JP7526416B2 - Accent estimation device and accent estimation method - Google Patents

Description

The present invention relates to an accent estimation device that outputs an accent to be assigned to the reading of a written text, an accent learning device, an accent estimation method, and an accent learning method.

Accent, which is the strength and pitch of the mora when a word is read aloud, plays an important role in making artificially created synthetic speech sound closer to natural human speech. A technique for estimating accents has been proposed that includes accent rules for shifting the accent of each word, and applies the accent rules to the structure of a compound word, thereby outputting the accent of the compound word (see, for example, Patent Document 1). There is also known a technique for outputting the prosody of reading from morphological analysis of input text, and modifying the prosody so that the speaking speed and intonation are natural, using a speech fragment sequence that matches the input text (see, for example, Patent Document 2).

Japanese Patent Application Publication No. 11-249866 JP 2008-139631 A

On the other hand, in Japanese, where compounds can be made by simply linking nouns, new words are being created every day, and the number of unknown words, such as words not registered in the dictionary used for estimating accents and compounds that are not linked in a way that conforms to the binding rules, has been steadily increasing in recent years. In addition, the number of words registered in the accent dictionary and the number of compounds that conform to the binding rules are limited, so the reality is that it is not possible to assign an appropriate accent to the reading of an unknown word. Even if a word is registered in the accent dictionary, it can be said that it is very difficult to estimate the appropriate accent when guessing the reading from the word written in kana, or when dealing with words that are assigned different accents depending on the context, in Japanese, where there are many homonyms, it is also very difficult to estimate the appropriate accent when dealing with words that are assigned different accents depending on the context. Furthermore, in Japanese, where it is easy to adopt foreign words, there are cases where the accent of a general katakana word is assigned to a foreign word, the accent of the original language is assigned to a foreign word, and cases where an accent that does not fit either of these is assigned, making it even more difficult to create an appropriate accent for the reading of an unknown word.

For example, when creating an accent for "World Heritage," which is not registered in the accent dictionary, the unknown word "World Heritage" is split into two morphemes, "World" and "Heris." If there is no combination rule that matches the two morphemes and the two morphemes are registered in the accent dictionary, the accent for "World Heritage" will simply be a concatenation of the accents of each morpheme. As a result, the correct accent would be "I\San (100)," but the incorrect accent "I/San (011)" is created, which is a flat accent. The same can be said for unknown words such as "Yakiniku Teishoku (grilled meat set meal)." Note that 0 and 1 here indicate the intonation pitch, with 0 being low and 1 being high. "\" and "/" indicate that the intonation gets lower and higher between the letters.

For example, when creating an accent for the new word "hadakata kaikaku" (work style reform) that is not registered in the accent dictionary, there is no combination rule that matches the new word, and "hadakata kaikaku" is split into three morphemes: "hada" (work), "kata" (way), and "kaikaku" (reform). And while the appropriate accent should be "ka\ikaku (1000)," the flat accent "ka/ikaku (0111)" is created.

For example, the accent of the reading inferred from the kana spelling "kougyou" differs between the kanji spelling "kougyou" (industry) and the kanji spelling "kougyou" (industry). The accent of the reading inferred from the kana spelling "ishi" also changes depending on whether the kanji spelling is "ishi" (will), "ishi" (doctor), or "ishi" (stone). Similarly, the accent of words used in place names and personal names, such as "Shibuya" and "Ueno," differs between the place names used in the context and the personal names used in the context. In addition, even among place names written in kana as "xxxxx-cho," some fall under the general accent combination rules, such as "ji/nbo\ucho (Jinbocho)" and "yu/ura\kucho (Yurakucho)," while others do not, such as "o/ote\machi (Otemachi)" and "ko/ujimachi (Kojimachi)." Additionally, words such as "croissant," "baumkuchen," and "basketball" are examples of words that use the accent of typical katakana words, with the accent placed on the third or fourth mora from the end of the katakana word. On the other hand, words such as "quartet," "swimming," and "advance" are given accents that are different from both the accent of typical katakana words and the accent of the original language.

The present invention aims to provide an accent estimation device, an accent estimation method, an accent learning device, and an accent learning method that can improve the accuracy of accent estimation.

The accent estimation device for solving the above problem includes an accent dictionary, which is a dictionary that associates spellings, readings, and accents, an analysis unit that analyzes a suspected spelling based on the accent dictionary and applies accents registered in the accent dictionary to the reading of the suspected spelling, including unknown words not registered in the accent dictionary, thereby outputting the dictionary accent of the suspected spelling, and an estimation unit that includes a machine-learned model that uses the spelling, reading, and dictionary accent as input values and outputs an appropriate accent for the spelling as an output value, and uses the suspected spelling, the reading of the suspected spelling, and the dictionary accent of the suspected spelling output by the analysis unit as input values of the model, and outputs the output value of the model as the accent of the suspected spelling.

