JP5191470B2 - Reading text set creation method, mass Japanese text database repair method, apparatus, and program - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To create a reading text set making the generation frequency of retake to the minimum even when an amateur reads the set. <P>SOLUTION: A method for creating the reading text set includes: a reading text set generation process; and a reading error patter deletion process. A reading text set generation part extracts a plurality of reading text set candidates including individual phonemic sequences by regarding a plurality of the phonemic sequences wished to be recorded as input from an extensive Japanese text database. A reading error pattern deletion part determines to refer to a reading error pattern database whether a reading error pattern is included in the individual reading text set candidates by regarding the plurality of the reading text set candidates as input, retrieves the reading text set candidate including the same phonemic sequence from the extensive Japanese text database again when including the reading error pattern, and outputs the reading text set candidates as the reading text set when not including the reading error pattern. <P>COPYRIGHT: (C)2011,JPO&amp;INPIT

Description

  The present invention relates to a method for creating a text set for reading, a method for revising a large-scale Japanese text database, and a device therefor, which are used when constructing a database of speech segments as speech elements in speech synthesis technology for converting text into speech , Related to the program.

  In a conventional speech synthesis system, text to be read out when constructing a speech unit database has been created using, for example, the following criteria disclosed in Non-Patent Document 1.

(1) All speech elements necessary for covering all Japanese texts are included. (2) A database can be constructed with as little text as possible. (3) A phoneme sequence that is continuous as long as possible is included in the database. (4) As many pitch (voice pitch) variations as possible are included. (5) To include as many phoneme duration variations as possible. Note that (3) to (5) are for the purpose of avoiding quality degradation of the synthesized speech.

  Patent Document 1 discloses a text selection device 600 that selects a combination of texts including keywords as candidate texts for the purpose of improving the quality of synthesized speech in a specific field. FIG. 6 shows a functional configuration of the text selection device 600 and its operation will be briefly described.

  The text selection device 600 includes a keyword list storage unit 2, a keyword count unit 6, a keyword count storage unit 5, a candidate text database storage unit 4, a text selection unit 8, and an output unit 10. The keyword list storage unit 2 stores keywords that are important in speech synthesis in advance. The candidate text database storage unit 4 stores a large amount of Japanese text as digital data. The keyword counting unit 6 searches the candidate text database storage unit 4 for candidate texts including keywords stored in the keyword list storage unit 2 and also counts the number of keywords included in the candidate texts. The keyword count storage unit 5 stores the text number of the retrieved candidate text and the number of keywords.

  The text selection unit 8 selects a combination of candidate texts including all the keywords stored in the keyword list storage unit 2 from the retrieved candidate texts. The output unit 10 outputs the selected candidate text.

Kaigai, Mizuno, Mano, “Examination of Database Design Method for Speech Synthesis”, Proceedings of the Acoustical Society of Japan, 3-2-10, pp.335-336 (September 2004)

JP 2007-334264 A

  The speech segment database is created by creating a Japanese text set for reading that includes the phoneme sequence of a desired speech segment using the method disclosed in Patent Document 1, for example, and then reading it out by a human. The The conventional text for constructing a speech segment database is based on the assumption that professional announcers and voice actors will read it out. Further, conventionally, it has been assumed that if the voice intended by the database builder is recorded, or if there is a reading error after listening with the ear after recording, it is re-recorded as needed (hereinafter referred to as retake). Therefore, the cost of constructing one speech segment database is very expensive.

  On the other hand, in recent years, there has been an increasing need to perform speech synthesis with various voices. The cost of constructing the diversified individual speech segment database needs to be low. However, the text for constructing a speech segment database created by a conventional method has a problem that the request cannot be answered.

  In addition, when trying to construct a speech segment database with voices of various people, that is, amateurs who are not professional announcers or voice actors, using the text for constructing a conventional speech segment database is likely to cause reading errors. There is also a problem that the reading error directly affects the quality degradation of the synthesized speech.

  The present invention has been made in view of such problems, and is intended to create a diversified individual speech segment database at a low cost, and to prevent speech deterioration that does not deteriorate the quality of speech synthesized speech even if read by an amateur. It is an object of the present invention to provide a method for creating a text set for reading, a method for modifying a large-scale Japanese text database, a device for the same, and a program for creating a text for constructing a fragment database.

