "prepare before coming",
The word string 3 “I'm urged to come” is stored in the RAM 3 as a first next candidate and a second next candidate, respectively. The display unit 5 displays on the display unit 5 the word string of “initial car” stored in the display area of the RAM 3 (step ST6, display step).

[0059] Initial candidate "I'm tired of driving"

When the user recognizes a result different from his / her intention by looking at the output of the display unit 5, the user designates a plurality of misrecognized character strings in the initial candidates using the keyboard 6 and the mouse 7 (step ST7, misrecognized characters). Column designation step). In the first embodiment, an erroneously recognized portion is specified in units of character strings by using the word string “I'm tired of a car” of an initial candidate that does not need to be changed.

At this point, since all the incorrectly recognized character strings are specified, the user can easily perform the correction operation. In the above example, since “car” and “urine” are incorrect, the “car” and “urine” are moved in the end direction by the keyboard 6 and the mouse 7 by the operation indicated by the arrow of the display unit (display means) 5A in FIG. (<> In the text indicates the range of misrecognized character strings).

Initial candidate "<car> means <urine>"

The range specification section 8 stores the start position and the end position of each specified erroneously recognized character string in the RAM 3 (step ST8). In other words, the character positions of <car> and <urine> in "<car> mean <urine>" are expressed in {3, 3} as {1, 1} and {3, 3} in the form of {start position, end position}. (If the end position is smaller than the start position, it indicates that there is a missing character before the start character. Designate by dragging the keyboard 6 or the mouse 7 toward the beginning of the sentence.)

The display changing section 9 reads the designated range of the misrecognized character string from the RAM 3 and updates the display area of the RAM 3 so that the display form is different from that of the part other than the misrecognized character string. The display unit 5 changes the display mode and displays the erroneously recognized character string according to the update of the display area of the RAM 3 (step ST9, display change step).

In the display section (display means) 5B of FIG. 4, the character positions {1, 1 {3, 3} of <car> and <urine> are stored in the RAM3.
And misrecognized character strings are shaded. Of course, without being limited to shaded display, reverse display,
A display mode such as a luminance change display, a color change display, and an animation display may be used. By doing so, it is possible to display an erroneously recognized character string according to the user's preference and use.

The candidate retrieval unit 10 reads the next candidate from the RAM 3 and differs only in the character string at the character position corresponding to the specified erroneously recognized character string (character has been changed), and corresponds to other than the erroneously recognized character string. A word string in which the character strings at the character positions are all the same (characters have not been changed) is searched from the next candidate and R
It is stored in the display area of AM3 (step ST10, candidate search step). In the first embodiment, the candidate search unit 10 reads the following next candidate from the RAM 3.

First Next Candidate “Prepare Before Coming” Second Next Candidate “Urgent until Coming”

The search condition 1 used in the search for the next candidate by the candidate search unit 10 is shown below (here, + is an operator indicating a combination of character strings).

Search condition 1 Character string where <car> is not at the same character position + "de" + Character string where <urine> is not at the same character position + "meaning"

In the case of this search, the character position is determined by referring to the number of characters in the syllable string in the case of Japanese speech recognition. The candidate search unit 10 changes only the character strings at the character positions corresponding to the misrecognized character strings <car> and <urine> of the initial candidates, and changes the characters corresponding to the character positions of “de” and “mean”. As for the column, a word sequence in which all characters are not changed is selected from the RAM 3 as a next candidate.

First, it is determined whether or not the first next candidate satisfies search condition 1. When there are a plurality of matching patterns, it is determined whether or not all the patterns satisfy the search condition 1. Comparing the initial candidate with the misrecognized character string to the first next candidate results in the following (inserting a space for comparison).

Initial candidate “<Urine> in <car>” <br> First candidate “Preparation before coming”

From the above comparison, the first next candidate shows that the designated erroneously recognized character strings <car> and <urine> have been changed to “come” and “prepared” respectively, and It can be seen that the character strings “de” and “mean” other than the recognition character string have not been changed. That is, the first next candidate satisfies the search condition 1.