The accent estimation method for solving the above problem includes: analyzing a presumed notation based on an accent dictionary, which is a dictionary that associates notations, readings, and accents; applying accents registered in the accent dictionary to the reading of the presumed notation, including readings of unknown words not registered in the accent dictionary, thereby outputting a dictionary accent of the presumed notation; and providing a machine-learned model that uses notations, readings, and dictionary accents as input values and outputs an appropriate accent for the notation as an output value, using the presumed notation, the reading of the presumed notation, and the dictionary accent of the presumed notation as input values of the model, and outputting the output value of the model as the accent of the presumed notation.

The accent learning device for solving the above problem includes an accent dictionary, which is a dictionary that associates spellings, readings, and accents, an analysis unit that analyzes a learning spelling based on the accent dictionary and applies accents registered in the accent dictionary to the readings of the learning spelling, including unknown words not registered in the accent dictionary, thereby outputting the dictionary accent of the learning spelling, and a learning unit that outputs a machine-learned model, in which the learning spelling, the reading of the learning spelling, and the dictionary accent of the learning spelling are input values of the model, and the learning unit trains the model by machine learning using an appropriate accent of the learning spelling as an ideal value for the output value of the model.

The accent learning method for solving the above problem includes analyzing a learning notation based on an accent dictionary, which is a dictionary that associates notations, readings, and accents, and applying accents registered in the accent dictionary to the readings of the learning notation, including readings of unknown words not registered in the accent dictionary, thereby outputting the dictionary accent of the learning notation, and outputting a machine-learned model, in which the learning notation, the reading of the learning notation, and the dictionary accent of the learning notation are used as input values for the model, and an appropriate accent of the learning notation is used as an ideal value for the output value of the model, and the model is trained by machine learning.

In the above accent estimation device, the analysis unit is a first analysis unit, the dictionary that associates notations with readings is a reading dictionary, and the device further includes a second analysis unit that analyzes the notation to be estimated based on the reading dictionary and outputs the reading of the notation to be estimated, and the estimation unit may use the reading of the notation to be estimated as an input value of the model as the reading of the notation to be estimated output by the second analysis unit.

In the above accent estimation device, the estimation unit may set a reading of the estimated spelling, which is used as an input value of the model, to an appropriate reading input by a user.
In the above accent estimation device, the analysis unit may analyze the estimated notation based on the accent dictionary and output a reading of the estimated notation, and the estimation unit may use the reading of the estimated notation that is to be used as an input value of the model as the reading of the estimated notation output by the analysis unit.

The present invention can assign appropriate accents to the pronunciation of unknown words, such as words not registered in an accent dictionary and compound words that are combined in a manner that does not conform to the combination rules, thereby improving the accuracy of accent estimation.

1 is a hardware block diagram showing the configuration of an accent estimation device according to an embodiment. FIG. 1 is a functional block diagram showing a configuration of an accent estimation device according to an embodiment. FIG. 1 is a schematic diagram showing a part of a machine learning model provided in an accent estimation device. FIG. 4 is a flowchart showing the flow of an accent estimation method performed by the accent estimation device. FIG. 1 is a functional block diagram showing the configuration of an accent learning device according to an embodiment of the present invention. 4 is a flow chart showing the flow of an accent learning method performed by the accent learning device.

With reference to Figures 1 to 6, an embodiment of an accent estimation device, accent learning device, accent estimation method, and accent learning method will be described.

[Estimation device]
As shown in Fig. 1, the accent estimation device 10 executes an accent estimation program to perform an accent estimation method. The accent estimation device 10 receives an estimation request 20Q from a user terminal 20. The accent estimation device 10 receives the estimation request 20Q and acquires an estimation notation 20D. The estimation notation 20D is a target for which an accent is to be estimated. The accent estimation device 10 includes the following four input values (A) to (D) in the input values of the machine learning model. The machine learning model outputs the accent of the estimation notation 20D using the four input values. The accent estimation device 10 provides the accent output by the machine learning model to the user terminal 20 as an answer 10A to the estimation request 20Q.
(A) Estimated notation 20D
(B) Reading of the presumed notation 20D (C) Part of speech of the presumed notation 20D (D) Dictionary accent of the presumed notation 20D (A) The presumed notation 20D is a string of characters written using letters, numbers, and symbols. The presumed notation 20D is a word that is the smallest unit that has a grammatical meaning, such as "World Heritage Sites". The presumed notation 20D is a sentence that has a syntactic function as a sequence of words, such as "The number of World Heritage Sites registered in Northern Europe is...". The presumed notation 20D is a sentence that expresses a single content as a sequence of sentences, such as "The number of World Heritage Sites registered in Northern Europe is 32".

