JP2006267319A - Support system for converting voice to writing, method thereof, and system for determination of correction part - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To make it possible to correctly and efficiently correct an erroneous recognition part of a voice recognition result in converting voice to writing using voice recognition. <P>SOLUTION: A display reproducing part 141 displays a recognition result 133 of a voice recognition device 102 on a display device 104 and also reproduces an original voice signal to output it from a loudspeaker 105. Discovering the erroneous recognition part of the voice recognition result, a user inputs a right answer character string for correcting the error from a keyboard 106. A correction part determining means 143 retrieves a character string part similar to the right answer character string in pronunciation inputted through the right answer input means 142 from the voice recognition result as a correction part. A recognition result correction means 144 replaces the determined correction part with the right answer character string. <P>COPYRIGHT: (C)2007,JPO&INPIT

Description

  The present invention relates to a speech transcription support apparatus and method thereof, and more particularly to speech writing for recognizing speech and converting it into text by correcting a recognition error included in a recognition result obtained as a result with a correct character string input from a keyboard. The present invention relates to a wake-up support apparatus and method.

  Although the speech recognition accuracy has been remarkably improved due to recent improvements in speech recognition technology, it is not possible to completely eliminate misrecognition due to the influence of ambient noise and the like. For this reason, in the transcription work using the speech recognition result, it is necessary to correct the misrecognized portion included in the speech recognition result.

  An example of a conventional technique for correcting a speech recognition result is described in Patent Document 1. In this prior art, when correcting the recognition result of the first input voice, the user utters the same phrase as the second voice. For example, when it is recognized that “the racket is a count” with respect to the first input voice “Do you want to buy a ticket”, the user repeats the same phrase “Do you want to buy a ticket”? On the device side, between the first and second voices, a similar part of feature information and a part that is not similar are detected, and at the same time, an emphasized utterance part is detected. The recognition result of the first input speech is excluded from the recognition candidates, and the second input speech portion that is uttered by emphasizing the dissimilar portion with the first input speech adopts the first recognition candidate of the second input speech. As a result, for example, when the speaker emphasizes the part of “I want to buy” in the second input voice, the recognition result of the second input voice is explained as “Do you want to buy a ticket?” ing.

  Another conventional technique for correcting a speech recognition result is described in Patent Document 2. In this prior art, when there is a misrecognized part in the recognition result of the first input voice, the user utters one or a plurality of words to replace the misrecognized part as the second input voice. On the device side, a speech recognition dictionary is selected from a plurality of corrected recognition results obtained by exhaustively replacing different portions of the recognition result of the first input speech with one or more words of the recognition result of the second input speech. The misrecognized portion is corrected by presenting only the listed items to the user and selecting one of them.

Another conventional technique for correcting a speech recognition result is described in Patent Document 3. In this prior art, a speech recognition error is found and corrected by a plurality of correction terminals. The user of each correction terminal points out an erroneous part of speech recognition by touching the touch panel, and corrects it by inputting a correct character string from the keyboard.
JP 2003-316386 A JP 2001-92493 A JP 2004-226910 A

  As described above, various technologies have been proposed as a technology for assisting the user in correcting the error part of the speech recognition result. However, as seen in Patent Document 1 and Patent Document 2, correction is performed by voice input. In the case of a method, there exists a subject that it is difficult to correct correctly. The reason for this is that the correct character string to be corrected is given by voice input, so even if it is emphasized and spoken, the correct character string may not be correctly input due to misrecognition, and a word (column that is not in the recognition dictionary) This is because voice recognition cannot be input in the first place.

  For this reason, when correct correction is performed, the correct character string must be input from the keyboard as shown in Patent Document 3, but in that case, both the designation of the misrecognized portion and the input of the correct character string must be performed. There is a problem that it needs to be done and takes time and effort.

  The present invention has been proposed in view of such circumstances, and an object of the present invention is to provide a speech transcription support apparatus and method that can correct a misrecognition portion of a speech recognition result correctly and efficiently. It is in.

  The first speech transcription support device according to the present invention includes a storage means for storing a speech recognition result obtained by subjecting a speech signal to speech recognition processing, and a correct answer for correcting a recognition error portion of the speech recognition result. Correct input means for inputting a character string from a keyboard; correction location determination means for searching a character string portion whose pronunciation is similar to the correct character string as a correction location in the speech recognition result; and the determined correction location And a recognition result correcting means for replacing with a correct character string.

  According to a second speech transcription support device of the present invention, in the first speech transcription support device, the correction location determination means includes speech data created from the correct character string and speech data corresponding to the speech recognition result. The character string portion whose pronunciation is similar to that of the correct character string is searched for by matching between the character strings.

  According to a third speech transcription support device of the present invention, in the second speech transcription support device, a speech standard pattern generated from the correct character string is used as speech data of the correct character string, and the speech recognition result The voice signal analysis result corresponding to the voice recognition result is used as voice data.

  According to a fourth speech transcription support device of the present invention, in the second speech transcription support device, a speech signal generated by speech synthesis from the correct character string is used as speech data of the correct character string, and the speech recognition is performed. A voice signal corresponding to the voice recognition result is used as the resulting voice data.

  According to a fifth speech transcription support device of the present invention, in the first speech transcription support device, the correction location determination means includes a character string that is a phoneme string or a syllable string created from the correct character string and the voice recognition. A character string portion similar in pronunciation to the correct answer character string is searched for by matching with a character string that is a phoneme string or a syllable string created from the result.

  According to a sixth speech transcription support device of the present invention, in the first speech transcription support device, the correction location determination means does not include a character string portion whose pronunciation is similar to the correct character string in the speech recognition result. In this case, a position with a high probability that the correct character string is inserted is searched as an insertion position using a language model in which word connection constraints are described, and the recognition result correcting means is set at the determined insertion position. The correct character string is inserted.

  According to a seventh speech transcription support device of the present invention, in the first speech transcription support device, the correction location determination means searches for a correction location according to the order of correct character strings input from a keyboard, and the recognition result The search range is successively narrowed based on the correction position corrected by the correction means.

  In the speech transcription support method of the present invention, the display / playback unit displays a speech recognition result obtained by performing speech recognition processing on the speech signal on the display device, and the correct input means recognizes the recognition error of the speech recognition result. The correct character string for correcting the part is input from the keyboard, and the correction part determining means searches for the character string part whose pronunciation is similar to the correct character string in the voice recognition result as the correction part, and the recognition result correcting means Is characterized in that the determined correction portion is replaced with the correct character string.

  The first correction location determination apparatus of the present invention converts a correct character string input from a keyboard into a phoneme string in order to correct a recognition error part of a voice recognition result obtained by performing voice recognition processing on a voice signal. A phoneme string conversion means for storing in the storage unit, and reading out the phoneme string of the correct character string from the storage unit, searching the standard pattern storage unit for storing the standard pattern for each phoneme, A phoneme standard pattern conversion means for converting to a standard pattern and storing it in the storage unit and a candidate for a correction location on the analysis result of the speech signal corresponding to the speech recognition result are comprehensively determined. Position determining means for storing position information in the storage unit; and reading out the standard pattern of the correct character string, the analysis result, and the position information of each correction location candidate from the storage unit, and the standard pattern of the correct character string Calculates the probability of outputting the analysis result part on the analysis result specified by the position information of the correction location candidate for each correction location candidate, and stores the probability calculation means for storing in the storage unit, and each correction from the storage unit And selecting means for reading out and comparing the probabilities calculated for each location candidate and selecting and outputting a correction location from the plurality of correction location candidates.

