JP2009204795A - Fundamental frequency estimation device, fundamental frequency estimation method, fundamental frequency estimation program, and storage medium - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To provide a device, a method and a program capable of estimating a fundamental frequency more approximate to interactive voice, and its storage medium. <P>SOLUTION: The fundamental frequency estimation device is comprised of a morpheme analysis dictionary, a category name dictionary, a text analysis part, an information re-forming part, a fundamental frequency summary value estimation part, and a detail fundamental frequency estimation part. The text analysis part analyzes an inputted sentence and outputs information such as a category name different from a the part of speech in a main word of an accent phrase and a relative pitch of the accent phrase as linguistic information. The information re-forming part forms the linguistic information in a predetermined form. The fundamental frequency summary value estimation part obtains a fundamental frequency summary value every accent phrase using with a regression model. The detail fundamental frequency estimation part adapts a temporal change in the fundamental frequency obtained from the linguistic information to the fundamental frequency summary value and outputs the fundamental frequency. <P>COPYRIGHT: (C)2009,JPO&INPIT

Description

  The present invention relates to a fundamental frequency estimation device, a fundamental frequency estimation method, a fundamental frequency estimation program, and a storage medium that estimate a fundamental frequency for each accent phrase from text information to be synthesized when speech synthesis is performed from text information.

The conventional fundamental frequency estimation method for speech synthesis is, as in Non-Patent Document 1 and Non-Patent Document 2, an average value, maximum value, start point height, and end point height of the basic frequency of accent phrases in units of accent phrases. Are estimated from the accent type, length, position and part-of-speech information of the corresponding and preceding and following accent phrases, and a detailed basic frequency is estimated based on that.
Abe Yasunobu and Sato Yamato, “Two-step control method of fundamental frequency based on speech segmentation model”, Journal of Acoustical Society, Vol. 49, No. 10, pp.682-690, 1993. Yoshinori Sagisaka, “On the prediction of global F0 shape for Japanese Text -to-speech”, Proceedings of ICASSP, pp.325-328, 1990.

  By the methods described in Non-Patent Documents 1 and 2, it is possible to synthesize a read-out speech for a newspaper article or the like. However, conversational voices spoken to people such as storytelling of children's stories, product promotions, and conversations differ greatly from the basic frequency of newspaper reading handled in the prior art. For this reason, information used for estimation in the prior art is not sufficient for estimation of the fundamental frequency in dialogue speech. For example, the combination of adverbs and adjectives, the conventional method only gives part of speech, so it was uniform fundamental frequency control in all word chains belonging to adjectives, but in dialogue speech that speaks to people, adjectives The frequency of the fundamental frequency differs greatly depending on the category different from the grammatical function of the part of speech, such as positive or negative meaning and the strength of the meaning. Is impossible. The present invention was invented in view of the above-described problems, and an object of the present invention is to provide an apparatus, a method, a program, and a storage medium capable of estimating a fundamental frequency that is closer to a dialog voice.

  The fundamental frequency estimation apparatus according to the present invention includes a morphological analysis dictionary, a category name dictionary, a text analysis unit, an information shaping unit, a fundamental frequency summary value estimation unit, and a detailed fundamental frequency estimation unit. The morphological analysis dictionary records parts of speech, readings, and accent types for words. The category name dictionary records a category name different from the part of speech for the word. For example, as a category name, a category name of a subcategory of part of speech for a word whose fundamental frequency summary value is known to be greatly different by a prior analysis (for example, a category name constituted by the meaning and the strength of the word) is there. The text analysis unit analyzes the input sentence and determines the part of speech of the beginning and end of the accent phrase, the accent type of the accent phrase, the reading of the accent phrase, the length of the accent phrase, the part of speech in the main word of the accent phrase, Outputs information on the different category names, the tone combination type between accent phrases, and the relative height of accent phrases as language information. However, any one of the category name different from the part of speech in the main word of the accent phrase or the relative height information of the accent phrase may be included in the language information. If the language information does not include a category name different from the part of speech of the main word of the accent phrase, the category name dictionary may not be provided. The information shaping unit shapes the language information output from the text analysis unit into a predetermined format. The fundamental frequency summary value estimation unit obtains a fundamental frequency summary value for each accent phrase from the formatted language information using the regression model. The regression model is a model that estimates the summary value of the fundamental frequency by using the information extracted by the information shaping unit as an input. The detailed fundamental frequency estimation unit adapts the temporal change of the fundamental frequency obtained from the language information to the fundamental frequency summary value and outputs the fundamental frequency.