(B) The reading of the presumed notation 20D is a character string indicating how to read the presumed notation 20D. The reading of the presumed notation 20D is a sequence of the smallest unit to which an accent can be applied. For example, if the presumed notation 20D is "World Heritage," the reading of the presumed notation 20D is "Sekai Isan." The smallest unit to which an accent can be applied is a string of one or more kana characters (mora), such as "se" or "ka." A string of one or more kana characters is a sound that can be expressed by a combination of one kana character (including the Japanese syllabary, the Japanese syllabary with voiced dots, the Japanese syllabary with semi-voiced dots (pa row), the long vowel mark "-", the short consonant "tsu", and the glottal stop "n"), or one kana character and one small kana character (aiweoweyowyayuyowa), and this unit is called a mora.

Accent is the strength and pitch of the beat when a word is read aloud or when the synthesized sound of the word is output. Accent includes intonation, which is the strength and pitch of the beat when a phrase or sentence is read aloud or when the synthesized sound of the phrase or sentence is output. Accent is the fundamental frequency pattern of the beat when a word is read aloud or when the synthesized sound of the word is output.

One accent is assigned to each beat, which is the basic unit of time into which the reading of a word is divided. Accents are written using, for example, the HL system, the linear system, or the numeric system. For example, when counting from the first letter of the three-letter word "Isan," where the pitch rises from the second letter onwards, the accents are written as "LHH," "I/San," "011," etc.

(C) The part of speech of the presumed notation 20D is the part of speech of the words that make up the presumed notation 20D. When the presumed notation 20D is written in Japanese, the parts of speech of the presumed notation 20D are, for example, verbs, adjectives, adjectival verbs, common nouns, proper nouns, pronouns, numerals, formal nouns, adverbs, conjunctions, interjections, auxiliary verbs, particles, adverbial particles, conjunction particles, final particles, prefixes, and suffixes. The parts of speech of the presumed notation 20D are each associated with the smallest unit that makes up the reading of the presumed notation 20D. For example, when the presumed notation 20D is "Isan," the reading of the presumed notation 20D is "Isan," and the parts of speech of the presumed notation 20D are associated with "I" as "common noun," "Sa" as "common noun," and "N" as "common noun."

(D) The dictionary accent of the presumed notation 20D is an accent that is applied to the reading of the presumed notation 20D based on the accent dictionary 13A. The dictionary accent of the presumed notation 20D is obtained by applying the reading of the word registered in the accent dictionary 13A to the reading of the presumed notation 20D. For example, suppose that the words registered in the accent dictionary 13A are "World" and "Isan" and "World Heritage" is not registered in the accent dictionary 13A. In this case, "Se\kai (100)" and "I/san (011)" are registered in the accent dictionary 13A, and the dictionary accent of "World Heritage" is "Se\kai I/san", which is a combination of these accents.

The accent dictionary 13A is a dictionary that associates the spelling of a word with the reading of the word, the part of speech of the word, and the accent of the word. For example, if the spelling of a word is "legacy," the reading of the word is "Isan," the part of speech of the word is "common noun" (three characters), and the accent of the word is "011" or "LHH." The accent dictionary 13A is used to assign accents to the reading of the estimated spelling 20D, i.e., to create dictionary accents.

The reading dictionary 13R is a dictionary that associates the notation of a word with the reading of the word and the part of speech of the word. For example, if the notation of a word is "ichi ni chin senshu", the reading of the word is "ichi jitsu senshu" and the part of speech of the word, "common noun" of eight characters, is associated with the word. The reading dictionary 13R is used to create the reading of the estimated notation 20D.

As mentioned above, the number of unknown words that are not registered in the accent dictionary 13A has been steadily increasing in recent years. In addition, the frequency with which the accent dictionary 13A is updated is actually much lower than the frequency with which the reading dictionary 13R is updated. If the reading of an unknown word is analyzed based on the accent dictionary 13A, it is possible to output the reading, but a word that originally has a single meaning is likely to be treated as a group of registered words that have a different meaning. For example, if the unknown word is written as "ichinichi chiaki," the reading is divided into the registered words "ichinichi" and "chiaki," and "ichinichi chiaki" is created as the reading of the writing. In contrast, if the reading of an unknown word is analyzed based on the reading dictionary 13R, the reading of the estimated notation 20D is more likely to be an appropriate reading than an analysis based on the accent dictionary 13A.

The input/output unit 21 of the user terminal 20 inputs the estimation notation 20D in accordance with an operation by the user. The input/output unit 21 of the user terminal 20 may also include a text conversion unit. The text conversion unit converts the user's voice input to the microphone into text information. The input/output unit 21 of the user terminal 20 inputs the text information converted by the text conversion unit as the estimation notation 20D.

The accent estimation device 10 includes a communication unit 11, a control unit 12, and a storage unit 13. The communication unit 11, the control unit 12, and the storage unit 13 are configured by one or more computers. The computers that configure the communication unit 11, the control unit 12, and the storage unit 13 are connected to a computer network. The computer that configures the communication unit 11, the computer that configures the control unit 12, and the computer that configures the storage unit 13 may be connected to separate closed communication networks that configure a wide area network, and may be connected to each other through gateways.

