JP4787686B2 - TEXT SELECTION DEVICE, ITS METHOD, ITS PROGRAM, AND RECORDING MEDIUM - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To select a candidate text to additionally be collected in an existent speech database storage section for synthesis so as to improve a synthesized speech. <P>SOLUTION: Candidate texts including keywords in a keyword list storage section 2 are retrieved from a candidate text database storage section 4. Here, the database storage section 4 is stored with a large amount of texts in advance and the storage section 2 is stored with important keywords in advance. Sets of components are generated from the retrieved candidate texts, hierarchical coverages are computed by referring to a frequency distribution table by the sets of components, and importances are generated from the hierarchical coverages, and a combination of candidate texts which has the highest importance, includes all the keywords, and is least in amount of text data is selected by greedy algorithm, dynamic programming, etc. <P>COPYRIGHT: (C)2008,JPO&amp;INPIT

Description

  In the present invention, in the state where a speech database storage unit required for speech synthesis in Japanese and other languages already exists, when recording additional speech data to be stored, the recording from which the speaker reads aloud A device for selecting a set of candidate texts to be recorded from a text database storage unit in which texts to be converted into speech are stored at the same time, the method, the program, and the recording medium About.

In the conventional speech synthesis technology, in recent years, as the cost of using a large-capacity storage device decreases and the calculation capacity of an electronic computer improves, speech data uttered over several tens of minutes to several hours can be stored as it is. In accordance with the input text to be synthesized and the prosody information, the speech unit is appropriately selected from the speech data stored in the storage device and connected as it is or according to the prosody information A speech synthesis method for synthesizing high-quality speech by deforming and connecting them is described in Patent Document 1, Non-Patent Document 1, and the like.
However, even if it becomes possible to store voice data for tens of hours in a large-capacity storage device, in order to create a voice database storage unit that is a source of voice synthesis, voice is recorded, In addition, it is necessary to construct a database storage unit for speech synthesis by performing a segmentation operation so that the speech can be used as a speech unit. In other words, in reality, the amount of speech that can be collected realistically is determined from the time and cost costs associated with the above work, so a speech synthesis database that can realize high-quality synthesized speech. How to build a storage unit in a short period of time and at a low cost was a big issue. For this reason, technologies for automating important segmentation tasks have been proposed for automatically constructing a speech synthesis database storage unit. If there is only data for speech synthesis, the database storage unit for speech synthesis is efficient. It is described in Non-Patent Document 2 and the like that it can be constructed automatically.

Also, as a method of collecting speech data, when synthesizing the input text, the speech synthesis database storage unit is set so that the probability that the speech unit to be used is recorded is maximized in terms of acoustics. A design method is described in Non-Patent Document 3 and the like. Also, Non-Patent Document 4 and the like describe a method of prosody multiplexing of utterances of the same content in order to avoid deterioration due to synthesis processing.
On the other hand, as a method for creating the text for recording, an appearance frequency table or an appearance rate of a unit of unit (for example, chain phoneme) included in a sentence corpus that generates linguistic data obtained by accumulating and organizing linguistic expressions used reliably is created. A method for generating a set of recording texts by sequentially selecting sentences with high scores from the sentence corpus using the total number of occurrences or the rate of occurrence of units included in each sentence in the sentence corpus as a selection criterion score Are described in Non-Patent Document 5, Non-Patent Document 6, and the like.
Japanese Patent No. 2761552 “Choose the best to modi fi cation the least: A new generation concatenative synthesis”, Proc. Eurospeech '99, 1999, pp. 2291-2294 “Phoneme labeling workbench using the HMM expert system” IEICE Technical, SP92-132, 1993 Chu, M .; Yang, H .; and Chang, E.A. , “Selecting Non-Uniform Units From a Very Large Corpus for Concatenative Speech Synthesizer”, ICASSP 2001, Vol. 2, SPEECH-L2.2, 2001. Hirota et al., “Design and Evaluation of Prosodic Multiple Databases”, Proc. 289-290, 2002-3 Jan P. H. van Santen, “Diagnostic perceptual experiments for text-to-speech system validation”, Proc. ICSLP92, pp555-558, 1992 M.M. Isogai, H .; Mizuno, K.M. Mano, “Recording Script Design for Corpus-Based TTS System Based on Coverage of Variant Phonetic Elements”, Proc ICASP2005, vol. 1, pp. 301-304, 2005

For speech synthesis that already exists, such as when there is a problem with the quality of synthesized speech when a document containing important keywords is synthesized, or when a particularly high quality of synthesized speech is required for text in a specific field. There are cases where it is desired to improve the quality of synthesized speech by recording speech in addition to the database storage unit.
However, in the conventional technology, since the text to be recorded is selected based on a predetermined unit such as a predetermined morpheme or 3-chain phoneme, speech synthesis is performed using an existing synthesis database storage unit. In this case, it is not considered whether or not keywords such as words and phrases that have a problem in the quality of the synthesized speech are to be recorded, and there is no guarantee that keywords that improve the quality of the synthesized speech can be recorded. Therefore, even if additional recording was performed, there was no guarantee that keywords would be recorded that would improve the quality of speech synthesis.

