JP2016045467A - Utterance evaluation device, utterance evaluation method and program - Google Patents

Abstract

PROBLEM TO BE SOLVED: To evaluate whether or not, there is a reading error in utterance.SOLUTION: A reading error creation part 161 of an utterance evaluation device 1 regards a word including a Chinese character out of words forming a sentence acquired from text data, as a processing object word, and then creates a way of reading of reading error. A sound production dictionary update part 162 associates the processing object word to which identification information is added, with the way of reading of reading error, and then registers, thereby updating a sound production dictionary. A language model update part 163 acquires association between words, from the text data, and based on association between words which is created by adding identification information to the processing object word included in the association between words, and a predetermined occurrence frequency applied to the created association between words, updates a language model. A recognition processing part 19 recognizes voice data based on the updated sound production dictionary and the language model. A user interface control part 14 outputs a reading error when the word to which identification information is added, is included in a result of voice recognition.SELECTED DRAWING: Figure 1

Description

  The present invention relates to an utterance evaluation device, an utterance evaluation method, and a program.

  There is a technique for automatically evaluating the quality of human speech (see, for example, Patent Document 1). In this technology, based on an acoustic model, a language model, and a phoneme duration model generated in advance from a native speaker's utterance, for example, utterance pronunciation, utterance speed, fluency of utterance from speech data For example, feature quantities that can be related to pronunciation are extracted. Then, the pronunciation is evaluated in sentence units and word units based on the extracted feature values.

JP 2006-84966 A

  In a conventional utterance evaluation device, pronunciation is evaluated based on a sentence to be uttered (correct sentence). Therefore, pronunciation cannot be evaluated unless the same sentence as the correct sentence is spoken. In addition, the evaluation cannot be performed without the utterance of the correct native speaker. Furthermore, the related art is an evaluation related to pronunciation, and does not evaluate whether the speaker's pronunciation is bad for the correct sentence or whether the speaker has mistaken reading. In the speech recognition process, there was no model that could recognize misreading.

  The present invention has been made in view of such circumstances, and provides an utterance evaluation apparatus, an utterance evaluation method, and a program that can evaluate whether or not there is a reading error in an utterance.

One aspect of the present invention is a pronunciation dictionary storage unit that stores a pronunciation dictionary in which a word and a reading of the word are associated with each other, a language model storage unit that stores a language model representing the ease of connection between words, and text data Among the words constituting the sentence acquired from the above, the word including kanji is set as a processing target word, and a reading error of the processing target word is generated based on a reading that can be taken by the kanji included in the processing target word By registering the reading error generation unit, the processing target word to which identification information indicating a reading error is added, and the reading error generated by the reading error generation unit with respect to the processing target word in association with each other A pronunciation dictionary updating unit that updates the pronunciation dictionary; and a connection between words from the text data, and the processing included in the acquired connection between the words A language model updating unit that updates the language model based on a connection between words generated by adding the identification information to a target word and a predetermined appearance frequency assigned to the generated connection between the words; and the pronunciation A recognition processing unit that recognizes speech data based on the pronunciation dictionary updated by the dictionary update unit and the language model updated by the language model update unit, and the identification information is added to the result of speech recognition by the recognition processing unit An utterance evaluation apparatus comprising: an output unit that outputs a reading error when a recorded word is included.
According to this invention, the utterance evaluation device generates a reading error by using a word including a kanji as a processing target word among words constituting a sentence acquired from text data, and adds the identification information to the processing target word. Are registered in the pronunciation dictionary in association with reading mistakes. When the utterance evaluation device acquires the connection between words from the text data, the utterance evaluation device adds identification information to the processing target word included in the acquired connection between words and adds words to each other that include a word that has been given a reading error. A connection is generated and given frequency of appearance is given. The utterance evaluation device updates the language model based on the generated connection between the words and the given appearance frequency. The utterance evaluation device recognizes speech data based on the updated pronunciation dictionary and language model, and outputs a reading error when the speech recognition result includes a word with identification information added.
Thereby, the utterance evaluation apparatus can notify a reading error when there is a reading error in the utterance indicated by the voice data.