Next, it is determined whether or not the second next candidate satisfies search condition 1. Comparing the initial candidate with the misrecognized character string and the second next candidate results in the following (inserting a space for comparison).

Initial candidate “<Urine> in <car>” Second candidate “Urium until coming”

From the above comparison, in the second next candidate, one of the misrecognized character strings <car> has been changed to “come”,
It can be seen that the character strings “de” and “I” other than the misrecognized character string have not been changed, but the other incorrectly recognized character string <urine> has not been changed. That is, the second next candidate does not satisfy the search condition 1.

As described above, the word string that matches the search condition 1 can be determined as the first next candidate “prepare before coming”, and the candidate search unit 10 displays the first next candidate on the RAM 3. The character string candidates {1, 3}, {5, 6} corresponding to the specified range of the erroneously recognized character string are stored in the RAM and stored in the area.
3 is stored. Then, the display unit 5 displays the first next candidate “prepare before coming” of the display area, and the display change unit 9 causes the {1, 3}, $ 5,
The display form of the part "come" or "prepared" corresponding to 6 @ is changed.

Correction result "I'll be ready for <Coming>"

When the user confirms that a correctly recognized word string has been obtained by looking at the correction result on the display unit 5, the user proceeds to the next voice input.

As has been seen above, in the first embodiment, erroneous recognition is not limited to word units, but is specified all at once at a plurality of locations in units of character strings. Since the next candidate is searched for the word string in which only the character string corresponding to the character position of the column has been changed, the erroneously recognized word string can be used without performing complicated delimiter changes or multiple correction operations. Can be easily corrected to a word string that is correctly recognized.

In order to better understand the operation of the candidate retrieval unit 10 using the retrieval condition 1, another utterance example will be described.

Suppose that the utterance example C corrects the word string "Give me the money" of the initial candidate which misrecognized "I'm sick of the money". Here, the misrecognized character string is specified in units of character strings, such as “<wife><mo><mo>”. At this time, the next candidates stored in the RAM 3 are the following four word strings.

The first candidate, "Give me the money made by my wife" The second candidate, "Give me the money that came to Tsu" The third candidate, "Send the north wind to Tsu" The fourth candidate, "Give me Tsu Send money! "

Initial candidate “<wife> got money” First first candidate “wife came money” In the first next candidate, <ki> is “come” Has been changed to However, the characters of the other <wife> and <koku> have not been changed, and “de” other than the misrecognized character string has been changed to “de”. As described above, since the search condition 1 is not satisfied, the candidate search unit 10 rejects the first next candidate.

The initial candidate “<wife> and <kin>” The second next candidate “the money that came to Tsu” “Kuma” and “<wife” > Has been changed to “Tsuma” and “Kai”, respectively. However, the remaining characters are not changed. As described above, since the search condition 1 is not satisfied, the candidate search unit 10 rejects the second next candidate.

[0086] Initial candidate "<Wife> and <Kita><Kokure>" Third candidate [Tsuma to north wind sending] In the third candidate, <Wife> and <Koku> Has been changed to "tsuma" and "send" respectively. However, the characters of the remaining misrecognized character string <ki> were changed to include "north" and "ta", and the character string "gold" other than the misrecognized character string was changed to "wind". I have. As described above, since the search condition 1 is not satisfied, the candidate search unit 10 rejects the third next candidate.

The initial candidate “<wife> was paid by <w>” The fourth next candidate was “send money that came to Tsu” The fourth candidate was <wife>, <w> The characters of and have been changed to "tsuma", "kuru" and "send", respectively. In addition, all the character strings “de”, “takin”, and “re” other than the erroneously recognized character string are held as they are without being changed. From the above, to satisfy the search condition 1,
The candidate search unit 10 stores the fourth next candidate in the display area of the RAM 3.

When there are a plurality of next candidates satisfying the search condition 1, the plurality of next candidates satisfying the search condition 1 may be displayed in descending order of likelihood.

