JP2010256498A - Conversion model generating apparatus, voice recognition result conversion system, method and program - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To facilitate generation of a conversion model according to each use in a plurality of use, determined by a purpose for converting voice recognition result. <P>SOLUTION: A conversion model generating apparatus has a learning data for each usage selecting means 501, that extracts a part where a voice recognition result text shown by voice recognition result data for learning is different from a formatted text for learning, using voice recognition result data for learning showing a result of voice recognition processing, with respect to certain voice data and the formatted text for learning showing a result of converting the voice data into a text with a predetermined purpose, and selects learning data for generating the conversion model according to predetermined individual use, based on appearance tendency of the extracted difference part; and a conversion model generating means 502, that generates the conversion model according to the individual use using the learning data selected by using the learning data for each usage selecting section. <P>COPYRIGHT: (C)2011,JPO&INPIT

Description

  The present invention relates to a conversion model generation device for generating a conversion model for performing conversion for the purpose of correcting speech recognition errors and shaping for improving readability, and a speech recognition result using the conversion model. The present invention relates to a speech recognition result conversion system, method and program thereof.

  Speech recognition technology is a technology that automatically transcribes spoken words as accurately as possible. One application that uses speech recognition technology is a device that creates sentences such as conference minutes and subtitles. For example, when creating a conference record, by recognizing speech data of a certain speaker's utterance, each phrase of a certain speaker's utterance is transcribed into characters, and a sentence of the conference record is generated.

  However, in many cases, sentences written using such a speech recognition technique include not only speech recognition errors. In addition, a sentence that is transcribed correctly word by phrase is redundant because it contains many colloquial expressions and the like, and there is a high possibility that the readability is low.

  In order to create an appropriate sentence such as a conference record, it is necessary to correct speech recognition errors and process text shaping to improve readability. For example, Patent Document 1, Patent Document 2, and Non-Patent Document 1 are prior art documents relating to a method for automatically performing such correction or shaping.

  Patent Document 1 describes a method of generating a deformation model using a speech recognition result and a corrected text for learning that has been corrected and shaped.

  Patent Document 2 discloses a phrase dictionary necessary for conversion using a learning database in which a sentence to be converted into speech or text and a sentence that has been converted into a language are paired. It describes a method for automatically creating rules and rules between phrases as much as possible. Further, as an example of the created conversion rule, there is a language conversion device that converts a spoken sentence (spoken expression) into a text sentence (sentence expression) such as a written word.

  Non-Patent Document 1 describes a method of generating a conversion model using a speech transcript that does not include a speech recognition error and a pre-formatted text for learning that has been shaped.

JP 2008-216341 A JP 2000-305930 A

Yuya Akita and Tatsuya Kawahara, “Automatic Spoken Speech Recognition Results for Producing Minutes”, Proceedings of the Acoustical Society of Japan, September 2008, p. 103-104

  However, as in the method described in Patent Document 1, the result of processing according to a predetermined application such as “correction of speech recognition error included in speech recognition result” or “text shaping such as improving readability” In the method of generating a conversion model using the correct text, which is the formatted text obtained, as learning data, there is a problem that a conversion model corresponding to each application cannot be generated when a plurality of uses are set.

  Note that in the method described in Patent Document 2, learning of a desired application in which a sentence that is a target of language conversion and a sentence that has been subjected to language conversion are paired in order to create a conversion rule for the desired application. Database is required. There is a problem that enormous effort is required to prepare the learning database for the desired application.

  Further, in the method described in Non-Patent Document 1, conversion according to the use of, for example, “text shaping to improve readability” is performed by using a speech transcription that does not include a speech recognition error instead of a speech recognition result. A model can be generated. However, there is a problem that enormous effort is required to prepare a voice transcription that does not include a voice recognition error.

  Therefore, the present invention provides a conversion model generation device capable of easily generating a conversion model corresponding to each application in a plurality of applications determined by what purpose the speech recognition result is converted. It is an object of the present invention to provide a speech recognition result conversion system, method and program for converting speech recognition results.

  The conversion model generation device according to the present invention includes learning speech recognition result data indicating a result of speech recognition processing for certain speech data, and learning shaped text indicating a result of converting speech data into text with a predetermined purpose, Is used to extract a different part between the speech recognition result text indicated by the learning speech recognition result data and the shaped text for learning, and is determined in advance based on the appearance tendency of the extracted different part A learning model for each application for selecting learning data for generating a conversion model for each application, and a conversion model for each application by using the learning data selected by the learning data for each application. And conversion model generation means for generating.

  In addition, the speech recognition result conversion system according to the present invention is preliminarily determined that text shaping for a certain purpose is one of individual uses when generating a conversion model. The speech recognition result text indicated by the learning speech recognition result data using the learning speech recognition result data indicating the result of the learning, and the learning formatted text indicating the result of converting the speech data into text with a predetermined purpose, and Extract different parts from the pre-formatted text for learning, and select the learning data to generate a conversion model according to each predetermined application based on the appearance tendency of the extracted different parts Using the learning data sorting means for each purpose and the learning data sorted by the learning data sorting means for each application, a conversion model for each purpose is generated. A conversion model generation unit that performs conversion processing on the given speech recognition result data using a conversion model that is generated by the conversion model generation unit and that is generated by the conversion model generation unit according to a text shaping application for a certain purpose. Voice recognition result shaping means for shaping the text to match the purpose is provided.

  In addition, the speech recognition result conversion system is preliminarily determined that text shaping for a certain purpose is one of individual uses when generating a conversion model, and the result of speech recognition processing is applied to certain speech data. The speech recognition result text indicated by the learning speech recognition result data and the learning shaping using the learning speech recognition result data shown and the learning formatted text showing the result of converting the speech data into text with a predetermined purpose -By-application learning that extracts different parts from the existing text and selects the learning data to generate a conversion model according to each predetermined application based on the appearance tendency of the extracted different parts Using the learning data selected by the data selection means and the application-specific learning data selection means, a conversion model that generates a conversion model for each application By performing conversion processing on the given speech recognition result data using the conversion model according to the text shaping application for a certain purpose generated by the data generation means and the conversion model generation means, Voice recognition result shaping means for shaping the matching text may be provided.

  In addition, the speech recognition result conversion system is defined in advance that the correction of speech recognition errors and the text shaping for a certain purpose are individually used when generating a conversion model. Speech recognition indicated by the speech recognition result data for learning using the speech recognition result data for learning indicating the result of the speech recognition processing and the preformatted text for learning indicating the result of converting the speech data into text with a predetermined purpose Learning data for extracting different parts between the result text and the pre-formatted text for learning, and generating a conversion model for each predetermined application based on the appearance tendency of the extracted different parts Using the conversion model according to the application to correct the speech recognition error generated by the learning data selection means by application and the conversion model generation means. A speech recognition result correcting means for correcting a speech recognition error included in the speech recognition result by performing a conversion process on the obtained speech recognition result data, and a text shaping for a certain purpose generated by the conversion model generating means A speech recognition result shaping unit may be provided that performs transformation processing on given speech recognition result data using a conversion model according to the application, thereby shaping the text to match a certain purpose. .