One aspect of the present invention is the utterance evaluation device described above, wherein the reading error generation unit is configured to process the word having a predetermined part-of-speech including a kanji among words constituting a sentence acquired from the text data. It is characterized by being a word.
According to this invention, the utterance evaluation device gives a reading error to a word having a predetermined part-of-speech word among words including kanji constituting the sentence indicated by the text data.
Thereby, the utterance evaluation apparatus can detect a reading error for a word having a predetermined part of speech.

One aspect of the present invention is the utterance evaluation device described above, wherein the language model update unit uses the text data for the connection between the words before adding the identification information to the generated connection between the words. An appearance frequency based on the calculated appearance frequency is assigned.
According to this invention, the utterance evaluation device sets the appearance frequency of a misread word based on the appearance frequency of a correctly read word.
As a result, the utterance evaluation device can make it easier to recognize misread words when speech recognition of utterances is performed even when the appearance probability of actual misread words is low.

  One aspect of the present invention is an utterance evaluation method executed by an utterance evaluation apparatus, wherein the word including kanji among words constituting a sentence acquired from text data is set as a processing target word, and is included in the processing target word A reading error generation step for generating a reading error of the word to be processed based on a reading that can be taken by the kanji, and a pronunciation dictionary in which the word and the reading of the word are associated with each other. A pronunciation dictionary updating step for updating the pronunciation dictionary by registering the added processing target word and the reading mistake generated in the reading error generation step in association with the processing target word; , Acquiring a connection between words from the text data, and identifying the processing target word included in the acquired connection between the words A language model update step for updating a language model representing the ease of connection between words based on the connection between words generated by adding a report and the predetermined appearance frequency assigned to the generated connection between the words; Recognition processing step for recognizing speech data based on the pronunciation dictionary updated in the pronunciation dictionary update step and the language model updated in the language model update step, and the identification in the result of speech recognition in the recognition processing step An utterance evaluation method comprising: an output step of outputting a reading error when a word to which information is added is included.

  According to one aspect of the present invention, a computer stores a pronunciation dictionary storage unit that stores a pronunciation dictionary in which a word and a reading of the word are associated with each other, and a language model storage unit that stores a language model that indicates the ease of connection between words. The word including kanji in the words constituting the sentence acquired from the text data is set as the processing target word, and the reading error of the processing target word is read based on the reading that the kanji included in the processing target word can take. A registration error generation unit that generates identification error information, a processing target word to which identification information indicating a reading error is added, and a reading method of the reading error generated by the reading error generation unit with respect to the processing target word. The pronunciation dictionary updating means for updating the pronunciation dictionary by acquiring the connection between words from the text data, and the acquired words A language model that updates the language model based on a connection between words generated by adding the identification information to the processing target word included in a connection and a predetermined appearance frequency assigned to the generated connection between the words Updating means, recognition processing means for recognizing speech data based on the pronunciation dictionary updated by the pronunciation dictionary updating means and the language model updated by the language model updating means, and a result of speech recognition by the recognition processing means Is a program for functioning as an utterance evaluation apparatus comprising: an output means for outputting a reading error when the identification information is added to a word.

  According to the present invention, it is possible to evaluate whether an utterance has a reading error.

It is a functional block diagram which shows the structure of the speech evaluation apparatus by one Embodiment of this invention. It is a flowchart which shows the whole process of the speech evaluation apparatus by the embodiment. It is a flowchart which shows the pronunciation dictionary and language model update process of the speech evaluation apparatus by the embodiment. It is a flowchart which shows the reading mistake indication process of the speech evaluation apparatus by the embodiment. It is a figure which shows the example of the reading mistake indication screen which the speech evaluation apparatus by the embodiment displays.

Hereinafter, embodiments of the present invention will be described in detail with reference to the drawings.
An utterance evaluation apparatus according to an embodiment of the present invention assigns a reading error to learning text data, and adapts a model used for speech recognition based on the given reading error, thereby detecting a reading error included in the utterance. Make it recognizable. Thereby, the utterance evaluation apparatus of this embodiment can evaluate whether there is a reading error in the speech recognition result. For example, the speech evaluation apparatus according to the present embodiment automatically points out a reading error using a voice recognition technology when an announcer, an actor, or the like reads a predetermined sentence such as a script.