  The second correction location determination apparatus of the present invention provides a correct character string input from a keyboard for correcting a recognition error portion of a speech recognition result obtained by performing speech recognition processing on a speech signal, and the speech recognition result. A phoneme string conversion means for converting each of the character strings into a phoneme string or a syllable string and storing it in the storage unit; and reading out the phoneme string or syllable string of the speech recognition result from the storage unit; A position determination unit that comprehensively obtains candidates for correction points and stores the position information of each correction point candidate in the storage unit, and the correct character string and the phoneme string or syllable string of the speech recognition result from the storage unit And the position information of each correction part candidate, and the phoneme string or syllable string part of the speech recognition result specified by the position information of the correction part candidate and the phoneme string or sound of the correct character string The similarity calculation means for calculating the similarity to the column and storing it in the storage unit, and reading out and comparing the similarity calculated for each correction location candidate from the storage unit, the plurality of correction location candidates And selecting means for selecting and outputting a correction portion from the inside.

"Action"
The misrecognized portion included in the recognition result of the speech recognition apparatus is unlikely to be completely different from the correct answer, and is similar to the correct answer. That is, the misrecognized part and the correct answer are often similar in pronunciation. The present invention pays attention to this point, and retrieves a character string portion whose pronunciation is similar to that of the correct character string as a corrected portion in the speech recognition result. The recognition result portion that is phonetically similar to the correct character string can be detected by voice signal similarity detection or phonological similarity detection. The former similarity detection of speech signals is based on speech signals such as speech standard patterns and speech waveform signals created from correct character strings, analysis results obtained by speech recognition devices in the process of obtaining speech recognition results, and original speech signals. Is possible by comparing The latter phonological similarity detection can be performed by converting the correct character string and the character string of the speech recognition result into a phoneme string and a syllable string, respectively, and comparing them.

  According to the present invention, when a correct character string of a misrecognized part is input with a keyboard in a transcription work using a speech recognition result, the misrecognized part is automatically detected and can be corrected to a correct character string, thereby improving work efficiency. Can be made.

  Next, a first embodiment of the present invention will be described in detail with reference to the drawings.

  Referring to FIG. 1, the speech transcription support device according to the first exemplary embodiment of the present invention includes a speech recognition device 102 that performs speech recognition processing on speech 101, a speech recognition result of the speech recognition device 102, and the like. A display device 104 for displaying a voice recognition result or the like to a user, a speaker 105 for reproducing and outputting the voice 101 as a voice recognition target for confirmation by the user, and voice recognition In order to correct an error part included in the result, the keyboard 106 for inputting a correct character string and a processing device 107 connected thereto are configured.

  The voice recognition device 102 includes a voice input unit 111, an analysis unit 112, and a collation unit 113, and a storage device 124 that stores a dictionary 121, a language model 122, and an acoustic model 123. In the dictionary 121, words to be recognized and their pronunciation are registered. The language model 122 describes connection restrictions between words to be recognized. An example of a language model is a word N-gram model that is widely used for speech recognition. The acoustic model 123 describes acoustic features of recognition units. As an example of the acoustic model, there is an HMM (Hidden Markov Model) widely used for speech recognition.

  The voice input unit 111 captures the voice 101 using a microphone or the like, and stores a voice signal (digital data of a voice waveform) 131 processed into a form that can be analyzed by the analysis unit 112 in the storage device 103.

  The analysis unit 112 receives the audio signal 131 from the storage device 103 and performs an analysis process on the audio signal 131 to extract features related to the frequency spectrum. Specifically, a discrete Fourier transform is performed using a fast Fourier transform algorithm over a time window to obtain a short-time spectrum. The analysis result 132 obtained by the analysis unit 112 is stored in the storage device 103 in association with the audio signal 131 by time data or the like so that it can be clearly seen which part of the audio signal 131 is the analysis result.

  The matching unit 113 uses the dictionary 121, the language model 122, and the acoustic model 123 stored in the storage device 124 to calculate the probability between the speech signal analysis result 132 input from the storage device 103 and the candidate word string. And the word string having the highest likelihood is output as a recognition result. The recognition result 133 obtained by the collating unit 113 is stored in the storage device 103 in association with the analysis result 132. Specifically, for each word or individual syllable or phoneme that is a component of the word, information indicating the correspondence with the analysis result 132, that is, which part of the analysis result from which word is added. And stored in the storage device 103.

  The processing device 107 is a device for the user to find and correct an error location included in the recognition result 133 of the speech 101 stored in the storage device 103. The display device 141, correct input means 142, correction location Determination means 143 and recognition result correction means 144 are provided.

  The display reproduction unit 141 reproduces the recognition result display unit 145 that displays the recognition result 133 stored in the storage device 103 on the display device 104 and the sound signal 131 stored in the storage device 103 and outputs the sound from the speaker 105. Reproduction means 146.

  The correct answer input unit 142 takes in a correct character string for correcting erroneous recognition from the keyboard 106.

  The correction part determination unit 143 searches the recognition result 133 stored in the storage device 103 as a correction part for a part whose pronunciation is similar to the correct character string taken in by the correct input unit 142. A search for a recognition result portion that is phonetically similar to the correct character string is performed by speech signal similarity detection or phonological similarity detection, as will be described later.

  The recognition result correction unit 144 automatically corrects the recognition result 133 stored in the storage device 103 based on the correct character string captured by the correct input unit 142 and the correction part determined by the correction part determination unit 143. Or prompt the user to make corrections.

  An example of a screen displayed on the display device 104 by the processing device 107 is shown in FIG. When the playback button 201 is turned on in the screen of FIG. 2, the display / playback unit 141 reads out and plays back the audio signal 131 stored in the storage device 103 by the audio playback unit 146, and outputs the playback audio from the speaker 105. In synchronization with this sound output, the recognition result display means 145 displays the recognition result 133 of the sound signal being reproduced in the recognition result display column 202. The playback button 201 also serves as a stop button. When it is turned on again, the playback operation is stopped.

  The recognition result display field 202 is composed of a plurality of lines, and the currently reproduced line is highlighted as indicated by a broken line, and is enlarged and displayed in the attention line display field 203, and the sound waveform image is displayed in the waveform display field 204. Is displayed. Around the playback button 201, there are a rewind button 205, a fast forward button 206, and a volume adjustment slider 207, which can rewind, fast forward, and adjust the volume of the playback position, respectively. The time displayed in the time display column 208 indicates the start / end time of the audio signal corresponding to the speech recognition result displayed in the attention line display column 203. The video display column 209 is a column for displaying a reproduced video when there is a video accompanying audio.

  Also, as a GUI component for correcting the recognition result, a correct character string input field 210, an OK button 211, a return button 212 and a feed button 213 that advance the attention line to the previous line or the next line, attention line There is a delete button 214 for deleting the entire recognition result. In addition, in the recognition result displayed in the attention line display field 203, the correction location determined by the correction location determination means 143 is highlighted and displayed as indicated by a broken line 215.

  In the screen of FIG. 2, a recognition result “From the beginning of President Hussein ’s hometown is low” is displayed in the attention line display field 203. This is a result of erroneously recognizing the “ticlit” portion as “low once” when recognized by the recognition device 102. Such a misrecognized portion is easily found by the user turning on the playback button 201 and listening to the original audio signal 131 with his / her ear and comparing it with the recognition result displayed in the attention line display field 203. be able to.

  Conventionally, the correction of such a misrecognized portion had to be specified by the user himself. However, in the present embodiment, “ticlit” is entered in the correct character string input field 210 as the correct character string. When input, the correction location determination means 143 makes a “low once” location that is pronouncedly similar to the correct character string “ticlit” among the recognition result character strings displayed in the target line display field 203. Is detected, and this detected portion is presented to the user as a corrected portion. For this reason, the user can save the trouble of designating the correction part himself / herself, and can efficiently perform the correction work. If the detected correction location is incorrect, the correct correction location can be specified on the attention line display field 203 with a mouse cursor or the like. In this case, the user himself / herself specifies the correction location from the beginning. The same as when you do.