  Processing in the text analysis unit will be described in detail. The text analysis unit includes morphological analysis means, accent phrase determination means, accent phrase reading estimation means, accent phrase accent type estimation means, accent phrase intertone combination estimation means, relative height calculation means, and category name assignment means. The morpheme analysis means generates a word string by decomposing the inputted sentence for each word, refers to the morpheme analysis dictionary, and obtains the part of speech, the reading, and the accent type for the word alone. The accent phrase determining means collects the word strings into accent phrases composed of one or more words serving as accent units. The accent phrase reading estimation means estimates the reading for each accent phrase and estimates the length of the accent phrase. The accent phrase accent type estimation means estimates an accent type for each accent phrase. The accent phrase-to-accent phrase type estimation means estimates the tone-joint type between adjacent accent phrases. The relative height calculation means calculates the relative height of each accent phrase using the tone combination type information. The category name assigning means assigns, for each accent phrase, a category name obtained by searching the category name dictionary to the main word in the accent phrase using only the surface layer of the word or a combination of the word surface layer and the part of speech as a key. To do. However, when a category name different from the part of speech for the word is not used as language information, the category name assigning means and the category name dictionary are unnecessary. When the relative height information of the accent phrase is not used as the language information, the relative height calculating means is not necessary.

  Since the estimation is performed by introducing a category name different from the part of speech for the word, it is possible to estimate the summary value of the fundamental frequency for each accent phrase more accurately than in the past. Further, since relative height information is used instead of position information, the correlation between the reproduction of the fundamental frequency summary value and the actual fundamental frequency summary value fluctuation increases. As a result, it is possible to estimate the fundamental frequency closer to the conversational voice as compared with the prior art.

  Now, an embodiment of the present invention will be described.

  FIG. 1 is a diagram illustrating a configuration example of the fundamental frequency estimation apparatus according to the first embodiment. FIG. 2 is a diagram illustrating an example of a processing flow of the fundamental frequency estimation method according to the first embodiment. The fundamental frequency estimation device 10 includes a morphological analysis dictionary 110, a category name dictionary 120, a text analysis unit 100, an information shaping unit 200, a fundamental frequency summary value estimation unit 400, and a detailed fundamental frequency estimation unit 500. Note that the morphological analysis dictionary 110 and the category name dictionary 120 may be provided outside the fundamental frequency estimation device 10.

  The morphological analysis dictionary 110 stores the part of speech, reading, and accent type of each word. FIG. 3 is a diagram illustrating an example of items in the category name dictionary of the first embodiment. The category name dictionary 120 stores a surface layer of words as search keys and a category name different from the part of speech for the word (for example, a category name constituted by the meaning of the word). The word surface layer means the word face. A word surface layer and part of speech may be stored as a search key.

  The text analysis unit 100 divides the synthesis target sentence into word strings, refers to the morphological analysis dictionary 110, and gives part-of-speech information, reading, and accent-type information to all words. Accent phrases are collected from the word string, and the reading and accent type of each accent phrase are determined. The tone combination type between accent phrases is determined, and the relative height of each accent phrase is calculated. Further, by using the category name dictionary 120, a category name obtained by searching the category name dictionary is assigned to the main word in the accent phrase using the surface of the word or the surface of the word and the part of speech as a key (S100). ).

  FIG. 4 is a diagram illustrating a configuration example of the text analysis unit according to the first embodiment. FIG. 5 is a diagram illustrating an example of the processing flow of the text analysis step according to the first embodiment. The text analysis unit 100 includes a morphological analysis unit 101, an accent phrase determination unit 102, an accent phrase reading estimation unit 103, an accent phrase accent type estimation unit 104, an accent phrase tone combination estimation unit 105, a relative height calculation unit 106, and a category name. It is comprised from the provision means 107. FIG.