  FIG. 1 is a functional block diagram showing a configuration of an utterance evaluation apparatus 1 according to an embodiment of the present invention, and only functional blocks related to the present embodiment are extracted and shown. The utterance evaluation device 1 can be realized by a computer device. As shown in the figure, the utterance evaluation apparatus 1 includes an acoustic model storage unit 11, a pronunciation dictionary storage unit 12, a language model storage unit 13, a user interface control unit 14, a text data acquisition unit 15, an update unit 16, and voice data acquisition. A unit 17, an acoustic feature amount extraction unit 18, and a recognition processing unit 19 are configured.

  The acoustic model storage unit 11 stores an acoustic model. The acoustic model is data in which a phoneme is associated with an acoustic feature amount of the phoneme. In the present embodiment, an HMM (Hidden Markov Model) acoustic model representing a statistical relationship between acoustic features and phonemes is used as the acoustic model. The pronunciation dictionary storage unit 12 stores a pronunciation dictionary. The pronunciation dictionary is data representing correspondence between words and readings. In the present embodiment, data representing the relationship between characters and words and their phonemes is used as the pronunciation dictionary. The language model storage unit 13 stores a language model. The language model is data representing the ease of connection between words. In the present embodiment, an n-gram language model representing the statistical probability of the order in which n elements (characters and words) appear in the language is used as the language model.

The user interface control unit 14 is an example of an output unit that outputs information, and displays a screen on a display device such as a display. The text data acquisition unit 15 acquires text data.
The update unit 16 includes a reading error generation unit 161, a pronunciation dictionary update unit 162, and a language model update unit 163. The misreading generation unit 161 acquires a sentence from text data, and if there is a word with a predetermined part-of-speech that includes kanji in the words constituting the acquired sentence, the misreading generation unit 161 assigns a reading error to the word. The pronunciation dictionary update unit 162 updates the pronunciation dictionary stored in the pronunciation dictionary storage unit 12 based on how to read the reading mistake given by the reading mistake generation unit 161. When the language model update unit 163 obtains the n-gram and the appearance frequency from the sentence acquired from the text data, the reading error generation unit 161 gives a reading error to the correct reading word included in the obtained n-gram. An n-gram replaced with the word is generated. The language model update unit 163 gives the appearance frequency based on the appearance frequency of the n-gram before the replacement with the word with the reading error to the n-gram generated by the replacement with the word with the reading error. The language model update unit 163 updates the language model stored in the language model storage unit 13 based on the generated n-gram and the given appearance frequency.

  The voice data acquisition unit 17 acquires voice data of an utterance. The acoustic feature amount extraction unit 18 extracts an acoustic feature amount from the voice data acquired from the voice data acquisition unit 17. The recognition processing unit 19 includes an acoustic model stored in the acoustic model storage unit 11, an updated pronunciation dictionary stored in the pronunciation dictionary storage unit 12, and an updated language model stored in the language model storage unit 13. , And the acoustic feature quantity extracted by the acoustic feature quantity extraction unit 18 is used to perform voice recognition of the voice data. The recognition processing unit 19 outputs the recognition result data in which the voice recognition result is set to the user interface control unit 14. When the recognition result data includes a misread word, the user interface control unit 14 causes the display device to display information notifying that the read error has occurred.

FIG. 2 is a flowchart showing the overall processing of the utterance evaluation apparatus 1.
First, the text data acquisition unit 15 of the utterance evaluation apparatus 1 acquires text data. The update unit 16 acquires a sentence from the text data acquired by the text data acquisition unit 15, and if the word constituting the acquired sentence includes a word with a predetermined part-of-speech including kanji, the word is misread Is granted. The updating unit 16 updates the pronunciation dictionary stored in the pronunciation dictionary storage unit 12 and the language model stored in the language model storage unit 13 based on how to read the given reading mistake (step S105). The update unit 16 adds identification information indicating that the word is misread to the word associated with the reading mistake and registers it in the pronunciation dictionary. Moreover, the update part 16 acquires n-gram from the sentence contained in text data. The update unit 16 reads the correct reading word by adding the reading error identification information to the word when the original word to which the reading error is given is included in the acquired n-gram. An n-gram replaced with an erroneous word is generated. The update unit 16 updates the language model with an n-gram that includes a misread word. In addition, about the n-gram containing the word of misreading, the update part 16 gives the appearance frequency based on the n-gram containing the correct word used as the origin which produced | generated the n-gram.