  Next, some examples of the present embodiment will be described in detail with a focus on the correction location determination unit 143 and the recognition result correction unit 144.

  Referring to FIG. 3, the correction location determination unit 143 according to the present embodiment stores the correct character string 301 input in the correct character string input field 210 and the recognition result displayed in the attention line display field 203. 103, the analysis result 132 memorize | stored in 103 is inputted, and for every correction part candidate set up comprehensively on the analysis result 132, the analysis result part of the correction part candidate and the sound signal similarity of the correct character string 301 A search unit 302 that calculates a score indicating a degree, and one or a plurality of correction points are selected from a plurality of correction point candidates obtained by the search unit 302 based on the similarity score, and the recognition result correction unit 144 It comprises a selection means 303 for notifying the correction location on the recognition result.

  The search unit 302 includes a standard pattern creation unit 311 and a score calculation / position detection unit 312.

  The standard pattern creation unit 311 creates a standard pattern of speech corresponding to the correct character string 301. Specifically, the correct character string 301 is converted into a phoneme string, phoneme standard patterns prepared in advance are arranged in that order, and a standard pattern of the character string is created. A configuration example of the standard pattern creation means 311 is shown in FIG.

  Referring to FIG. 4, the standard pattern creation unit 311 includes a first storage unit 401, a second storage unit 402, an input unit 403, a phoneme string conversion unit 404, a phoneme standard pattern conversion unit 405, and a third storage unit 406. The The input unit 403 inputs the correct character string 301 and stores it in the first storage unit 401. The phoneme string conversion unit 404 inputs the characters in the correct character string 301 from the first storage unit 401 in order from the top, and searches the third storage unit 406 that stores in advance a dictionary describing words and their pronunciations, that is, phoneme sequences. Then, each character is converted into a phoneme string and stored in the first storage unit 401 as a phoneme string 411. The phoneme standard pattern conversion unit 405 inputs the phonemes in the phoneme string 411 from the first storage unit 401 in order from the head, and corresponds to the phoneme from the second storage unit 402 storing the standard pattern for each phoneme. By retrieving the phoneme standard pattern and repeating the process of outputting it to the first storage unit 401, a standard pattern 412 corresponding to the correct character string 301 is generated on the first storage unit 401. Here, the phoneme standard pattern stored in advance in the second storage unit 402 uses the same type of pattern as the analysis result 132 (for example, the above-described HMM) so that the similarity with the analysis result 132 can be calculated.

  For example, in the case of “ticlit” input in the correct character string input field 210 in FIG. 2, “ticlit” is converted into a phoneme string “tikuri Qt to” as shown in FIG. The phoneme standard patterns are arranged in that order, and a standard pattern representing “ticlit” is created.

  The score calculation / position detection means 312 is stored in the storage device 103 in correspondence with the standard pattern 412 of the correct character string created by the standard pattern creation means 311 and the recognition result displayed in the attention line display field 203. With the analysis result 132 as an input, the probability (score) that the standard pattern outputs a correct character string when a certain analysis result part is given is calculated and output. A configuration example of the score calculation / position detection means 312 is shown in FIG.

  Referring to FIG. 6, the score calculation / position detection unit 312 includes a storage unit 601, a first input unit 602, a second input unit 603, a position determination unit 604, and a probability calculation unit 605. The first input unit 602 reads the standard pattern 412 of the correct character string created by the standard pattern creation unit 311 from the first storage unit 401 of FIG. 4 and stores it in the storage unit 601. The second input unit 603 reads the analysis result 132 corresponding to the recognition result displayed in the attention line display field 203 from the storage device 103 and stores it in the storage unit 601. The position determination unit 604 reads the analysis result 132 from the storage unit 601, comprehensively obtains candidates for correction locations on the analysis result 132, and sets the correction location candidate position 611, the score 612, and the processed flag 613 in each row. A list 614 in which a set is set is generated and stored in the storage unit 601. At this time, each score 612 is NULL, and the processed flag 613 indicates unprocessed. The probability calculation unit 605 reads out each row of the list 614 from the storage unit 601 one by one, reads out the analysis result part indicated by the position 611 of the corrected portion candidate of the read out row from the analysis result 132 of the storage unit 601, and stores the storage unit 601. The probability (score) that the standard pattern 412 read out from the output of the read analysis result part is calculated and written in the score 612, and the processed flag 613 is set as processed. The process ends when all the flags 613 in the list 614 have been processed.

  The selection unit 303 includes a score comparison unit 313 and a position conversion unit 314 as shown in FIG.

  The score comparison unit 313 reads the list 614 created by the search unit 302 from the storage unit 601, compares the score 612 with a preset threshold value for each row, and corrects a corrected portion candidate for a row having a score equal to or higher than the threshold value. The position 611 is selected as a correction location and output to the position conversion unit 314. Here, the position 611 of the correction part candidate having a score equal to or higher than a preset threshold is selected as the correction part, but the position 611 of the correction part candidate having the highest score in the list 614 is selected as the correction part. It may be.

  Based on the correspondence between the analysis result 132 and the recognition result 133 in the storage device 103, the position conversion unit 314 converts the correction part on the analysis result 132 output from the score comparison unit 313 to the correction part on the recognition result 133. And output to the recognition result correction means 144.

  A configuration example of the recognition result correction unit 144 is shown in FIG. Referring to FIG. 7, the recognition result correction unit 144 includes a storage unit 701, a first input unit 702, a second input unit 703, a third input unit 704, a correction location presentation unit 705, and a character string operation unit 706. . The first input unit 702 reads the correct character string 301 from, for example, the first storage unit 401 in FIG. 4 and stores it in the storage unit 701. The second input unit 703 reads the recognition result 133 displayed in the attention line display field 203 in FIG. 2 from the storage device 103 and stores it in the storage unit 701. The third input unit 704 inputs the correction position 711 on the recognition result output from the position conversion unit 314 in the selection unit 303 of the correction point determination unit 143 and stores it in the storage unit 701. The correction location presenting means 705 reads the correction location 711 from the storage unit 701, highlights and displays the portion indicated by the correction location 711 in the recognition result displayed in the attention line display field 203 of FIG. Present corrections to the person. When the OK button 211 in FIG. 2 is turned on, the character string operation unit 706 reads the correct character string 301, the recognition result 133, and the correction position 711 from the storage unit 701, and the correction on the recognition result 133 indicated by the correction position 711. The character string of the place is replaced with the correct character string 301, and the recognition result after the replacement is written back to the storage device 103, and at the same time, the display in the attention line display column 203 in FIG. When the user changes the correction position on the attention line display field 203 by a mouse operation or the like, the third input unit 703 updates the correction position 711 in the storage unit 701 with the changed correction position.

  Next, for example, the voice recognition result used when creating a speech transcript of “from President Hussein's hometown Tikrit” is displayed in the attention line display column 203 of FIG. Refer to the flowchart of FIG. 8 for the operation after the user inputs “ticlit” in the correct character string input field 210 so that the user can change “low once” to “ticlit”. To explain.

