JP2008203717A - Text sentence selecting method for corpus-based speech synthesis, and program thereof and device thereof - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To obtain a text sentence selecting method for corpus-based speech synthesis which evaluates a quality index of a text sentence set while searching for suitable text sentences from an expanded range, and is small in computational amount. <P>SOLUTION: The text sentence selecting method for corpus-based speech synthesis includes: a step of providing a thesaurus dictionary 110 containing synonyms of an object category of speech synthesis, selecting one arbitrary sentence from a text corpus, and selecting a text sentence (selected sentence) such that when the text is added to a work text sentence set, the corpus quality becomes maximum; a step of generating a similar sentence group by substituting synonyms that the thesaurus dictionary holds for words and phrases constituting the selected sentence; a re-decision step of selecting one arbitrary sentence from a group of sentences similar to the selected sentence and finding a text sentence (best sentence) such that when the sentence is added to the work text sentence set, the corpus quality becomes maximum; and a step of adding the best sentence obtained in the re-decision step to the work text sentence set. <P>COPYRIGHT: (C)2008,JPO&INPIT

Description

  The present invention relates to a method for selecting a suitable text sentence set used in corpus-based speech synthesis from a text corpus, a program therefor, and an apparatus therefor.

Conventionally, “a sentence set automatic generation method, apparatus, program, and storage medium for automatically generating a sentence set that can efficiently collect variable words that frequently appear by processing a part of a sentence are provided. As a technique for the purpose of "using a task sentence corpus in which the text of a specific task that is a candidate for the selected text is stored in the task sentence corpus storage unit 1, and a word list 2 of words specific to the specific task, The frequency of occurrence of a word in the word list 2 in the task sentence corpus is obtained, the corresponding word portion in the task sentence corpus is replaced with a word symbol, a symbol sentence corpus is obtained, and a combination of candidate texts from the symbol sentence corpus is set as a symbol text set. The sentence set automatic generation method which embeds the word symbol part included in the text set in the order of the appearance frequency of the words in the order of the appearance frequency of the words corresponding to the appearance order of the word symbols in the text set. Is proposed (Patent Document 1).
In addition, “a sentence set for reading a sentence (text) necessary for creating a speech database used for speech synthesis is efficiently generated. As a technique for the purpose of the above, “for each sentence in the sentence corpus, the total value of the appearance rate of basic segments with the three-chain phoneme as a basic unit is calculated as a score (step 1-1). Except for the basic segments already stored in the sentence set, the duplicates of a plurality of basic segments of the same form in the same sentence are not included in the score calculation. If there is one sentence with the highest score (step 1-4), this is stored in the sentence set, and if there are more than one sentence, various sentences in units of extended segments, such as syllables with environment and morphemes with environment, are used. The total value of the appearance rate of the extended segment is calculated as a score by the same method as described above. Various extended segment scores for the same sentence are weighted to form a composite score, and the sentence with the highest composite score is stored in the sentence set. The task dependency can be increased by setting the weighting coefficient for load addition. Is proposed (Patent Document 2).

JP 2004-347955 A (summary) JP 2004-246140 A (summary)

In the technique described in Patent Document 1, a range of searching for a suitable text sentence is expanded by replacing a part of the text sentence stored in the task sentence corpus with a word list. However, when expanding the search range, there has been a problem that the quality index of the text sentence set such as the phoneme balance of the entire text sentence set and the number of phoneme appearances is not considered.
On the other hand, the technique described in Patent Document 2 evaluates the quality index of the text sentence set by score calculation, but does not extend the search range as in Patent Document 1.

Therefore, it is conceivable to perform a quality index evaluation of a text sentence set as described in Patent Document 2 for an extended search range as described in Patent Document 1.
However, the expanded search range is increased by a combination of the number of replaceable words and the replacement word section. For example, when the number of replacement words = 10 and the word interval = 3, 10 × 10 × 10 = 1000 extended text sentence sets can be created for one text sentence before expansion. In this case, if the text corpus population before expansion is 100,000 sentences, the text sentence set after expansion will be 1 billion sentences.