  After updating the pronunciation dictionary and the language model, the voice data acquisition unit 17 acquires the input voice data. The acoustic feature quantity extraction unit 18 extracts an acoustic feature quantity from the voice data acquired by the voice data acquisition unit 17. The recognition processing unit 19 uses the acoustic feature amount extracted by the acoustic feature amount extraction unit 18, the acoustic model stored in the acoustic model storage unit 11, and the pronunciation dictionary and language model updated in step S105. Recognize the input voice data. The recognition processing unit 19 outputs the recognition result data in which the voice recognition result is set to the user interface control unit 14. When the recognition result data includes a misread word to which identification information is added, the user interface control unit 14 causes the display device to display information notifying that there is a misread (step S110).

FIG. 3 is a flowchart showing the pronunciation dictionary and language model update processing of the utterance evaluation apparatus 1. This figure shows the details of the pronunciation dictionary and language model update processing in step S105 of FIG.
First, the user interface control unit 14 displays a screen for inputting text data such as a script or a manuscript on the display device. The text data acquisition unit 15 acquires text data input by an input unit (not shown) that performs keyboard input, file input, and the like, and outputs the text data to the update unit 16 (step S205). The update unit 16 updates the pronunciation dictionary based on the input text data by the processing from step S210 to step S250.

  When there is a sentence that has not yet been acquired in the text data (step S210: NO), the reading error generation unit 161 of the update unit 16 acquires one sentence that has not yet been acquired from the text data (step S215). The reading error generation unit 161 performs morphological analysis on the sentence acquired in step S215, divides the sentence into words, and acquires the part of speech of the word (step S220). If the reading error generation unit 161 determines that the word cannot be acquired from the sentence acquired in step S215 as a result of the morphological analysis (step S225: NO), the processing from step S210 is repeated.

  If the reading error generation unit 161 determines that a word can be acquired from the sentence acquired in step S215 (step S225: YES), the reading error generation unit 161 determines whether the sentence includes a noun word (step S230). When the reading error generation unit 161 determines that the acquired sentence does not include the noun word (step S230: NO), the reading error generation unit 161 repeats the processing from step S210.

  If the reading error generation unit 161 determines that the acquired sentence includes a noun word (step S230: YES), the reading error generation unit 161 decomposes the noun into characters one by one (step S235). The misreading generation unit 161 repeats the processing from step S210 when determining that kanji is not included in the character obtained by decomposing the noun (step S240: NO).

  On the other hand, if the misreading generation unit 161 determines that kanji is included in the character obtained by decomposing the noun (step S240: YES), the misreading is given to the noun including the kanji. (Step S245). Hereinafter, a noun word including kanji is referred to as a “processing target word”.

  For example, the phonetic reading and knot reading data of each kanji are stored in advance in a storage unit (or pronunciation dictionary storage unit 12) (not shown) provided inside or outside of the speech evaluation apparatus 1. These sound readings and knot reading data may be data represented by phonemes. The reading error generation unit 161 reads the kanji sound reading and kanji reading data included in the processing target word from the storage unit, thereby obtaining the kanji reading. For each word to be processed, the reading error generation unit 161 creates a reading error reading method using the kanji readings included in the processing word as follows.

  That is, the misreading generation unit 161 creates all combinations of readings obtained for kanji included in the processing target word, and sets the reading of the processing target word. The reading error generation unit 161 reads the correct reading of the processing target word from the pronunciation dictionary stored in the pronunciation dictionary storage unit 12. The reading error generation unit 161 determines reading errors other than the correct reading among the readings generated for the processing target word.

  For example, in the case of the noun “Kigata”, the readings of “Elephant” are “Sho”, “Elephant”, “Kun” and “Katado”, and “Seki” is read as “Kata”. And the cautionary reading “how” is read out. The misreading generation unit 161 changes “Shiseki”, “Shōkata”, “Elephant”, “Elephant”, etc. from all combinations of “Elephant” reading and “gata” reading. As a way to read "". The reading error generation unit 161 makes a reading error other than the correct reading “Kisakata” of “Kisakata” read from the pronunciation dictionary stored in the pronunciation dictionary storage unit 12.