  The morpheme analysis unit 101 applies morpheme analysis to the composition target sentence, breaks it down into words constituting the sentence, and gives a part of speech, a reading, and an accent type for each word with reference to the morpheme analysis dictionary 110 (S101). ). In addition, the utilization type | mold and utilization form for every word may be provided simultaneously, and you may utilize in the subsequent part which comprises an apparatus. The accent phrase determination unit 102 collects accent phrases using the word string that is the output result of the morpheme analysis unit 101. Further, when the accent phrase is determined, the part of speech of the first word and the last word of the accent phrase are determined (S102). The accent phrase reading estimation means 103 determines the accent phrase reading and the length of the accent phrase from the word reading as the output result of the morpheme analysis means 101 and the accent phrase information as the output result of the accent phrase determination means 102 ( S103). In the determination, word accent type information may be used. The accent phrase accent type estimation unit 104 determines the accent type of the accent phrase from the accent type of the word that is the output result of the morpheme analysis unit 101 and the accent phrase information that is the output result of the accent phrase determination unit 102 (S104). Note that word reading information may be used in the determination. The accent phrase tone combination estimation means 105 determines the tone combination type with the preceding accent phrase (S105). Note that the tone combination type with the subsequent accent phrase may be determined independently, or may be determined simultaneously with the tone combination type with the preceding accent phrase. In this case, the fundamental frequency summary value estimation unit 400 may estimate the summary value using the tone combination type with the subsequent accent phrase alone or simultaneously with the tone combination type with the preceding accent phrase. The relative height calculation means 106 calculates the relative height of the accent phrase by using the tone combination type information with the preceding accent phrase (S106). The relative height information is that the weak coupling of the tonal coupling type is considered as rising, the strong coupling is regarded as falling, the height of the accent phrase at the beginning is 0, and the next coupling type is weak coupling, If 1 is added and if it is a strong coupling, -1 is added to express the relative height of the accent phrase. The height is calculated within the range of the accent phrase sequence sandwiched by any two of the sentence head, pose, and sentence end. It is also possible to calculate the relative height by repeating the range setting by the same method until the end of the sentence with one range starting from the beginning of the sentence until the next weak coupling starts. The category name assigning means 107 searches the category name dictionary 120 for each accent phrase using the surface layer of the main word of the accent phrase as a key, and assigns a category name constituted by the meaning of each word (S107). The category names include the adjectives “fun” and “bright”, the adjectives and “noble” nouns derived from the adjective verbs, the positive “P” as a category name that expresses positiveness, the adjectives “sad” “ “Night” as a category name that expresses negativeness, and “*” as a category name that does not correspond to either of the adjectives and nouns derived from adjective verbs such as “Tsurai” Give one or more category names. In the category name dictionary, a category name different from the part of speech for the word, a category name of a subcategory of the part of speech for the word, or a category name constituted by the meaning and strength of the word may be recorded as the category name. Good. Therefore, the category name dictionary may be provided with classes obtained by further subdividing the three types of classes “P”, “N”, and “*” of the embodiment, and associating words belonging to the respective classes. Is possible.

The information shaping unit 200 shapes the information given to the explanatory variable of the regression model 300 into a predetermined format (S200). Specifically, the information shaping unit 200 determines the beginning and end of each accent phrase from the accent phrase, one or more preceding accent phrases and one or more subsequent accent phrases from the result of the text analysis unit 100. The word part-of-speech information and accent phrase accent type information, accent phrase length information, and a category name different from the part of speech of the main word are extracted. In addition, the tone combination type from the preceding accent phrase to the accent phrase and the relative height information of the accent phrase calculated from the tone combination type are extracted. For example, when retrieving information of the accent phrase, one preceding accent phrase, and one succeeding accent phrase,
(Part of speech information for the first word of the preceding accent phrase,
Part of speech information at the end of the preceding accent phrase,
Accent type information of the preceding accent phrase,
Information on the length of the preceding accent phrase,
A category name different from the part of speech of the main word of the preceding accent phrase, the part of speech information of the first word of the accent phrase,
Part of speech information at the end of the accent phrase,
Accent type information of the corresponding accent phrase,
Information on the length of the accent phrase,
A category name different from the part of speech in the main word of the accent phrase,
Part-of-speech information for the first word in the following accent phrase,
Part-of-speech information for the last word of the following accent phrase,
Accent type information for subsequent accent phrases,
Information about the length of the following accent phrase,
A category name that is different from the part of speech in the main word of the following accent phrase,
Tone coupling type from the preceding accent phrase to the accent phrase,
Information on the relative height of the accent phrase calculated from the tone combination type)
Is output as a shaping result, and information is passed to the fundamental frequency summary value estimation unit 400. The accent type of the accent phrase is an integer value indicating the position of the accent nucleus, and the length of the accent phrase is also an integer value. Note that if the same unit is always used to measure the position and length of the accent nucleus, a mora may be used as a unit, and a phoneme or syllable may be used as a unit. FIG. 6 is a diagram illustrating an example of the formatted language information output from the information shaping unit 200 according to the first embodiment. Note that it is possible to use all the information of each accent phrase or a part thereof.