  In addition, the conversion model generation method according to the present invention includes learning speech recognition result data indicating the result of speech recognition processing for certain speech data, and learning preformatted indicating the result of converting the speech data into text for a predetermined purpose. The text is used to extract a different portion between the speech recognition result text indicated by the learning speech recognition result data and the learning formatted text, and is determined in advance based on the appearance tendency of the extracted different portions. Using the learning data selected in the learning data selection step by use and the learning data selection step by use to select the learning data for generating a conversion model according to the individual application, according to the individual application A conversion model generation step of generating a conversion model.

  Further, the speech recognition result conversion method according to the present invention includes learning speech recognition result data indicating the result of speech recognition processing on certain speech data, and learning shaping indicating the result of converting speech data into text for a predetermined purpose. A different portion between the speech recognition result text indicated by the learning speech recognition result data and the learned pre-formatted text is extracted, and predetermined based on the extracted appearance tendency of the different portion. Depending on the individual application, using the learning data selection step for selecting learning data to generate a conversion model according to each individual application and the learning data selected in the learning data selection step for each application A conversion model generation step that generates a converted conversion model, and a conversion model that is generated in the conversion model generation step according to the individual application. Te, characterized in that it comprises a speech recognition result conversion step of performing a conversion process on the speech recognition result given.

  The conversion model generation program according to the present invention indicates to a computer the learning speech recognition result data indicating the result of speech recognition processing for certain speech data, and the result of converting the speech data into text for a predetermined purpose. Extracting the different parts between the speech recognition result text indicated by the speech recognition result data for learning and the pre-formatted text for learning using the learning pre-formed text, and based on the appearance tendency of the extracted different parts , Using learning data selected by use-specific learning data selection processing for selecting learning data for generating a conversion model corresponding to a predetermined individual use and learning data selection processing for each application, A conversion model generation process for generating a conversion model according to the application is executed.

  Further, the speech recognition result conversion program according to the present invention provides a computer with learning speech recognition result data indicating the result of speech recognition processing on certain speech data, and a result obtained by converting the speech data into text with a predetermined purpose. Based on the appearance tendency of the extracted different parts, the difference between the speech recognition result text indicated by the speech recognition result data for learning and the formatted text for learning is extracted using Individual learning data selected by the application-specific learning data selection process for selecting the learning data for generating the conversion model according to the predetermined individual application and the learning data selection process for each application. The conversion model generation process that generates a conversion model according to the usage of the conversion model and the conversion model according to the individual application generated by the conversion model generation process Using Le, characterized in that to execute a speech recognition result conversion process for performing conversion processing on the speech recognition result given.

  ADVANTAGE OF THE INVENTION According to this invention, the production | generation of the conversion model according to each use in the some use defined by what kind of purpose the speech recognition result is converted can be performed easily. Further, by using the conversion model, it is possible to obtain text that has been subjected to conversion processing according to each application.

It is a block diagram which shows the structural example of the speech recognition result conversion model production | generation apparatus of the 1st Embodiment of this invention. It is a flowchart which shows an example of operation | movement of the speech recognition result conversion model production | generation apparatus 100 of 1st Embodiment. It is explanatory drawing which shows the example of the speech recognition result text contained in the learning speech recognition result data D14, and the learning preformatted text D15. It is explanatory drawing which shows the example of a word graph. It is explanatory drawing which shows the example of various data at the time of applying the "text shaping for obtaining a better speech translation result" use. It is explanatory drawing which shows the example of various data at the time of applying the "text shaping for speech understanding" use. It is a block diagram which shows the structural example of the speech recognition result correction system of 2nd Embodiment. It is explanatory drawing which shows the example of the speech recognition result data for correction, and the corrected text. It is a block diagram which shows the structural example of the speech recognition result shaping system of 3rd Embodiment. It is explanatory drawing which shows the example of the speech recognition result data for shaping, and the shaped text. It is a block diagram which shows the structural example of the speech recognition result conversion system of 4th Embodiment. It is explanatory drawing which shows the example of the speech recognition result data for correction | amendment / formatting, and the corrected / formatted text. It is a block diagram which shows the outline | summary of this invention. It is a block diagram which shows the outline | summary of the speech recognition result conversion system of this invention. It is a block diagram which shows the outline | summary of the speech recognition result conversion system of this invention. It is a block diagram which shows the outline | summary of the speech recognition result conversion system of this invention.

Embodiment 1. FIG.
Hereinafter, embodiments of the present invention will be described with reference to the drawings. FIG. 1 is a block diagram illustrating a configuration example of the speech recognition result conversion model generation device according to the first embodiment of this invention. A speech recognition result conversion model generation apparatus 100 shown in FIG. 1 includes a conversion model learning data selection unit 11, a conversion model generation unit 12, a conversion model storage unit 13, a learning speech recognition result storage unit 14, and a learning shaping unit. And a completed text storage unit 15.

  The conversion model learning data selection unit 11 and the conversion model generation unit 12 are realized by, for example, a CPU that operates according to a program provided in the speech recognition result conversion model generation device 100. The conversion model storage unit 13, the learning speech recognition result storage unit 14, and the learning shaped text storage unit 15 are realized by, for example, a storage device included in the speech recognition result conversion model generation device 100.

  The conversion model storage unit 13 stores the conversion model generated by the conversion model generation unit 12 described later. In the present embodiment, the “conversion model” is information indicating a rule or method of conversion performed on a speech recognition result in order to achieve a predetermined purpose such as correction of a speech recognition error or shaping for improving readability. Say. Hereinafter, each conversion model stored in the conversion model storage unit 13 is referred to as a conversion model D13. The conversion model D13 may be a set of information indicating conversion rules, for example.

  The learning speech recognition result storage unit 14 stores learning speech recognition result data. The learning speech recognition result data is information indicating a speech recognition result for certain speech data used as learning data in the present apparatus. Hereinafter, each learning speech recognition result data stored in the learning speech recognition result storage unit 14 is referred to as learning speech recognition result data D14. The speech recognition result data D14 for learning includes, for example, speech recognition result text, N best, word graph, and words included in the speech recognition result obtained by performing speech recognition processing on conference speech sound data using a speech recognition engine. Part of speech or semantic information, acoustic likelihood, reliability, and the like.

  The learning formatted text storage unit 15 stores the learning formatted text. The learning formatted text is a transcription text obtained by correcting or / and shaping the speech recognition result indicated by the learning speech recognition result data D14 (hereinafter referred to as correction / shaping). Hereinafter, each of the learning shaped text stored in the learning shaped text storage unit 15 is referred to as a learning shaped text D15. The learning-formatted text D15 may be, for example, a text such as a conference record in which a certain speaker's utterance is appropriately corrected and shaped. Note that the learning-formatted text D15 is not particularly conscious of the usage, and may be data created for a certain final purpose. Further, in the present embodiment, it is assumed that a plurality of text sentences created under the same conditions are included, such as a conversion result of the entire conference record.

  Strictly, the learned formatted text D15 does not necessarily have to be a speech recognition result corrected and shaped. For example, it may be a text that has been transcribed for a certain final purpose from the input speech that is the basis of the speech recognition result without going through the speech recognition process. That is, it is sufficient that the information is equivalent to the information obtained as a result of the conversion process performed for a predetermined final purpose with respect to the voice recognition result.

