JPH11344998A - Method, device for setting reading and metrical information and storage medium in which reading and metrical information setting program is stored - Google Patents

Abstract

PROBLEM TO BE SOLVED: To perform voice synthesis which is natural and easily listened to by improving accuracy of reading and metrical information. SOLUTION: A word dictionary to authorize a word by the length in morpheme analysis S1 and to generate no degradation in impartation S2 of reading and meter is held and utilized based on the Miyazaki system. Rules of impartation of an accent like connection of accent phrases between segmentations, etc., are added in impartation S3 of the accent. A multi-stage setting method to stepwise impart a pause is introduced in impartation S4 of the pause.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a method and apparatus for setting prosody information and a storage medium storing a program for setting prosody information, and more particularly to providing reading prosody information to a Japanese text with high accuracy. The present invention relates to a reading prosody information setting method and apparatus for outputting a natural and easy-to-hear synthesized speech, and a storage medium storing a reading prosody information setting program.

[0002]

2. Description of the Related Art Japanese text-to-speech synthesis is used for reading, accenting (setting accent phrases that are basic units in prosody and assigning accent types), and pauses for Japanese text mixed with kanji and kana. Set the prosody information,
Based on these, a speech waveform is generated and a synthesized speech is output.
In order to output natural and easy-to-hear synthesized speech, it is necessary to correctly set the reading prosody information.

[0003] Since the reading is obtained by reading a word recognized by morphological analysis, the accuracy of the morphological analysis is directly linked to the accuracy of the reading. However, it is difficult to register all of the readings of the numerical expression (11 books → juyippon) and rendaku (child + room → kodomobeya) as words, so readings are given according to rules. For information on reading numerical expressions, see "Miyazaki:" Numerical Reading Rules for Japanese Text-to-Speech Conversion, "IPSJ Transactions, Vol. 25, No.6. For rendaku, see" Sato: " The rendaku rules, Miyaji et al., Lecture Japanese and Japanese Language Education 2, Japanese Voice / Phonology, Meiji Shoin, etc.] have almost established the method.

[0004] Regarding the setting of the accent type of an accent phrase, see "Sakazaka, Sato:" Accent Rules for Japanese Word Chains ", IEEJ Transactions, Vol.J66-D, No.7 attached word accent binding rules, compound The method is established by the word (independent word) accent combination rule and inter-phrase accent combination rule. Many methods have been proposed for setting accent phrase boundaries and poses.

[0005] As a heuristics-based method, Hakota and Sato: "Tone rules in sentence speech synthesis" using the structure of dependency, IEICE Transactions, Vol.
3-D, No. 9, "Kai, Hirose, Fujisaki:" Prosodic Rules for Synthesis of Japanese Sentence Speech ", Using Syntactic Information such as Right Branching Boundaries, Journal of the Acoustical Society of Japan, No. 50, Vol. ].
In addition, the method of setting poses by a statistical method is a method using dependency information, "Miki, Sakazaka:" Study of Pose Insertion Rule Using Local Phrase Structure, "IEICE Transactions, Vol. J79-D-II, No. 9] has been proposed.

Also, "Suzuki, Saito:" N-Phrase Structure Analysis for Japanese Text-to-Speech Synthesis and Prosody Control Based on It "," Transactions of the Institute of Electronics, Information and Communication Engineers, Vol.J78 D II, No.2 "
Estimating local dependency structure using part-of-speech information of N clauses, and "Fujio, Sagisaka, Higuchi:" Prediction of prosodic phrase boundaries and pose positions using stochastic context-free grammar ", IEICE Transactions on Information and Communication , Vol.J80D II, No. 1], learns dependency using stochastic context-free grammar with a part-of-speech sequence as input, and sets accent phrase boundaries and poses.

Also, based on a morphological analysis based on a multi-stage analysis method, "Miyazaki and Oyama:" Language processing method for Japanese sentence speech output "Transactions of Information Processing Society of Japan, Vol. 27, NO. After summarizing the method of Miyazaki's paper referred to in
Miyazaki method) has been proposed. Here, the morphological analysis by the multi-stage analysis method is based on the local brute force method that tests all combinations of words within a tentatively set range (provisional clause) by focusing on differences in character types, For effective compound word analysis, homomorphic word discrimination, etc., the analysis is partially deep. In particular, in the case of a compound word, the semantic connection between the words constituting the compound word is analyzed by dependency. In the semantic dependency analysis in this compound word, the number, proper noun, affix, verbal noun, non-verbal noun, etc.
Using four types of dependency rules (Miyazaki: "Automatic compound word segmentation using dependency analysis", IPSJ Transactions,
Vol.25, No.6).

According to this method, accents and pauses in a phrase can be set with high precision by using semantic dependency relations in a compound word.

[0009]

However, the above-described conventional method of setting the boundaries of accent phrases and poses assumes that linguistic information such as dependency is known. Is not mentioned. In addition, the above-mentioned conventional "Suzuki, Saito:" N-phrase structure analysis for Japanese text-to-speech synthesis and prosodic control based on it ", IEICE Transactions, Vol.J80D, II, No.1
Does not refer to the processing within a clause, and the method by Fujio, Sagisaka, and Higuchi raises a problem in that the accuracy of the setting within the clause due to an incorrect prediction of the structure within the clause is a problem.