  A candidate text database storage unit that stores a large amount of candidate text as digital data, and a keyword list storage unit in which important keywords in speech synthesis are stored in advance, and a candidate that includes a keyword in the keyword list storage unit The text is searched from the candidate text database storage unit, the number of keywords included in the candidate text is counted, and all keywords in the keyword list storage unit are included from the searched candidate texts. A candidate text combination is selected, and the selected candidate text combination is output.

  According to the above configuration, in the case where it is desired to improve the quality of synthesized speech in a specific field by additionally recording the voice to the already existing synthesized speech database storage unit, it is important for this specific field. A keyword list storage unit in which keywords are stored in advance is provided, and a combination of texts including all keywords in the keyword list storage unit is selected as candidate text to be recorded in the database storage unit for synthesized speech. You can choose. For this reason, there is no omission of important keywords, and the quality of synthesized speech can be guaranteed.

Example 1
FIG. 1 shows a functional configuration of Embodiment 1 of the present invention, and FIG. 2 shows a processing flow.
The keyword list storage unit 2 stores important keywords in speech synthesis in advance. In the candidate text database storage unit 4, for example, a large amount of Japanese text is stored as digital data. Here, for example, the candidate text database storage unit may store text as a read-out manuscript, text related to a task assumed to be used for speech synthesis, text that is likely to need voice output, or A large number of candidate texts including at least one keyword stored in the keyword list storage unit 2 may be collected and stored in the candidate text database storage unit 2.
The candidate text database storage unit 4 may be a learning database storage unit.
The keyword counting unit 6 searches the candidate text database storage unit 4 for candidate texts containing the keywords stored in the keyword list storage unit 2 (step S101). The keyword counting unit 6 also counts the number of keywords in the retrieved candidate text (step S102). This counting is performed without duplication.

FIG. 3 shows a specific example of the contents stored in the keyword list storage unit 2. In this case, the keyword list storage unit 2 stores “Nippon Telegraph and Telephone”, “Co., Ltd.”, “Hanako Sato”, “Taro Yamada”, “Sales”, and the like. If one of the candidate texts is “Taro Yamada, sales representative of Nippon Telegraph and Telephone Corporation, contacted the sales office”, “Nippon Telegraph and Telephone” “Co.” “Sales” “Taro Yamada” “Sales” ”And five keywords appear, but since the third and fifth keywords are the same“ business ”, the number of candidate text keywords is counted as four.
The text number of the candidate text retrieved by the keyword counting unit 6 and the number of keywords are stored in the keyword counting storage unit 5.
The text selection unit 8 selects a combination of candidate texts including all the keywords in the keyword list storage unit from the retrieved candidate texts. Hereinafter, an example of a method for selecting a combination of candidate texts by the text selection unit 8 will be described.

First, the text selection unit 8 selects a text number of a candidate text including the largest number of keywords in the keyword count storage unit 5, and takes out the candidate text corresponding to the text number from the candidate text database storage unit 4 (step S103). . The extracted candidate text is stored in the candidate text storage unit 7 in the text selection unit 8 (step S104). The keyword count storage unit 5 may store search candidate text and the number of keywords of the candidate text. In this case, it is not necessary to retrieve the search candidate text from the candidate text database storage unit 4 by the text selection unit 8. The same applies to the following explanation embodiments.
The text selection unit 8 excludes keywords included in the selected candidate text from the keyword list storage unit 2 (step S106). The control unit 9 in the text selection unit 8 then excludes keywords, searches for candidate text, counts keywords, and selects search candidate text for the most keywords until the keyword in the keyword list storage unit 2 is empty (step S108). The storage in the candidate text storage unit 7 is repeated.
When the keyword in the keyword list storage unit 2 becomes empty, the candidate text in the candidate text storage unit 7 is output from the output unit 10 as a combination of candidate texts (step S10). Selection of a candidate text set by the above-described repetitive control has an effect of reducing the data amount of the candidate text combination. In addition, the processing method of the text selection part 8 is not restricted to this. In the above description, candidate texts including important keywords are searched. However, the number of important keywords included in each candidate text may be counted. Furthermore, although there are a plurality of the same important keywords in the same candidate text, they are counted as one, but they may be counted in duplicate.