  The reading text set creation method of the present invention includes a reading text set generation process and a reading error pattern deletion process. In the reading text set generation process, the reading text set generation unit extracts a plurality of reading text set candidates including individual phoneme strings from a large amount of Japanese text database by inputting a plurality of phoneme strings to be recorded. In the reading error pattern deletion process, the reading error pattern deletion unit receives the plurality of reading text set candidates as input, and whether or not each reading text candidate includes a reading error pattern (a pattern that is easily read by an amateur). Is read with reference to the reading error pattern database.If the reading error pattern is included, another text candidate including the same phoneme string is re-searched from the above large Japanese text database. Output other text candidates as one text.

  According to the reading text set creation method of the present invention, it is determined whether or not a reading error pattern is included in each reading text candidate in the reading error pattern deletion process, and the text including a pattern that is easy to read is deleted. As a result, it is possible to create a text set for reading that does not include sentences that are easily misread. Therefore, even if an amateur reads the text set for reading, the occurrence frequency of retake can be minimized. Thereby, it is possible to create a text set for reading which is suitable for constructing various speech segment databases at low cost.

The figure which shows the function structural example of the text set preparation apparatus 100 for reading of this invention. The figure which shows the operation | movement flow of the text set preparation apparatus 100 for reading. The figure which shows the function structural example of the reading error pattern deletion part 40. FIG. The figure which shows the operation | movement flow of the reading error pattern deletion part. The figure which shows the procedure which produces the mass Japanese text database which does not contain the sentence which is easy to read. The figure which shows the function structure of the text selection apparatus 600 disclosed by patent document 1. FIG.

  Embodiments of the present invention will be described below with reference to the drawings. The same reference numerals are given to the same components in a plurality of drawings, and the description will not be repeated.

  FIG. 1 shows an example of the functional configuration of a text set creation device 100 for reading according to the present invention. The operation flow is shown in FIG. The reading text set creation device 100 includes a reading text set generation unit 20, a mass Japanese text database 30, a reading error pattern deletion unit 40, and a reading error pattern database 50.

  The reading text set creation apparatus 100 is realized by reading a predetermined program into a computer composed of, for example, a ROM, a RAM, a CPU, and the like, and executing the program by the CPU.

  The reading text set generation unit 20 extracts a plurality of reading text set candidates including individual phoneme strings from the mass Japanese text database 30 by using a plurality of phoneme strings to be recorded as input (step S20). The mass Japanese text database 30 is basically the same as the candidate text database 4 described in the prior art in that it records the text set candidates for reading, and it records, for example, several years of text in newspapers. .

  Table 1 shows examples of phoneme strings and text set candidates for reading.

ここで、（・）内の音素は、収録したい音素列の前後の音韻環境を表す。例えば、「（Ｉ）ＨＯ（ｎ）」の場合、収録したい音素列は「ＨＯ」であり、その前環境が「Ｉ」、かつ後環境が「ｎ」で発声されたものを表している。「＃」は文頭、文末を表す。「^＊」は母音の無声化を表す。/はポーズを表す。

Here, the phonemes in (.) Represent the phoneme environment before and after the phoneme string to be recorded. For example, in the case of “(I) HO (n)”, the phoneme string to be recorded is “HO”, which indicates that the previous environment is “I” and the subsequent environment is “n”. “#” Represents the beginning and end of a sentence. “ ^* ” Indicates devoicing of vowels. / Represents a pause.

  The reading text set generation unit 20 extracts a text set candidate for reading “red chili” from the mass Japanese text database 30 in order to record the phoneme string “#AK (A)”. The simplest example of the number of phoneme strings to be recorded is about 200 when recording only the syllables of “A, I, U, E,…”, and it is about considering the pre- and post-phoneme environment. About 6000. Further, as one method for improving the quality of the speech synthesized sound, there is a method of lengthening one phoneme string. In that case, the number of phoneme strings is infinite. As described above, the number of phoneme strings varies depending on the speech synthesis system.

  The text set generator for reading 20 generates a plurality of text set candidates for reading corresponding to the number of phoneme strings to be recorded required by the speech synthesis system.