[0010] Also, in "Miyazaki:" Automatic division of compound words using dependency analysis "" based on the conventional multi-stage analysis method described above, the recognition accuracy for case particle equivalents is low. It is mentioned as a problem, and the word recognition unit of morphological analysis is recognized in short units, such as using the stem and the ending of the verb as one word, so the pause is given especially. There is a problem of making mistakes. For example, when giving a pose, the part-of-speech information of the word before and after the accent phrase boundary is important, but immediately after the accent phrase boundary of “exercise on the ground” (/ = accent phrase boundary), “motion: sa-variant noun” "+" To "do: sa variable verb conjugation form", and the verb "exercise" is originally treated as a noun movement, and a pose is given by mistake.

As described above, the improvement of the morphological analysis accuracy is as follows.
The accuracy of morphological analysis is improved by directly performing word recognition in units of length, which is directly linked to improvement in accuracy of reading prosody information. However, simply identifying a long unit word cannot cope with a case where an accent phrase boundary occurs in the long unit word. In addition, conventionally, in many methods for setting prosodic information, at present, the accuracy of recognition is not sufficient and costly dependency analysis is used between phrases. However, errors caused by errors in dependency analysis, There is an error in the accent and pause assignment in the phrase.

Also, in the Miyazaki method, the accuracy of giving a pose is low, and no accent connection between phrases is performed.
In compound words, parts of speech with a high degree of independence (such as part-of-speech) are always treated as independent words, resulting in an accent error. The present invention has been made in view of the above points, and based on the Miyazaki method, in morphological analysis, certified in units of long, and
In reading prosody assignment, hold word dictionary information to prevent degrading, use it, add rules for adding accents such as interphrase accent phrase joining, etc. A reading prosody information setting method and apparatus capable of improving the accuracy of reading prosody information and performing natural and easy-to-hear speech synthesis by introducing a multistage setting method to be provided, and a storage medium storing a reading prosody information setting program The purpose is to provide.

[0013]

FIG. 1 is a diagram for explaining the principle of the present invention. The present invention (Claim 1) is to input a Japanese text mixed with kanji and kana, perform text analysis for adding reading prosody information, and perform speech generation for generating a synthesized speech based on the reading prosody information. In the reading prosody information setting method for text analysis in a speech synthesis method for outputting speech, a morphological analysis based on a multistage analysis method is performed using a word dictionary for setting word recognition and reading prosody information (step 1), and the morphological analysis is performed. In addition to the information for word recognition in step (1), using a word dictionary and using the result of the morphological analysis, reading (step 2), accenting (step 3), and pose are performed (step 4).

The present invention (claim 2) relates to a plurality of accent phrases corresponding to a case where an accent phrase boundary exists in a compound word registered with one word when performing morphological analysis based on a multistage analysis method. Accent phrase information that can store information, registered as one word, constituent word information for expanding into short words corresponding to the case where an accent phrase boundary exists inside and including an adjunct word, registered as one word Although there is no accent phrase boundary inside the compound word, a word dictionary having the number of words to indicate that it is a long unit word is used, and in the word dictionary, the stem endings of the separately registered words are set to 1
It is recognized as a unitized word as a case particle equivalent.

According to the present invention (claim 3), when giving an accent, accent connection between phrases having strong connection is performed. According to the present invention (claim 4), when performing a pose assignment, a pose assignment based on a symbol, a pose assignment based on a date and time expression / quantity expression, a pose assignment based on a combination of specific words, a pose assignment based on a binding force, and a correction of the pose assignment. Are given by a multi-step setting method consisting of

FIG. 2 is a diagram showing the principle of the present invention. According to the present invention (claim 5), when a Japanese text mixed with kanji and kana is input, text analysis means for providing reading prosody information, and speech synthesis for generating a synthesized speech based on the reading prosody information. Means for setting prosody information in text analysis in a speech synthesis system having means, a text analysis means 10 for performing morphological analysis based on a multistage analysis method, and a morphological analysis means for identifying words. A word dictionary 15 having information for setting reading prosody information in addition to the information, and a reading giving means 1 for giving a reading by using a result of the morphological analysis means 11.
2 and accent giving means 13 for giving an accent
And a pose giving means 14 for giving a pose.

According to the present invention (claim 6), the word dictionary 15
As information of the word dictionary, accent phrase information that can store information on a plurality of accent phrases corresponding to a case where an accent phrase boundary exists in a compound word registered with one word, registered with one word, Constituent word information for expanding into short unit words corresponding to the case where an accent phrase boundary exists within it, a compound word registered with one word without an accent phrase boundary inside, but a long unit word In the morphological analysis means 11, the morphological analysis means 11 converts the stem endings of the separately registered words into one word, and recognizes the case particle equivalent word as a collocation as one unit. Having means.