  The pronunciation dictionary update unit 162 additionally registers a reading error of each processing target word created by the reading error generation unit 161 in step S245 in the pronunciation dictionary (step S250). The pronunciation dictionary update unit 162 assigns identification information indicating a reading error to the processing word when the processing word and the phoneme of the reading error of the processing target word are associated with the pronunciation dictionary and registered in the pronunciation dictionary. To do. For example, when registering the word “Kigata” and the misreading “Shiseki” in association with each other, identification information is added to the word “Kisakata”. In the present embodiment, “*” is added to a word as identification information to indicate a misread word. For example, a misread word generated from the word “Kigata” is “* Kigata”. Note that identification information is not given to words that are registered in the pronunciation dictionary in association with correct readings.

  Misread words registered in the pronunciation dictionary are usually not included in the language model learning data. Therefore, in the language model, words with a low appearance probability such as 0% are assigned to misread words. In this case, it is difficult to obtain a misread word as a recognition result in speech recognition. Therefore, next, through the processing from step S255 to step S280, the language model update unit 163 gives an appearance frequency higher than the actual to n-gram using a misread word and adds it to the language model. In this embodiment, it is assumed that a reading error occurs with the same appearance probability as the correct reading method, and the language model is adapted to the reading error.

  The language model update unit 163 acquires an n-gram that is a chain of n words based on a sentence included in the text data acquired by the text data acquisition unit 15 for updating the language model, and the n-gram of the n-gram is updated. The appearance frequency is calculated (step S255). If all the acquired n-grams have not been extracted (step S260: NO), the language model update unit 163 extracts one n-gram that has not yet been extracted (step S265). When the language model update unit 163 determines that the extracted n-gram does not include a word that is a generation source of a misread word (step S270: NO), the process from step S260 is repeated.

  If the language model update unit 163 determines that the n-gram extracted in step S265 includes the word that is the source of the misread word (step S270: NO), the language model update unit 163 includes the misread word. An n-gram is created (step S275). Specifically, the language model update unit 163 adds identification information indicating misreading to the word that is the generation source of the misread word from the extracted n-gram and adds a new n -Create a gram. The language model update unit 163 may calculate the occurrence frequency of the misread word to which the identification information is added, for example, the same appearance frequency as the word from which the misread word is generated, or a predetermined calculation for the appearance frequency. Appearance frequency obtained by performing is given. For example, if n = 2 and the n-gram extracted in step S265 is “Kisakata” → “no” (appearance frequency a), the language model update unit 163 adds identification information to “Kisakata” and adds n− The gram “* Kigata” → “no” (appearance frequency a) is generated. The language model update unit 163 repeats the processing from step S260.

  Then, the language model update unit 163 determines that all the n-grams acquired in step S255 have been extracted (step S260: YES). The language model update unit 163 updates the language model stored in the language model storage unit 13 using the n-gram acquired in step S255 and the n-gram generated in step S275 in the same manner as in the related art. To do.

  Note that the language model updating unit 163 adds a predetermined appearance frequency when creating a new n-gram by adding identification information indicating a reading error to words constituting the n-gram extracted from the text data. It may be given.

FIG. 4 is a flowchart showing a misreading indication process of the utterance evaluation apparatus 1. This figure shows the details of the misread indication process in step S110 of FIG.
The voice data acquisition unit 17 waits for voice data input (step S305: NO). When utterance voice data is input to the utterance evaluation device 1, the voice data acquisition unit 17 acquires the input voice data (step S305: YES). The utterance may be an utterance obtained by reading a script or a manuscript indicated by the text data acquired in step S205 in FIG. 3, or an utterance obtained by reading other sentences.

  The acoustic feature quantity extraction unit 18 extracts an acoustic feature quantity from the voice data acquired by the voice data acquisition unit 17. The recognition processing unit 19 performs speech recognition processing using the acoustic feature amount read from the acoustic feature amount extraction unit 18, the acoustic model, the pronunciation dictionary, and the language model as in the conventional technique. That is, the phoneme series is obtained by comparing the time-series acoustic feature quantity extracted by the acoustic feature quantity extraction unit 18 with the acoustic model stored in the acoustic model storage unit 11. The recognition processing unit 19 collates the phoneme sequence in the obtained phoneme sequence with the pronunciation dictionary stored in the pronunciation dictionary storage unit 12 to obtain a word string corresponding to the phoneme sequence. The recognition processing unit 19 uses the language model stored in the language model storage unit 13 to obtain the appearance probability of the obtained word string. The recognition processing unit 19 outputs recognition result data in which the word string having the highest appearance probability is set as the voice recognition result to the user interface control unit 14 (step S310).