Example 2
Next, FIG. 4 shows a functional configuration of the second embodiment, and FIG. 5 shows a processing flow. In the second embodiment, the text selection unit 8 includes the candidate text storage unit 7 and the control unit 9 described in the first embodiment, an importance generation unit 12, an importance storage unit 14, and a maximum importance set selection unit 16. , And a component generation unit 18. Note that the same reference numerals are given to portions that perform the same processing. In the following second to fifth embodiments, the same reference numerals are assigned to the same functional components, and the duplicate description is omitted.
The keyword counting unit 6 searches for candidate texts including keywords from the candidate text database storage unit 4, and the text numbers of the searched candidate texts are stored in the search keyword count storage unit 5. The component generation unit 18 in the text selection unit 8 sequentially extracts text numbers from the search keyword count storage unit 5 and extracts candidate texts corresponding to the text numbers from the candidate text database storage unit 4 (step S2).
The component generation unit 18 performs, for example, morphological analysis / reading, which is a known technique (step S4), and determines word boundaries, word part of speech, and word reading. Further, the syllable / phoneme series corresponding to the reading of the word is converted (step S6). With respect to the candidate text searched using the analyzed morpheme, syllable / phoneme, the component generation unit 18 performs analysis based on at least one layer in the acoustic and / or linguistic hierarchical structure of the spoken language. A set of one or more components is generated for each of the retrieved candidate texts (step S8).

Specifically, for example, if one candidate text in the candidate text database storage unit 4 is a sentence “This is a beautiful flower”, a morpheme is first generated by the above morphological analysis for the linguistic hierarchy. In this case, the morphemes are “this” “ha” “beautiful” “flower” “is”. Next, if a combination of two adjacent morphemes is defined as one chain morpheme, the chain morpheme is “this is” “is beautiful” “beautiful flower” “is a flower”. Next, if a combination of three adjacent morphemes is defined as one three-chain morpheme, it is “this is beautiful”, “is a beautiful flower”, and “is a beautiful flower”. In addition, the subject and predicate are generated from the morpheme, and this is “This is a beautiful flower”.
On the other hand, with respect to the acoustic hierarchy, the phoneme series conversion results in “K” “O” “R” “E” “W” “A” “K” “I” “R” “E” “I”. “N” “A” “H” “A” “N” “A” “D” “E” “S” “U” are generated, followed by “K” “R” “W” as syllables. “Ki” “Le” “I” “Na” “Ha” “Na” “De” “Su” are generated. Next, using the next two syllables as a chain syllable, “Kore” “Rewa” “Waki” “Kire” “Rei” “Ina” “Naha” “Hana” “Nade” “Death” is generated, and next Three syllables are made into three chain syllables, and “Colewa”, “Rewaki”, “Wakire”, “Beautiful”, “Reina”, “Inaha”, “Inaha”, “Nahana”, “Hanade”, “Nades” are generated. Here, each of the generated items is stored in the component storage unit 20 as a hierarchical structure as shown in FIG. As shown in FIG. 6, each generated by the component generation unit 18 is defined as a component, and each of the morpheme, chain morpheme, three chain morpheme, subject / predicate, phoneme, syllable, chain syllable, and three chain syllable is defined. A layer is defined, and a set of components for a certain layer is defined as a component set. In addition, about a layer, these are enumerated exemplarily and are not restricted to these. Further, at least one of the linguistic hierarchy and the acoustic hierarchy may be considered, and at least one of these may be considered for the layer. The generation of the hierarchical structure can also be performed manually.

Returning to FIG. 4, the generated set of components is temporarily stored in the component storage unit 20, and then read out and input to the importance generation unit 12. The importance generation unit 12 generates an importance for each searched candidate text (step S12).
Here, the importance is a value indicating the importance in selecting the candidate text, that is, a value indicating the importance of recording the candidate text in an existing speech synthesis database storage unit (not shown). Is defined as Therefore, if the candidate text with high importance is selected, the voice is recorded, and added to the existing voice synthesis database storage unit, the quality of the synthesized voice can be further improved.
In this embodiment, in general sentences, the proper noun part is often highly important, and the particle part is often less important. Based on this concept, the proper nouns such as “Taro Yamada” and “Tokyo”, which are constituent elements, are set to high numerical values, for example “10”, and the constituent elements “to” and “ha” are based on this idea. For example, “1” is set as a low numerical value as the importance level, and the importance levels of the other components of the part of speech are set as “5”, for example. Then, the importance is multiplied by the number of parts of speech, and the sum thereof is calculated as the importance of the candidate text.

For example, when the candidate text extracted by the component generation unit 18 is “I go to Tokyo and Osaka”, the configuration is “I”, “ha”, “Tokyo”, “to”, “Osaka”, “to”, “go”. Elements (morphemes) are generated (step S8), and these components are stored as one component set in the component storage unit 20 and then input to the importance generation unit 12. The importance generation unit 12 generates the importance of the candidate text with “Tokyo” and “Osaka” as two proper nouns and “ha”, “to”, and “to” as three particles. In this case, the importance of the candidate text is 23 (= 2 × 10 + 3 × 1). Note that the concept of importance is not limited to this.
Importance is generated for each retrieved candidate text (step S12), and the text number and importance of the candidate text are paired and stored in the importance storage unit 14. Further, instead of the text number, the candidate text and the importance of the text may be stored as a set in the importance storage unit 14.
The maximum importance set selecting unit 16 selects a combination of searched candidate texts having the maximum importance. An example of a method for selecting a combination of candidate texts by the maximum importance set selecting unit 16 will be described below.
First, text numbers or candidate texts are arranged in descending order of importance, candidate texts are selected from the top, and selected until all keywords in the keyword list storage unit are included. The selected combination of candidate texts is temporarily stored in the candidate text storage unit 7 and output by the output unit 10. Note that the method of selecting a combination of candidate texts by the maximum importance set selecting unit 16 is not limited to this.