According to the present invention (claim 7), the accent imparting means 13 includes a means for performing an accent connection between strongly connected phrases. In the present invention (claim 8), the pose assigning means 14 performs the assignment of a pose based on a symbol, the assignment of a date and time expression / quantity expression, the assignment of a pose by a combination of specific words, the assignment of a pose based on a binding force, and the correction of the assignment of a pose. Means for giving a pose by a multi-stage setting method.

According to the present invention (claim 9), when a Japanese text mixed with kanji or kana is input, a text analysis process for providing reading prosody information and a speech generation for generating a synthesized speech based on the reading prosody information are provided. A storage medium storing a reading prosody information setting program in text analysis in a speech synthesis program having a speech synthesis process to be performed, wherein the text analysis process uses a word dictionary for setting word recognition and reading prosody information. In addition to the morphological analysis process that performs morphological analysis based on the analysis method and information for certifying words by morphological analysis, using a word dictionary,
It has a reading giving process for giving a reading using the result of the morphological analysis process, an accent giving process for giving an accent using a word dictionary, and a pause giving process for giving a pose.

According to the present invention (claim 10), in the morphological analysis process, accent phrase information capable of storing information on a plurality of accent phrases corresponding to a case where an accent phrase boundary exists in a compound word registered with one word. , Word information that is registered in one word, contains adjuncts, and expands into short words corresponding to the presence of accent phrase boundaries inside the compound word, and accents inside compound words registered in one word There is no phrase boundary, but using a word dictionary having the number of words to indicate that it is a long unit word, the word dictionary separates the stem endings of the separately registered words into one word, Has a process to certify as a set of collocations.

The present invention (claim 11) includes, in the accent giving process, a process of performing accent connection between strongly connected phrases. The present invention (claim 12)
In the pose assignment process, a symbol-based pose assignment, a date / time expression / quantity expression pose assignment, a combination of specific words, a pose assignment based on a binding force,
The method includes a process of giving a pose by a multi-step setting method including a correction of a pose.

According to the present invention, based on the above-mentioned Miyazaki system, word dictionary information for preventing grading in morphological analysis and for avoiding degradation in grammar analysis is stored and used. The accuracy of reading prosody information is improved by adding rules for adding accents, such as accenting phrases between phrases, etc. It is possible to perform easy speech synthesis. That is, by introducing a multi-stage analysis method in text analysis, it becomes possible to set an appropriate pause reflecting the structure of a compound word.

Thus, it is possible to add reading prosody information to a Japanese text with high accuracy and to output a natural and easy-to-hear synthesized speech.

[0024]

FIG. 3 shows the configuration of a reading prosody information setting device according to the present invention. The reading prosody information setting device shown in the figure performs a morphological analysis based on the reading prosody information when a Japanese text is input, and sets the reading prosody information in addition to information for word recognition in the morphological analysis. A text dictionary 10 and a speech generator 20 that hold a word dictionary having the information of, and perform reading, accenting, and posing using the result of the morphological analysis.

The text analysis unit 10 includes the morphological analysis unit 1
1, a reading section 12, an accent section 13, a pose section 14, and a word dictionary 15. When a Japanese text is input, the morphological analyzer 11 reads the word dictionary 1
5, the morphological analysis based on the multi-stage analysis method is performed, and the result is transferred to the reading giving unit 12.

The morphological analysis unit 11 converts the stem endings of the idioms registered separately in the word dictionary 15 into one word,
A case particle such as "about" is recognized as a group of collocations. The word dictionary 15 includes the morphological analyzer 11
In addition to the constituent word information having the head length, part of speech, and good length of each constituent word for word recognition used in the word recognition, a plurality of words corresponding to a case where an accent phrase boundary exists in a compound word registered with one word are provided. Accent phrase information that can store information on accent phrases, registered word, composed word information for expanding to short words corresponding to the case where an accent phrase boundary exists inside, and 1 Although there is no accent phrase boundary inside a compound word registered as a word, it has the number of words to indicate that it is a long unit word.

The reading giving section 12 gives readings using the above-mentioned Miyazaki method. The accent imparting unit 13 sets the accent phrase boundary and the accent type based on the Miyazaki method using the semantic dependency information in the compound word and the word dictionary 15. In addition, at the time of adding an accent, accent connection between phrases having strong connection is performed. The pause assigning unit 14 assigns pose information by introducing a multi-step setting method in which a pause is set step by step from within a compound word or the like to an accent phrase boundary and an accent type. The multi-stage setting method includes, in pose assignment, a pose assignment based on a symbol, a date assignment, a pose assignment in a numerical expression, a pose assignment based on a combination of specific words, a pose assignment based on a binding force,
It consists of the correction of the pose.

As the pose information, four-stage pose information of no pose, small pose, medium pose, and large pose is set. The voice generation unit 20 includes the text analysis unit 10
The reading prosody information added with the reading, accent, and pose information output from the pause giving unit 14 is obtained, and the synthesized speech is output by performing speech synthesis.