  The user interface control unit 14 determines whether the word string indicated by the recognition result data received from the recognition processing unit 19 includes a word with identification information added (step S315). When the user interface control unit 14 determines that the speech recognition result indicated by the recognition result data does not include the word with the identification information added (step S315: NO), the utterance evaluation apparatus performs the processing from step S305. repeat. When the user interface control unit 14 determines that the speech recognition result indicated by the recognition result data does not include the word with the identification information added (step S315: NO), the utterance evaluation apparatus performs the processing from step S305. repeat. On the other hand, when the user interface control unit 14 determines that the speech recognition result indicated by the recognition result data includes a word to which the identification information is added (step S315: YES), the user interface control unit 14 displays a reading error on the display device. To notify the user (step S320). The utterance evaluation device 1 repeats the processing from step S305.

  FIG. 5 is a reading error indication screen displayed on the display device by the user interface control unit 14 of the speech evaluation apparatus 1. As shown in the figure, the user interface control unit 14 displays the speech recognition result indicated by the recognition result data output from the recognition processing unit 19 and, if the speech recognition result includes a misread word, the reading error. Display sentences that contain the word. In the figure, the last sentence (read mistake sentence) and the previous sentence (read mistake history) among the sentences detected as including misread words are displayed.

  As described above, the utterance evaluation apparatus 1 extracts words that may cause reading mistakes from the text data for learning, changes the notation of the words, and gives a reading method of reading mistakes (reading error candidate pronunciation strings). To add to the pronunciation dictionary. Furthermore, the utterance evaluation apparatus 1 additionally generates an n-gram in which the words constituting the n-gram extracted from the text data are replaced with words that are given a reading error, and updates the language model. As a result, the utterance evaluation apparatus 1 outputs the changed notation word when the utterance includes a reading error. Therefore, when a word included in the learning text data is not uttered with correct reading, it is possible to detect a reading error. In addition, since the utterance of a specific speaker is not used as learning data, it is possible to detect misreading for various utterances.

  The utterance evaluation apparatus 1 described above can automatically point out a speaker's reading error automatically for a predetermined content such as a manuscript, a script, or a textbook. For example, the utterance evaluation apparatus 1 reads the text aloud after updating the pronunciation dictionary and the language model in advance using the text as learning text data. When there is a reading error, the utterance evaluation device 1 displays that fact. This allows you to check whether there are any mistakes in your reading when pre-reading on the broadcast site or stage site, or to check in advance whether there is an error in reading aloud at the education site. Can lead to the transmission of correct information and smooth work and classes.

In a language-oriented profession, eliminating misreading is a major challenge. For example, an announcer maintains a word collection of easily misread words, and how to reduce misreading is a serious concern. Therefore, by using the utterance evaluation apparatus of the present embodiment, it is possible to discover reading mistakes in advance, leading to broadcasting correct information.
In addition to the misreading of the announcer, the reading of the actor's script can be practiced and recorded smoothly by practicing in advance using the speech evaluation apparatus of the present embodiment.
In addition, reading aloud in Japanese is an indispensable educational method in educational settings. At that time, by using the utterance evaluation apparatus of the present embodiment as if it were a game, it is possible to reduce the delay in the course progression due to a simple reading error of kanji and the shame caused by the student's reading error.

  The utterance evaluation apparatus 1 described above has a computer system inside. The operation process of the utterance evaluation apparatus 1 is stored in a computer-readable recording medium in the form of a program, and the above processing is performed by the computer system reading and executing this program. The computer system here includes a CPU, various memories, an OS, and hardware such as peripheral devices.

Further, the “computer system” includes a homepage providing environment (or display environment) if a WWW system is used.
The “computer-readable recording medium” refers to a storage device such as a flexible medium, a magneto-optical disk, a portable medium such as a ROM and a CD-ROM, and a hard disk incorporated in a computer system. Furthermore, the “computer-readable recording medium” dynamically holds a program for a short time like a communication line when transmitting a program via a network such as the Internet or a communication line such as a telephone line. In this case, a volatile memory in a computer system serving as a server or a client in that case, and a program that holds a program for a certain period of time are also included. The program may be a program for realizing a part of the functions described above, and may be a program capable of realizing the functions described above in combination with a program already recorded in a computer system.