  When the correct character string is input to the correct character string input field 210, the correct answer input unit 142 of the processing device 107 captures the correct character string and transmits it to the correction location determination unit 143 (step S801 in FIG. 8). As shown in FIG. 5, the standard pattern creation means 311 in the search means 302 of the correction location determination means 143 converts “ticlit” into a phoneme string “tikuri Qt to” and prepares a phoneme standard pattern prepared in advance. Are arranged in that order to create a standard pattern expressing “ticlit” (step S802). Next, the score calculation / position detection means 312 receives the analysis result corresponding to the standard pattern of “Tikrit” and the recognition result “From the time when President Hussein's hometown is low” in the analysis result 132 held in the storage device 103. And the probability (score) of outputting a correct character string (ticlit) when each analysis result part is given to the standard pattern is calculated (step S803). In this case, the standard pattern has the highest score in the analysis result of the part uttered as “ticlit” (the part that is erroneously recognized as “low once” in the recognition result), and the other part has a low score. .

  Next, the score comparison unit 313 of the selection unit 303 performs comparison with a score prepared in advance as a threshold value, and the position conversion unit 314 determines the position on the recognition result corresponding to the analysis result portion having a score equal to or higher than the threshold value. Output as the correction position. As the threshold value, a value is set such that if the score is equal to or higher than that value, the portion can be regarded as a correct character string. Or you may make it output the positional information on the part in order with a high score. Here, it is assumed that there is only one score that satisfies the threshold, and only the portion that is uttered “ticlit” (the misrecognition “low once” portion).

  The recognition result correction unit 144 focuses the recognition result portion “low once” corresponding to the correction position output by the selection unit 303 among the recognition results displayed in the attention line display field 203, and corrects the correction position to the user. Is presented (steps S804 and S805). When the user turns on the OK button 211, the recognition result correcting unit 144 replaces the focused recognition result portion (the erroneous recognition result “low once”) with the correct character string “ticlit” (step). S806).

  In addition, when there are a plurality of scores satisfying the threshold or when arranging in the order of the highest score, the conversion may be performed while confirming with the user whether or not the conversion may be sequentially performed in the order of the highest score. .

  Next, the effect of the present embodiment will be described.

  If a correct character string is entered, a recognition result part whose pronunciation is similar to that of the correct character string is automatically set as the correction part, so that the user himself does not have to specify the correction part. The burden can be reduced.

  Since the correct character string is input from the keyboard instead of voice input, the correct character string of the misrecognized part can be given accurately, and a correct answer can be given even for a word (string) whose correct character string is not in the recognition dictionary. . In addition, anyone can make correction work, not just the original voice speaker.

  Since the search for a portion whose pronunciation is similar to the correct character string is limited to a part of the recognition result (the portion displayed in the target line display field 203), the correction portion can be detected quickly and accurately. .

  The correction location determination means 143 of the present embodiment converts the correct character string into a voice standard pattern, and determines the similarity of the voice signal between the standard pattern and the analysis result 132 on the storage device 103. Thus, it is also possible to determine a correction location by generating a speech waveform from the correct character string and determining phonological similarity between the speech waveform and the speech signal 131 on the storage device 103.

  In addition, the correction point determination unit 143 according to the present embodiment reads the analysis result to be compared with the standard pattern of the correct character string from the storage device 103, but the same as the analysis unit 112 from the audio signal 131 stored in the storage device 103. It is also possible for the correction location determination means 143 itself to generate and use the analysis result in the processing. Alternatively, it is also possible to create and use a standard voice pattern from a voice recognition result character string as if a standard pattern was created from a correct character string. According to these configurations, the speech recognition apparatus 102 does not need to output the analysis result 132 that is an intermediate result of the speech recognition processing to the storage device 103.

  In addition, the recognition result correction unit 144 of the present embodiment presents the correction location determined by the correction location determination unit 143 to the user and corrects the recognition result after obtaining confirmation from the user. It is also possible to omit the presentation and confirmation by the user, and the recognition result correction unit 144 may automatically replace the determined correction part with the correct character string.

  In this embodiment, the user inputs the correct answer only for the misrecognized part. However, the user may input a character string composed of the misrecognized part and the character before or after it as the correct character string. it can. For example, in the above-described example, it is possible to input a part that has been correctly recognized, such as “from hometown Tikrit”. This has the advantage that the determination of the correction location becomes more accurate. In this case, the recognition result correcting means 144 corrects the entire “from hometown low once” to “from hometown Tikrit” or the part of the correct answer string that does not match the recognition result “low once” to “ticlit. It is arbitrary whether to correct it.

  Referring to FIG. 9, the correction location determination means 143 in the present embodiment receives the correct character string 301 input in the correct character string input field 210 and the recognition result 133 displayed in the attention line display field 203 as inputs. Search means 902 for calculating a score indicating the phonological similarity between the recognition result portion of the correction location candidate and the correct character string 301 for each correction location candidate comprehensively set on the recognition result 133, and the search means A selection unit 903 that selects one or a plurality of correction points from the plurality of correction point candidates obtained in 902 based on the similarity score, and notifies the recognition result correction unit 144 of the correction points on the recognition result. It is configured.

  The search unit 902 includes a character string creation unit 911 and a score calculation / position detection unit 912.

  The character string creating unit 911 receives the correct character string 301 and the recognition result 133 as input, and converts both into a phoneme string so that the correct character string 301 and the recognition result 133 can be matched. A configuration example of the character string creating unit 911 is shown in FIG.

  Referring to FIG. 10, the character string creating unit 911 includes a first storage unit 1001, a second storage unit 1002, a first input unit 1003, a second input unit 1004, and a phoneme string conversion unit 1005. The first input unit 1003 inputs the correct character string 301 and stores it in the first storage unit 1001. The second input unit 1004 inputs the recognition result 133 displayed in the attention line display field 203 and stores it in the first storage unit 1001. The phoneme string conversion unit 1005 inputs the characters in the correct character string 301 from the first storage unit 1001 in order from the top, and searches the second storage unit 1002 that stores in advance a dictionary describing words and their pronunciations, that is, phoneme sequences. Then, each character is converted into a phoneme string and stored in the first storage unit 1001 as a phoneme string 1011 of a correct character string. Further, the phoneme sequence conversion unit 1005 inputs characters in the recognition result 133 from the first storage unit 1001 in order from the top, searches the dictionary of the second storage unit 1002 to convert each character into a phoneme sequence, One storage unit 1001 stores the recognition result as a phoneme string 1012.

  The score calculation / position detection unit 912 receives the phoneme string 1011 of the correct character string created by the character string creating unit 911 and the phoneme string 1012 of the recognition result as input, and correct characters for each part of the phoneme string 1012 of the recognition result The similarity (score) with the phoneme sequence 1011 is calculated and output. A configuration example of the score calculation / position detection means 912 is shown in FIG.

  Referring to FIG. 11, the score calculation / position detection means 912 includes a storage unit 1101, a first input unit 1102, a second input unit 1103, a position determination unit 1104, and a similarity calculation unit 1105. The first input unit 1102 reads the phoneme string 1011 of the correct character string created by the character string creating unit 911 from the first storage unit 1001 of FIG. 10 and stores it in the storage unit 1101. The second input unit 1103 reads out the phoneme string 1012 of the recognition result created by the character string creating unit 911 from the first storage unit 1001 of FIG. 10 and stores it in the storage unit 1101. The position determination unit 1104 reads out the phoneme string 1012 of the recognition result from the storage unit 1101, comprehensively obtains candidates for correction points on the phoneme string 1012, and has processed the position 1111 and the score 1112 of the correction point candidates in each row. A list 1114 in which a set with the flag 1113 is set is generated and stored in the storage unit 1101. At this time, each score 1112 is NULL, and the processed flag 1113 indicates unprocessed. The similarity calculation unit 1105 reads out each row of the list 1114 from the storage unit 1101 one by one, and reads out the phoneme sequence portion indicated by the position 1111 of the corrected portion candidate of the read row from the phoneme sequence 1012 of the recognition result of the storage unit 1101. The similarity (score) between the correct character string read from the storage unit 1101 and the phoneme string 1011 is calculated by the DP matching method or the like, written to the score 1112, and the processed flag 1113 is set to processed. The process ends when all the flags 1113 in the list 1114 are processed.