Hereinafter, the processing of the above configuration will be described in detail. FIG. 4 is a diagram for explaining the reading prosody information setting method of the present invention. In the morphological analysis (step 110) in the text analysis process (step 100), the morphological analysis is performed using the word dictionary 15 based on the Miyazaki method of multi-stage analysis. However, the following four items were extended to the morphological analysis unit 11 and the word dictionary 15 as long unit recognition corresponding to the reading prosody provision process (step 120).

The first is the unification of the language stem. In the word dictionary 15, words used regularly are separated into an invariant part (eg, similar) and a changing part (eg, ru, re, ro) and registered as one word each. In addition, there is a case where the sa-variable noun (exercise) forms a part of the sa-variable verb (moving), but only the sa-varin type noun (exercise) and the sa-varun inflected form (do) are registered on the word dictionary 15. However, it is not registered as a word (exercise) that summarizes the stem ending. However, since the part-of-speech information before and after the accent phrase boundary is important in the pause assigning unit 13, after the word recognition, the "verbal invariant part + variable part" and the "verbal noun + conjugative ending" are combined into one. Integrate into words. Thus, for example, "exercise" is treated as a one-word adjective and "substantially" is treated as a one-word adjective. Since the stem ending does not adversely affect the addition of accents and poses, it does not hold special information for giving prosody.

The second is to deal with the addition of prosody to compound words such as proper nouns that have accent phrase boundaries inside. The unit of the word registered in the word dictionary 15 is, in principle, a short unit word such as a base or an affix, and a long unit word such as a compound word is regarded as a short unit word combination. However, there are some exceptions that are registered in long units, such as compound words and some proper nouns, which cannot synthesize the meaning of long unit words from short unit words. Here, in a compound word of a long unit, an accent phrase boundary may be set at a specific position inside the compound word.

(Example) First prize meditation → med / first prize / treasure emblem (/: accent phrase boundary) However, if “first meditation first treasure emblem” is set as one word, an accent phrase boundary cannot be set inside the word. Therefore, accent phrase information, which is one of the word dictionary information, is stored for a maximum of three accent phrases. The accent phrase information includes the number of mora for each accent, the reading length (the length in reading notation), and accent type information. For example, the accent phrase information of the short unit word “Japan” is as follows: the first accent phrase: the number of mora = 3, the reading length = 3, the accent type = 2, the second and third accent phrases: none, and the long unit word “kun” The first accent phrase: (3 accent phrases) is the first accent phrase: the number of mora = 2, reading length = 2, accent type = 1 (Isao) The second accent phrase: the number of mora = 4, reading length = 4, accent type = 0 (first class) Third accent phrase: Mora number = 6, reading length = 7, accent type = 3 (Zuihosho). With this information, the accent phrase boundary can be correctly set as “Issue / First class / Ruihosho”.

The third is recognition of collocations. In morphological analysis using the Miyazaki method of multi-stage analysis, in order to improve the recognition accuracy of case particle equivalents (about him) and auxiliary verbs (speaking), which were relatively low in recognition accuracy, "particle + auxiliary words" (T + I) and case particle equivalents (N + T + T) are registered as collocations for recognition. However, since the accent phrase may be divided in these collocations, the collocation is regarded as one word at the time of (ka'reni / tsuiite /: accent phrase boundary, ': accent position), accent, and pause. There is a problem to deal with. Here, for a compound word that has an accent phrase boundary inside in units of length, it was handled by storing accent phrase information for up to three accent phrases.
Words that make up collocations almost always include adjuncts, and therefore cannot be handled with accent phrase information. This is because, as described in Sakazaka and Sato "Japanese Word Chain Accent Rules", adjuncts have different accent combination styles for each part of speech of the independent word in front, and use this combination style to cyclically This is because the accent type is determined. For this reason, constituent word information having the heading length, part of speech, and reading length of each constituent word is stored in the word dictionary 15, and a word that is recognized as one word by morphological analysis is subjected to a dictionary search based on the constituent word information. , Accents, and poses are expanded in short units. As described above, when the accent and the pose are given, the information expanded in short units is used, so that the accent phrase boundary and the accent type can be correctly set.

(Example) Collocation: Teshima (te + i + masi) Te: Headline length = 1, part of speech = connected particle, reading length = 1 i: Headline length = 1, part of speech = one-stage verb stem, reading length = 1 Masashi: Headline length = 1, Part of speech = Conjunctive verb form (assertion), Reading length = 1 The fourth is to deal with reading prosody in compound words that do not have accent phrase boundaries inside. A compound word without an accent phrase boundary inside is not a problem if the compound word appears alone, but if it is concatenated with other words to form a compound word, Wrong setting of accent phrase boundaries to be handled in the same way.

(Example) "Exchange rate" is one word (long unit registration)
In the case of morphological analysis result: exchange rate + flash report + service (3 words, actually 4 words) → × “exchange rate report service” and no accent phrase boundary are set. (○ exchange rate / flash report service) Therefore, the word dictionary information has the number of words representing the number of words constituting the registered word. For example, in "Exchange rate (2) + flash report (1) + service (1)" (the number in parentheses indicates the number of words), the number of authorized words in the morphological analysis is three, but the number of words is four depending on the number of words. It can be seen that the word is a compound word, and the accent phrase boundary can be set correctly as “Exchange rate / breaking service”.