The process of evaluating the quality index for all such a large number of text sentence sets and selecting a suitable text sentence set (for example, 1000) requires a very large amount of calculation.
Therefore, a text sentence set selection method for corpus-based speech synthesis, which evaluates the quality index of a text sentence set while searching for a suitable text sentence from the expanded search range, and has a small amount of computation, its program, and A device was desired.

A text sentence set selection method for corpus-based speech synthesis according to the present invention includes:
A method for selecting a text sentence set for corpus-based speech synthesis from a text corpus,
A storage means for storing a thesaurus that holds synonyms in the target category of speech synthesis;
First, an initialization step to empty the working text statement set,
A selection step of selecting an arbitrary sentence from the text corpus and selecting from the text corpus a text sentence (hereinafter referred to as a selected sentence) that maximizes the corpus quality when the sentence is added to the working text sentence set;
A thesaurus replacement step of creating a similar sentence group by replacing a phrase constituting the selected sentence in the selection step with a synonym held by the thesaurus dictionary;
Redetermination for selecting a text sentence (hereinafter referred to as the best sentence) that maximizes the corpus quality when an arbitrary sentence is selected from the selected sentence and similar sentence groups and the sentence is added to the working text sentence set. Steps,
A sentence adding step of adding the best sentence obtained by the redetermination step to the working text sentence set;
It has a design condition judging step that repeats from the selection step to the sentence addition step until the corpus quality of the added work text sentence set satisfies the corpus design condition.

  According to the text sentence set selection method for corpus-based speech synthesis according to the present invention, the quality index of the text sentence set is evaluated and the amount of calculation is reduced while maintaining a wide range for searching for a suitable text sentence. Therefore, it contributes to shortening of the search time, downsizing of the apparatus, reduction of memory, and cost reduction.

Embodiment 1 FIG.
FIG. 1 is a functional block diagram of a text sentence set selection apparatus 100 for corpus-based speech synthesis according to Embodiment 1 of the present invention.
The text sentence set selection device 100 includes a thesaurus dictionary 110, a work text sentence set recording unit 120, an initialization unit 130, a selection unit 140, a thesaurus replacement unit 150, a re-determination unit 160, a sentence addition unit 170, and a design condition determination unit 180. Prepare.
The thesaurus dictionary 110 holds a list of synonyms in the target category of speech synthesis.
The working text sentence set recording unit 120 stores a working text sentence set at an intermediate stage during the text sentence set selection process.
The initialization unit 130 empties the work text sentence set in the work text sentence set recording unit 120 at the start of the text sentence set selection process.
The selection unit 140 selects an arbitrary sentence from the text corpus, and selects a text sentence (hereinafter referred to as a selected sentence) that maximizes the corpus quality when the sentence is added to the working text sentence set from the text corpus. .
The thesaurus replacement unit 150 creates a similar sentence group by replacing the words constituting the selected sentence selected by the selection unit 140 with synonyms held in the thesaurus dictionary 110.
The re-determination unit 160 selects an arbitrary sentence from the selected sentence and the previous similar sentence group, and adds the sentence to the work text sentence set recorded in the work text sentence set recording unit 120. A text sentence with the highest quality (hereinafter referred to as the best sentence) is obtained.
The sentence adding unit 170 adds the best sentence obtained by the re-determination unit 160 to the work text sentence set recorded in the work text sentence set recording unit 120.
The design condition determination unit 180 repeats from the selection step to the sentence addition step until the corpus quality of the work text sentence set recorded in the work text sentence set recording unit 120 satisfies the corpus design condition. On the other hand, when the corpus design quality is satisfied, the work text sentence set recorded in the work text sentence set recording unit 120 is output as a text sentence set.
The text corpus 200 stores a text sentence mother set that is a basis of a text sentence set for corpus-based speech synthesis.