The communication unit 11 includes a communication device for connecting to the user terminal 20. The communication unit 11 is connected to the user terminal 20, for example, through a computer network within an organization. Alternatively, the communication unit 11 is connected to the user terminal 20 through the Internet, a satellite network, a mobile phone network, Wi-Fi (registered trademark), etc.

The communication unit 11 accepts input of an estimation request 20Q from the user terminal 20. The communication unit 11 inputs the accepted estimation request 20Q to the control unit 12. The communication unit 11 accepts an acquisition command input by the control unit 12 and executes acquisition of an estimation notation 20D. The communication unit 11 inputs the acquired estimation notation 20D to the control unit 12.

The control unit 12 includes a central processing unit and a memory. The control unit 12 is not limited to a unit that processes all of the various processes for accent estimation using software. The control unit 12 may include, for example, dedicated hardware (ASIC) that executes at least some of the various processes. In other words, the control unit 12 may be configured as a circuit that includes one or more dedicated hardware circuits such as an ASIC, one or more processors that operate according to a knowledge retrieval program, or a combination of these. Note that a circuit that includes a dedicated hardware circuit, a processor, or a combination of these may constitute two or more functional units.

The control unit 12 has a machine learning model for reading estimation. The control unit 12 functions as a first analysis unit and a second analysis unit. The machine learning model for reading estimation is a deep learning model equipped with a neural network. The machine learning model for reading estimation is a deep learning model that has been machine-trained to use the estimated notation 20D and the like as input values and the reading of the estimated notation 20D as an output value. The control unit 12 performs an analysis such as a morphological analysis of the estimated notation 20D based on the reading dictionary 13R, and outputs the reading of the estimated notation 20D and the parts of speech of the words that make up the estimated notation 20D.

The control unit 12 applies a reading registered in the accent dictionary 13A to the reading of the estimated notation 20D, thereby applying an accent to the reading of the estimated notation 20D. The control unit 12 applies an accent to the reading of the estimated notation 20D, including unknown words that are not registered in the accent dictionary 13A.

The control unit 12 is equipped with a machine learning model for accent estimation. The control unit 12 functions as an estimation unit. The machine learning model for accent estimation is a deep learning model equipped with a neural network. The machine learning model for accent estimation handles (A) the notation, (B) the reading of the notation, (C) the parts of speech of the words that make up the notation, and (D) the dictionary accent of the notation as a set of input values. The machine learning model for accent estimation is a deep learning model that has been machine-trained to use the above (A) to (D) as input values and output an appropriate accent for the notation included in the input values.

The storage unit 13 includes a computer-readable storage medium such as a hard disk. The storage unit 13 readably stores an accent dictionary 13A and a reading dictionary 13R. The storage unit 13 stores an accent estimation program that is read by the control unit 12. The storage unit 13 readably stores a machine learning model. Note that the accent dictionary 13A, the reading dictionary 13R, and the machine learning model may be stored by the control unit 12.

The user terminal 20 is a computer such as a desktop, laptop, tablet, or smartphone, and includes an input/output unit 21. The input/output unit 21 includes an interface for inputting an estimation request 20Q and an estimation notation 20D. The input/output unit 21 includes an interface for outputting the reading and accent of the estimation notation 20D. The input/output unit 21 includes input devices such as a touch panel, keyboard, mouse, and microphone, and output devices such as a display and speaker.

[module]
2, the accent estimation device 10 functions as a reading estimation module 14 and an accent estimation module 15. The accent estimation device 10 functions as the reading estimation module 14 and the accent estimation module 15 by the control unit 12 executing an accent estimation program.

The reading estimation module 14 analyzes the estimated notation 20D acquired through the communication unit 11. The reading estimation module 14 analyzes the estimated notation 20D based on the reading dictionary 13R, and uses the analysis results to output the reading of the estimated notation 20D and the parts of speech of the words that make up the estimated notation 20D.

The reading estimation module 14 performs, for example, a morpheme analysis of the estimated notation 20D. The morpheme analysis includes word division for dividing the estimated notation 20D into words, part-of-speech processing for assigning a part of speech to each word, and conjugation processing for assigning a conjugation to each word. The morpheme analysis also includes part-of-speech processing and conjugation processing for substrings and compound words within a word, regarding adjacent character strings as words that are morphemes. In addition to the morpheme analysis, the reading estimation module 14 performs stemming processing for returning the divided words to their stem form by referring to the parts of speech of the divided words, or lemmatization processing for returning the divided words to their entry words in a dictionary or the like by referring to the conjugations of the divided words.

The reading estimation module 14 includes a reading estimation model, which is a machine learning model. The reading estimation model is a model that learns the association between the learning input value and the learning reading so as to output a reading that matches the result of the analysis of the estimated notation 20D. The result of the analysis of the estimated notation 20D is, for example, the estimated notation 20D, the words contained in the estimated notation 20D, the parts of speech of the words, the conjugations of the words, and the types of characters.