  The conversion model learning data selection means 11 uses the speech recognition result text included in the learning speech recognition result data D14 and the learning formatted text D15 to extract different portions of the two texts, and displays the different tendencies. Using this, the learning data for generating the conversion model is selected, and the selected learning data is provided to the conversion model generating means 12.

  Here, the extraction of the different parts of both texts (speech recognition result text and pre-formatted text for learning) is, for example, partial characters of parts that are different between the two texts by aligning (comparing) the two texts. Processing for obtaining a column may be used. The different tendency is, for example, the number of appearances of a set of partial character strings of different portions in a speech recognition result text included in a certain learning speech recognition result data D14 and a shaped text D15 for learning (that is, Appearance probability in the entire learning formatted text D15).

  The conversion model generation means 12 generates a conversion model D13 using the learning data of the conversion model selected by the conversion model learning data selection means 11. For example, the conversion model generation unit 12 selects different portions between the speech recognition result text included in the learning speech recognition result data D14 selected by the conversion model learning data selection unit 11 and the learning formatted text D15. A model may be generated in which each set of partial character strings is a conversion rule. In such a case, a set of conversion rules indicated by using a set of partial character strings of different portions of both texts may be generated as the conversion model D13.

  Each storage unit may be realized by an external storage device. In such a case, data input / output means may be provided.

  Next, the operation of this embodiment will be described. FIG. 2 is a flowchart showing an example of the operation of the speech recognition result conversion model generation device 100 of the present embodiment. In the present embodiment, two types of applications for conversion of speech recognition results are given: “correction of speech recognition error” (use 1) and “text formatting to improve readability” (use 2). And This means that the learning formatted text D15 is a speech recognition result that has undergone conversion related to correction processing and conversion related to shaping processing (or text corresponding to the conversion result). . The purpose of text formatting is not particularly limited. As in this example, it may be for improving readability, or for converting spoken language expressions into sentence language expressions. Further, for example, text shaping for obtaining a better speech translation result or text shaping for speech understanding may be used.

  As shown in FIG. 2, in the speech recognition result conversion model generation device 100 according to the present embodiment, learning speech recognition result data D14 and learning shaped text D15 are first input to the conversion model learning data selection unit 11. (Step S11). Note that the conversion model learning data selection unit 11 may read necessary data from the learning speech recognition result storage unit 14 or the learning shaped text storage unit 15 upon receiving a notification to start processing.

  FIG. 3 is an explanatory diagram illustrating an example of the speech recognition result text included in the learning speech recognition result data D14 and the learning formatted text D15 corresponding to the speech recognition result data D14 (that is, the same phrase). . In FIG. 3, the learning speech recognition result text is “E / Blood //” (“/” represents a word boundary), whereas the learning formatted text is “Marriage / It is shown that

  When the learning speech recognition result data D14 and the learning shaped text D15 to be processed are input, the conversion model learning data selection unit 11 selects the learning speech recognition result data D14 and the learning shaped text D15. Are used to select learning data for generating a conversion model (step S12). Here, different portions of the learning speech recognition result data D14 and the learning shaped text D15 are selected as learning data for generating a conversion model corresponding to the application.

  Specifically, the conversion model learning data selection means 11 is different from the learning speech recognition result text “Eh / Blood / I am doing” and the learning preformatted text “Marriage / doing / doing”. Extract the part. Here, we extracted three different parts: (1) Delete “Eh”, (2) Replace “Blood stain” with “Marriage”, and (3) Replace “I do” with “I am”. To do. Then, based on the set of the different parts, the learning data is selected for each application as learning data for generating a conversion model according to the application.

  As an example of selection of learning data, the case where the formatted text D15 for learning whose final purpose is the creation of a conference record is used will be described below.

  In many cases, minutes recorded using speech recognition technology contain speech recognition errors. In addition, sentences that are transcribed correctly word by phrase contain many colloquial expressions and the like, so they are redundant and have low readability. Therefore, in order to create a sentence such as an appropriate conference record, a text shaping process is performed to correct speech recognition errors and improve readability.

  In the example shown in FIG. 3, the process of replacing “blood stain”, which is a voice recognition error, with “marriage” (corresponding to the different part (2) described above) is a voice recognition error correction process. Also, the process of deleting unnecessary words “e” and replacing the spoken phrase “I am” with the word expression “I am” (corresponding to the different parts (1) and (3) above) Is the shaping process. Here, a conversion model corresponding to each application is generated for two applications, ie, “correction of speech recognition error” and “shaping to improve readability”. For this purpose, the conversion model learning data sorting unit 11 sorts each set of different parts into one based on correction processing and one based on shaping processing. As one of the selection methods, an example in which the number of appearances of different parts is used will be described.

  For example, since the formatting process is performed throughout the entire proceedings, the number of occurrences of different parts such as “Delete“ Eh ”” or “Replace“ I do ”” with “I am” is inevitably growing. On the other hand, the number of occurrences of different parts due to speech recognition errors such as “Replace“ blood stain ”with“ marriage ”” is due to a variety of factors that cause speech recognition errors. Is smaller than the number of occurrences of colloquial expressions such as “replace with“ doing ”” to sentence expressions. In this way, learning data for generating a conversion model corresponding to each application is selected based on the number of appearances in which a difference is caused by the conversion process.

  For example, the conversion model learning data selection unit 11 may extract a different part for each sentence from the conversion result of the entire minutes, and obtain the number of appearances in the whole minutes for each extracted conversion process. . Then, the number of appearances obtained for each conversion process is compared with a predetermined threshold provided for each use, and when it is determined that the number of appearances is within the applicable range indicated by the threshold, the number of appearances The learning-formatted text that has been subjected to the conversion process (that is, the learning-formatted text D15 corresponding to a sentence having a different part for which the conversion process for which the number of appearances has been calculated is performed) You may select as learning data for producing | generating the conversion model corresponding to a use.

  For example, when selecting learning data for a “correction of voice recognition error” application, if the number of appearances of a certain conversion process is smaller than a threshold value determined for the application, the conversion process is performed. The pre-formatted text for learning may be selected as learning data for the application. Further, for example, when learning data for use in “shaping to improve readability” is selected, if the number of appearances of a certain conversion process is larger than a threshold value set for the use, the conversion process is performed. The learned preformatted text may be selected as learning data for the application. In addition, using one threshold, if the number of appearances (or appearance probability) is smaller than the threshold, it is selected as learning data for “correction of speech recognition error”, and if it exceeds the threshold, “enhance readability” It is also possible to select as learning data for “shaping for use”.

  Next, the learning model generation means 12 generates a conversion model D13 using the learning data selected by the conversion model learning data selection means 11 (step S13). For example, the conversion model generation unit 12 may generate, as the conversion model D13, information indicating a different part included in the selected learning data as a conversion rule. In the above example, the conversion model D13 for the purpose of “correction of speech recognition errors” is included in at least the learned formatted text D15 selected as learning data for “correction of speech recognition errors”. Information including (2) “replace“ blood stain ”with“ marriage ”” as a conversion rule may be generated. Further, for example, as the conversion model D13 for the purpose of “shaping for improving readability”, the difference included in the learning formatted text D15 selected as the learning data for at least “shaping for improving readability” is used. Information including (1) “deletion of“ e ”” and (3) “replace“ is done ”with“ is ”” as conversion rules may be generated.