Example 3
In this embodiment, a candidate text that includes a component that does not exist or rarely exists in the database storage unit of the utterance text corresponding to the speech stored in the existing synthesized speech database storage unit is highly important. I believe that. Generally, when speech synthesis is performed using an existing speech synthesis database storage unit and the speech quality is poor, the speech text database storage unit corresponding to the speech synthesis database storage unit is rarely present. This is because the quality of the synthesized speech can often be improved by recording candidate texts that are not composed components.
In this embodiment, a tier coverage calculation unit 22, a tier coverage storage unit 24, and a frequency distribution storage unit 30 are further added to the text selection unit 8 described in the first and second embodiments. For the convenience of explanation, an utterance text database storage unit 26 and a component list storage unit 28 are added.
The utterance text database storage unit 26 stores text corresponding to speech stored in the existing speech synthesis database storage unit as digital data. In the constituent element list storage unit 28, for example, the hierarchical structure of FIG. 6 is generated and stored as the constituent element list storage unit for all the utterance texts in the utterance text database storage unit 26 by the above-described constituent element generation unit 18 or manually. Has been.

In this embodiment, the concept of the coverage of components is used. The coverage is defined as the ratio of a certain component to all the components included in the utterance text database storage unit 26. That is, a component with a low coverage rate is relatively not included in the utterance text database storage unit 26, and the speech of candidate text including a component with a low coverage rate is recorded in the existing synthesized speech database storage unit. By doing so, it is possible to synthesize higher quality speech. The component layers generated by the component generation unit 18 are the same as the component layers included in the component list storage unit.
For example, for the phoneme layer, R is the total number of appearances of all phonemes included in all texts stored in the utterance text database storage unit 26, and r is the total number of target phonemes x. Then, the coverage of phoneme x can be obtained by r / R. The coverage is obtained for every component, and a frequency distribution table is created for all layers of text included in the utterance text database storage unit 26 and stored in the frequency distribution storage unit 30.
FIG. 7 shows a specific example of the frequency distribution table for the phoneme layer. The phoneme environment of a certain phoneme means the phoneme immediately before the phoneme, and the phoneme environment of a phoneme means the phoneme immediately after the phoneme. “#” Means silence. And it rearranges in order with a high coverage, for example.
Here, FIG. 7 includes two phonemes “A” of order 1 and phonemes “A” of order 2, but the front phoneme environment is the same, but the rear phoneme environment is the same. Since the phoneme “A” of the order 1 is different from the point “#” and the phoneme “A” of the order 2 is “S”, these are treated as different.

Note that the acoustic hierarchy here is “phoneme”, “syllable”, “chained syllable”, “three-chained syllable”, but the environment for each element of these layers is not a prephoneme environment or a postphoneme environment, Other sounds may be used. For example, the pre-environment and post-environment for a three-chain syllable may be any of phonemes, syllables, chain syllables, and three-chain syllables. In addition, the frequency distribution table for each layer of the acoustic hierarchy may be configured so that either the previous environment or the rear environment is considered or not both, and the frequency distribution table for each layer of the linguistic hierarchy. Does not consider both the pre-environment and the post-environment.
Returning to FIG. 4, the set of components generated by the component generation unit 18 from the candidate text is temporarily stored in the component storage unit 20, and then the set of these components is stored in the hierarchical coverage rate calculation unit 22. Entered. For each constituent element set for each retrieved candidate text, the hierarchical coverage calculation unit 22 obtains the coverage for the constituent element with reference to the frequency distribution storage unit 30 and obtains the sum of these coverage ratios.
For example, in the case of the above candidate text “This is a beautiful flower”, the hierarchical coverage calculation unit 22 refers to the frequency distribution table of FIG. For example 1 claws of the front phoneme environment silence "K", the rear phoneme environment is "O", since the "K" in order 3, the coverage is a _3.
Then, the coverage is obtained for all the remaining phonemes, the coverage is also obtained for the constituent elements of the other layers, and the hierarchical coverage is obtained for each layer using the obtained coverage (see FIG. Step S10 in step 5).