The reading estimation model includes a neural network. The reading estimation module 14 expresses the results of analysis of the estimated notation 20D, the words contained in the estimated notation 20D, the parts of speech of the words, the conjugations of the words, the types of characters, and the like, in a vector space. The reading estimation module 14 uses the results of the above-mentioned analysis as input values and the readings of the estimated notation 20D as output values. The output of readings using the reading dictionary 13R and the reading estimation model makes it easier to assign appropriate readings to unknown words contained in the estimated notation 20D.

The number of words registered in the reading dictionary 13R is very small compared to the constituent words of the estimated notation 20D, which may contain many new words and compound words. The number of words registered in the accent dictionary 13A is even smaller than the number of words registered in the reading dictionary 13R. Morphological analysis, which regards partial strings and compound words as words, makes it possible to improve the accuracy of readings when using words registered in the reading dictionary 13R. Furthermore, when creating accents using words registered in the accent dictionary 13A, it makes it possible to improve the accuracy of accents by the same amount that the accuracy of readings is improved.

The reading estimation module 14 may perform, for example, syntactic analysis of the estimated notation 20D as a preprocessing step. Syntactic analysis includes dependency analysis, which identifies relationships between words and assigns roles to the words. Syntactic analysis includes phrase structure analysis, which identifies the phrase structure of a phrase, which is a string of words, and assigns roles to the phrases. When syntactic analysis is performed as a preprocessing step, for example, it becomes possible to assign final particles with high accuracy to the morphemes that make up the estimated notation 20D. Furthermore, when final particles that have a significant effect on accents, such as interrogative, emphatic, and exclamatory, are assigned, it becomes possible to create more appropriate accents when creating accents using the readings output by the reading estimation module 14.

The accent estimation module 15 creates a dictionary accent for (D) the estimated notation 20D. When creating the dictionary accent, the accent estimation module 15 refers to (B) the reading of the estimated notation 20D output by the reading estimation module 14, (C) the parts of speech of the words that make up the estimated notation 20D, and the accent dictionary 13A. The accent estimation module 15 matches the reading of the estimated notation 20D to the readings registered in the accent dictionary 13A, thereby creating the dictionary accent. At this time, the accent estimation module 15 divides an unknown word included in the estimated notation 20D into multiple words, and matches the readings of the divided words to the readings registered in the accent dictionary 13A, thereby creating the accent of the unknown word.

For example, when the reading of the estimated notation 20D is "sekai isan," the accent estimation module 15 determines that the reading is an unknown word that is not registered in the accent dictionary 13A. Then, using words registered in the accent dictionary 13A, the accent estimation module 15 concatenates "100" associated with "sekai" and "110" associated with "isan." As a result, the accent estimation module 15 outputs the dictionary accent.

For example, when the reading of the estimated notation 20D is "ichijitsusenshuu," the accent estimation module 15 determines that the reading is an unknown word that is not registered in the accent dictionary 13A. Then, using words registered in the accent dictionary 13A, the accent estimation module 15 concatenates "1000" associated with "ichijitsu," "10" associated with "sen," and "00" associated with "shuu." In this way, the accent estimation module 15 outputs the dictionary accent.

The accent estimation module 15 includes an accent estimation model, which is a machine learning model. The accent estimation model is a deep learning model that learns the association between the learning input values and the learning teacher accents so as to output an accent that matches the meaning of the above four input values (A) to (D). The accent estimation model includes a neural network. The accent estimation module 15 represents the above four input values (A) to (D) in a vector space. The accent estimation module 15 inputs the above four input values (A) to (D) into the accent estimation model, thereby creating an appropriate accent for the estimated notation 20D.

[Machine learning model]
3, the accent estimation model provided in the accent estimation module 15 includes a neural network. The neural network includes an input layer 18A consisting of a plurality of input nodes NA, an intermediate layer 18B (hidden layer) consisting of a plurality of intermediate nodes NB, and an output layer 18C consisting of a plurality of output nodes NC.

A neural network is a model that applies a predetermined operation to an input value to calculate an output value. In FIG. 3, for the convenience of showing the entire accent estimation model, only the characters 20Qs that appear in the reading of the estimated notation 20D and the dictionary accents 20Ps that are the result of analysis based on the accent dictionary 13A are shown as encoded input values. Also, an example is shown in which the intermediate layer 18B is a single layer. The intermediate layer 18B may be two or more layers, and multiple other nodes may be interposed between the intermediate nodes NB that constitute the intermediate layer 18B and the output nodes NC that constitute the output layer 18C.

The edge E connecting the input node NA and the intermediate node NB, and the edge E connecting the intermediate node NB and the output node NC each have their own weights. The weight determined by each edge E is multiplied by the input value of the node that is the input end of that edge E, and the result of the multiplication is passed to the node that is the output end of that edge E.