  The selection unit 903 includes a score comparison unit 913 and a position conversion unit 914 as shown in FIG.

  The score comparison unit 913 reads the list 1114 created by the search unit 902 from the storage unit 1101, compares the score 1112 with a preset threshold value for each row, and corrects a corrected portion candidate for a row having a score equal to or higher than the threshold value. Position 1111 is selected as a correction location and output to the position conversion unit 914. Here, the position 1111 of the correction part candidate having a score equal to or higher than a preset threshold is selected as the correction part. However, the position 1111 of the correction part candidate of the line with the highest score in the list 1114 is selected as the correction part. It may be.

  The position conversion unit 914 recognizes a correction location on the phoneme string of the recognition result 133 output from the score comparison unit 913 based on the correspondence between the character string constituting the recognition result 133 and the phoneme string of the recognition result 133. The result 133 is converted into a correction location on the character string and output to the recognition result correction means 144.

  Other components such as the recognition result correcting means 144 are the same as those in the first embodiment.

  Next, for example, the speech recognition result used when creating a speech transcript of “Japan and Korea is” is “Japan and Recommendation”, and the user changes “Recommendation” to “Korea”. The operation after “Korea” is input as the correct character string will be described with reference to the flowchart of FIG.

  When the correct character string is input to the correct character string input field 210, the correct answer input unit 142 of the processing device 107 captures the correct character string and transmits it to the correction location determination unit 143 (step S1201 in FIG. 12). The character string creation means 911 in the search means 902 of the correction location determination means 143 uses the correct character string “Korea” as the phoneme string “ka N koku” and the recognition result “Japan and recommendation” as the phoneme string niho N / to. / ka N koku / wa ”(step S1202) Next, the score calculation / position detection means 912 receives the phoneme sequence of“ Korea ”and the phoneme sequence of“ Japan and recommendation ”, and the latter The phoneme sequence is matched with the former phoneme sequence and a score indicating the similarity is calculated (step S1203) .In this case, the “kaNkoku” portion (recognition in the “Japan and Recommendations” phoneme sequence) is recognized. As a result, the score of “parts misrecognized as“ recommendation ”) is the highest, and the other parts have low scores.

  Next, the score comparison unit 913 of the selection unit 903 compares the score prepared as a threshold value in advance (step S1204), and the position conversion unit 914 corresponds to the phoneme string portion of the recognition result that has a score equal to or higher than the threshold value. The position of the recognition result on the character string is output as the correction position. Thereafter, the same operation as in the first embodiment is performed (steps S1205 and S1206).

  In the above example, the correct character string and the wrong recognition phoneme string are the same in the case of correcting what was misrecognized as a homonym, but by using a method such as DP matching, It is also possible to automatically detect misrecognized parts with insertion, omission and replacement. For example, if the speech recognition result when creating a transcript of the speech “Yesterday Koizumi Prime Minister Election” is “Yesterday Koizumi Award is Election”, and “Award” is changed to “Prime Minister”, the correct answer string The phoneme sequence is “sh u sh oo”, the phoneme sequence of the recognition result is “kinoo / koizumi / ju sh oo / wa / se Nky o / no”, and the result of DP matching is “ju sh oo” Since the matching score is the highest compared to other parts, the part that has been misrecognized as “win” can be detected as the misrecognized part of “Prime Minister”, that is, the corrected part.

  As described above, according to the present embodiment, the correct portion to be replaced with the correct character string is automatically detected by matching the phoneme sequence of the correct character string with the phoneme sequence of the recognition result. The same effect can be obtained.

In this embodiment, in order to detect a portion similar in pronunciation to the correct character string from the recognition result, the phoneme sequence of the correct character string is matched with the phoneme sequence of the recognition result. Each recognition result may be converted into a syllable string, and matching may be performed between the syllable strings. For example, taking “Home of President Hussein” as an example, the phoneme sequence and syllable sequence are as follows.
○ Phoneme sequence… husei N daitoo ry oonoko ky oo
○ Syllable strings ... Fusindai no Ryoko

  Referring to FIG. 13, the correction location determination unit 143 in the present embodiment is different from that in the first embodiment in that the selection unit 303 includes an estimation unit 1301.

  In the present embodiment, the score comparison means 313 of the selection means 303, when there is a score output by the score calculation / position detection means 312 that is equal to or greater than the threshold value, is similar to the first embodiment, the position conversion unit 314. Through this, the correction position is notified to the recognition result correction means 144, but when there is no score equal to or higher than the threshold, the estimation means 1301 is activated.

  The estimation means 1301 uses the correct character string input from the correct answer input means 142 and the language model prepared in advance to check the connection restriction between the correct character string and the word in the recognition result, and correct character string A position with a high probability of being inserted is detected on the recognition result. Then, the estimation unit 1301 notifies the recognition result correction unit 144 of the detected insertion position. The recognition result correction unit 144 presents the insertion position to the user by a method such as displaying a cursor at the notified insertion position in the recognition result displayed in the attention line display field 203 of FIG. When the user turns on the OK button 211, the recognition result correcting unit 144 inserts the correct character string at the insertion position of the recognition result on which the cursor is placed.

  A configuration example of the estimation unit 1301 is shown in FIG. Referring to FIG. 14, the estimation unit 1301 includes a first storage unit 1401, a second storage unit 1402, a first input unit 1403, a second input unit 1404, a position determination unit 1405, a connectability calculation unit 1406, and a selection unit 1407. Consists of. The first input unit 1403 reads the correct character string input to the correct character string input field 210 from, for example, the first storage unit 401 in FIG. 4 and stores the read character string in the first storage unit 1401. The second input unit 1404 reads the recognition result 133 displayed in the attention line display field 203 from the storage device 103 and stores it in the first storage unit 1401. The position determination unit 1405 reads the recognition result 133 from the first storage unit 1401, comprehensively obtains candidates for insertion positions on the recognition result 133, and inserts position candidates 1411, scores 1412, and processed flags in each row. A list 1414 in which a set with 1413 is set is generated and stored in the first storage unit 1401. In general, a word boundary position is a candidate for an insertion location. At this time, each score 1412 is NULL, and the processed flag 1413 indicates unprocessed.

  The connectability calculation unit 1406 reads out each row of the list 1414 from the first storage unit 1401 one by one, and is indicated by the position 1411 of the insertion location candidate of the read row in the recognition result 133 read out from the first storage unit 1401. A language model stored in advance in the second storage unit 1402 as a score indicating the possibility of connection with words before and after the insertion position when the correct character string 301 read from the first storage unit 1401 is inserted at the position. 1421 is calculated and written in the score 1412, and the processed flag 1413 is set to processed. When all the flags 1413 in the list 1414 have been processed, the processing of the connectability calculation unit 1406 ends.

  As the language model 1421, bigram, trigram, or a longer chain probability can be used. Here, bigram is the probability of the next occurrence of word B under the condition that word A is observed, and trigram is under the condition that word A and word B are continuously observed, The probability that the word C appears next. This probability is used as the score value.

  Finally, the selection unit 1407 sequentially reads a pair of the insertion position candidate position 1411 and the score 1412 in the list 1414 from the first storage unit 1401, and specifies the pair of the insertion position candidate position 1411 and the score 1412 having the maximum score. When the score of the pair is higher than a predetermined threshold value, the recognition result correcting unit 144 is notified of the position 1411 of the pair of insertion location candidates as the insertion position 1431.

  Other components are the same as those in the first embodiment.