The hierarchy coverage is the total sum of the coverage of the constituent elements included in the target candidate text for each hierarchy. That is, in a layer J of a candidate text I, there are N (IJ) components, and when the coverage of the component k is C _IJk , the hierarchical coverage C _IJ is obtained by the following equation (1). I can do it.
C _IJ = Σ _{k = 1} ^{N (IJ)} C _IJk (1)
For example, the constituent elements of the morpheme layer of the candidate text “This is a beautiful flower” are “this” “ha” “beautiful” “flower” “is”. When the candidate text has the tier coverage in the morpheme layer as shown in FIG. FIG. 8 shows, for example, that the coverage of the morpheme “this” is 7.25 × 10 ⁻⁵ .
Therefore, 7.25 × 10 ⁻⁵ + 2.83 × 10 ⁻⁴ + 5.84 × 10 ⁻⁶ + 1.43 × 10 ⁻⁵ + 6.93 × 10 ⁻⁴ ≈1.07 × 10 ⁻³ .
Next, the importance generation unit 12 calculates the sum of the hierarchical coverages for each retrieved candidate text, and if this sum is large, a value that is small is determined as the importance. Here, after obtaining the weighted sum of the hierarchical coverage, the reciprocal is obtained.
That severity S _I of the candidate text I, the number of layers to L, and the W _J and weighting coefficients of the hierarchical J, can be obtained by the following equation (2).
S _I = 1 / Σ _{J = 1} ^L W _J・ C _IJ・ ・ ・ ・ (2)
For example, for the above candidate text “This is a beautiful flower”, as shown in FIG. 9, hierarchical coverage of the subject / predicate, 3 chain morpheme, chain morpheme, morpheme, 3 chain syllable, chain syllable, syllable, phoneme When the rates are b, c, d, e, f, g, h, a, the weighting factors are W _b , W _c , W _d , W _e , W _f , W _g , W _h , W, respectively. _{If a} , then the importance of this candidate text can be determined as follows.
1 / (b · W _b + c · W _c + d · W _d + e · W _e + f · W _f + g · W _g + h · W _h + a · W _a )

Here, the (weighted) sum of the hierarchical coverage is obtained, and if this value becomes large, the reason why the value that becomes small becomes the importance is the (weighted) sum of the hierarchical coverage of the candidate text. Indicates that the candidate text can be reconstructed from the components included in the utterance text database storage unit 26 corresponding to the existing speech synthesis database storage unit. If the (weighted) sum of the hierarchical coverage is low, the candidate text is composed of components that rarely exist in the utterance text database storage unit 26. Therefore, if a candidate text including a highly important component that is the reciprocal of the (weighted) sum of the hierarchical coverage is recorded and added to the speech synthesis database storage unit created from the utterance text database storage unit 26, the synthesized speech Quality can be improved. This is because, as described above, the synthesized speech is recorded by adding candidate texts including components that are rarely present in the existing speech synthesis database storage unit and adding them to the existing speech synthesis database storage unit. This is because the quality of the product can be improved in many cases.
As for the weighting factor, when the keyword counting unit 6 makes the retrieved candidate text for an application depending on a certain field, the linguistic hierarchical structure hierarchy is obtained from the weighting coefficient for the hierarchical coverage of the acoustic hierarchical structure. What is necessary is just to enlarge the weighting coefficient with respect to a coverage. This is because, when the retrieved candidate text is intended for use depending on a certain field, the morpheme generated from the candidate text often includes words or the like depending on the specific field. Conversely, if the retrieved candidate text is to be used for applications that do not depend relatively on a certain field, the weighting factor for the hierarchical coverage of the linguistic hierarchical structure is made smaller than the weighting factor for the hierarchical coverage of the acoustic hierarchical structure. do it. If weighting is not considered, all the weighting factors may be set to “1”.
In the above description, the importance is generated by taking the reciprocal of the weighted sum. However, the importance is not limited to this. For example, it is possible to add a minus to the weighted sum to make the importance.

Example 4
In this embodiment, an example of a specific processing flow of the maximum importance set selecting unit 16 described in the first to third embodiments will be described. Note that the processing of the maximum importance set selecting unit 16 described below uses a greedy algorithm (step S14 in FIG. 5) which is a known technique. The maximum importance group selection unit 16 includes a most important text selection unit 32, a keyword removal unit 34, and a repetition control unit 36. FIG. 10 shows the flow of processing of the greedy algorithm.
For example, the importance storage unit 14 may store the importance in the format shown in FIG. For example, a case where the constituent element generation unit 18 generates constituent elements for a chain morpheme, a chain syllable, and a phoneme will be described. For each candidate text including a keyword, a candidate text number such as “7” and the number of keywords For example, K ₁ , each layer name, a set of components for each layer, a hierarchical coverage for each layer, and importance are stored. This storage format is the same in the fifth embodiment below.