  When calculating the number of appearances by conversion process, not only count the parts with the same conversion process for different parts, but the same if the parts where the different parts are extracted, the parts of speech that are replaced or deleted, etc. are the same. A method of counting as one conversion process by considering it as a conversion method is also conceivable.

  In the above example, the number of appearances is used as the learning data selection method. However, other methods are also possible. For example, it is possible to use a word graph included in the learning speech recognition result data D14. FIG. 4 is an explanatory diagram showing an example of a word graph. FIG. 4 shows a word graph in which a plurality of recognition result candidates are expressed in a graph when the voice data “E / Marriage / Do / Do” is input. In the example shown in FIG. 4, the recognition result candidate for the voice data “E / marriage / do / do” is “e / marry / do / do” and “e / blood / ”And the first candidate is“ Eh / Blood / I ’m doing it ”.

  In this example, the conversion model learning data selection unit 11 uses the word graph included in the learning speech recognition result data D14 and the learning formatted text D15 to learn to generate a conversion model according to the application. Select data. The conversion model learning data selection unit 11 compares, for example, two recognition result candidates indicated by the word graph included in the learning speech recognition result data D14 and the learning formatted text D15, and extracts different portions. The learning data of the conversion model corresponding to the application is selected from the extracted set of different parts. For example, in the example shown in FIG. 4, two recognition result candidates “E / Blood / I am doing”, “E / Marriage / I am doing”, and the corresponding formatted text D15 for learning “ The differences between “marriage / do / do” are (1) deletion of “e”, (2) “blood stain” replaced with “marriage”, and (3) “doing” You may extract three of “Replace”.

  Hereinafter, similarly to the example of the number of appearances, the selection method of this example will be specifically described using a case where the minutes are used as the learning-formatted text D15 as an example.

  In the speech recognition result data, each word of speech data input by the speech recognition process is transcribed into characters, so unnecessary words “e” and colloquial expression “setsu” that are inappropriate for the minutes are also included. If there is no recognition error, the text representation is output as it is. Therefore, the recognition result candidates included in the word graph, which is an example of the speech recognition result, include the unnecessary word “e” and the colloquial expression “setsu”, while the text of the minutes includes those words. There may be differences such as no. Focusing on this difference, among the sets of partial character strings of different parts, a set including a partial character string that is not included in the word graph is determined as data related to the shaping process, and other data (the different part is the word graph) Learning data for generating a conversion model corresponding to the application is acquired by determining that the group including the partial character string included in is data related to the correction process.

  In the case of the sorting method using the word graph, the learning formatted text D15 as the sorting source is not the data created under the same condition such as the data of the entire conference record, but the data sorted for each sentence. It may be.

  Further, in this example, an example in which whether or not a partial character string of a different part is included in the word graph is used as a selection criterion for learning data, for example, the part of speech or semantic information of a word included in the speech recognition result, It is also possible to obtain the likelihood that different partial character strings can exist in the word graph using acoustic likelihood, language likelihood, reliability, etc., and select them based on the likelihood.

  For example, in the example shown in FIG. 4, only the word notation of the speech recognition result is shown. For each word, the part of speech and the semantic information, the acoustic likelihood that is the acoustic likelihood of the word, the language of the word Information such as language likelihood, which is accuracy, and reliability, which is accuracy of the word, may be given as information.

  Using these pieces of information, it is only necessary to determine the likelihood of whether or not different character strings exist in the word graph. For the word part of speech and semantic information, for example, the likelihood is set for each part of speech (or semantic information), and the likelihood of the word part of speech (or semantic information) included in a partial character string of a different part is equal to or less than a predetermined threshold value. In the case of the case, even if there is a partial character string of the different part in the word graph, it is treated as not existing, and the learning data according to the desired application is selected according to the presence or absence of the result Also good.

  For acoustic likelihood, language likelihood, and reliability, the values themselves may be used as likelihoods to determine whether or not they are present. For example, learning data corresponding to a desired application may be selected on the assumption that a reliability of 0.1 or less does not exist in the word graph.

  A plurality of likelihoods of these pieces of information may be used. For example, if the numerical value obtained by adding the likelihood of each information is equal to or less than the reference value, a process such as not existing in the word graph may be performed. Furthermore, it is also possible to use a statistically learned model that uses one or more of these pieces of information to output the likelihood of whether or not there are different partial character strings in the word graph. For example, the existence likelihood may be calculated and output using a maximum entropy model defined as the following equation (1).

  In Expression (1), x is a speech recognition result such as a word notation, its part of speech, and semantic information. Further, y is a label indicating whether or not there is a partial character string of a different part in the word graph. P (y | x) is the likelihood of y with respect to x, that is, the likelihood of whether or not there are different partial character strings in the word graph. Λ is a set of model parameters. Φ (y, x) is a set of feature value values of the model, and is information such as word part of speech or a combination thereof. Zx is a normalization term. As a statistically learned model, a neural network, a hidden Markov model, a support vector machine, or the like can be used.

  In the above examples, the word graph has been described as an example, but the same determination method can be implemented even in the case of N best. Specifically, the process selected based on whether or not it is included in the word graph may be selected based on whether or not it is included in the N best.

  In the above example, the learning data selection process and the process of generating the conversion model have been described by taking the example of the “text shaping for improving readability” application used for the minutes and subtitles as the application for the shaping process. In addition to this, shaping processing is also performed for selection of learning data and generation of conversion models for “speech translation” for machine translation using speech recognition results and “speech understanding” for machine operation Is possible.

  The present invention can be applied to, for example, “text shaping for obtaining better speech translation results”. Speech translation is a technology that recognizes a certain utterance and translates the result into another language by machine translation. For example, a Japanese utterance “I want to go to the station” is input, and the utterance is speech recognized and machine-translated to obtain an English output “I want to go to the station.”. Hereinafter, in order to obtain a good speech translation result, a case where a fixed sentence of machine translation is used will be described as an example.

  FIG. 5 is an explanatory diagram showing an example of various data when the “text shaping for obtaining a better speech translation result” application is applied. As shown in Fig. 5, for example, an utterance “I want to go to the station” is input as an input voice related to the standard sentence of machine translation “I want to go to the station”, and the speech recognition result is “ Eh / I want to go to / to the tree. " Also, as a preformatted text D15 for learning that includes at least one better speech translation result, machine translation for the speech recognition result “Eh / I want to go to the tree” Suppose that the fixed sentence “I want to go to the station” is input as it is corrected and shaped.

  The conversion model learning data selection unit 11 aligns the learning-formatted text D15 and the learning speech recognition result data by aligning the learning-formatted text D15 and the learning speech recognition result data, similarly to the generation process of the generation model that uses text shaping for preparing the minutes. However, it is only necessary to extract “deletion of“ e ””, “replace“ to wood ”with“ to station ””, and “replace“ I want to go ”with“ want to go ”” as different parts.

  Then, as in the case of text formatting for the minutes, a predetermined sorting process is performed based on the number of appearances and whether it is included in the word graph (or N best). In this example, as a result, “replace“ in the wood ”with“ in the station ”” is determined as the speech recognition error correction processing, and other different parts are shaped and processed to obtain better speech translation results. Determined.