  The fundamental frequency summary value estimation unit 400 sets the formatted language information as an explanatory variable of the regression model 300, and uses the value of the dependent variable estimated by the regression model 300 as a summary value of the fundamental frequency for each accent phrase. It passes to the estimation part 500 (S400). The regression model 300 is an arbitrary model such as a regression tree or quantification type I that can be estimated using categorical information as explanatory variables and real values as dependent variables. The coefficients and structure of the regression model 300 are calculated using learning data in which the language information shaped by the information shaping unit 200 and the summary value of the fundamental frequency at each accent phrase are paired. To do. The summary value includes an average value, a median value, a dynamic range, a maximum value, and the like. The regression model 300 obtains one or more of the average value, median value, and dynamic range as summary values. The fundamental frequency summary value estimation unit 400 constructs a dedicated regression model 300 for each summary value and performs estimation individually. The processing of the information shaping unit 200 and the fundamental frequency summary value estimation unit 400 is performed for all accent phrases for each accent phrase (S210).

  The detailed fundamental frequency estimation unit 500 calculates a syllable position where the fundamental frequency rises and a syllable position where the fundamental frequency begins to descend from the accent type of the accent phrase, and obtains a line segment connecting linear segments between them. The basic frequency is changed with time (S500). This is the simplest example, and the temporal change of the fundamental frequency can be determined using other information and description methods (such as a line segment connecting nonlinear line segments as a temporal change of the fundamental frequency). It may be estimated. After estimating the temporal change in the fundamental frequency, the temporal change is moved up and down, expanded and contracted so that the summary value of the temporal change matches the summary value output from the fundamental frequency summary value estimation unit 400.

  FIG. 7 is an example in which the detailed fundamental frequency estimation unit according to the first embodiment performs detailed fundamental frequency estimation based on the average value. When the summary value is the average value or median of the fundamental frequency in the accent phrase, the average value or median value of the temporal change and the average value or median value estimated by the fundamental frequency summary estimation unit 400 are matched. Translate temporal changes up and down. Let the value of the temporal change after translation be a detailed fundamental frequency. FIG. 8 is an example in which the detailed fundamental frequency estimation unit of the first embodiment performs detailed fundamental frequency estimation based on the dynamic range. When the summary value is the dynamic range of the fundamental frequency in the accent phrase, the temporal change so that the minimum value and the maximum value of the temporal change coincide with the minimum value and the maximum value estimated by the fundamental frequency summary estimation unit 400. Scale up and down. The value of the temporal change after scaling is taken as the detailed fundamental frequency.

  As described above, the fundamental frequency estimation apparatus according to the first embodiment introduces a category name different from the part of speech for a word, and uses a relative height information, so that the fundamental frequency closer to the dialogue voice can be obtained. Estimation is possible.

[Modification]
In the first embodiment, the text analysis unit 100 includes a morphological analysis unit 101, an accent phrase determination unit 102, an accent phrase reading estimation unit 103, an accent phrase accent type estimation unit 104, an accent phrase tone combination estimation unit 105, and a relative height calculation. Although it comprises means 106 and category name assigning means 107, the relative height calculating means 106 may not be provided. Hereinafter, only the difference from the first embodiment will be described.