The work text sentence set recording unit 120, the initialization unit 130, the selection unit 140, the thesaurus replacement unit 150, the re-determination unit 160, the sentence addition unit 170, and the design condition determination unit 180 are configured by hardware such as a circuit device. It can also be configured as software executed on an arithmetic device such as a microcomputer or CPU.
As a configuration example of the thesaurus dictionary 110, it is conceivable to store a file in which a synonym list is recorded in a storage device having a relatively large capacity such as an HDD (Hard Disk Drive). The thesaurus dictionary 110 may be provided outside the text sentence set selection device 100.
Similarly, the text corpus 200 can be configured by storing a file in which a text sentence is recorded in a storage device having a relatively large capacity such as an HDD.

FIG. 2 conceptually illustrates the operation of the text sentence set selection apparatus 100 according to the first embodiment.
In FIG. 2, a text corpus 200 stores a set of text sentences represented by small circles. Further, the range of the text sentence obtained by converting the phrases constituting each text sentence with the synonyms stored in the thesaurus dictionary 110 is represented by a broken-line great circle.

As described above, when selecting a text sentence set suitable for corpus-based speech synthesis from a set of text sentences stored in the text corpus 200, the text sentences after synonym conversion by the thesaurus dictionary 110 are processed. If the search range is also used, the amount of calculation may become enormous.
Therefore, in the first embodiment, the most suitable text sentence is selected in advance before the synonym conversion by the thesaurus dictionary 110, and only the range in which the text sentence is synonymous converted by the thesaurus dictionary 110 is set as the search range.
By doing in this way, the amount of calculation can be reduced significantly compared with the case where the entire range after synonym conversion by the thesaurus dictionary 110 is used as the search range. In addition, since the most suitable text sentence is selected at the stage before synonym conversion by the thesaurus dictionary 110, it is considered that a suitable text sentence exists in the vicinity, and a certain quality is ensured.

Here, prior to the description of the operation flow, a supplementary explanation will be given regarding the quality index of the text sentence set.
As a quality index of the text sentence set used here, the number and coverage of each phoneme unit (for example, vowel / a /, / i /, / u /, / e /, / o /) included in the text sentence The number of acoustic performances such as the number of the chain phoneme units (for example, / aa /, / ai /) and the coverage thereof can be considered.
Furthermore, as shown in Patent Document 1, a coverage based on a complicated hierarchical definition may be used, or other suitable acoustic performance amount and linguistic performance amount may be used in combination.

  FIG. 3 is an operation flow of the text sentence set selection apparatus 100 according to the first embodiment. Hereinafter, each step will be described.

(S330) Initialization Step The initialization unit 130 empties the work text sentence set in the work text sentence set recording unit 120 at the start of the text sentence set selection process.
(S340) Selection Step The selection unit 140 selects an arbitrary sentence from the text corpus and selects a text sentence (hereinafter referred to as a selected sentence) that maximizes the corpus quality when the sentence is added to the working text sentence set. Select from a text corpus.
(S350) Thesaurus replacement step The replacement unit 150 creates a similar sentence group by replacing the words constituting the selected sentence selected by the selection unit 140 with synonyms held by the thesaurus dictionary 110.
(S360) Re-determination step The re-determination unit 160 selects any one sentence from the selected sentence and the previous similar sentence group, and the work text sentence set recorded in the work text sentence set recording unit 120 is selected. The text sentence (hereinafter referred to as the best sentence) that maximizes the corpus quality when it is added to is obtained.
(S370) Sentence adding step The sentence adding unit 170 adds the best sentence obtained by the re-determination unit 160 to the work text sentence set recorded in the work text sentence set recording unit 120.
(S380) Design Condition Determination Step The design condition determination unit 180 repeats the selection step to the statement addition step until the corpus quality of the work text statement set recorded in the work text statement set recording unit 120 satisfies the corpus design condition. . On the other hand, when the corpus design quality is satisfied, the work text sentence set recorded in the work text sentence set recording unit 120 is output as a text sentence set.

  The output destination of the text sentence set may be a storage device provided outside the text sentence set selection device 100, a network interface, a memory area provided separately, and the like. The output format may be text data itself or only pointer information.