As described above, according to the first embodiment, the voice from the microphone 1 is recognized to generate a plurality of syllable strings, and the syllable string is generated using the n-gram dictionary in the n-gram dictionary storage unit 4. A speech recognition unit 2 that generates a plurality of word strings from a syllable string, selects them as initial candidates and a plurality of next candidates, and stores the initial candidates in a display area of the RAM 3, and outputs a word string stored in the display area of the RAM 3 to a user. When the erroneously recognized character string is designated by the display unit 5 and the keyboard 6 or the mouse 7, a word string in which only the character string corresponding to the character position of the erroneously recognized character string is changed to the next candidate. Search from and R
A candidate search unit 10 for storing in the display area of the AM3 is provided. Therefore, even if a plurality of erroneously recognized character strings exist, the erroneously recognized word string can be corrected by a single operation even if a plurality of erroneously recognized character strings exist. Can be easily corrected.

Further, according to the first embodiment, the display change unit 9 for displaying a misrecognized character string from any of the display modes of the shaded display, the inverted display, the luminance change display, the color change display, and the animation display. Is provided, it is possible to obtain an effect that an erroneously recognized character string can be displayed according to a user's preference or use.

Embodiment 2 In the first embodiment, only the misrecognized character string is designated. However, it is actually difficult to accurately designate only the misrecognized character string. For example, in the case of the utterance example B, the place where the misrecognition range should be designated only for <urine> includes the part of "meaning" which is not mistaken as shown in the display unit (display means) 5C of FIG. Often, it is specified as>. Therefore, in the search condition 1 according to the first embodiment, a word string that is correctly recognized cannot be obtained because “me” is at the same character position. In the second embodiment, assuming such a case, a description will be given of a method of obtaining a correctly recognized word string by a search condition 2 which is an extension of the search condition 1.

FIG. 6 is a diagram showing a configuration of a speech recognition apparatus according to Embodiment 2 of the present invention. The same or corresponding components as those in FIG. 1 are denoted by the same reference numerals. In FIG. 6, reference numeral 11 denotes a candidate search unit (candidate search means) for searching for a next candidate word string according to search condition 2 described later.

FIG. 7 is a flowchart showing a voice recognition method according to the second embodiment of the present invention. Operations that are the same as or correspond to those in FIG. 2 are denoted by the same reference numerals. When the operations from step ST1 to step ST9 are performed in the same manner as in the first embodiment, the candidate search unit 11 reads the word string of the next candidate from the RAM 3 and, according to the search condition 2, sets 1 for each erroneously recognized character string. A word string that has a difference of characters or more and all of the same character strings other than the misrecognized character string are searched for and stored in the display area of the RAM 3 (step ST11, candidate search step).

A specific description will be given using utterance example B. FIG.
In the case of the display unit 5C, the specified misrecognition character string is {1, 1}, {3, 4}, and the place where only <urine> should be designated is mistakenly designated as <urinary>. I have.

Modified initial candidate "<Uri> in <car>"

The word string of the next candidate at this time is the same as in the first embodiment. Since "me" is at the same character position, a word string of correct recognition cannot be obtained according to the search condition 1, but a word satisfying the next search condition 2 obtained by expanding the search condition 1 with respect to the above-mentioned next candidate is used. The candidate search unit 11 searches.

Search condition 2 Character string not identical to <car> + “de” + character string not identical to <urinary> + “do”

Search condition 2 means that even if the same character string is included at the same character position as the misrecognized character string, if at least one character has been changed, it is regarded as a character change. . The first next candidate is as follows when compared with the initial candidate. A character string in which <Uri> includes the correct character “U”, but <Uri ”is not the same as“ Preparation ” Since the parts other than the misrecognized character string are the same, the search condition 2 is satisfied.

[0099] Initial candidate "<Uri> in <car>" First candidate "Prepare before coming"

The second next candidate is as follows when compared with the initial candidate. Since the <urinary> is a completely identical character string, it does not satisfy the search condition 2.