  The present invention can also be applied to, for example, “text shaping for speech understanding”. Speech understanding is a technique for recognizing a certain utterance, understanding the intention of the utterance using the recognition result, and performing processing according to the intention. For example, it is used for the operation of a car navigation system using voice recognition technology. Hereinafter, a case where a fixed sentence provided for speech understanding is used for speech understanding will be described as an example.

  FIG. 6 is an explanatory diagram showing an example of various data when the “text shaping for speech understanding” application is applied. As shown in FIG. 6, for example, it is assumed that there is “I want to go to Kyoto” (a fixed sentence that sets the destination of the car navigation system in Kyoto) as a fixed sentence for voice understanding of the car navigation system. Similarly to the application to speech translation, the utterance “I want to go to Kyoto” is input as the input speech related to the standard sentence “I want to go to Kyoto” for speech understanding of the car navigation system, and the speech recognition result As a result, “Eh / I want to go to today / I want to go”. Also, as a preformatted text D15 for learning, which is at least one of the purpose of understanding speech, the standard text of the car navigation system “Kyoto in Kyoto” for the speech recognition result “Eh / Today / I want to go” Suppose you input something that has been modified and shaped to "/ I want to go".

  The conversion model learning data selection means 11 aligns the learning-formatted text D15 and the learning speech recognition result data in the same manner as the generation process of the generation model that uses text shaping for the preparation of conference minutes and speech translation as one use. By extracting “Deleting“ E ””, “Replacing“ Today ”with“ Kyoto ””, and “Replacing“ I want to go ”with“ I want to go ”” as different parts Good.

  Then, in the same manner as in the case of text formatting for conference record creation and speech translation, a predetermined selection process is performed based on the number of appearances and whether or not it is included in the word graph (or N best). In this example, as a result, “Replace“ Today and Toni ”with“ Kyoto ”” is determined as the speech recognition error correction processing, and other differences are determined as the shaping processing for speech understanding. .

  As described above, according to the present embodiment, based on the learning speech recognition result data D14 and the learning-prepared text D15, the conversion model learning data selection unit 11 performs “speech recognition included in the speech recognition result”. The learning data according to the conversion purpose such as “correction of error” and “text shaping such as improving readability” is automatically selected, and the conversion model generation means 12 performs conversion according to each application using each learning data. Configured to generate a model. Thereby, it is possible to easily generate a conversion model corresponding to each application such as “correction of a speech recognition error included in a speech recognition result” and “text shaping such as improving readability” with less effort.

  In general, the process of “text formatting such as improving readability” has a characteristic that it greatly depends on the type of conference record to be generated and the creator of the conference record. On the other hand, the processing of “correction of speech recognition error included in speech recognition result” does not depend on them. Therefore, it is possible to create a conversion model by extracting only independent conversion processes, or to create a conversion model by extracting only dependent conversion processes. For example, “a speech recognition error included in a speech recognition result” The conversion model corresponding to the “correction” application can be easily applied even when the type of minutes and the creator of the minutes are different. In such cases, automatic correction of speech recognition errors can be efficiently performed. It becomes possible.

  As an example of the processing result of “text formatting to improve readability” differ depending on the purpose of the sentence creation and the creator, for example, if it is a minutes of a government assembly such as the Diet, utterance to prevent the deliberation of the intention of the member of the Diet It is common to accurately describe the contents of the remarks without summarizing their remarks. On the other hand, in a conference record of a company or the like, it is often the case that only the gist summarizing the points of a conference is described rather than describing the content of a statement accurately.

  In addition, if the creators of the minutes are different, the summary of the meeting points is often different for each creator. In addition, not only the gist but also the expression may be different, such as “Iso Masune” and “Naru Tone”, so that the processing content for the speech recognition result is also different.

  As described above, although the processing content for the speech recognition result varies depending on the type of the minutes and the creator, according to the present invention, the processing independent of the purpose of the document creation and the creator (in this example, “correction of voice recognition error”). Process) and the process that depends on the purpose of the document creation and the creator (in this example, the “text shaping” process) are discriminated based on a predetermined standard, and data relating to each process is automatically selected as learning data. Therefore, a conversion model corresponding to each process can be created.

  If a conversion model is generated without distinguishing between “independent processing” and “dependent processing” data as before, the generated conversion model is inevitably converted depending on the purpose of creating the document and the creator. Become a model. Such a conversion model is difficult to apply when the given learning data is different from the purpose of document creation and the creator. On the other hand, in the present embodiment, data relating to each process is selected by distinguishing between “independent processes” and “dependent processes”, and a conversion model is generated using data extracted corresponding to each process. For example, a conversion model generated from learning data extracted corresponding to “independent processing” can be applied even when the document used as learning data is different in purpose and creator. Thereby, even if it is difficult to apply with the conventional conversion model, it is possible to reduce the effort of sentence creation using the speech recognition technology by using the conversion model of “independent processing”.

  For example, a conversion model for “speech recognition error correction” generated by using the minutes of the government assembly for the learning-ready text can be used for making minutes of a general company.

  Further, for example, if learning speech recognition result data D14 is given for each different pattern, and the difference between the conversion models generated as a result is taken, a conversion model that performs text shaping depending on the purpose of the document creation and the creator is generated. It is also possible.

Embodiment 2. FIG.
Next, a second embodiment of the present invention will be described. In this embodiment, a speech recognition result conversion model for correcting a speech recognition result error is used as an application example using the conversion model generated by the speech recognition result conversion model generating apparatus described in the first embodiment. A speech recognition result correction system will be described. FIG. 7 is a block diagram illustrating a configuration example of the speech recognition result correction system according to the present embodiment. The speech recognition result correction system 200 shown in FIG. 7 includes each component (11 to 15) of the speech recognition result conversion model generation device 100 shown in FIG. 1, a correction speech recognition result storage unit 21, and speech recognition result correction means. 22 and a corrected text storage unit 23.

  Note that the conversion model storage unit 13 in the present embodiment only needs to store at least a conversion model for correcting speech recognition errors. In FIG. 7, in order to clearly indicate that a conversion model for correcting speech recognition errors is included, the conversion model storage unit 13A is distinguished. Note that a conversion model corresponding to another application may be included. In such a case, for example, an identifier that can specify which is a conversion model for correcting speech recognition errors may be added.

  In this embodiment, the conversion model learning data selection unit 11 selects only learning data for generating a conversion model for correcting speech recognition errors, and the conversion model generation unit 12 uses the learning data to You may perform limited implementation, such as producing | generating only the conversion model for recognition error correction.

  The correction speech recognition result storage unit 21 stores correction speech recognition result data. The correction speech recognition result data is information indicating a speech recognition result obtained by speech recognition of certain speech data using a certain speech recognition engine, which is a correction target in the present system.

  The speech recognition result correction means 22 corrects (converts) the correction speech recognition result data using the conversion model D13 for correcting speech recognition errors stored in the conversion model storage unit 13A, and generates corrected text. . For example, the speech recognition result correcting unit 22 defines a conversion rule that “replaces“ blood stain ”with“ marriage ”” as an example of a conversion model generated for the purpose of correcting a speech recognition error. Scans the input correction speech recognition result data for the location to which the conversion rule is applied, and if it is included, performs processing to replace “blood stain” with “marriage” By doing so, the corrected text may be generated.