  As described above, according to the first embodiment, before a text sentence is converted by the thesaurus dictionary 110, one suitable text sentence is searched in advance, and synonym conversion by the thesaurus 110 is performed for only that one sentence. Therefore, the amount of calculation can be greatly reduced as compared with the case where the entire range after conversion by the thesaurus dictionary 110 is set as the search range.

Embodiment 2. FIG.
In the second embodiment of the present invention, an operation example that can further reduce the amount of calculation of the first embodiment will be described. Note that the configuration of the text sentence set selection apparatus 100 is the same as that of FIG.

FIG. 4 conceptually illustrates the operation of the text sentence set selection apparatus 100 according to the second embodiment. In the second embodiment, the search range is not the entire range obtained by synonymous conversion of the text sentence with the thesaurus dictionary 110, but only a part of the range is set as the search range.
In FIG. 4, what is represented by a triangle indicates a text sentence to be searched out of an area represented by a broken-line great circle. In FIG. 2 described in the first embodiment, the entire search area within the area represented by the broken circle is the search range. However, in the second embodiment, as shown by the triangle in FIG. Narrow down.

FIG. 5 is a specific example of the conceptual diagram described in FIG.
When converting a text sentence using the thesaurus dictionary 110, first, a conversion target is determined from words or phrases constituting the original text sentence. For example, when the original sentence before conversion is “this invention relates to a method for selecting a text sentence”, a noun phrase such as “text sentence” or a verb phrase such as “related” among them. Is the main conversion target.
In Embodiment 1, since all of these are conversion targets, for example, when there are six syllable parts to be converted and there are ten synonyms for each conversion target part, the original text is converted and derived. There can be 10 6 text sentences.

On the other hand, in the second embodiment, only noun words are converted. Therefore, there are fewer derived text sentences. For example, if there are three noun phrases in the six syllable parts to be converted, the number of derived text sentences is less than half.
The reason for converting the noun word / phrase is that it is suitable for creating a variation of the text sentence. However, the present invention is not limited to this. Depending on the text sentence or the contents of the thesaurus dictionary 110, the verb phrase may be converted. It may be a target.

The operation flow of the text sentence set selection apparatus 100 in the second embodiment is basically the same as that described in FIG. 3 of the first embodiment. However, since the number of derived sentences by synonym conversion decreases as described with reference to FIGS. 4 to 5, the amount of calculation in steps S350 to S360 can be reduced.
This contributes to a reduction in search time and a reduction in the size, memory, and cost of the text sentence set selection apparatus 100. Moreover, since many synonyms stored in the thesaurus dictionary 110 are mainly related to noun phrases, it is possible to secure sufficient variations of text sentences by converting noun phrases and obtaining derived sentences.

Embodiment 3 FIG.
FIG. 6 conceptually illustrates the operation of the text sentence set selection apparatus 100 according to Embodiment 3 of the present invention.
In the second embodiment, only noun phrases are subject to synonym conversion, but in this third embodiment, the phrase having the longest number of characters or syllables is subject to synonym conversion. The longest of all the conversion target phrases such as noun phrases and verb phrases may be the target, or the longest of the noun phrases alone may be the target.

The operation flow of the text sentence set selection apparatus 100 in the third embodiment is basically the same as that described in FIG. 3 of the first embodiment. However, as described with reference to FIG. 6, the number of derived sentences by synonym conversion is significantly reduced, so that the amount of calculation in steps S350 to S360 can be greatly reduced.
This contributes to a reduction in search time and a reduction in the size, memory, and cost of the text sentence set selection apparatus 100. In addition, since the search range is determined with the longest word or phrase having a relatively large influence on the phoneme environment or the like as the conversion target, the quality of the search result can be maintained at a certain level.

It is a functional block diagram of the text sentence set selection apparatus 100 for corpus-based speech synthesis according to the first embodiment. The operation | movement of the text sentence set selection apparatus 100 which concerns on Embodiment 1 is demonstrated notionally. It is an operation | movement flow of the text sentence set selection apparatus 100 which concerns on Embodiment 1. FIG. The operation | movement of the text sentence set selection apparatus 100 which concerns on Embodiment 2 is demonstrated notionally. FIG. 5 is a specific example of the conceptual diagram described in FIG. 4. The operation | movement of the text sentence set selection apparatus 100 which concerns on Embodiment 3 is demonstrated notionally.