  The reading error pattern deletion unit 40 receives a plurality of reading text set candidates as input, determines whether or not each reading text set candidate includes a reading error pattern with reference to the reading error pattern database 50, and reads If an error pattern is included, a text set candidate for reading that includes the same phoneme string is re-searched from the large-scale Japanese text database 30. (Step S40).

  Table 2 shows examples of reading error patterns recorded in the reading error pattern database 50.

読み誤りパターン削除部４０は、表２に例示した単語等を含む読み上げ用テキスト候補を、読上げ用テキストセットから排除する。この結果、読み誤りパターン削除部４０から出力される読み上げ用テキストセットには、読み誤りのパターンが含まれないので、素人が読んでもリテイクの発生の少ない読み上げ用テキストセットを作成することが出来る。

The reading error pattern deletion unit 40 excludes the reading text candidates including the words and the like illustrated in Table 2 from the reading text set. As a result, the reading text set output from the reading error pattern deletion unit 40 does not include a reading error pattern, so that it is possible to create a reading text set with less retake even when read by an amateur.

  Note that the reading text set generation unit 20 has the same function as the text selection unit 8 of the conventional text selection device 600. The large-volume Japanese text database 30 is the same as the candidate text database storage unit 4 of the prior art as described above. Therefore, the feature of the first embodiment is that the reading error pattern deletion unit 40 and the reading error pattern database 50 are provided.

  FIG. 3 shows a more specific functional configuration example of the reading error pattern deletion unit 40 which is a feature of the first embodiment, and will be described in more detail. FIG. 4 shows the operation flow. The reading error pattern deletion unit 40 includes a single sentence extraction unit 41, a reading error pattern matching unit 42, a database re-search unit 43, another sentence substitution determination unit 44, a meaningless word text set generation unit 45, and a sentence output. Means 46.

  The single sentence extraction unit 41 extracts one reading text candidate (hereinafter, reading is omitted) from the reading text set candidate generated by the reading text set generation unit 20 (step S41a).

  The extracted one text candidate is input to the reading error pattern matching means 42. The reading error pattern matching unit 42 determines whether or not the one text candidate includes a reading error pattern with reference to the reading error pattern database 50 (step S42). When a reading error pattern is not included in one text candidate (NO in step S42b), the one text candidate is output to the outside from the sentence output means 46 as one text (step S46). The output may be recorded in the memory with the one text or may be printed by a printer (not shown).

  When the reading error pattern is included in the one text candidate (YES in step S42b), the reading error pattern matching unit 42 includes the phoneme string to be recorded included in the one text candidate with respect to the database re-search unit 43. A control signal for re-searching other text candidates is output. Upon receiving the instruction (control signal), the database re-search means 43 re-searches other text candidates including the same phoneme string from the large-volume Japanese text database 30 (step S43). This re-search is started from the middle of the mass Japanese text database 30. That is, assuming that the target phoneme string is, for example, the 100th text in the mass Japanese text database 30, the re-search is performed from the 101st text.

  The re-searched other text candidates are input to the reading error pattern matching unit 42, and it is determined whether or not they include a reading error pattern. When the re-searched text candidate does not include a reading error pattern, the other text candidate is output from the sentence output means 46 as one text.

  If another re-searched text candidate includes a reading error pattern, the database re-search means 43 re-searches the other text candidate again. This re-searching operation is repeated until the reference of all the texts recorded in the mass Japanese text database 30 is completed.

  If a desired phoneme string cannot be extracted by referring to all the texts recorded in the mass Japanese text database 30, the meaningless word text generation means 45 is instructed to generate meaningless word text (NO in step S44). . The meaningless word text is, for example, text including a meaningless word such as “This is SESU” when the phoneme “SES” is not found.

  The meaningless word text generation means 45 generates one text which is a meaningless word text including the phoneme string to be recorded and outputs it to the outside from the sentence output means 46 (step S46). The operations in steps S41 to S46 described above are repeated until completion for all of the plurality of text set candidates for reading (YES in step S41b).