In addition to the word dictionary information described above, the word dictionary 15 holds a lot of information for setting reading prosody information in addition to the information used for word recognition. FIG.
Shows an example of the word dictionary information for reading prosody information of the present invention. In the type shown in the drawing, the common information includes common information used in all processes for setting prosody information, and the reading provision information includes information for correcting the reading of the numerical expression and for forming rendaku. The information for accent phrase boundary setting is various flags for representing the internal structure of a compound word, and the flag for prosodic features. The information for accent type setting is various flags for accent features, and the information for pause setting is It has a particle pose attribute that classifies particles syntactically.

In the reading providing process in step 120, the Miyazaki method is used as it is. In the accent giving process of step 130, the accent phrase boundary and the accent type are set based on the Miyazaki system using the semantic dependency information in the compound word and the word dictionary information of FIG. here,
In the Miyazaki method, accent connection between phrases is not performed (all phrase boundaries are accent phrase boundaries). This means that even if accents between phrases are not
Strong accent phrases (no pauses,
This is because it is determined that suppressing the subsequent accent phrase = strong coupling) is practically sufficient. However, between the following strongly connected phrases, an accentual connection is performed because an experimental result was obtained in which unnaturalness was felt even as a strong connection.・ Instructive adverb + verb: Ex: so (instructive adverb) + think (verb) → accent phrase:
I think so ・ Adverb + noun (≠ compound): Example: This (adjective) + meeting (noun) → accent phrase: this meeting cf. This (adjective) + international (noun) + conference (noun)
→ Accent phrase: This / international conference In setting accent phrase boundaries in compound words, in the Miyazaki method, highly independent time phrases were always independent accent phrases, but Japanese Patent Application No. 7-115925 “Synthetic speech” Based on the method described in "Output Method and Apparatus", the accent phrase boundary of the time particle is set.

Example: New Year (time particle) + program (general noun) + (case particle) → accent phrase: New Year program today (time particle) + program (general noun) + (case particle) → accent phrase: In today / program In addition, there is a similar phenomenon in independent word accent combining as in secondary accent in adjunct accent combining. For example, if the preceding word has an accent such as "Kanagawa", such as Kanagawa etc .: Kana 'gawa / to' Osaka etc .: Osaka, then it has a secondary accent. Therefore, in the independent word, a word having a secondary accent attribute is set as in the case of the accessory word, and the secondary accent is set.

In the step 140 of giving a pause, a multi-step setting method for setting a pause step by step from within a compound word or the like is introduced. In the multi-stage setting method, first, time / quantity expressions, proper noun compounds, etc. can be handled independently, and the structure is accent phrase boundaries and punctuation marks in local structures obtained from semantic dependency information in compounds. Set the presence or absence of a pause at an accent phrase boundary where a pause always enters (does not enter), such as immediately after. Here, the presence or absence of a pause is set at all accent phrase boundaries, but for continuous accent phrase examples in which a pause is not set continuously for a certain mora length or longer (pause unset section), the binding force is used. To make a pose. The value of the binding power is 10 steps from 1 to 10 (the larger the value, the stronger the binding power), and it is impossible to depend on the accent phrase before and after the accent phrase boundary (eg, colloquial form <accent phrase boundary) > Noun), the value is extremely small, and conversely there is a high possibility of dependency (eg: verb adjunct form <accent phrase boundary> noun)
So we set it large.

By introducing this multi-step setting method,
Even in a compound word composed of multiple accent phrases, which was inaccurate in the conventional method, it is possible to set an appropriate pause reflecting the internal structure, and it is possible to combine without using dependency analysis between phrases Since the dependency structure is approximately estimated by the force, and the pose is given stepwise according to the strength of the coupling force and the number of mora in the section where the pose is not set, the pose can be set with sufficient accuracy for practical use. This multi-stage analysis method includes five steps shown in FIG. Here, there are 4 poses: no pose, small pose, medium pose, and large pose.
Suppose that you want to set stage pose information.

FIG. 6 is a flow chart of the pose assignment based on the multi-step setting method of the present invention. Step 141) In the pose assignment based on the symbol, if the end of the accent phrase is a symbol such as a punctuation mark, a pause is set immediately after the accent phrase boundary. In the case of a period, an exclamation point, or a question mark, set a large pose, a reading mark, a medium pose for open parentheses, and a small pose for other symbols other than the middle point.

Step 142) In the pause assignment of the date / time expression / quantity expression, a pause is assigned to the date / time expression and the quantity expression. The date and time expression classifies the expression into a year element, a month element, and a day element, and the presence / absence of a pause is given by the appearance pattern. Example: [Small pose] Year element [Small pose] Month element [No pose]
Day element: → Meeting is [small pose] 1998 [small pose]
From June [no pause] 1st to [small pose] month element [no pause] day element: → Meeting is [small pose] June 1st [no pause] 1st day. ) + (Sign) + numeral + (numerical number) + (affix) where () indicates omission possible. Is a defined expression. In accordance with the pattern of this numerical expression and the part of speech of the accent phrase before and after the numerical expression, the presence or absence of the pause in the numerical expression and before and after it is set.