  Next, for example, the speech recognition result used when creating a transcript of the voice “just a little drunk driving” is “just a little driving”, and the dropout error (“drinking” is dropped) ) Will be described with reference to the flowchart of FIG. 15 after the user has input “drinking” that has been dropped into the correct character string input field 210 of FIG.

  When the correct character string is input to the correct character string input field 210, the correct input means 142 of the processing device 107 captures the correct character string and transmits it to the correction location determining means 143 (step S1501 in FIG. 15). The standard pattern creation means 311 in the search means 302 of the correction location determination means 143 converts “drinking” into a phoneme string “i N sh u”, arranges phoneme standard patterns prepared in advance in that order, and “drinks” Is created (step S1502). Next, the score calculation / position detection means 312 inputs the standard pattern of “drinking” and the analysis result corresponding to the recognition result “just a little driving” in the analysis result 132 held in the storage device 103. The probability (score) of outputting a correct character string (drinking) when each analysis result part is given to the standard pattern is calculated (step S1503), and the score comparison means 313 of the selection means 303 is prepared as a threshold value in advance. Comparison with the score is made (step S1504). However, in this case, there is no analysis result portion having a score equal to or higher than the threshold value.

  In such a case, the selection means 303 uses the language model 1421 by the estimation means 1301 to check the possibility that a drop error has occurred, and for each insertion position candidate in the recognition result, A score (probability) at which “drinking” is inserted is calculated (step S1505), and it is determined whether or not the maximum score is equal to or greater than a predetermined threshold (step S1506). If it is equal to or greater than the threshold value, the insertion position candidate having the maximum score is determined as the insertion position and notified to the recognition result correction means 144. In this case, it is assumed that the position immediately before “driving” is determined as the insertion position and notified to the recognition result correcting means 144.

  The recognition result correction unit 144 presents the insertion position to the user by displaying a cursor at the insertion position portion output from the selection unit 303 among the recognition results displayed in the attention line display field 203 (step S1507). ). When the user turns on the OK button 211, the recognition result correcting unit 144 inserts the correct character string “drinking” into the portion of the recognition result on which the cursor is hit (immediately before “driving”) (step S1508).

  In the present embodiment, the correction portion is detected using the standard pattern of the correct character string as in the first embodiment, but the correction portion is detected using a character string such as a phoneme string of the correct character string as in the second embodiment. It is also possible to search.

  According to the present embodiment, in addition to the effects of the first and second embodiments, the correct character string is inserted using the information of the language model for the corrected portion that could not be detected by the search using the standard pattern or the character string. The position to be done can be inferred.

  Next, a second embodiment of the present invention will be described in detail with reference to the drawings.

  Referring to FIG. 16, the speech transcription support device according to the second exemplary embodiment of the present invention is provided with a plurality of correct answer inputs instead of the correct answer input means 142, the correction location determining means 143, and the recognition result correcting means 144 of the processing device 107. The second embodiment is different from the first embodiment shown in FIG. 1 in that a means 1602, a plurality of correction location determination means 1603, and a recognition result correction means 1604 are provided.

  The plurality of correct answer input means 1602 collectively takes in the correct character strings of the plurality of erroneous recognition portions from the keyboard 106 and sends them one by one to the plurality of correction location determination means 1603. Specifically, since there are a plurality of misrecognized portions in the recognition result displayed in the attention line display field 203 of FIG. 2, the user sets each correct character string in the order of the misrecognition that appears, for example, Are input to the correct answer character string input field 210 at a time, and the correct answer character strings are sent one by one to the multiple correction location determination means 1603 in order from the first correct answer character string.

  Each time the correct character string is sent from the multiple correct answer input means 1602, the multiple correction location determination means 1603 is displayed in the attention line display field 203 of FIG. 2 in the same manner as in each example of the first embodiment. Of the displayed recognition results, a portion where the correct character string and the pronunciation are most similar is determined as a correction portion and notified to the recognition result correction means 1604.

  The recognition result correction unit 1604 automatically recognizes the recognition result 133 stored in the storage device 103 based on the correct character string captured by the plurality of correct input units 1602 and the correction part determined by the plural correction part determination unit 1603. Or urge the user to correct it. Each time correction of one erroneously recognized portion is completed, a correction completion notification 1601 indicating the corrected final position is sent to a plurality of correction location determination means 1603.

  When receiving the correction completion notification 1601, the multiple correction location determination means 1603 determines the correction location based on the next correct character string sent from the multiple correct answer input means 1602. At that time, the plurality of correction location determination means 1603 performs matching with the correct character string only in the portion after the correction final position notified by the correction completion notification 1601 in the recognition result displayed in the attention line display field 203. Take.

  According to the present embodiment, the same effects as those of the first embodiment can be obtained, and correct answers for each of a plurality of misrecognitions can be arranged and input together in the order of the misrecognitions that have appeared. In the correction location determination means 1603, the correct answer after the correct input order is matched with the correct character string only for the part after the position where the correct input order is first and the correction is confirmed by the recognition result correction means 1604. The processing time required for matching can be reduced.

  Next, an example of the present embodiment will be described in detail with a focus on the multiple correction location determination unit 1603 and the recognition result correction unit 1604.

  Referring to FIG. 17, the plurality of correction location determination means 1603 in the present embodiment includes the correct character string 301 input to the correct character string input field 210 sent one by one from the multiple correct answer input means 1602, and the attention line display field 203. The analysis result 132 stored in the storage device 103 corresponding to the recognition result displayed on the screen and the correction completion notification 1601 from the recognition result correction means 1604 are input, and the data after the last correction position indicated by the correction completion notification 1601 are input. Search means 1702 for calculating a score indicating a speech signal similarity between the analysis result portion of the correction part candidate and the correct character string 301 for each correction part candidate comprehensively set on the analysis result 132, and the search Based on the similarity score, one or more correction locations are selected from among the plurality of correction location candidates obtained by the means 1702, and the recognition result correction means 160 is selected. It is composed of a selection means 1703 for notifying the correction point on the recognition result.

  The search unit 1702 includes a standard pattern creation unit 1711 and a score calculation / position detection unit 1712.

  The standard pattern creation unit 1711 has the same basic configuration as the standard pattern creation unit 311 of the first embodiment illustrated in FIG. 4, but each time a correction completion notification 1601 is received from the recognition result correction unit 1604, a plurality of correct answers are input. The difference is that the correct character string 301 in the next order is received from the means 1602 and its standard pattern is created.

  The score calculation / position detection means 1702 is an analysis stored in the storage device 103 corresponding to the standard pattern of the correct character string created by the standard pattern creation means 1711 and the recognition result displayed in the attention line display field 203. The result 132 and the correction completion notification 1601 are input, and the probability (score) that the standard pattern outputs the correct character string when a certain analysis result portion after the final correction position indicated by the correction completion notification 1601 is given is calculated. And output. A configuration example of the score calculation / position detection means 1712 is shown in FIG.

  Referring to FIG. 18, the score calculation / position detection unit 1712 is the same as the first embodiment shown in FIG. 6 in that a third input unit 1801 is added and the function of the position determination unit 1802 is changed. This is different from the score calculation / position detection means 312. The third input unit 1801 inputs a correction completion notification 1601 sent from the recognition result correction unit 1604 and stores the final correction position 1811 included in this notification in the storage unit 601. The position determination unit 1802 reads the final correction position 1811 from the storage unit 601, reads the analysis result portion after the final correction position 1811 in the analysis result 132 of the storage unit 601, and becomes a candidate for a correction location on the analysis result portion. Are generated, and a list 614 in which a set of a correction location candidate position 611, a score 612, and a processed flag 613 is set in each row is generated and stored in the storage unit 601. Other configurations and operations are the same as the score calculation / position detection means 312 of FIG.