  Relative height information is not included in the language information output from the text analysis unit and the information shaped by the information shaping unit 200. The fundamental frequency summary value estimation unit 400 sets the formatted language information that does not include the relative height information as the explanatory variable of the regression model 300, and sets the value of the dependent variable estimated by the regression model 300 for each accent phrase. The summary value of the fundamental frequency is passed to the detailed fundamental frequency estimation unit 500 (S400).

  As described above, the fundamental frequency estimation apparatus according to the modified example can also estimate the fundamental frequency closer to the dialogue voice by introducing a category name different from the part of speech for the word.

  FIG. 9 is a diagram illustrating a configuration example of the text analysis unit when the category name assigning unit and the category name dictionary according to the second embodiment are not provided. FIG. 10 is a diagram illustrating an example of the flow of processing in the text analysis step when the category name assigning unit and the category name dictionary according to the second embodiment are not provided. Hereinafter, only the difference from the first embodiment will be described. A configuration example of the fundamental frequency estimation apparatus according to the second embodiment will be described with reference to FIG. 1, and a diagram illustrating an example of a process flow of the fundamental frequency estimation method according to the second embodiment with reference to FIG. 2.

  The fundamental frequency estimation device 10 includes a morphological analysis dictionary 110, a text analysis unit 100 ', an information shaping unit 200, a fundamental frequency summary value estimation unit 400, and a detailed fundamental frequency estimation unit 500. In the second embodiment, the category name dictionary 120 of FIG. 1 is not provided. The text analysis unit 100 ′ divides the synthesis target sentence into word strings, refers to the morphological analysis dictionary 110, and gives part of speech information, reading, and accent type information to all words. Accent phrases are collected from the word string, and the reading and accent type of each accent phrase are determined. The tone combination type between accent phrases is determined, and the relative height of each accent phrase is calculated (S100 '). The text analysis unit 100 ′ includes a morphological analysis unit 101, an accent phrase determination unit 102, an accent phrase reading estimation unit 103, an accent phrase accent type estimation unit 104, an accent phrase tone combination estimation unit 105, and a relative height calculation unit 106. Is done. The category name assigning means 107 is not provided. Information on category names other than the part of speech of the main word is not included in the linguistic information output from the text analysis unit or the information shaped by the information shaping unit 200. The basic frequency summary value estimation unit 400 sets formatted language information that does not include category name information other than the part of speech of the main word as an explanatory variable of the regression model 300, and accentuates the value of the dependent variable estimated by the regression model 300. As a summary value of the fundamental frequency for each phrase, it is passed to the detailed fundamental frequency estimation unit 500 (S400).

  As described above, the fundamental frequency estimation apparatus according to the second embodiment can also estimate the fundamental frequency closer to the dialog voice by using the information on the relative height.

It is a figure which shows the structural example of the fundamental frequency estimation apparatus of Example 1. FIG. It is a figure which shows the example of the flow of a process of the fundamental frequency estimation method of Example 1. FIG. It is a figure which shows the example of an item of the category name dictionary of Example 1. FIG. FIG. 3 is a diagram illustrating a configuration example of a text analysis unit according to the first embodiment. It is a figure which shows the example of the flow of a process of the text analysis step of Example 1. FIG. It is a figure which shows the example of the formatted language information output from the information shaping part of Example 1. FIG. At the beginning of a sentence (SS in the figure) or the end of a sentence or pause (PP in the figure) where the accent type or accent phrase length cannot be defined, “-” is given as a non-applicable symbol. In the detailed fundamental frequency estimation unit of the first embodiment, detailed fundamental frequency estimation is performed based on an average value. This is an example in which a detailed fundamental frequency estimation unit according to the first embodiment performs detailed fundamental frequency estimation based on a dynamic range. It is a figure which shows the structural example of the text analysis part in the case of not providing the category name provision means and category name dictionary of Example 2. It is a figure which shows the example of the flow of a process of the text analysis step in the case of not providing the category name provision means and category name dictionary of Example 2.