Example: [Small pose] Prefix classifier [Small pose] Accent phrase of two or more numbers (+ classifier) → [Small pose] No. [Small pose] 100/20/3 times [Small pose] Prefix Classifier [No pause] One number (+
Classifier) Accent phrase → [Small pose] No. [No pause]
Three times Step 143) In the pause assignment using a combination of specific words, word information such as the part of speech of two or three consecutive accent phrases is referred to, and the presence or absence of a pause at the boundary where the pause should be (or should not) be given at all times is determined. Set.

Example: Position [Small Pose] Name → President [Small Pose] Yamada / Taro, etc. (Auxiliary particle) [Small Pose] Words other than diacritics → [Small Pose] Trial hearing, etc. starts. Step 144) In the pose assignment based on the bonding strength, the presence or absence of a pause is determined for the accent phrase boundary for which the presence or absence of the pause is not set in steps 141 to 143 by using the bonding power obtained by the parts of speech of the words before and after the accent phrase boundary. Set.

To set the binding strength, words immediately before the accent phrase boundaries are classified by parts of speech as shown in classifications P1 to P11 shown in FIG.
11 and the immediately following words are classified into 7 of the classifications N1 to N7 shown in FIG.
Classify by type. These are the syntactic properties of parts of speech (eg, P
3 = a word for modifying the union) and independence. For this reason, in the classification of the immediately preceding word, the particles are subdivided into five types according to the particle pose attribute of the word dictionary information (the particles are classified according to their syntactic properties and independence).

The bonding force is Pi × Nj (i = 1 to 11, j
= 1 to 7) (however, the maximum value of the binding force (10) is unconditionally given to the boundaries of words classified from collocations). The values of this combination matrix are the syntactic properties of Pi and Nj (possible or impossible), their independence, and, for possible combinations, their prosodic features. (Emphasis on the mora length of accent phrases and multiple accent phrases in compound words). Here, only the combination matrix of Pi × N7 will be specifically described with reference to FIG.

The words immediately before the accent phrase boundary are P2, P3,
In the case of P11, the phrase immediately before the boundary of the accent phrase is likely to have a continuous modification relationship with the immediately following phrase.
As shown in FIG. 10, the binding power is changed according to conditions such as the accent phrase immediately after. Item No. 1 in FIG. 9 shows a case in which the immediately succeeding word is a compound word composed of a plurality of accent phrases. In this case, even if the adnominal modification relationship is established, the structure of the accent phrase unit is a right-branched structure. Therefore, the bonding force is minimized. For example, "last year / enacted / (A) / man and woman / employment opportunity / equal law" (/: accent phrase boundary) indicates that the phrase unit structure is "enacted" → "man and woman employment opportunity law". However, the structure of the accent phrase unit is
"Equal method", and (A) is a right branching boundary.

Since this bonding force is also used in the correction of the pose addition in step 145, it is also obtained for the accent phrase boundaries for which “no pause” has already been set in steps 141 to 143 (however, the pause is not required). Do not set the presence or absence). However, for these accent phrase boundaries, 10 is added to the obtained binding power. This is because the presence / absence of the pause set in steps 141 to 143 is considered to be more reliable than the presence / absence of the pause set in step 144.

For an accent phrase boundary for which no pause is set, a small pause is given to the accent phrase boundary where the binding strength = 1, and no pause is given to the other accent phrase boundaries. Step 145) In the correction of the provision of a pose, when a pose is not set to a certain mora or more (a section in which this pose is not set is referred to as a section in which a pose is not set), a pose is set using the binding force. Here, the pause is given stepwise as shown in the flowchart of FIG. 11 according to the length of the section in which the pause is not set and the strength of the coupling force set in step 144.

Hereinafter, the flowchart of FIG. 11 will be described. Step 1451) It is determined whether or not the pause-unset section length is longer than 20 moras. If it is longer, the flow shifts to step 1452. If the number is less than 20 mora, the process is terminated (the correction of the pose is not performed). Step 1452) It is determined whether the pause-unset section length is 30 mora or less. If it is less than 30 mora, the process proceeds to step 1455;
Go to 53.

Step 1453) From the fifth to the fifth mora from the head of the pause unset section to the fifth mora from the end,
It is determined whether there is an accent phrase boundary having a binding power of 6 or less. If the condition is satisfied, the flow shifts to step 1456; otherwise, the flow shifts to step 1454. Step 1454) Second from beginning of pause unset section
A small pose is set at the accent phrase boundary where the bonding strength is minimum between the last mora and the second mora from the end, and the process ends.

Step 1455) Between the 5th mora from the head of the pause-unset section and the 5th mora from the end,
It is determined whether there is an accent phrase boundary having a binding strength of 3 or less. If the condition is satisfied, the flow shifts to step 1456; otherwise, the process ends. Step 1456) Fifth from beginning of pause unset section
A small pose is set at the boundary of the accent phrase with the smallest binding strength between the last five moora from the last moora, and the process ends.