  Although the reading error pattern matching (step S42a) has been described by the method of detecting word matching, a reading error pattern may be detected by applying a certain rule. For example, vowel devoicing is known to occur at “I” and “U” sandwiched between unvoiced consonants or sentence ending symbols. If “I” and “U” sandwiched between these unvoiced consonants or sentence ending symbols are used as a rule, and if one text set candidate matches the rule, pattern matching (YES in step S42b) may be performed.

  By using such a rule, the amount of data recorded in the reading error pattern database 50 can be reduced, and the detection accuracy of the reading error pattern can be improved. Note that both the words and the rules may be recorded in the reading error pattern database 50.

  In the first embodiment, an example has been described in which, after a plurality of reading text set candidates are once created, text that is likely to cause reading errors is deleted from the plurality of reading text set candidates. Conversely, a text set for reading may be created using a modified large-volume Japanese text database 30 'that does not include text that is easily misread.

  FIG. 5 shows a procedure for creating a modified large-volume Japanese text database 30 '. The single sentence extraction unit 51 extracts one text including a desired phoneme sequence from the large-volume Japanese text database 30 (step S51). The reading error pattern matching unit 53 performs matching with the reading error pattern recorded in the reading error pattern database 50 to determine whether or not the reading error pattern is included in one text (step S53). If it does not match the reading error pattern (NO in step S54), the text registration unit 55 records the one text in the modified large-volume Japanese text database 30 '(step S55). Note that the illustration of the single sentence extraction unit 51, the reading error pattern matching unit 53, and the text registration unit 55 is omitted.

  The operations in steps S51 to S55 are repeated until the last sentence in the mass Japanese text database 30 is extracted and there are no more sentences to be extracted (YES in step S52). In FIG. 5, step S53 is referred to as a reading error pattern matching process, but this process is substantially the same as the process performed in the reading error pattern deletion 40 of the text set creation apparatus 100 for reading.

  Through the above procedure, a modified large-volume Japanese text database 30 'that does not include a text set that is easy to read can be created. By using this modified version of the large-scale Japanese text database 30 ', it is possible to create a text for reading that can minimize the occurrence frequency of retake even with the conventional text selection device 600.

  According to the method and apparatus for creating a text set for reading according to the present invention, it is possible to create a text for reading that does not include a sentence that is easy to read. The reading text created by the method of the present invention can minimize the frequency of occurrence of retake even by an amateur, and contributes to the construction of a low-cost and diverse speech segment database.

Note that the processes described in the above method and apparatus are not only executed in time series according to the order of description, but may also be executed in parallel or individually as required by the processing capability of the apparatus that executes the processes. Good.

  Further, when the processing means in the above apparatus is realized by a computer, the processing contents of functions that each apparatus should have are described by a program. Then, by executing this program on the computer, the processing means in each apparatus is realized on the computer.

  The program describing the processing contents can be recorded on a computer-readable recording medium. As the computer-readable recording medium, for example, any recording medium such as a magnetic recording device, an optical disk, a magneto-optical recording medium, and a semiconductor memory may be used. Specifically, for example, as a magnetic recording device, a hard disk device, a flexible disk, a magnetic tape or the like, and as an optical disk, a DVD (Digital Versatile Disc), a DVD-RAM (Random Access Memory), a CD-ROM (Compact Disc Read Only). Memory), CD-R (Recordable) / RW (ReWritable), etc., magneto-optical recording medium, MO (Magneto Optical disc), etc., semiconductor memory, EEP-ROM (Electronically Erasable and Programmable-Read Only Memory), etc. Can be used.

  The program is distributed by selling, transferring, or lending a portable recording medium such as a DVD or CD-ROM in which the program is recorded. Further, the program may be distributed by storing the program in a recording device of a server computer and transferring the program from the server computer to another computer via a network.

  Each means may be configured by executing a predetermined program on a computer, or at least a part of these processing contents may be realized by hardware.