Initial candidate “<Uri> in <car>” Second candidate “Uri” until coming

Therefore, the candidate search unit 11 sets the RA
The first next candidate is stored in the display area of M3, and is displayed by the display unit 5. As described above, the candidate search unit 11 searches for the next candidate according to the search condition 2, so that even when the specified incorrectly recognized character string includes a correct character string,
It is possible to robustly correct a misrecognized word string to a correctly recognized word string.

In the case of utterance example C, the same processing can be performed. That is, it can be understood from the following comparison that even if an erroneously recognized character string is specified as the initial candidate word string as shown below, the fourth next candidate is searched from the search condition 2.

[0104] Initial candidate "<wife> <it> gave <money>" First candidate "wife came, give me money" Second candidate, "wife came, give me money" Third Next candidate "Send north wind to Tsu" Send fourth "Send gold to Tsu"

As described above, according to the second embodiment, when the character string of the next candidate corresponding to the character position of the misrecognized character string is not the same as the misrecognized character string, the character change of search condition 2 is performed. Since the definition is extended, even if a correct character string is included in the specified erroneously recognized character string, the erroneously recognized word string can be robustly corrected.

Embodiment 3 FIG. 8 is a diagram showing a configuration of a voice recognition device according to Embodiment 3 of the present invention. Figure 1
The same or corresponding components as in FIG. 6 are denoted by the same reference numerals. In FIG. 8, reference numeral 12 extracts the next candidate character string corresponding to the character position of the misrecognized character string from the next candidate satisfying the search condition 2, and displays the next candidate character string in accordance with the initial candidate misrecognized character string. A next candidate character string display unit (next candidate character string display means), and a next candidate character string selection unit (next candidate character string selection means) for correcting an erroneously recognized character string of an initial candidate into a next candidate character string by selecting a candidate. It is.

FIG. 9 is a flowchart showing a voice recognition method according to Embodiment 3 of the present invention. The same or corresponding operations as in FIGS. 2 and 7 are denoted by the same reference numerals.
When the operations from step ST1 to step ST11 are performed in the same manner as in the first and second embodiments, only the next candidate character string corresponding to the erroneously recognized character string is replaced with the next candidate character from the next candidate searched by the candidate search unit 11. The column display unit 12 displays it on the display unit (step ST12, next candidate character string display step).

In other words, the next candidate character string display section 12 displays each next candidate character string “come” which is a changed part of each erroneously recognized character string.
14A and "ready" 14B are read from the first next candidate in the RAM 3, and are displayed in accordance with the erroneously recognized character string as shown in a display unit (display means) 5D of FIG. When each of the displayed next candidate character strings “come” 14A and “prepared” 14B is selected with a mouse or a keyboard, the next candidate character string selection unit 13 is initialized as shown in a display unit (display unit) 5E in FIG. The candidate is corrected (step ST13, next candidate character string selecting step).

As described above, according to the third embodiment, the next candidate character strings "come" 14A and "prepare" 14B, which are the changed portions of each erroneously recognized character string, are displayed in the next candidate character string display section. 12 is displayed in accordance with the erroneously recognized character string, and the next candidate character string selecting unit 13 is selected by selecting the next candidate character string.
Corrects the initial candidate, so that an effect is obtained that the correction operation can be performed while comparing the incorrectly recognized character string with the next candidate character string. In particular, when a plurality of word strings satisfying the search condition are searched from the next candidate, the method according to the third embodiment is effective.

In the third embodiment, the second embodiment is used.
However, the candidate search unit 10 of the first embodiment may be used instead of the candidate search unit 11.

Further, each step of the voice recognition method shown in each embodiment may be replaced with each procedure and programmed to be recorded on a computer-readable recording medium, and the same effect can be obtained. .

[0112]

According to the present invention, a syllable string is generated by recognizing an input voice, a plurality of word strings corresponding to the syllable string are generated using an n-gram dictionary, and an initial candidate and a plurality of next word strings are generated. Speech recognition means for selecting candidates and storing initial candidates in a display area, display means for displaying a word string stored in the display area to a user, and misrecognized character strings in the initial candidates displayed on the display means When specified, the character string corresponding to the character position of the misrecognized character string is changed from the next candidate to a word string in which all the characters have been changed. Even if there are a plurality of erroneously recognized character strings, it is possible to easily correct the erroneously recognized word strings by a single correction operation without being aware of the word unit delimiter.