The process shown in FIG. 11 is performed recursively even in the pause-unset section after the correction of the pose is performed, until the pause is no longer provided. According to the flowchart of FIG. 11, the pause-unset section length is 20 mora or more and 3
If the length is less than 0 mora and the length is not unnatural even if no pause is inserted, a pause is given only when there is an accent phrase boundary with a weak binding force (3 or less), and the pause unset section length Is more than 30 mora, and if the length exceeds an unnatural length without a pose, a pose is always set and a two-stage correction of giving a pose is performed.

[0054]

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS Embodiments of the present invention will be described below with reference to the drawings. FIG. 12 shows a specific example of reading prosody information setting according to an embodiment of the present invention. First, step 110
Is divided into words. here,
Although “radioactivity” is recognized as one word, the number of words in the word dictionary information = 2, which indicates that the word is composed of two words. In addition, "about" is also recognized as one word, but according to the constituent word information of the word dictionary information,
It is expanded into three words, "ni / about / te".

Next, by giving the reading in step 120,
A reading shown in parentheses is given (in FIG. 12, hiragana portions are omitted in parentheses, but readings are similarly given). Here, “30 days”, which is a numerical expression, is given a reading according to a reading provision rule of a numerical value shown in the Miyazaki system, and other readings are given by reading word dictionary information. Next, step 1
By giving 30 accents, the accent phrase boundary and the accent type of the accent phrase are set.

Here, since "equal" having a secondary accent characteristic follows "measures" having a three-type accent, it is divided into two accent phrases "measures / equals". Also, using the information that the number of words of “radioactive” = 2, the accent phrase boundary is correctly set to “radioactive / waste”. Next, a pose is given by the pose assignment in step 140.

First, in step 141, a middle pose is given to the position [A] by giving a pose based on the symbol. Next, in step 142, “last month / thirty days” is a month element + day element in the pose assignment of date / time expression / quantity expression, so [small pose] last month [no pause]
Set to thirty days.

In step 143, there is no pause assignment based on a combination of specific words in this example, so that the description is omitted. Then, in step 144, a small pose is given to the position [C] where the binding force is 1 and no pause is given to the other accent phrase boundaries for which the presence or absence of the pose is not set, by giving a pose based on the binding force.

Further, by the correction of the pose addition in step 145, the section in which the pose is not set, "The method of joint investigation in the Sea of Japan and the like and the future countermeasures were discussed." Step 14 of 11
The process proceeds from 51 → step 1452 → step 1453 → step 1456, and in step 1456, a small pose is given to [D] having the minimum binding force 3.

In the section where no pause has been set after the pause has been set in [D], “Methods of joint investigation in the Sea of Japan etc.” and “Consultations on future countermeasures, etc.” are both 2
Although there is no accent phrase boundary between 0 mora and less than 30 mora, but no binding force of 3 or less, no more poses are given. Similarly, in the pause unset section “related to radioactive waste dumping in the ocean” (23 mora), the flow shifts from step 1451 to step 1452 to step 1455 in FIG. 11 and there is no accent phrase boundary having a binding force of 3 or less. Do not give a pose.

4 and 6 used in the above embodiment.
The operation shown in FIG. 11 and the device configuration shown in FIG. 3 are constructed as a program and stored in a disk device connected to a computer used as a reading prosody information setting device, a portable storage medium such as a floppy disk or a CD-ROM. In addition, the present invention can be easily realized by installing the present invention when implementing the present invention.

The present invention is not limited to the above embodiment, but can be variously modified and applied within the scope of the claims.

[0063]

As described above, according to the present invention, it is possible to set reading prosody information with high accuracy using a morphological analysis result based on a multistage analysis method without performing dependency analysis between phrases. become.

[Brief description of the drawings]

FIG. 1 is a diagram for explaining the principle of the present invention.

FIG. 2 is a principle configuration diagram of the present invention.

FIG. 3 is a configuration diagram of a reading prosody information setting device of the present invention.

FIG. 4 is a diagram for explaining a reading prosody information setting method of the present invention.

FIG. 5 is an example of word dictionary information for reading prosody information of the present invention.

FIG. 6 is a flowchart of pose assignment based on the multi-stage setting method of the present invention.

FIG. 7 is a diagram showing a part of speech classification of the immediately preceding word according to the present invention.

FIG. 8 is a diagram showing a part of speech classification of a word immediately after the present invention.

FIG. 9 is a Pi × N7 combination matrix of the present invention.

FIG. 10 is a view showing a bonding force of P2, P3, and P11 × N7 of the present invention.

FIG. 11 is a flowchart of a process of correcting pose assignment according to the present invention.

FIG. 12 is a specific example of reading prosody information setting according to an embodiment of the present invention.