Claims (9)

A reading text set generation unit that extracts a plurality of reading text set candidates including individual phoneme strings from a large amount of Japanese text database by inputting a plurality of phoneme strings to be recorded as input,
The reading error pattern deletion unit receives the plurality of reading text set candidates as input, determines whether or not each reading text candidate includes a reading error pattern, and reads the reading error pattern database. If it contains, another text candidate containing the same phoneme string is re-searched from the large Japanese text database, and if it does not contain a reading error pattern, the other text candidate is output as one text. Process,
To create a text set for reading. In the reading text set creation method according to claim 1,
The reading error pattern deletion process is as follows:
A sentence extraction means for extracting one text candidate from the plurality of reading text set candidates,
The reading error pattern matching means determines whether or not the one text candidate includes a reading error pattern by referring to the reading error pattern database, and if the reading error pattern does not include the one text candidate, A reading error pattern matching step for outputting as text and re-searching other text candidates including the same phoneme sequence from the mass Japanese text database when the reading error pattern is included,
A database re-searching step for performing the re-searching and inputting the re-searched other text candidates to the reading error pattern matching unit;
The alternative sentence substitution determination means determines whether or not the target phoneme string can be replaced with the output text when the target phoneme string cannot be searched again in the database re-search step. Another text alternative determination step for instructing generation of a meaningless word to the meaningless word text generation unit when the output text is set as one text corresponding to the target phoneme string and it is determined that substitution is impossible,
A meaningless word text generation means for generating one text in which the target phoneme string is replaced with a meaningless word;
A sentence output means for outputting the one text, and a sentence output step;
A method for creating a text set for reading. In the reading text set creation method according to claim 1 or 2,
The reading error pattern database is a register of easily misread words and / or reading error rules, and the reading error pattern deletion process is performed by adding the reading error pattern database to words constituting the one text set candidate. Determining whether or not the read error word or the read error rule is included, and outputting the one text set candidate as one text when the read error word is not included. How to create a text set for reading. A sentence extraction process for extracting one text including a desired phoneme sequence from a large amount of Japanese text database;
A reading error pattern matching process in which the reading error pattern is not included in the one text or the reading error pattern recorded in the reading error pattern database is matched;
A text registration process in which the one text is recorded in the modified mass Japanese text database when the one text does not match the reading error pattern;
A large-scale Japanese text database repair method including A text set generator for reading, which extracts a plurality of text set candidates for reading including individual phoneme strings from a large amount of Japanese text database by inputting a plurality of phoneme strings to be recorded,
Using the plurality of reading text set candidates as input, whether or not each reading text set includes a reading error pattern is determined by referring to the reading error pattern database. Re-search other text candidates including a column from the above-mentioned large-scale Japanese text database, and if it does not include a reading error pattern, a reading error pattern deletion unit that outputs the other text candidates as one text,
A text set creation device for reading. In the reading text set creation device according to claim 5,
The reading error pattern deletion unit
One sentence extracting means for extracting one text candidate from the plurality of reading text set candidates,
It is determined whether the one text candidate contains a reading error pattern by referring to the reading error pattern database. If the reading error pattern is not included, the one text candidate is output as one text. A reading error pattern matching means for re-searching other text candidates including the same phoneme sequence from the mass Japanese text database when including a pattern;
Database re-search means for performing the re-search and inputting the re-searched other text candidates to the reading error pattern matching means;
When it is impossible to re-search the target phoneme string in the database re-search step, it is determined whether or not the target phoneme string can be replaced with the output text. One text corresponding to the phoneme string, and another sentence substitution determining means that instructs the meaningless word text generating means to generate a meaningless word when it is determined that substitution is impossible,
Meaningless word text generation means for generating one text in which the target phoneme sequence is replaced with a meaningless word;
Sentence output means for outputting the one text, and
An apparatus for creating a text set for reading. In the reading text set creation device according to claim 5 or 6,
The reading text set creation apparatus, wherein the reading error pattern database is registered with words and / or reading error rules that are easily misread. A sentence extractor for extracting a single text including a desired phoneme sequence from a large amount of Japanese text database;
A reading error pattern matching unit in which the reading error pattern is not included in the one text or the reading error pattern recorded in the reading error pattern database is matched;
A text registration unit in which the one text is recorded in the modified large-volume Japanese text database when the one text does not match the reading error pattern;
Mass Japanese text database repair device equipped with   A program for causing a computer to execute the reading text set creation method according to any one of claims 1 to 3 or the large-scale Japanese text database modification method according to claim 4.

Images