The intermediate node NB adds up the values passed to it. The output node NC also adds up the values passed to it. The intermediate node NB applies the value added up at the intermediate node NB to an activation function determined by the intermediate node NB, and outputs an output value of the intermediate node NB. The output node NC applies the value added up at the output node NC to an activation function determined by the output node NC, and outputs an output value of the output node NC.

The activation function determined by each node NB, NC is, for example, a softmax function, a logistic function, a hyperbolic tangent function, or a Rectifier Linear Unit. The weight determined by each edge E is optimized by learning the neural network.

The estimation model provided by the accent estimation module 15 may include a convolutional neural network. The convolutional neural network handles, for example, (A) the estimated notation 20D, (B) the reading of the estimated notation 20D, (C) the parts of speech of the words that make up the estimated notation 20D, and (D) the dictionary accent assigned to the reading of the estimated notation 20D as two-dimensional inputs.

A convolutional neural network has three intermediate layers: a convolutional layer, a pooling layer, and a fully connected layer. The convolutional layer performs convolution processing on the input. The pooling layer performs pooling processing on the input. The fully connected layer connects each node that makes up the fully connected layer to all nodes that make up the layer immediately before the fully connected layer.

The convolution process applies a filter to the entire two-dimensional input. The filter condenses the input features in a two-dimensional region of a given size. The filter is updated by training the convolutional neural network. That is, the convolution process condenses the input features in the entire two-dimensional input and outputs them to the next layer.

The pooling process follows the convolution process. Pooling is applied to the two-dimensional input to extract important elements for each two-dimensional region of a given size. The important elements can be the maximum value in the two-dimensional region of a given size, or the average value in the two-dimensional region of a given size.

[Estimation method]
The accent estimation method performed by the accent estimation device 10 includes a reading estimation process, a fitting process, and an accent estimation process.

As shown in FIG. 4, in the reading estimation process, the reading estimation module 14 first acquires the estimated notation 20D from the user terminal 20 and acquires the reading dictionary 13R from the storage unit 13. Next, the reading estimation module 14 analyzes the acquired estimated notation 20D, inputs the analysis results into the reading estimation model, and outputs (B) the reading of the estimated notation 20D and (C) the part of speech of the word included in the reading (step S11).

In the matching process, the accent estimation module 15 first obtains the reading and part of speech of the estimated notation 20D from the reading estimation module 14, and obtains the accent dictionary 13A from the memory unit 13. Next, the accent estimation module 15 analyzes the obtained reading and part of speech of the estimated notation 20D, applies an accent to the reading of the estimated notation 20D, and outputs (D) the dictionary accent (step S12).

In the accent estimation process, the accent estimation module 15 first inputs the estimated notation 20D, the reading and part of speech of the estimated notation 20D that are the results of the reading estimation process, and the dictionary accent that is the result of the matching process, into the accent estimation model. As a result, the accent estimation module 15 outputs the accent 13P of the estimated notation 20D (step S13).

[Learning device]
The accent learning device 30 differs from the accent estimation device 10 in that it performs learning processing by modifying the machine learning model provided in the accent estimation device 10, but other configurations are common to the accent estimation device 10. Below, configurations that differ from the accent estimation device 10 will be described, and configurations that are common to the accent estimation device 10 will be assigned the same reference numerals and descriptions thereof will be omitted.

5, the accent learning device 30 executes an accent learning program to carry out an accent learning method. The accent learning device 30 accepts a learning request from the user terminal 20. The accent learning device 30 acquires learning input values 20R. The learning input values 20R are made up of the following four pieces of data (A1) to (D1).
(A1) Training notation (B1) Training notation reading (C1) Training notation part of speech (D1) Training notation reading teacher accent

The training notation is a string of characters written using letters, numbers, and symbols. The training notation includes notations that include unknown words and notations that do not include unknown words. The reading of the training notation is a string of characters that indicates the appropriate reading of the training notation. It is the appropriate part of speech of the words that make up the reading of the training notation. The teacher accent of the reading of the training notation is the appropriate accent that is applied to the appropriate reading of the training notation.

The control unit 12 refers to the accent dictionary 13A and assigns an accent to the reading of the learning notation based on (B1) the reading of the learning notation and (C1) the parts of speech of the words that make up the learning notation. The control unit 12 outputs (D2) the pre-learning dictionary accent through the accent assignment. At this time, the control unit 12 divides the unknown word included in the learning notation into multiple words and strings together the accents of the divided words to determine the accent of the unknown word.

The control unit 12 includes a machine learning model for reading learning. The machine learning model for reading learning is a deep learning model equipped with a neural network. The machine learning model for reading learning uses, as input values, (A1) a learning notation and the results of analysis such as morphological analysis using the learning notation. The machine learning model for reading learning uses, as ideal values for output values, (B1) the reading of the learning notation and (C1) the parts of speech of the words that make up the learning notation.

The control unit 12 uses a learning method such as backpropagation to make the deep learning model learn relationships such as weights between input values and output values. The control unit 12 outputs a reading estimation model, which is a learned model, as a result of the reading learning process.