[Explanation of symbols]

 DESCRIPTION OF SYMBOLS 10 Text analysis means, text analysis part 11 Morphological analysis means, morphological analysis part 12 Reading provision means, reading provision part 13 Accent provision means, Accent provision part 14 Pose provision means, Pose provision part 15 Word dictionary 20 Voice generation part

 ──────────────────────────────────────────────────の Continued on the front page (72) Inventor Eiji Ohara 3-19-2 Nishishinjuku, Shinjuku-ku, Tokyo Nippon Telegraph and Telephone Corporation

Claims (12)

[Claims] 1. A method of inputting Japanese text mixed with kanji and kana, performing text analysis for adding reading prosody information, performing speech generation for generating synthesized speech based on the reading prosody information, and outputting synthesized speech. In a method for setting prosody information in text analysis in a speech synthesis method, a morphological analysis based on a multi-stage analysis method is performed using a word dictionary for setting word recognition and reading prosody information, and information for word recognition in the morphological analysis. A reading prosody information setting method, wherein reading, accenting, and pausing are performed using the result of the morphological analysis using the word dictionary. 2. A morphological analysis based on the multi-stage analysis method, wherein an accent phrase capable of storing information on a plurality of accent phrases corresponding to a case where an accent phrase boundary exists in a compound word registered with one word. Information, which is registered in one word, includes an accessory word, and constituent word information for expanding into a short unit word corresponding to the case where an accent phrase boundary exists therein,
Although there is no accent phrase boundary inside a compound word registered as one word, the word dictionary having the number of words to indicate that it is a long unit word is used, and the word dictionary is registered separately in the word dictionary. 2. The reading prosody information setting method according to claim 1, wherein the stem ending of the word is made into one word, and the case particle equivalent is recognized as a unit as a collocation. 3. The method according to claim 1, wherein, when the accent is given, accent connection between phrases having strong connection is performed.
How to set the reading prosody information. 4. The method according to claim 1, further comprising the steps of: providing a pose based on a symbol; providing a pose based on a date and time expression / quantity expression; providing a pose based on a combination of specific words; 2. The reading prosody information setting method according to claim 1, wherein a pause is assigned by a step setting method. 5. When a Japanese text mixed with kanji or kana is input, the system includes a text analysis unit for providing reading prosody information and a speech synthesis unit for generating a speech based on the reading prosody information. A reading prosody information setting device for text analysis in a speech synthesis system, wherein the text analysis unit includes a morphological analysis unit that performs a morphological analysis based on a multistage analysis method, and information for certifying a word by the morphological analysis unit.
A word dictionary having information for setting reading prosody information; a reading giving means for giving a reading by using a result of the morphological analysis means; an accent giving means for giving an accent; and a pose giving for giving a pose. Reading prosody information setting device. 6. The word dictionary according to claim 1, wherein said word dictionary stores information on a plurality of accent phrases corresponding to a case where an accent phrase boundary exists in a compound word registered with one word. , Word information that is registered in one word, contains adjuncts, and expands into short words corresponding to the presence of accent phrase boundaries inside the compound word, and accents inside compound words registered in one word Although there is no phrase boundary, it has a number of words to indicate that it is a long unit word, and the morphological analysis unit converts the stem endings of the vocabulary words registered separately in the word dictionary into one word, 6. The reading prosody information setting device according to claim 5, further comprising a unit that recognizes the equivalent word as a collocation unit. 7. The reading prosody information setting device according to claim 5, wherein said accent providing means includes means for performing accent connection between phrases having strong connection. 8. A multi-stage setting comprising: a pose assignment based on a symbol, a pose assignment based on a date / time expression and a quantity expression, a pose assignment based on a combination of specific words, a pose assignment based on a joint force, and a correction of the pose assignment. 6. The reading prosody information setting device according to claim 5, further comprising means for giving a pause by a method. 9. A text analysis process for providing reading prosody information when a Japanese text mixed with kanji and kana is input, and a speech synthesis process for generating a speech based on the reading prosody information. A storage medium storing a reading prosody information setting program in text analysis in a speech synthesis program, wherein the text analysis process uses a word dictionary for setting word recognition and reading prosody information based on a multi-stage analysis method. A morphological analysis process of performing the following, in addition to the information for word recognition in the morphological analysis, using the word dictionary, a reading giving process of giving a reading using the result of the morphological analysis process, Accenting process of giving an accent using a pose, and Pose giving process of giving a pose Storage medium storing read prosodic information setting program characterized in that it comprises a. 10. The morphological analysis process includes: accent phrase information that can store information on a plurality of accent phrases corresponding to a case where an accent phrase boundary exists in a compound word registered with one word. , Component word information for expanding into short words corresponding to the case where an accent phrase boundary exists, including adjuncts,
Although there is no accent phrase boundary inside a compound word registered as one word, the word dictionary having the number of words to indicate that it is a long unit word is used, and the word dictionary is registered separately in the word dictionary. 10. The storage medium storing the prosodic information setting program according to claim 9, comprising a process of converting the stem ending of the word into one word, and certifying the case particle equivalent word as a group of collocations. 11. The storage medium storing the prosody information setting program according to claim 9, wherein said accent providing process includes a process of performing accent connection between phrases having strong connection. 12. The multi-stage setting includes a pause assignment process based on a symbol, a pause assignment based on a date / time expression and a numerical expression, a pose assignment based on a combination of specific words, a pose assignment based on a joint force, and a pose assignment correction. 10. The method of claim 9, further comprising the step of:
A storage medium storing the described prosody information setting program.