According to the present invention, when the character string of the next candidate corresponding to the character position of the misrecognized character string is not the same as the misrecognized character string, the candidate retrieving means expands and defines the character change. Therefore, even when a correct character string is included in the specified erroneously recognized character string, an effect that the erroneously recognized word string can be robustly corrected can be obtained.

According to the present invention, the display changing means for displaying the erroneously recognized character string is provided in any of the display modes of shaded display, inverted display, luminance change display, color change display, and animation display. In addition, an effect is obtained that an erroneously recognized character string can be displayed according to the user's preference and use.

According to the present invention, the next candidate character string corresponding to the character position of the erroneously recognized character string is extracted from the next candidate searched by the candidate search means, and is matched with the erroneously recognized character string together with the initial candidate of the display unit. Since next candidate character string display means for displaying the next candidate character string and next candidate character string selecting means for correcting the erroneously recognized character string of the initial candidate to the next candidate character string by the user's candidate selection are provided, The effect is obtained that the correction operation can be performed while visually checking the correspondence between the recognized character string and the next candidate character string.

According to the present invention, a syllable string is generated by recognizing an input voice, and a plurality of word strings corresponding to the syllable string are generated using an n-gram dictionary to be used as an initial candidate and a plurality of next candidates. A voice recognition step of selecting and storing an initial candidate in a display area; a display step of displaying a word string stored in the display area to a user; and a misrecognized character string in the initial candidate displayed in the display step. And a candidate search step of searching a next candidate for a word string in which only the character string corresponding to the character position of the misrecognized character string is changed from the next candidate and storing the word string in the display area. This makes it possible to easily correct an erroneously recognized word string with a single correction operation even when there are multiple erroneously recognized character strings without being aware of word-unit delimiters. That.

According to the present invention, in the candidate search step, when the character string of the next candidate corresponding to the character position of the misrecognized character string is not the same as the misrecognized character string, it is extended to define a character change. Therefore, even when a correct character string is included in the specified erroneously recognized character string, an effect that the erroneously recognized word string can be robustly corrected can be obtained.

According to the present invention, a display change step of displaying an erroneously recognized character string is provided in any of the display modes of shaded display, inverted display, luminance change display, color change display, and animation display. In addition, an effect is obtained that an erroneously recognized character string can be displayed according to the user's preference and use.

According to the present invention, the next candidate character string corresponding to the character position of the erroneously recognized character string is extracted from the next candidate searched in the candidate search step, and is matched with the erroneously recognized character string together with the initial candidate of the display unit. Since a next candidate character string displaying step of displaying the next candidate character string and a next candidate character string selecting step of correcting the erroneously recognized character string of the initial candidate to the next candidate character string by the user's selection of the candidate are provided. The effect is obtained that the correction operation can be performed while visually checking the correspondence between the recognized character string and the next candidate character string.

According to the present invention, a syllable string is generated by recognizing an input voice, and a plurality of word strings corresponding to the syllable string are generated using an n-gram dictionary to be an initial candidate and a plurality of next candidates. A voice recognition procedure for selecting and storing initial candidates in a display area, a display procedure for displaying a word string stored in the display area to a user, and a misrecognized character string in the initial candidates displayed in the display procedure are designated. And a candidate search step of searching a next candidate for a word string in which only the character string corresponding to the character position of the misrecognized character string has been changed from the next candidate and storing it in the display area. Therefore, even if there are a plurality of erroneously recognized character strings, it is possible to easily correct the erroneously recognized word strings by a single correction operation without being conscious of the delimitation in word units.