  FIG. 8 is an explanatory diagram showing examples of correction speech recognition result data and corrected text generated based thereon. As shown in FIG. 8, for example, when correction speech recognition result data “blood stains” is input, “married” is generated as the corrected text.

  The corrected text storage unit 23 stores the corrected text generated by the voice recognition result correction unit 22, that is, the text obtained as a result of applying the conversion model for correcting the voice recognition error to the correction voice recognition result data. To do.

  Note that the method for generating a conversion model for correcting speech recognition errors may be the same as in the first embodiment. In the present embodiment, the speech recognition result correction unit 22 is realized by a CPU or the like that operates according to a program. The correction speech recognition result storage unit 21 and the corrected text storage unit 23 are realized by a storage device such as a memory. Each of these means may be implemented as an apparatus different from the speech recognition result conversion model generation apparatus 100 or may be implemented as one apparatus.

  As described above, according to the present embodiment, the conversion process performed on the learning data given to the speech recognition result conversion model generation device 100 is different from the conversion process desired for creating a sentence for the correction speech recognition result data. Even when a thing (for example, conversion processing for text shaping) is included, automatic correction of speech recognition errors can be performed efficiently.

Embodiment 3. FIG.
Next, a third embodiment of the present invention will be described. In this embodiment, as an application example using the conversion model generated by the speech recognition result conversion model generation device described in the first embodiment, speech recognition result shaping using a speech recognition result conversion model for the purpose of text shaping. The system will be described. FIG. 9 is a block diagram illustrating a configuration example of the speech recognition result shaping system of the present embodiment. A speech recognition result shaping system 300 shown in FIG. 9 includes each component (11 to 15) of the speech recognition result conversion model generation device 100 shown in FIG. 1, a shaping speech recognition result storage unit 31, and speech recognition result shaping means. 32 and a formatted text storage unit 33.

  Note that the conversion model storage unit 13 in the present embodiment only needs to store at least a conversion model for text shaping. In FIG. 9, in order to clearly indicate that a conversion model for text shaping is included, the conversion model storage unit 13 </ b> B is distinguished and described. Note that a conversion model corresponding to another application may be included. In such a case, for example, an identifier that can identify which is a conversion model for text formatting may be added.

  In this embodiment, the conversion model learning data selection unit 11 selects only learning data for generating a conversion model for text shaping, and the conversion model generation unit 12 uses the learning data for text shaping. A limited implementation may be performed, such as generating only the conversion model.

  The shaping speech recognition result storage unit 31 stores shaping speech recognition result data. The speech recognition result data for shaping is information indicating a speech recognition result obtained by performing speech recognition on certain speech data using a certain speech recognition engine, which is a shaping target in this system.

  The speech recognition result shaping means 32 shapes (converts) the shaping speech recognition result data using the text shaping conversion model D13 stored in the conversion model storage unit 13B, and generates a shaped text. For example, the speech recognition result shaping means 32 defines a conversion rule that “replaces“ I do ”” as “is” as an example of a conversion model generated for the purpose of text formatting to improve readability. If it is included, the input voice recognition result data for correction is scanned for the part to which the conversion rule is applied. A preformatted text may be generated by performing a process of replacing “mass”.

  FIG. 10 is an explanatory diagram showing examples of the speech recognition result data for shaping and the formatted text generated based on the data. As shown in FIG. 10, for example, when shaping speech recognition result data “blood stain” is input, “blood stain” is generated as the shaped text.

  The shaped text storage unit 33 stores the shaped text generated by the speech recognition result shaping unit 32, that is, the text obtained as a result of applying the text shaping conversion model to the shaping speech recognition result data.

  Note that a method for generating a conversion model for text shaping may be the same as in the first embodiment. In the present embodiment, the speech recognition result shaping means 32 is realized by a CPU or the like that operates according to a program. The shaping speech recognition result storage unit 31 and the shaped text storage unit 33 are realized by a storage device such as a memory. Each of these means may be implemented as an apparatus different from the speech recognition result conversion model generation apparatus 100 or may be implemented as one apparatus.

  As described above, according to the present embodiment, the conversion process performed on the learning data given to the speech recognition result conversion model generation device 100 is different from the conversion process desired for creating a sentence for the shaping speech recognition result data. Even if a thing (for example, conversion processing for correcting speech recognition errors) is included, desired text shaping can be performed efficiently.

Embodiment 4 FIG.
Next, a fourth embodiment of the present invention will be described. In the present embodiment, as an application example using the conversion model generated by the speech recognition result conversion model generation device described in the first embodiment, a speech recognition result conversion model for the purpose of correcting a speech recognition result error; A speech recognition result conversion system using a speech recognition result conversion model for text shaping will be described. FIG. 11 is a block diagram illustrating a configuration example of the speech recognition result conversion system according to the present embodiment. A speech recognition result conversion system 400 shown in FIG. 11 includes components (11 to 15) of the speech recognition result conversion model generation device 100 shown in FIG. 1, a correction / shaping speech recognition result storage unit 41, and a speech recognition result. The correction unit 22, the speech recognition result shaping unit 32, and the converted text storage unit 43 are provided.

  Note that the conversion model storage unit 13 in the present embodiment only needs to store at least a conversion model for correcting speech recognition errors and a conversion model for text shaping. In FIG. 11, in order to clearly indicate that a conversion model for correcting speech recognition errors and a conversion model for text shaping are included, the conversion model storage unit 13A and the conversion model storage unit 13B are distinguished from each other. However, these may be realized by one conversion model storage unit 13. In such a case, for example, an identifier that can identify the conversion model for which application may be added.

  FIG. 12 is an explanatory diagram showing examples of correction / shaping speech recognition result data and corrected / shaped text. As shown in FIG. 12, for example, when corrected and recognizing speech recognition result data “blood stains” is input, firstly, “I am married” is generated as the corrected text, By giving the voice recognition result shaping means 32 as the voice recognition result data for shaping, “married” is generated as the corrected and shaped text.

  The converted text storage unit 43 converts the voice recognition error to the corrected and shaped text generated by the voice recognition result correcting unit 22 and the voice recognition result shaping unit 32, that is, correction / shaping voice recognition result data. The text obtained as a result of applying the model and the conversion model for text formatting is stored.

  In the present embodiment, an example in which two conversion models are used together is shown. However, the user may select which conversion model to use. For example, the voice recognition result correction means 22 and the voice recognition result generation means 32 are provided with an instruction means for issuing an instruction for conversion processing using a conversion model, and the instruction means performs voice according to the conversion model instructed by the user. A control instruction may be issued to the recognition result correction unit 22 and the voice recognition result generation unit 32. Note that not only selection for speech recognition error correction and text shaping can be selected, but also a method of selecting one model designated by the user from a plurality of shaping conversion models is possible.

  Next, the outline of the present invention will be described. FIG. 13 is a block diagram showing an outline of the present invention. A conversion model generation apparatus 500 shown in FIG. 13 includes learning data sorting means 501 for each use and conversion model generation means 502.