Explanation of symbols

DESCRIPTION OF SYMBOLS 10 Fundamental frequency estimation apparatus 100 Text analysis part S100 Text analysis step 110 Morphological analysis dictionary 120 Category name dictionary 200 Information shaping part S200 Information shaping step 300 Regression model 400 Fundamental frequency summary value estimation part S400 Fundamental frequency summary estimation step 500 Detailed fundamental frequency estimation Part S500 Detailed fundamental frequency estimation step 101 Morphological analysis means S101 Morphological analysis sub-step 102 Accent phrase determination means S102 Accent phrase determination sub-step 103 Accent phrase reading estimation means S103 Accent phrase reading estimation sub-step 104 Accent phrase accent type estimation means S104 Accent phrase Accent type estimation sub-step 105 Accent inter-phrase tone combination type estimation means S105 Accent type inter-phrase tone combination type estimation sub-step 106 Relative height calculation means S106 Relative height calculation substep 107 Category name assignment means S107 Category name assignment substep

Claims (12)

A morphological analysis dictionary that records parts of speech, readings, and accent types for words;
A category name dictionary that records a category name different from the part of speech for the word,
Parses the input sentence, and the part of speech of the beginning and end of the accent phrase, the accent type of the accent phrase, the reading of the accent phrase, the length of the accent phrase, the category name different from the part of speech in the main word of the accent phrase, A text analysis unit that outputs tone-linked information between accent phrases as linguistic information;
An information shaping unit for shaping the language information into a predetermined format;
A fundamental frequency summary value estimator that obtains a fundamental frequency summary value for each accent phrase from the formatted language information using a regression model;
A detailed fundamental frequency estimating unit that adapts a temporal change of the fundamental frequency obtained from the language information to the fundamental frequency summary value and outputs the fundamental frequency; and
The text analysis unit
Morphological analysis means for decomposing an inputted sentence for each word to generate a word string, referring to the morphological analysis dictionary, each part of speech, reading, and obtaining an accent type of the word alone;
An accent phrase determining means for collecting the word string into an accent phrase composed of one or more words serving as an accent unit;
An accent phrase reading estimation means for estimating the reading of each accent phrase and estimating the length of the accent phrase;
An accent phrase accent type estimating means for estimating an accent type for each accent phrase;
An interaccent phrase concatenation type estimation means for estimating a tone combination type between adjacent accent phrases;
For each accent phrase, category name giving means for giving a category name obtained by searching the category name dictionary to the main word in the accent phrase using only the surface layer of the word or a combination of the word surface layer and the part of speech as a key When,
A fundamental frequency estimation apparatus comprising: The fundamental frequency estimation device according to claim 1,
The language information includes information on the relative height of accent phrases,
The text analysis unit
Relative height calculation means for calculating the relative height of each accent phrase using the tone combination type information,
A fundamental frequency estimation apparatus comprising: A morphological analysis dictionary that records parts of speech, readings, and accent types for words;
Parses the input sentence, and the part of speech of the beginning and end of the accent phrase, accent phrase accent type, accent phrase reading, accent phrase length, tone combination between accent phrases, and accent phrase relative A text analysis unit that outputs height information as language information;
An information shaping unit for shaping the language information into a predetermined format;
A fundamental frequency summary value estimator that obtains a fundamental frequency summary value for each accent phrase from the formatted language information using a regression model;
A detailed fundamental frequency estimating unit that adapts a temporal change of the fundamental frequency obtained from the language information to the fundamental frequency summary value and outputs the fundamental frequency; and
The text analysis unit
Morphological analysis means for decomposing an inputted sentence for each word to generate a word string, referring to the morphological analysis dictionary, each part of speech, reading, and obtaining an accent type of the word alone;
An accent phrase determining means for collecting the word string into an accent phrase composed of one or more words serving as an accent unit;
An accent phrase reading estimation means for estimating the reading of each accent phrase and estimating the length of the accent phrase;
An accent phrase accent type estimating means for estimating an accent type for each accent phrase;
An interaccent phrase concatenation type estimation means for estimating a tone combination type between adjacent accent phrases;
A relative height calculating means for calculating a relative height of each accent phrase using the tone-binding type information;
A fundamental frequency estimation apparatus comprising: The fundamental frequency estimation device according to claim 1 or 2,
The category name dictionary is
A fundamental frequency estimation device that records category names of subcategories of parts of speech for words. The fundamental frequency estimation device according to claim 1 or 2,