According to the present invention, in the candidate search procedure, when the character string of the next candidate corresponding to the character position of the misrecognized character string is not the same as the misrecognized character string, it is extended and defined as character change. Therefore, even when a correct character string is included in the specified erroneously recognized character string, an effect that the erroneously recognized word string can be robustly corrected can be obtained.

According to the present invention, a display change procedure for displaying an erroneously recognized character string is provided in any of the display modes of shaded display, inverted display, luminance change display, color change display, and animation display. In addition, an effect is obtained that an erroneously recognized character string can be displayed according to the user's preference and use.

According to the present invention, the next candidate character string corresponding to the character position of the misrecognized character string is extracted from the next candidate searched in the candidate search procedure, and is matched with the misrecognized character string together with the initial candidate of the display unit. A next candidate character string display procedure for displaying the next candidate character string and a next candidate character string selection procedure for correcting the misrecognized character string of the initial candidate to the next candidate character string according to the user's candidate selection are provided. The effect is obtained that the correction operation can be performed while visually checking the correspondence between the recognized character string and the next candidate character string.

[Brief description of the drawings]

FIG. 1 is a diagram showing a configuration of a speech recognition device according to a first embodiment of the present invention.

FIG. 2 is a flowchart showing a speech recognition method according to Embodiment 1 of the present invention.

FIG. 3 shows n−g in the case of n = 1 and n = 2
It is a figure showing an example of a ram dictionary.

FIG. 4 is a diagram in which character strings “car” and “urine” are designated as erroneously recognized character strings.

FIG. 5 is a diagram in which character strings “car” and “urine” are designated as erroneously recognized character strings.

FIG. 6 is a diagram showing a configuration of a voice recognition device according to a second embodiment of the present invention.

FIG. 7 is a flowchart showing a voice recognition method according to Embodiment 2 of the present invention.

FIG. 8 is a diagram showing a configuration of a voice recognition device according to a third embodiment of the present invention.

FIG. 9 is a flowchart showing a voice recognition method according to Embodiment 3 of the present invention.

FIG. 10 is a diagram in which respective next candidate character strings “come” and “ready” are displayed in accordance with the respective erroneously recognized character strings <car> and <urgency>.

FIG. 11 is a diagram showing a configuration of a conventional voice recognition device.

FIG. 12 is a diagram showing a state in which a misrecognized word “recognition” by a conventional speech recognition device is corrected to a correct recognition word “recognition”.

[Explanation of symbols]

1 microphone (voice recognition unit), 2 voice recognition unit (voice recognition unit), 3 random access memory (voice recognition unit), 4 n-gram dictionary storage unit (voice recognition unit),
5, 5A, 5B, 5C, 5D, 5E display unit (display means), 6 keyboard, 7 mouse, 8 range designation unit,
9 display change section (display change means), 10, 11 candidate search section (candidate search means), 12th candidate character string display section (next candidate character string display means), 13th candidate character string selection section (next candidate character string) Selection means).

 ──────────────────────────────────────────────────続 き Continuing from the front page (72) Inventor Yuzo Maruta 2-3-2 Marunouchi, Chiyoda-ku, Tokyo Within Mitsubishi Electric Corporation (72) Inventor Shinobu Arai 2-3-2 Marunouchi, Chiyoda-ku, Tokyo Rishi Electric Co., Ltd. F term (reference) 5D015 HH12 HH16 KK03 LL04 LL08 5D045 AB02

Claims (12)