The most important text selection means 32 selects the candidate text having the highest importance generated by the importance generation unit 12 from the candidate text database storage unit 4 and stores it in the candidate text storage unit 7 (step S16). The keyword removal unit 34 removes all keywords included in the selected candidate text from the keyword list storage unit 2 (step S18).
Then, the control unit 36 repeatedly performs the processes of the keyword counting unit 6, the component generation unit 18, the importance generation unit 12, the most important text selection unit 32, and the keyword removal unit 34 described above (step S20). This iterative process is repeated until the keywords included in the set of all candidate texts selected by the most important text selection means 32 match all the first keywords in the keyword list storage unit.
The importance used in the processing of the maximum importance set selecting unit 16 by this greedy algorithm is not limited to the case of obtaining by the method shown in the third embodiment, but may be obtained by the method shown in the second embodiment.

Example 5
In this embodiment, the processing flow of the maximum importance set selecting unit 16 is processed using a dynamic programming method (step S22 in FIG. 5) which is a known technique. Dynamic programming is an algorithm that selects a value that maximizes or minimizes a certain value while satisfying a certain condition. Details of dynamic programming are described in “R. Bellman“ Dynamic programming ”Princeton University Press 1957”.
In this case, the maximum importance set selecting unit 16 includes a data amount measuring unit 38 and a dynamic programming executing unit 40.
First, the data amount measuring means 38 measures the data amount for all candidate texts stored in the importance storage unit 14. As indicated by a broken line in FIG. 11, the importance storage unit 14 stores the candidate text in a format to which the data amount is added. And the dynamic programming execution means 40 selects a candidate text by the following process.
The dynamic programming execution means 40 uses the dynamic programming to maximize the sum of the importance levels of the combinations of retrieved candidate texts, includes all the keywords in the keyword list storage unit 2, and Select a combination of candidate texts that minimizes the total amount of data.

Here, whether or not all the keywords in the keyword list storage unit 2 are included may be determined by determining whether or not the number of keywords in the keyword list storage unit 2 is equal to the number of keywords in the selected candidate text. Here, when calculating the sum of the number of keywords included in the keyword candidate text, the same keyword is counted as one. If they are equal, all the keywords in the keyword list storage unit 2 are included. It is also possible to execute a dynamic programming method by setting a threshold value for the total amount of data, and repeatedly execute the dynamic programming method by gradually lowering the threshold value for the total data amount until the condition is not satisfied. It is done.
In addition to the above embodiments, the text selection device / method according to the present invention is not limited to the above-described embodiments, and can be appropriately changed without departing from the spirit of the present invention. In addition, the processing described in the above text selection device / method is not only executed in time series in the order described, but also executed in parallel or individually as required by the processing capability of the device that executes the processing. It is good.
When the processing function in the text selection device is realized by a computer, the processing content of the function that the text selection device should have is described by a program. Then, by executing this program on a computer, the processing functions in the text selection device are 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, and 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-Programmable-Ready), 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. Furthermore, the program may be distributed by storing the program in a storage device of the server computer and transferring the program from the server computer to another computer via a network.

A computer that executes such a program first stores, for example, a program recorded on a portable recording medium or a program transferred from a server computer in its own storage device. When executing the process, the computer reads a program stored in its own recording medium and executes a process according to the read program. As another execution form of the program, the computer may directly read the program from a portable recording medium and execute processing according to the program, and the program is transferred from the server computer to the computer. Each time, the processing according to the received program may be executed sequentially. A configuration in which the above-described processing is executed by a so-called ASP (Application Service Provider) type service that realizes a processing function only by an execution instruction and result acquisition without transferring a program from the server computer to the computer. It is good. Note that the program in this embodiment includes information that is used for processing by an electronic computer and that conforms to the program (data that is not a direct command to the computer but has a property that defines the processing of the computer).
In this embodiment, the language model creation apparatus is configured by executing a predetermined program on a computer. However, at least a part of the processing contents may be realized by hardware.

The block diagram which shows the specific structural example of Example 1 of this invention. The flowchart which shows the flow of the main processes of Example 1 of this invention. A specific example of the keyword list storage unit 2. The block diagram which shows the specific structural example of Examples 2-6 of this invention. The flowchart which shows the flow of the main processes of Examples 2-6 of this invention. The figure which shows the hierarchical structure of the component produced | generated by the component generation part 18. FIG. A specific example of a frequency distribution table for a phoneme layer in the frequency distribution storage unit 4. The example which shows the coverage of each component in the layer of a morpheme. The figure which shows the hierarchy coverage about each of the subject and predicate in a linguistic hierarchy, 3 chain morpheme, a chain morpheme, a morpheme, and 3 chain syllables, a chain syllable, a syllable, and a phoneme in an acoustic hierarchy. The flowchart which shows the flow of a process at the time of using a greedy algorithm in the maximum importance group selection part 16. FIG. In Examples 4-6, the figure which shows the specific memory format in the importance memory | storage part 14. FIG.