  The embodiment of the present invention has been described in detail with reference to the drawings. However, the specific configuration is not limited to this embodiment, and includes designs and the like that do not depart from the gist of the present invention.

DESCRIPTION OF SYMBOLS 1 Speech evaluation apparatus 11 Acoustic model memory | storage part 12 Pronunciation dictionary memory | storage part 13 Language model memory | storage part 14 User interface control part 15 Text data acquisition part 16 Update part 17 Speech data acquisition part 18 Acoustic feature-value extraction part 19 Recognition processing part 161 Reading mistake Generation unit 162 Pronunciation dictionary update unit 163 Language model update unit

Claims (5)

A pronunciation dictionary storage unit that stores a pronunciation dictionary in which a word and a reading of the word are associated;
A language model storage unit for storing a language model representing the ease of connection between words;
Of the words constituting the sentence acquired from the text data, the word including the kanji is set as the processing target word, and the reading of the processing target word is read based on the reading that the kanji included in the processing target word can take. A reading error generation section to generate,
The pronunciation dictionary is updated by associating and registering the processing target word to which identification information indicating a reading error is added and the reading error generated by the reading error generation unit with respect to the processing target word. Pronunciation dictionary update part,
A connection between words is acquired from the text data, and a connection between words generated by adding the identification information to the processing target word included in the acquired connection between words and a connection between the generated words A language model update unit that updates the language model based on the given frequency of appearance;
A recognition processing unit that recognizes speech data based on the pronunciation dictionary updated by the pronunciation dictionary update unit and the language model updated by the language model update unit;
An output unit that outputs a reading error when a word to which the identification information is added is included in a result of speech recognition by the recognition processing unit;
An utterance evaluation apparatus comprising: The misreading generation unit sets the word of a predetermined part-of-speech including the kanji among the words constituting the sentence acquired from the text data as the processing target word.
The utterance evaluation apparatus according to claim 1. The language model update unit gives an appearance frequency based on the appearance frequency calculated from the text data for the connection between the words before adding the identification information to the generated connection between the words,
The utterance evaluation apparatus according to claim 1 or 2, characterized by the above. An utterance evaluation method executed by the utterance evaluation device,
Of the words constituting the sentence acquired from the text data, the word including the kanji is set as the processing target word, and the reading of the processing target word is read based on the reading that the kanji included in the processing target word can take. A reading error generation step to generate,
The processing target word in which identification information indicating a reading error is added to the pronunciation dictionary in which the word and the reading of the word are associated, and the reading error generated in the reading error generation step for the processing target word A pronunciation dictionary update step of updating the pronunciation dictionary by registering the readings in association with each other;
A connection between words is acquired from the text data, and a connection between words generated by adding the identification information to the processing target word included in the acquired connection between words and a connection between the generated words A language model update step for updating a language model representing the ease of connection between words based on the given frequency of appearance;
A recognition processing step for recognizing speech data based on the pronunciation dictionary updated in the pronunciation dictionary update step and the language model updated in the language model update step;
An output step of outputting a reading error when a word to which the identification information is added is included in a result of speech recognition in the recognition processing step;
An utterance evaluation method characterized by comprising: Computer
Pronunciation dictionary storage means for storing a pronunciation dictionary in which a word and a reading of the word are associated with each other;
A language model storage means for storing a language model representing ease of connection between words;
Of the words constituting the sentence acquired from the text data, the word including the kanji is set as the processing target word, and the reading of the processing target word is read based on the reading that the kanji included in the processing target word can take. A reading error generation means to generate,
The pronunciation dictionary is updated by associating and registering the processing target word to which identification information indicating a reading error is added and the reading mistake generated by the reading error generation unit with respect to the processing target word. Pronunciation dictionary update means;
A connection between words is acquired from the text data, and a connection between words generated by adding the identification information to the processing target word included in the acquired connection between words and a connection between the generated words Language model updating means for updating the language model based on the given frequency of appearance;
Recognition processing means for recognizing speech data based on the pronunciation dictionary updated by the pronunciation dictionary update means and the language model updated by the language model update means;
An output means for outputting a reading error when a word to which the identification information is added is included in a result of speech recognition by the recognition processing means;
A program for functioning as an utterance evaluation apparatus.

Images