[Claims] 1. A speech recognition apparatus for displaying a word string as a speech recognition result of an input speech to a user and correcting a misrecognized character string included in the word string. Generate a sequence, n-gr
a voice recognition unit that generates a plurality of word strings corresponding to the syllable strings using an am dictionary, selects initial candidates and a plurality of next candidates, and stores the initial candidates in a display area; Display means for displaying the generated word string to a user, and when a misrecognized character string in the initial candidate displayed on the display means is designated, only a character string corresponding to the character position of the misrecognized character string A candidate retrieving means for retrieving a word string in which all characters have been changed from the next candidate and storing it in the display area. 2. The method according to claim 1, wherein the candidate searching means defines the character string of the next candidate corresponding to the character position of the misrecognized character string as a character change when the character string is not the same as the misrecognized character string. Item 2. The speech recognition device according to Item 1. 3. A shaded display, an inverted display, a brightness change display,
3. The speech recognition apparatus according to claim 1, further comprising a display change unit that displays an erroneously recognized character string in one of a color change display and an animation display. 4. A next candidate character string corresponding to the character position of the erroneously recognized character string is extracted from the next candidate searched by the candidate search means, and the next candidate is displayed together with the initial candidate of the display unit in accordance with the erroneously recognized character string. Next candidate character string display means for displaying a character string, and next candidate character string selection means for correcting the misrecognized character string of the initial candidate to the next candidate character string by user selection. Claim 1 or Claim 2
The speech recognition device according to the above. 5. A speech recognition method for displaying to a user a word string that is a speech recognition result of an input speech and correcting a misrecognized character string included in the word string. Generate a sequence, n-gr
a voice recognition step of generating a plurality of word strings corresponding to the syllable strings using an am dictionary, selecting the word strings as initial candidates and a plurality of next candidates, and storing the initial candidates in a display area; A display step of displaying the generated word string to a user; a misrecognized character string designating step in which a misrecognized character string in the initial candidate displayed in the display step is designated; and a character position of the misrecognized character string. And a candidate searching step of searching a next candidate for a word string in which only the character string corresponding to is changed from the next candidate and storing the word string in the display area. 6. In the candidate search step, when the character string of the next candidate corresponding to the character position of the misrecognized character string is not the same as the misrecognized character string, the character string is extended by defining a character change. Item 6. The voice recognition method according to Item 5. 7. A shaded display, an inverted display, a brightness change display,
7. The voice recognition method according to claim 5, further comprising a display change step of displaying an erroneously recognized character string in one of a color change display mode and an animation display mode. 8. A next candidate character string corresponding to the character position of the erroneously recognized character string is extracted from the next candidate searched in the candidate searching step, and the next candidate is displayed together with the initial candidate of the display unit in accordance with the erroneously recognized character string. A next candidate character string displaying step of displaying a character string; and a next candidate character string selecting step of correcting the misrecognized character string of the initial candidate to the next candidate character string by selecting a candidate. The speech recognition method according to claim 5. 9. A computer-readable recording medium on which a speech recognition program for displaying a word string as a speech recognition result of an input speech to a user and correcting an erroneously recognized character string included in the word string is recorded. A syllable string is generated by speech recognition of the input speech, and n-gr
a voice recognition procedure for generating a plurality of word strings corresponding to the syllable strings using an am dictionary, selecting the word candidates as initial candidates and a plurality of next candidates, and storing the initial candidates in a display area; A display procedure for displaying the word string thus displayed to the user, a misrecognition string specification procedure for specifying a misrecognition string in the initial candidate displayed in the display procedure, and a character position of the misrecognition string. And a candidate search procedure for searching the next candidate for a word string in which only the character string corresponding to is changed from the next candidate and storing the word string in the display area. Medium. 10. In the candidate search procedure, when the character string of the next candidate corresponding to the character position of the misrecognized character string is not the same as the misrecognized character string, the character string is extended and defined as character change. Item 10. A computer-readable recording medium recording the voice recognition program according to Item 9. 11. A display changing procedure for displaying a misrecognized character string in any one of a display mode of shading display, reverse display, luminance change display, color change display, and animation display. Or claim 10
A computer-readable recording medium on which the speech recognition program described above is recorded. 12. A next candidate character string corresponding to the character position of the misrecognized character string is extracted from the next candidate searched in the candidate search procedure, and the next candidate is displayed together with the initial candidate of the display unit in accordance with the misrecognized character string. A next candidate character string displaying step of displaying a character string; and a next candidate character string selecting step of correcting the misrecognized character string of the initial candidate into the next candidate character string by selecting a candidate. Claim 9 or Claim 1
A computer-readable recording medium on which the speech recognition program described in Item 0 is recorded.