  The learning data sorting unit 501 for each use (for example, the conversion model learning data sorting unit 11) is a learning speech recognition result data (for example, learning speech recognition result data D14) indicating the result of speech recognition processing for certain speech data. A speech recognition result text indicated by the learning speech recognition result data using a learning formatted text (for example, learning shaped text D15) indicating a result of converting the speech data into a text with a predetermined purpose, and Extract different parts from the pre-formatted text for learning, and select the learning data to generate a conversion model according to each predetermined application based on the appearance tendency of the extracted different parts .

  The conversion model generation unit 502 (for example, the conversion model generation unit 12) uses the learning data selected by the application-specific learning data selection unit 501 to generate a conversion model corresponding to an individual application.

  The application-specific learning data selection unit 501 may use, for example, the number of appearances of combinations of partial character strings existing in different parts as a reference for selecting a conversion model.

  In addition, for example, when the number of occurrences of combinations of partial character strings existing in the different parts is larger than a predetermined threshold, the data having the different parts is converted to text formatting and conversion usage. It may be selected as learning data for the conversion model. Further, for example, when the number of appearances of the combination of partial character strings existing in the different part is smaller than a predetermined threshold for the extracted different part, the data including the different part is converted into the voice included in the voice recognition result. You may select as the data for learning of the conversion model which uses correction of recognition error for conversion.

  In addition, the application-specific learning data selection unit 501 uses, for example, a partial character string existing in the speech recognition result text indicated by the learning speech recognition result data as a reference for selecting a conversion model, in a different part. The determination result as to whether or not it exists in the finished text may be used.

  For example, in the extracted different part, when the partial character string existing in the learning formatted text does not exist in the learning speech recognition result data, the data having the different part is converted using text shaping as a conversion It may be selected as model learning data. Also, for example, in the extracted different portion, when a partial character string existing in the learning formatted text is also present in the learning speech recognition result data, the data having the different portion is included in the speech recognition result. You may select as the learning data of the conversion model which uses correction of a speech recognition error for conversion.

  Further, the learning data selecting means 501 for each use includes, for example, N best included in the learning speech recognition result data, word graph, or part of speech or semantic information of a word included in the speech recognition result, acoustic likelihood, language likelihood, Different portions may be extracted based on the likelihood that can exist as speech recognition result text, which is indicated by any one or more of the reliability levels.

  Further, the application-specific learning data selection unit 501 extracts, for each different part, whether or not the conversion process for generating the different part is a text creation use or a process depending on the creator. By discriminating on the basis of the appearance tendency, the learning data for generating a conversion model corresponding to at least a speech recognition error correction application or a text shaping application for a certain purpose determined as an individual application is selected. Also good.

  FIG. 14 is a block diagram showing an outline of the speech recognition result conversion system of the present invention. A speech recognition result conversion system 600 illustrated in FIG. 14 includes a speech recognition result correction unit 601 in addition to the stepwise learning data selection unit 501 and the conversion model generation unit 502 illustrated in FIG.

  The voice recognition result correction means 601 (for example, the voice recognition result correction means 22) generates a conversion model by preliminarily determining that correction of a voice recognition error is one of individual uses when generating a conversion model. The speech recognition error included in the speech recognition result is corrected by performing conversion processing on the given speech recognition result data using a conversion model corresponding to the application for correcting the speech recognition error generated by the means 502. .

  FIG. 15 is a block diagram showing another configuration example of the speech recognition result conversion system of the present invention. A speech recognition result conversion system 600 illustrated in FIG. 15 includes a speech recognition result shaping unit 602 in addition to the stepwise learning data selection unit 501 and the conversion model generation unit 502 illustrated in FIG.

  The speech recognition result shaping means 602 (for example, the speech recognition result shaping means 32) converts the text shaping for a certain purpose in advance as one of the individual uses when generating the conversion model. By converting the given speech recognition result data using the conversion model generated by the model generation unit 502 according to the text shaping application for a certain purpose, the text matching the certain purpose is obtained. Perform shaping.

  As illustrated in FIG. 16, the speech recognition result conversion system 600 may include a speech recognition result correction unit 601 and a speech recognition result shaping unit 602.

  In such a case, it is possible to obtain text data in which the speech recognition error included in the speech recognition result is corrected and the text is shaped into text that matches a certain purpose.

  The present invention relates to a speech recognition device that converts speech data into text using speech recognition technology, a text shaping device that creates a conference record from speech data or speech recognition results, and a speech translation device that allows speech understanding to return a desired response. It can also be suitably applied to car navigation.

DESCRIPTION OF SYMBOLS 100 Speech recognition result conversion model production | generation apparatus 11 Conversion model learning data selection means 12 Conversion model production | generation means 13 Conversion model memory | storage part 13A Speech recognition error correction conversion model memory | storage part 13B Text shaping conversion model memory | storage part 14 Learning voice recognition result memory | storage Unit 15 Reformed text storage unit for learning 200 Speech recognition result correction system 21 Speech recognition result storage unit for correction 22 Speech recognition result correction unit 23 Corrected text storage unit 300 Speech recognition result shaping system 31 Speech recognition result storage unit for shaping 32 Speech recognition result shaping means 33 Shaped text storage section 400 Speech recognition result conversion system 41 Correction / shaping voice recognition result storage section 43 Converted text storage section 500 Conversion model generation means 501 Application-specific learning data selection means 502 Conversion model generation means 600 Voice recognition result conversion system 601 speech recognition result correction unit 602 speech recognition result formatting means

Claims (20)