The category name dictionary is
A fundamental frequency estimation device, wherein category names composed of meanings and intensities of words are recorded. A morphological analysis dictionary that records parts of speech, readings, and accent types for words;
A category name dictionary that records a category name different from the part of speech for the word,
A fundamental frequency estimation method for estimating a fundamental frequency using
The text analysis unit analyzes the input sentence, and the part of speech of the beginning and end of the accent phrase, the accent type of the accent phrase, the reading of the accent phrase, the length of the accent phrase, the part of speech in the main word of the accent phrase Is a text analysis step for outputting information on the combination of tones between different category names and accent phrases as linguistic information,
An information shaping step in which the information shaping unit shapes the language information into a predetermined format;
A fundamental frequency summary value estimation unit for obtaining a fundamental frequency summary value for each accent phrase from the formatted language information using a regression model;
A detailed fundamental frequency estimating unit adapted to adapt a temporal change of the fundamental frequency obtained from the language information to the fundamental frequency summary value and output a fundamental frequency; and
The text analysis step includes:
Decomposing the input sentence for each word to generate a word string, referring to the morpheme analysis dictionary, each part of speech, reading, morpheme analysis substep to obtain the accent type of the word alone,
An accent phrase determination sub-step for collecting the word string into an accent phrase composed of one or more words as an accent unit;
An accent phrase reading estimation sub-step that estimates the reading of each accent phrase and estimates the length of the accent phrase;
An accent phrase accent type estimation substep for estimating an accent type for each accent phrase;
An interaccent phrase combination estimation sub-step for estimating a tone combination between adjacent accent phrases;
For each accent phrase, a category name giving sub that gives a category name obtained by searching the category name dictionary to the main word in the accent phrase using only the surface layer of the word or a combination of the word surface layer and the part of speech as a key Steps,
A fundamental frequency estimation method comprising: The fundamental frequency estimation method according to claim 6, wherein
The language information includes information on the relative height of accent phrases,
The text analysis step includes:
A relative height calculation sub-step for calculating a relative height of each accent phrase using the tone combination type information,
A fundamental frequency estimation method characterized by comprising: A morphological analysis dictionary that records parts of speech, readings, and accent types for words,
A fundamental frequency estimation method for estimating a fundamental frequency using
The text analysis unit analyzes the input sentence, and the part of speech of the beginning and end of the accent phrase, the accent type of the accent phrase, the reading of the accent phrase, the length of the accent phrase, the tone combination type between the accent phrases and the accent A text parsing step that outputs the relative height information of the phrase as linguistic information;
An information shaping step in which the information shaping unit shapes the language information into a predetermined format;
A fundamental frequency summary value estimation unit for obtaining a fundamental frequency summary value for each accent phrase from the formatted language information using a regression model;
A detailed fundamental frequency estimation step for adapting a temporal change of the fundamental frequency obtained from the language information to the fundamental frequency summary value and outputting a fundamental frequency; and
The text analysis step includes:
Decomposing the input sentence for each word to generate a word string, referring to the morpheme analysis dictionary, each part of speech, reading, morpheme analysis substep to obtain the accent type of the word alone,
An accent phrase determination sub-step for collecting the word string into an accent phrase composed of one or more words as an accent unit;
An accent phrase reading estimation sub-step that estimates the reading of each accent phrase and estimates the length of the accent phrase;
An accent phrase accent type estimation substep for estimating an accent type for each accent phrase;
An interaccent phrase combination estimation sub-step for estimating a tone combination between adjacent accent phrases;
A relative height calculation sub-step for calculating a relative height of each accent phrase using the tone combination type information;
A fundamental frequency estimation method comprising: A fundamental frequency estimation method according to claim 6 or 7,
The category name dictionary is
A fundamental frequency estimation method characterized by recording a category name of a subcategory of parts of speech for a word. A fundamental frequency estimation method according to claim 6 or 7,
The category name dictionary is
A fundamental frequency estimation method characterized in that a category name composed of the meaning and strength of a word is recorded. 6. A fundamental frequency estimation program for causing a computer to operate as the fundamental frequency estimation apparatus according to claim 1. A computer-readable storage medium storing the fundamental frequency estimation program according to claim 11.

Images