  As shown in FIG. 17, the selection unit 1703 includes a score comparison unit 1713 and a position conversion unit 1714. These means are the same as the score comparison means 313 and the position conversion unit 314 in the selection means 303 of the first embodiment shown in FIG.

  A configuration example of the recognition result correcting unit 1604 is shown in FIG. Referring to FIG. 19, the recognition result correcting unit 1604 is different from the recognition result correcting unit 144 of the first embodiment illustrated in FIG. 7 in that a correction completion notification unit 1901 is added. The correction completion notification unit 1901 receives the correction position of the recognition result from the character string operation unit 706, and converts the correction position on the recognition result into the correction position on the analysis result 132 according to the correspondence relationship between the recognition result 133 and the analysis result 132. Then, a correction completion notification 1601 including the converted correction position is sent to a plurality of correction location determination means 1603. Other configurations and operations are the same as those of the recognition result correcting unit 144 of FIG.

  Next, for example, the voice recognition result used when creating a transcript of the speech “Natural tonic or topical cold medicine not available” becomes “Demand tonic diamond topical cold medicine not available”. In order for the user to change “demand” to “nutrition” and “diamond” to “agent”, the operation after inputting “nutrient” in the correct character string input field 210 of FIG. This will be described with reference to the flowchart of FIG.

  When the correct character string is input to the correct character string input field 210, the plurality of correct answer input means 1602 of the processing device 107 fetches the correct character string (step S2001 in FIG. 20). Is set as the first processing target, and is transmitted to the multiple correction location determination means 1603 (step S2002). The standard pattern creation means 1711 in the search means 1702 of the multiple correction location determination means 1603 converts “nourishment” into a phoneme string “jiyoo”, arranges phoneme standard patterns prepared in advance in that order, and expresses “nourishment”. A standard pattern to be created is created (step S2003). Next, the score calculation / position detection unit 1712 inputs the standard pattern of “nourishment”, and since there is no final correction position notified by the correction completion notification 1601 at this time, the analysis result held in the storage device 103 Probability that correct character string (nourishment) is output when all recognition results in 132, “Demanded tonic diamond external cold medicine not available” are input and each part of the input analysis result is given to the standard pattern (Score) is calculated (step S2004). In this case, the standard pattern has the highest score in the analysis result of the part uttered as “nourishment” (the part erroneously recognized as “demand” in the recognition result), and the other part has a low score. Next, the score comparison unit 1713 of the selection unit 1703 compares the score prepared as a threshold value in advance (step S2005), and the position conversion unit 1714 recognizes the recognition result corresponding to the analysis result portion having a score equal to or higher than the threshold value. The upper position is output as the correction position. Here, it is assumed that there is only one score that satisfies the threshold, and only the portion that is pronounced “nourishment” (the misrecognized “demand” portion).

  The recognition result correction unit 1604 focuses on the recognition result portion “demand” corresponding to the correction position output by the selection unit 1703 among the recognition results displayed in the attention line display field 203, and sets the correction position to the user. Present (step S2006). When the user turns on the OK button 211, the focused recognition result portion (misrecognition result “demand”) is replaced with the correct character string “nourishment” (step S2007). Then, the recognition result correction unit 1604 sends a correction completion notification 1601 including the position on the analysis result 132 corresponding to the position immediately after “nourishment” after replacement (immediately before “strong”) to the plurality of correction point determination unit 1603. Then, the score calculation / position detection means 1712 of the multiple correction location determination means 1603 updates the final correction position (step S2008).

  The multiple correct answer input means 1602 sets the correct character string “diamond” in the next order as a processing target (steps S2009 and S2010), and the standard pattern creation means 1711 in the search means 1702 of the multiple correction location determination means 1603 "Is converted into a phoneme string" daiya "and its standard pattern is created (step S2003). Next, the score calculation / position detection means 1712 includes a standard pattern of “diamond” and a recognition result “analysis cold medicine for external use that is not available” in the analysis result 132 held in the storage device 103. Probability (score) of outputting a correct character string (nourishment) when each analysis result portion is given to the standard pattern with “tonic diamond external cold medicine” after the final correction position notified by the correction completion notification 1601 as the analysis result target ) Is calculated (step S2004). In this case, the standard pattern has the highest score in the analysis result of the part uttered as “agent” (the part erroneously recognized as “diamond” in the recognition result), and the other part has a low score. Next, the score comparison unit 1713 of the selection unit 1703 performs comparison with the threshold value, and the position conversion unit 1714 outputs the position on the recognition result corresponding to the analysis result portion having a score equal to or higher than the threshold value as the correction position. Here, it is assumed that a portion (a misrecognition “diamond” portion) uttered “agent” is output as a correction position.

  The recognition result correction unit 1604 focuses the recognition result portion “diamond” corresponding to the correction position output by the selection unit 1703 out of the recognition results displayed in the attention line display field 203, and sets the correction position to the user. Present (step S2006). When the user turns on the OK button 211, the recognition result correcting unit 1604 replaces the focused recognition result portion (misrecognition result “diamond”) with the correct character string “agent” (step S2007). ). Then, a correction completion notice 1601 including the position on the analysis result 132 corresponding to the position immediately after “agent” after replacement (immediately before “outside”) is sent to a plurality of correction location determination means 1603 (step S2008). At this point, since the processing of all correct character strings “nourishment” and “agent” input by the multiple correct answer input means 1602 is completed, the processing of FIG. 20 ends. At this time, the final correction position held by the score calculation / position detection means 1712 is set to NULL.

  Next, the effect of the present embodiment will be described.

  When there are a plurality of misrecognized parts in the recognition result displayed in the attention line display field 203, if each correct character string is input in the order of the erroneous recognition that has appeared, the correct character string in order from the first correct character string. Since the recognition result portion having a similar pronunciation is automatically set as the correction location, the user himself / herself can save the trouble of specifying the correction location, and the burden on the user can be reduced. In addition, since the search for the corrected portion by the second and subsequent correct character strings is performed on the portion after the corrected portion determined by the correction immediately before that, the corrected portion can be determined accurately and in a short time.

  The multiple correction location determination means 1603 of the present embodiment converts the correct character string into a voice standard pattern and determines the acoustic similarity between the standard pattern and the analysis result 132 on the storage device 103. Various modifications similar to those described in the Example section are possible.

  Further, in this embodiment, the correction portion is determined by matching using the standard pattern. However, as in the second embodiment, the correction character string and the recognition result are converted into a phoneme string or a syllable string, and matching is performed. May be determined.

  Although the embodiments and examples of the present invention have been described above, the present invention is not limited to the above examples, and various other additions and modifications can be made. Further, the processing device 107 or the correction location determination means 143 and the recognition result correction means 144 constituting the processing apparatus 107 can be realized by a computer and a program as well as the functions of the processing apparatus 107 and hardware. The program is provided by being recorded on a computer-readable recording medium such as a magnetic disk or a semiconductor memory, and is read by the computer at the time of starting up the computer, etc. And the processing device 107 in the embodiment and the example or the correction part determination unit 143 and the recognition result correction unit 144 constituting the processing unit 107 are caused to function.