Using the learning speech recognition result data indicating the result of speech recognition processing for certain speech data and the learning formatted text indicating the result of converting the speech data into text with a predetermined purpose, the learning speech A different portion between the speech recognition result text indicated by the recognition result data and the learning-prepared text is extracted, and based on the appearance tendency of the extracted different portion, the predetermined application is used. A learning data sorting means for each use for sorting learning data for generating a conversion model,
A conversion model generation device comprising: conversion model generation means for generating a conversion model corresponding to the individual application using the learning data selected by the application-specific learning data selection means. The conversion model generation device according to claim 1, wherein the use-specific learning data selection unit uses the number of appearances of a combination of partial character strings existing in different parts as a reference for selecting a conversion model. The use-specific learning data selection means performs text shaping on the data having the different parts when the number of occurrences of the combination of the partial character strings existing in the different parts is larger than a predetermined threshold. The conversion model generation apparatus according to claim 2, wherein the conversion model is selected as learning data for a conversion model to be used for conversion. The application-specific learning data selection means, for the extracted different parts, when the number of occurrences of the combination of the partial character strings existing in the different parts is smaller than a predetermined threshold, The conversion model generation apparatus according to claim 2 or 3, wherein the correction of the speech recognition award included in the data is selected as learning data for a conversion model for conversion purposes. The learning data selecting means for each application uses a partial character string existing in the speech recognition result text indicated by the learning speech recognition result data in the different parts as a reference for selecting the conversion model in the learning formatted text. The conversion model production | generation apparatus of any one of Claims 1-4 using the determination result of whether to do. The use-specific learning data selection means is one of N best, word graph included in the learning speech recognition result data, or part of speech or semantic information of words included in the speech recognition result, acoustic likelihood, language likelihood, or reliability. The conversion model generation apparatus according to any one of claims 1 to 5, wherein a different portion is extracted based on a likelihood that may exist as speech recognition result text indicated by one or more. The learning data sorting means for each use, when the partial character string existing in the learning shaped text is not present in the learning speech recognition result data in the extracted different part, the data having the different part is subjected to text shaping. The conversion model generation apparatus according to any one of claims 1 to 6, wherein the conversion model is selected as learning data for a conversion model to be used for conversion. The use-specific learning data selection means recognizes the data having the different parts in the extracted different parts when the partial character strings existing in the learning formatted text are also present in the learning voice recognition result data. The conversion model generation device according to any one of claims 1 to 7, wherein correction of a speech recognition error included in a result is selected as learning data for a conversion model for use in conversion. Based on the appearance tendency of the extracted different parts, the learning data sorting means for each use determines whether the conversion process for generating the different parts is a process that depends on the purpose of creating the text or the creator. The learning data for generating a conversion model corresponding to at least a speech recognition error correction application or a text shaping application for a certain purpose determined as an individual application is selected by discrimination. The conversion model generation device according to any one of 8. It has been determined in advance that the correction of speech recognition errors will be one of the individual uses when generating a conversion model,
Using the learning speech recognition result data indicating the result of speech recognition processing for certain speech data and the learning formatted text indicating the result of converting the speech data into text with a predetermined purpose, the learning speech Extracting the different part between the speech recognition result text indicated by the recognition result data and the learning shaped text, and according to the predetermined individual use based on the appearance tendency of the extracted different part Learning data selecting means for each purpose for selecting learning data for generating a converted model,
Using the learning data selected by the application-specific learning data selecting means, a conversion model generating means for generating a conversion model according to the individual application,
Speech recognition included in the speech recognition result by performing conversion processing on the given speech recognition result data using a conversion model according to the application for correcting the speech recognition error generated by the conversion model generating means A speech recognition result conversion system comprising speech recognition result correction means for correcting an error. It is stipulated that text formatting for a certain purpose is one of the individual uses when generating a conversion model,
Using the learning speech recognition result data indicating the result of speech recognition processing for certain speech data and the learning formatted text indicating the result of converting the speech data into text with a predetermined purpose, the learning speech Extracting the different part between the speech recognition result text indicated by the recognition result data and the learning shaped text, and according to the predetermined individual use based on the appearance tendency of the extracted different part Learning data selecting means for each purpose for selecting learning data for generating a converted model,
Using the learning data selected by the application-specific learning data selecting means, a conversion model generating means for generating a conversion model according to the individual application,
By using the conversion model according to the text shaping application for the certain purpose generated by the conversion model generating means, the given voice recognition result data is converted to meet the certain purpose. A speech recognition result conversion system comprising speech recognition result shaping means for shaping into text. It is determined in advance that correction of speech recognition errors and text formatting for a certain purpose will be used individually when generating a conversion model,
Using the learning speech recognition result data indicating the result of speech recognition processing for certain speech data and the learning formatted text indicating the result of converting the speech data into text with a predetermined purpose, the learning speech Extracting the different part between the speech recognition result text indicated by the recognition result data and the learning shaped text, and according to the predetermined individual use based on the appearance tendency of the extracted different part Learning data selecting means for each purpose for selecting learning data for generating a converted model,
Speech recognition included in the speech recognition result by performing conversion processing on the given speech recognition result data using a conversion model according to the application for correcting the speech recognition error generated by the conversion model generating means A speech recognition result correcting means for correcting an error;
By using the conversion model according to the text shaping application for the certain purpose generated by the conversion model generating means, the given voice recognition result data is converted to meet the certain purpose. A speech recognition result conversion system comprising speech recognition result shaping means for shaping into text. Using the learning speech recognition result data indicating the result of speech recognition processing for certain speech data and the learning formatted text indicating the result of converting the speech data into text with a predetermined purpose, the learning speech A different portion between the speech recognition result text indicated by the recognition result data and the learning-prepared text is extracted, and based on the appearance tendency of the extracted different portion, the predetermined application is used. A learning data selection step for each purpose for selecting learning data for generating a conversion model,
A conversion model generation step of generating a conversion model corresponding to the individual application using the learning data selected in the application-specific learning data selection step. The conversion model generation method according to claim 13, wherein the number of occurrences of combinations of partial character strings existing in different parts is used as a reference for selecting a conversion model in the learning data selection step according to use. In the use-specific learning data selection step, as a reference for selecting the conversion model, a partial character string existing in the speech recognition result text indicated by the speech recognition result data for learning is changed into a preformatted text for learning in a different part. The conversion model generation method according to claim 13 or 14, wherein a determination result of whether or not it exists is used. In the learning data selection step by use, any of N best, word graph, part of speech or semantic information of words included in the speech recognition result data for learning, speech likelihood, speech likelihood, language likelihood, reliability The conversion model generation method according to any one of claims 13 to 15, wherein a different portion is extracted based on a likelihood that may be present as speech recognition result text indicated by one or more. Based on the appearance tendency of the extracted different parts, whether or not the conversion process for generating the different parts is a process that depends on the purpose of creating the text or the creator, The learning data for generating a conversion model corresponding to at least a speech recognition error correction application or a text shaping application for a certain purpose determined as an individual application is selected by discrimination. The conversion model generation method according to any one of 16. Using the learning speech recognition result data indicating the result of speech recognition processing for certain speech data and the learning formatted text indicating the result of converting the speech data into text with a predetermined purpose, the learning speech A different portion between the speech recognition result text indicated by the recognition result data and the learning-prepared text is extracted, and based on the appearance tendency of the extracted different portion, the predetermined application is used. A learning data selection step for each purpose for selecting learning data for generating a conversion model,
A conversion model generation step for generating a conversion model according to the individual application using the learning data selected in the application-specific learning data selection step;
A speech recognition result conversion step including a speech recognition result conversion step of performing conversion processing on a given speech recognition result using a conversion model generated in the conversion model generation step according to an individual application. Method. On the computer,
Using the learning speech recognition result data indicating the result of speech recognition processing for certain speech data and the learning formatted text indicating the result of converting the speech data into text with a predetermined purpose, the learning speech A different portion between the speech recognition result text indicated by the recognition result data and the learning-prepared text is extracted, and based on the appearance tendency of the extracted different portion, the predetermined application is used. A learning data selection process for each use for selecting learning data for generating a conversion model,
A conversion model generation program for executing a conversion model generation process for generating a conversion model according to the individual application using the learning data selected in the application-specific learning data selection process. On the computer,
Using the learning speech recognition result data indicating the result of speech recognition processing for certain speech data and the learning formatted text indicating the result of converting the speech data into text with a predetermined purpose, the learning speech A different portion between the speech recognition result text indicated by the recognition result data and the learning-prepared text is extracted, and based on the appearance tendency of the extracted different portion, the predetermined application is used. A learning data selection process for each use for selecting learning data for generating a conversion model,
A conversion model generation process that generates a conversion model according to the individual application using the learning data selected in the application-specific learning data selection process;
A speech recognition result conversion program for executing a speech recognition result conversion process for performing a conversion process on a given speech recognition result using a conversion model generated by the conversion model generation process according to an individual application. .

Images