Explanation of symbols

  DESCRIPTION OF SYMBOLS 100 Text sentence set selection apparatus, 110 Thesaurus dictionary, 120 Work text sentence set recording part, 130 Initialization part, 140 Selection part, 150 Thesaurus replacement part, 160 Re-determination part, 170 sentence addition part, 180 Design condition judgment part, 200 Text corpus.

Claims (7)

A method for selecting a text sentence set for corpus-based speech synthesis from a text corpus,
A storage means for storing a thesaurus that holds synonyms in the target category of speech synthesis;
An initialization step to empty the working text statement set first,
A selection step of selecting an arbitrary sentence from the text corpus and selecting from the text corpus a text sentence (hereinafter referred to as a selected sentence) that maximizes the corpus quality when the sentence is added to the working text sentence set;
A thesaurus replacement step of creating a similar sentence group by replacing a phrase constituting the selected sentence in the selection step with a synonym held by the thesaurus dictionary; and
Re-determination step of selecting a text sentence (hereinafter referred to as the best sentence) that maximizes the corpus quality when an arbitrary sentence is selected from the selected sentence group and the similar sentence group and the sentence is added to the working text sentence set. When,
A sentence adding step of adding the best sentence obtained by the redetermination step to the working text sentence set;
Text sentence set selection for corpus-based speech synthesis characterized by having a design condition judgment step that repeats from the selection step to the sentence addition step until the corpus quality of the added work text sentence set satisfies the corpus design condition Method. In the thesaurus replacement step,
The method for selecting a text sentence set for corpus-based speech synthesis according to claim 1, wherein only noun words or phrases are set as replacement targets among the phrases constituting the text sentence selected in the selection step. In the thesaurus replacement step,
3. The corpus-based speech synthesis for corpus-based speech synthesis according to claim 1, wherein, among words and phrases constituting the text sentence selected in the selection step, those having the maximum number of characters or syllables are targeted for replacement. Text sentence set selection method. A text sentence set selection program for corpus-based speech synthesis, which causes a computer to execute the text sentence set selection method for corpus-based speech synthesis according to any one of claims 1 to 3. An apparatus for selecting a text sentence set for corpus-based speech synthesis from a text corpus,
Storage means for storing a thesaurus dictionary holding synonyms in the target category of speech synthesis;
First, an initialization part that makes the working text sentence set empty,
A selection unit that selects an arbitrary sentence from the text corpus and selects a text sentence (hereinafter referred to as a selected sentence) that maximizes the corpus quality when the sentence is added to the working text sentence set;
A thesaurus replacement unit that creates a similar sentence group by replacing a phrase constituting the selected sentence with a synonym held by the thesaurus dictionary;
A re-determination unit that selects an arbitrary sentence from the selected sentence and a group of similar sentences and obtains a text sentence (hereinafter referred to as the best sentence) that maximizes the corpus quality when the sentence is added to the working text sentence set. When,
A sentence adding unit for adding the best sentence obtained by the re-determination unit to a working text sentence set;
A design condition determination unit that repeats the processing of the selection unit, the thesaurus replacement unit, the re-determination unit, and the sentence addition unit until the corpus quality of the added work text sentence set satisfies a corpus design condition. A text sentence set selection device for featured corpus-based speech synthesis. The thesaurus replacement part is:
The text sentence set selection device for corpus-based speech synthesis according to claim 5, wherein only a noun word or phrase is a replacement target among words or phrases constituting the text sentence selected by the selection unit. The thesaurus replacement part is:
7. The corpus-based speech synthesis for corpus-based speech synthesis according to claim 5, wherein, among the words and phrases constituting the text sentence selected by the selection unit, those having the maximum number of characters or syllables are to be replaced. Text sentence set selection device.

Images