Claims (10)

A candidate text database storage unit in which a large amount of candidate text is stored as digital data,
A keyword list storage unit in which important keywords in speech synthesis are stored in advance;
A keyword counting unit that counts the number of the important keywords included in each of the plurality of candidate texts, stored in the candidate text database storage unit ;
From among the counted candidate text, and the text selection unit for selecting a combination of candidate text,
Have
The text selection part
A text selection means for selecting a candidate text having the largest number of the important keywords from the counted candidate texts, and storing the selected candidate text in a candidate text storage unit;
Keyword removing means for removing keywords included in the candidate text selected by the text selecting means from the keyword list storage unit;
Repeat control means for sequentially controlling the keyword counting section, the text selection means, and the keyword removal means until the keywords in the keyword list storage section are empty,
Text selection device according to claim <br/> contain. A candidate text database storage unit in which a large amount of candidate text is stored as digital data,
A keyword list storage unit in which important keywords in speech synthesis are stored in advance;
A keyword counting unit for searching candidate texts including the important keyword from the candidate text database storage unit;
A text selection unit for selecting a combination of candidate texts from the searched candidate texts;
Have
The text selection part
For each of the retrieved candidate texts , a component generation unit that generates a set of components of a predetermined hierarchy in the acoustic and / or linguistic hierarchical structure of the spoken language;
For each of the retrieved candidate texts, an importance generation unit that calculates the sum of the importance given to each component included in the set of component and generates the sum as the importance of the candidate text ; ,
The most important text selection for selecting the candidate text having the highest importance of the candidate text from the searched candidate texts not stored in the candidate text storage unit and storing the selected candidate text in the candidate text storage unit Means,
Keyword removing means for removing keywords included in the candidate text selected by the most important text selecting means from the keyword list storage unit;
The most important text selection means and the keyword removal means until the keywords included in the set of candidate texts stored in the candidate text storage unit match all of the first keywords in the keyword list storage unit. Repetitive control means for sequentially and repetitively controlling;
A text selection device comprising: A candidate text database storage unit in which a large amount of candidate text is stored as digital data,
A keyword list storage unit in which important keywords in speech synthesis are stored in advance;
A keyword counting unit for searching candidate texts including the important keyword from the candidate text database storage unit;
A text selection unit for selecting a combination of candidate texts from the searched candidate texts;
Have
The text selection part
A component generation unit that generates a set of one or more components by analyzing each of the retrieved candidate texts based on at least one layer in the acoustic and / or linguistic hierarchical structure of the spoken language; ,
Percentage of the constituent element in the constituent element set corresponding to at least one constituent element set to be generated with respect to the utterance text corresponding to the speech stored in the speech synthesis database storage unit A frequency distribution storage unit in which the coverage ratio indicating is stored;
For each component set for each of the retrieved candidate texts, a coverage with respect to the component is obtained by referring to the frequency distribution storage unit, and the sum of these coverages is used as a hierarchical coverage to obtain a hierarchical coverage calculation unit When,
An importance generation unit that obtains the sum of the hierarchical coverages for each of the searched candidate texts, and obtains a value that is small as the importance if the sum is large;
The most important text selection for selecting the candidate text having the highest importance of the candidate text from the searched candidate texts not stored in the candidate text storage unit and storing the selected candidate text in the candidate text storage unit Means,
Keyword removing means for removing keywords included in the candidate text selected by the most important text selecting means from the keyword list storage unit;
The most important text selection means and the keyword removal means until the keywords included in the set of candidate texts stored in the candidate text storage unit match all of the first keywords in the keyword list storage unit. Repetitive control means for sequentially and repetitively controlling;
A text selection device comprising: A candidate text database storage unit in which a large amount of candidate text is stored as digital data,
A keyword list storage unit in which important keywords in speech synthesis are stored in advance;
A keyword counting unit for searching candidate texts including the important keyword from the candidate text database storage unit;
A combination of candidate texts including all keywords in the keyword list storage unit by preferentially selecting candidate texts with high importance calculated for each candidate text from the searched candidate texts A text selection unit that stores a combination of the selected candidate texts in the candidate text storage unit,
Have
The text selection part
A component generation unit that generates a set of one or more components by analyzing each of the retrieved candidate texts based on at least one layer in the acoustic and / or linguistic hierarchical structure of the spoken language; ,
Percentage of the constituent element in the constituent element set corresponding to at least one constituent element set to be generated with respect to the utterance text corresponding to the speech stored in the speech synthesis database storage unit A frequency distribution storage unit in which the coverage ratio indicating is stored;
For each component set for each of the retrieved candidate texts, a coverage with respect to the component is obtained by referring to the frequency distribution storage unit, and the sum of these coverages is used as a hierarchical coverage to obtain a hierarchical coverage calculation unit When,
An importance generation unit that obtains the sum of the hierarchical coverages for each of the searched candidate texts, and obtains a value that is small as the importance if the sum is large;
A text selection device comprising: A candidate text database storage unit in which a large amount of candidate text is stored as digital data,
A keyword list storage unit in which important keywords in speech synthesis are stored in advance;
Use