The control unit 12 is equipped with a machine learning model for accent learning. The machine learning model for accent learning is a deep learning model equipped with a neural network. The machine learning model for accent learning uses (A1) the learning notation, (B1) the reading of the learning notation, (C1) the part of speech of the constituent words of the learning notation, and (D2) the pre-learning dictionary accent as input values. The machine learning model for accent learning uses (D1) the teacher accent as the ideal value for the output value.

The control unit 12 uses a learning method such as backpropagation to make the model learn the weighting relationship between the input value and the output value. The control unit 12 outputs the accent estimation model, which is the trained model, as a result of the accent learning process.

For example, the control unit 12 inputs an input value to the input layer 18A to find the output value of the intermediate layer 18B and the output value of the output layer 18C. Next, the difference between the output value of the output layer 18C and the ideal value is found as the error of the output layer 18C. In addition, the error of the output layer 18C and the output value in the intermediate layer 18B are used to find the error of the intermediate layer 18B. Next, the weight of the edge E connecting the output layer 18C and the intermediate layer 18B is adjusted so that the error of the output layer 18C is reduced. In addition, the weight of the edge E connecting the intermediate layer 18B and the input layer 18A is adjusted so that the error of the intermediate layer 18B is reduced. The control unit 12 then repeats the above process until the error of the output layer 18C becomes sufficiently small.

[Learning method]
The accent learning method performed by the accent learning device 30 includes a reading learning process, an application process, and an accent learning process.

As shown in FIG. 6, in the reading learning process, first, the control unit 12 acquires (A1) the learning notation from the user terminal 20, and acquires the reading dictionary 13R from the storage unit 13. Next, the control unit 12 analyzes the acquired learning notation based on the reading dictionary 13R. Then, the control unit 12 uses (A1) the learning notation and the result of the analysis of the learning notation as input values, and (B1) the reading of the learning notation and (C1) the parts of speech of the words that make up the learning notation as ideal values, and causes the machine learning model for reading learning to perform learning. As a result, the control unit 12 outputs a reading estimation model, which is a learned model.

In the matching process, first, the control unit 12 (A1) takes the learning notation and the result of the analysis of the learning notation as input values, and outputs the reading of the learning notation and the parts of speech of the words that make up the learning notation. Next, the control unit 12 retrieves the accent dictionary 13A from the storage unit 13, and uses the reading of the analyzed learning notation and the parts of speech of the analyzed words to match accents to the reading of the learning notation based on the accent dictionary 13A. This causes the control unit 12 to output (D2) the pre-learning dictionary accent (step S22). Note that, like the output of the dictionary accent, the control unit 12 divides the unknown word included in the learning notation into multiple words, and strings together the accents of the divided words to form the accent of the unknown word.

In the accent learning process, the control unit 12 takes as input values (A1) the learning notation, the analyzed reading of the learning notation, the analyzed part of speech of the word, and (D2) the pre-learning dictionary accent. Next, the control unit 12 causes the machine learning model for accent learning to perform learning using (D1) the teacher accent as the ideal value. As a result, the control unit 12 outputs the accent estimation model, which is a learned model (step S23).

As described above, according to the above embodiment, the following effects can be obtained.
(1) Whether the estimated notation 20D includes a new word or a compound word formed by an unknown combination rule, the accent registered in the accent dictionary 13A is applied to the reading of the unknown word. Then, even if the accent applied to the reading of the unknown word is not an appropriate accent, an estimation is performed by the accent estimation model to bring the accent closer to an appropriate accent. As a result, even if the estimated notation 20D includes an unknown word, it is possible to create an accent suitable for the reading of the estimated notation 20D, and therefore it is possible to improve the accuracy of accent estimation.

(2) The estimated notation 20D includes various information related to the vocabulary of the unknown word, the structure of the unknown word, the origin of the unknown word, the function of the unknown word, synonyms of the unknown word, related words of the unknown word, associative words of the unknown word, and similarity to words registered in the dictionary. In addition to the above information, the reading of the estimated notation 20D includes various information such as information related to the phonology of the unknown word and information related to the number of mora of the unknown word. The dictionary accent of the estimated notation 20D is information that can determine the relationship between an appropriate accent and an accent by matching using the accent dictionary 13A. Estimation using the accent estimation model using the above (A) to (D) as input values can reflect these various pieces of information in the creation of the accent, making it possible to improve the accuracy of accent estimation.

(3) In the accent estimation process, the reading and part of speech of the estimated notation 20R estimated by the reading estimation module 14 are included in the input value. This makes it possible to improve the accuracy of the reading and part of speech of the estimated notation 20R, and therefore further improve the accuracy of the accent estimation.