It is a block diagram of a 1st embodiment of the present invention. It is a figure which shows an example of the screen displayed on a display apparatus in the 1st Embodiment of this invention. It is a block diagram of the correction location determination means in Example 1 of this invention. It is a block diagram of the standard pattern preparation means in the correction location determination means in Example 1 of this invention. It is a figure which shows the example of creation of the voice standard pattern of a correct character string. It is a block diagram of the score calculation and position detection means in the correction location determination means in Example 1 of this invention. It is a block diagram of the recognition result correction means in Example 1 of the present invention. It is a flowchart which shows operation | movement of Example 1 of this invention. It is a block diagram of the correction location determination means in Example 2 of this invention. It is a block diagram of the character string preparation means in the correction location determination means in Example 2 of this invention. It is a block diagram of the score calculation and position detection means in the correction location determination means in Example 2 of the present invention. It is a flowchart which shows operation | movement of Example 2 of this invention. It is a block diagram of the correction location determination means in Example 3 of this invention. It is a block diagram of the estimation means in the correction location determination means in Example 3 of this invention. It is a flowchart which shows operation | movement of Example 3 of this invention. It is a block diagram of the 2nd Embodiment of this invention. It is a block diagram of the multiple correction location determination means in Example 4 of this invention. It is a block diagram of the score calculation and position detection means in the multiple correction location determination means in Example 4 of the present invention. It is a block diagram of the recognition result correction means in Example 4 of the present invention. It is a flowchart which shows operation | movement of Example 4 of this invention.

Explanation of symbols

DESCRIPTION OF SYMBOLS 101 ... Voice 102 ... Voice recognition device 103 ... Storage device 104 ... Display device 105 ... Speaker 106 ... Keyboard 107 ... Processing device 111 ... Speech input means 112 ... Analysis means 113 ... Collation means 121 ... Dictionary 122 ... Language model 123 ... Acoustic model 124 ... Storage device 131 ... Audio signal 132 ... Analysis result 133 ... Recognition result 141 ... Display / playback unit 142 ... Correct input means 143 ... Correction location determination means 144 ... Recognition result correction means 145 ... Recognition result display means 146 ... Voice reproduction means

Claims (17)

Storage means for storing a speech recognition result obtained by performing speech recognition processing on the speech signal;
Correct input means for inputting a correct character string for correcting a recognition error part of the speech recognition result from a keyboard;
Correction location determination means for searching for a character string portion whose pronunciation is similar to the correct character string as a correction location in the speech recognition result;
A speech transcription support device, comprising: a recognition result correction unit that replaces the determined correction portion with the correct character string.   The correction location determination means searches for a character string portion whose pronunciation is similar to that of the correct character string by taking a match between the voice data created from the correct character string and the voice data corresponding to the voice recognition result. The speech transcription support apparatus according to claim 1, wherein:   The speech standard pattern generated from the correct character string is used as the speech data of the correct character string, and the analysis result of the speech signal corresponding to the speech recognition result is used as the speech data of the speech recognition result. The voice transcription support device according to claim 2.   The speech signal generated by speech synthesis from the correct character string is used as the speech data of the correct character string, and the speech signal corresponding to the speech recognition result is used as the speech data of the speech recognition result. Item 2. The voice transcription support device according to Item 2.   The correction location determination means performs matching between a character string that is a phoneme string or syllable string created from the correct character string and a character string that is a phoneme string or syllable string created from the speech recognition result, 2. The voice transcription assisting apparatus according to claim 1, wherein a character string portion whose pronunciation is similar to that of the correct character string is searched.   The correction location determination means inserts the correct character string using a language model describing word connection constraints when a character string portion whose pronunciation is similar to the correct character string does not exist in the speech recognition result. 2. The voice transcription according to claim 1, wherein a position having a high probability is searched as an insertion position, and the recognition result correcting means inserts the correct character string at the determined insertion position. Support device.   The correction location determination means searches for a correction location according to the order of correct character strings input from the keyboard, and sequentially narrows the search range based on the correction position corrected by the recognition result correction means. The speech transcription support device according to claim 1. The display / playback unit displays a voice recognition result obtained by performing voice recognition processing on the voice signal on the display device,
The correct input means inputs a correct character string for correcting the recognition error part of the speech recognition result from the keyboard,
The correction location determination means searches for a character string portion whose pronunciation is similar to the correct character string in the speech recognition result as a correction location,
A speech transcription support method, wherein a recognition result correction means replaces the determined correction portion with the correct character string. A computer comprising storage means for storing a speech recognition result obtained by subjecting a speech signal to speech recognition processing;
Correct answer input means for inputting a correct character string for correcting a recognition error part of the speech recognition result from a keyboard;
A correction location determination means for searching a text string portion whose pronunciation is similar to the correct character string as a correction location in the speech recognition result;
Recognition result correction means for replacing the determined correction portion with the correct character string;
Program to function as.   The correction location determination means searches for a character string portion whose pronunciation is similar to that of the correct character string by taking a match between the voice data created from the correct character string and the voice data corresponding to the voice recognition result. The program according to claim 9, wherein:   The speech standard pattern generated from the correct character string is used as the speech data of the correct character string, and the analysis result of the speech signal corresponding to the speech recognition result is used as the speech data of the speech recognition result. The program according to claim 10.   The speech signal generated by speech synthesis from the correct character string is used as the speech data of the correct character string, and the speech signal corresponding to the speech recognition result is used as the speech data of the speech recognition result. Item 10. The program according to item 10.   The correction location determination means performs matching between a character string that is a phoneme string or syllable string created from the correct character string and a character string that is a phoneme string or syllable string created from the speech recognition result, The program according to claim 9, wherein the program searches for a character string portion whose pronunciation is similar to that of the correct character string.   The correction location determination means inserts the correct character string using a language model describing word connection constraints when a character string portion whose pronunciation is similar to the correct character string does not exist in the speech recognition result. 10. The program according to claim 9, wherein a position having a high probability is searched as an insertion position, and the recognition result correcting means inserts the correct character string at the determined insertion position.   The correction location determination means searches for a correction location according to the order of correct character strings input from the keyboard, and sequentially narrows the search range based on the correction position corrected by the recognition result correction means. The program according to claim 9. A phoneme string conversion means for converting a correct character string input from a keyboard into a phoneme string and storing it in a storage unit in order to correct a recognition error portion of a voice recognition result obtained by performing a voice recognition process on the voice signal; ,
A phoneme sequence of the correct character string is read from the storage unit, a standard pattern storage unit that stores a standard pattern for each phoneme is searched, the phoneme sequence is converted into a speech standard pattern, and the phoneme stored in the storage unit Standard pattern conversion means;
A position determination unit that comprehensively obtains candidates for correction points on the analysis result of the voice signal corresponding to the voice recognition result, and stores position information of each correction point candidate in the storage unit;
The standard pattern of the correct character string, the analysis result, and the position information of each correction part candidate are read from the storage unit, and the standard pattern of the correct character string is specified by the position information of the correction part candidate on the analysis result A probability calculating means for calculating the probability of outputting the analysis result part for each correction location candidate and storing it in the storage unit;
A correction unit comprising: a selecting unit that reads out and compares the probabilities calculated for each correction part candidate from the storage unit, and selects and outputs a correction part from the plurality of correction part candidates. Location determination device. In order to correct a recognition error portion of a speech recognition result obtained by performing speech recognition processing on a speech signal, each of a correct character string input from the keyboard and the character string of the speech recognition result is converted into a phoneme string or a syllable string. Phoneme string conversion means for converting and storing in the storage unit;
The phoneme sequence or syllable string of the speech recognition result is read from the storage unit, candidates for correction locations on the phoneme sequence or syllable sequence are comprehensively determined, and position information of each correction location candidate is stored in the storage unit Positioning means;
The correct character string and the phoneme string or syllable string of the speech recognition result and the position information of each correction location candidate are read from the storage unit, and the phoneme sequence of the speech recognition result specified by the position information of the correction location candidate or Calculating a similarity between a part on a syllable string and a phoneme string or a syllable string of the correct character string, and storing the similarity in the storage unit;
And selecting means for reading out and comparing the degree of similarity calculated for each correction location candidate from the storage unit, and selecting and outputting a correction location from the plurality of correction location candidates. Correction point determination device.

Images