A keyword counting process in which the keyword counting means counts the number of the important keywords included in each of the plurality of candidate texts stored in the candidate text database storage unit ;
Text selection means, from among the counted candidate text, a combination of candidate text, and the text selection step of selecting,
I have a,
The text selection process is
A candidate text selection process of selecting the candidate text having the largest number of the important keywords from the counted candidate text and storing the selected candidate text in the candidate text storage unit;
A keyword removal process for removing keywords included in the candidate text selected by the text selection means from the keyword list storage unit;
Until the keyword in the keyword list storage unit is empty, a repetitive control process in which the keyword counting unit, the text selecting unit, and the keyword removing unit are sequentially repetitively controlled,
A method for selecting text , comprising : A candidate text database storage unit in which a large amount of candidate text is stored as digital data,
A keyword list storage unit in which important keywords in speech synthesis are stored in advance;
Use
A keyword counting process in which the keyword counting means searches the candidate text database storage unit for candidate text containing the important keyword ;
A text selection process in which the text selection means selects a combination of candidate texts from the searched candidate texts;
Have
The text selection process is
A component generation process in which the component generation means generates a set of components in a predetermined hierarchy in the acoustic and / or linguistic hierarchical structure of the spoken language for each of the retrieved candidate texts;
For each of the retrieved candidate texts, the importance generation means calculates the sum of the importance assigned to each component included in the set of the components, and obtains the sum as the importance of the candidate text. Importance generation process ,
Among the searched candidate text is not stored in the candidate text storage unit, the most important text importance of the candidate text selects the largest candidate text, and stores the candidate selected text to the candidate text storage unit The selection process,
A keyword removal process for removing keywords included in the candidate text selected by the most important text selection means from the keyword list storage unit;
The most important text selection means and the keyword removal means until the keywords included in the set of candidate texts stored in the candidate text storage unit match all of the first keywords in the keyword list storage unit. A repetitive control process for sequentially repetitive control;
A method for selecting text, comprising: A candidate text database storage unit in which a large amount of candidate text is stored as digital data,
A keyword list storage unit in which important keywords in speech synthesis are stored in advance;
Use
A keyword counting process in which the keyword counting means searches the candidate text database storage unit for candidate text containing the important keyword ;
A text selection process in which the text selection means selects a combination of candidate texts from the searched candidate texts;
Have
The text selection process is
A component generation process for generating a set of one or more components for each of the retrieved candidate texts by analysis based on at least one layer in the acoustic and / or linguistic hierarchical structure of the spoken language; ,
Percentage of the constituent element in the constituent element set corresponding to at least one constituent element set to be generated with respect to the utterance text corresponding to the speech stored in the speech synthesis database storage unit A frequency distribution storage process in which the coverage ratio indicating is stored,
Hierarchical coverage calculation process for obtaining the coverage for each constituent element for each retrieved candidate text by referring to the frequency distribution storage unit and using the sum of these coverages as the hierarchical coverage When,
An importance generation process for obtaining the sum of the hierarchical coverages for each of the searched candidate texts, and obtaining a value that is small if the sum is large,
The most important text selection for selecting the candidate text having the highest importance of the candidate text from the searched candidate texts not stored in the candidate text storage unit and storing the selected candidate text in the candidate text storage unit Process,
A keyword removal process for removing keywords included in the candidate text selected by the most important text selection means from the keyword list storage unit;
The most important text selection means and the keyword removal means until the keywords included in the set of candidate texts stored in the candidate text storage unit match all of the first keywords in the keyword list storage unit. A repetitive control process for sequentially repetitive control;
A method for selecting text, comprising: A candidate text database storage unit in which a large amount of candidate text is stored as digital data,
A keyword list storage unit in which important keywords in speech synthesis are stored in advance;
Use
A keyword counting process for searching candidate texts containing the important keywords from the candidate text database storage unit;
A combination of candidate texts including all keywords in the keyword list storage unit by preferentially selecting candidate texts with high importance calculated for each candidate text from the searched candidate texts A text selection process for storing the selected combination of candidate texts in the candidate text storage unit, and
Have
The text selection process is
A component generation process for generating a set of one or more components for each of the retrieved candidate texts by analysis based on at least one layer in the acoustic and / or linguistic hierarchical structure of the spoken language; ,
Percentage of the constituent element in the constituent element set corresponding to at least one constituent element set to be generated with respect to the utterance text corresponding to the speech stored in the speech synthesis database storage unit A frequency distribution storage process in which the coverage ratio indicating is stored,
Hierarchical coverage calculation process for obtaining the coverage for each constituent element for each retrieved candidate text by referring to the frequency distribution storage unit and using the sum of these coverages as the hierarchical coverage When,
An importance generation process for obtaining the sum of the hierarchical coverages for each of the searched candidate texts, and obtaining a value that is small if the sum is large,
A method for selecting text, comprising: A text selection program for causing a computer to execute each step of the text selection method according to any one of claims 5 to 8 . A computer-readable recording medium on which the text selection program according to claim 9 is recorded.

Images