(4) The training notation includes various information related to the vocabulary of the unknown word, the structure of the unknown word, the origin of the unknown word, the function of the unknown word, synonyms of the unknown word, related words of the unknown word, associative words of the unknown word, and similarity to words registered in the dictionary. In addition to the above information, the reading of the training notation includes various information such as information related to the phonology of the unknown word and information related to the number of mora of the unknown word. The dictionary accent of the training notation 20D is information that can determine the association between an appropriate accent and an accent by applying it using the accent dictionary 13A. Learning using a machine learning model with the above (A1) to (C1) as input values and the above (D1) as an ideal value can reflect these various information in the association between the unknown word and its accent, making it possible to improve the accuracy of accent estimation.

(5) In the accent learning process, the input value includes the reading of the analyzed learning notation and the part of speech of the analyzed word. In other words, the reading learning process and the accent learning process are learned as a series of processes using a common learning notation, and the learning results of the reading learning process are reflected in the results of the accent learning process. Therefore, in a form in which the reading estimation process and the accent estimation process are performed as a series of processes, it is possible to further improve the accuracy of accent estimation.

The above embodiment can be modified as follows.
The accent estimation device 10 can also use the accent dictionary 13A as the dictionary used in the reading estimation process. As described above, if the dictionary used in the reading estimation process is the reading dictionary 13R, it is possible to improve the accuracy of the reading of the estimated notation 20D. On the other hand, if the dictionary used in the reading estimation process is the accent dictionary 13A, it is possible to reduce the number of dictionaries provided in the accent estimation device, thereby simplifying the device configuration.

The accent estimation device 10 can also obtain an appropriate reading of the estimated notation 20D from the user terminal 20, along with the estimated notation 20D. In this case, the accent estimation device 10 may be configured to omit the function of the reading estimation module 14. If the accent estimation device 10 is configured to use the appropriate reading input by the user to estimate the accent, it is possible to further improve the accuracy of the accent estimation.

-The accent learning device 30 can also use the accent dictionary 13A to perform reading learning processing for the learning spelling. In other words, the accent learning device 30 can also analyze the learning spelling using the accent dictionary 13A.

- The accent learning device 30 can also use the appropriate reading of the learning notation (B1) and the appropriate part of speech of the learning notation (C1) as input values for the accent learning process.

The accent learning device 30 can also use the appropriate learning spelling reading input by the user as an input value for the accent learning process.
When estimating an accent, the accent estimation device 10 may use the above (A), (B), and (D) as input values of a machine learning model and omit the above (C) from the input values of the machine learning model. When learning an accent, the accent learning device 30 may use the above (A1), (B1), and (D1) as input values of a machine learning model and omit the above (C1) from the input values of the machine learning model.

The accent estimation device 10 may further include a dictionary update unit that updates the accent dictionary 13A. The dictionary update unit registers in the accent dictionary 13A the estimated notation 20D, the reading of the estimated notation 20D, the parts of speech of the words that make up the estimated notation 20D, and the accents estimated by including these in the input, in association with each other.

- The accent estimation device 10 may be provided with a separate input unit for inputting the estimated notation 20D to the accent estimation device 10. In this case, the accent estimation device 10 may further include a synthesis unit that converts the reading of the estimated notation 20D into sound using the accent of the estimated notation 20D.

- The answer provided by the accent estimation device 10 to the user terminal 20 may be data for vocalizing the reading of the estimated notation 20D on the user terminal 20 using the accent of the estimated notation 20D.

10: Accent estimation device 13A: Accent dictionary 13R: Reading dictionary 20D: Estimation notation 20R: Learning input value 20Ps: Dictionary accent 30: Accent learning device

Claims (2)

an analysis unit that analyzes a putative notation based on an accent dictionary, and applies accents registered in the accent dictionary to the reading of the putative notation, including unknown words that are not registered in the accent dictionary, thereby outputting a dictionary accent of the putative notation;
a model trained by machine learning to treat a group of registered words as separate words having a single collective meaning, with a notation, a reading, and a dictionary accent as input values and an appropriate accent for the notation as an output value; and an estimation unit that uses the estimated notation, the reading of the estimated notation, and the dictionary accent of the estimated notation output by the analysis unit as input values of the model, and outputs the output value of the model as the accent of the estimated notation,
The accent estimation device, wherein the estimation unit uses an appropriate reading input by a user as the reading of the estimated spelling that is used as an input value of the model. a dictionary that associates spellings, readings, and accents is an accent dictionary, the putative spelling is analyzed based on the accent dictionary, and accents registered in the accent dictionary are applied to the reading of the putative spelling, including readings of unknown words that are not registered in the accent dictionary, thereby outputting a dictionary accent of the putative spelling; and
a model trained by machine learning that treats a group of registered words as separate words having a single collective meaning, receives a spelling, a reading, and a dictionary accent as input values, and outputs an appropriate accent for the spelling as an output value, the estimated spelling, the reading of the estimated spelling, and the dictionary accent of the estimated spelling as input values of the model, and outputs the output value of the model as the accent of the estimated spelling,
The accent estimation method uses an appropriate reading input by a user as the reading of the estimation spelling that is used as an input value of the model.