JP4441380B2 - Speech synthesis apparatus, speech synthesis method, and speech synthesis program - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To actually improve quality of a synthesis speech generated from an input text. <P>SOLUTION: A text classification part 40 assigns the input text to a prescribed category, and a sub-cost weighing determining part 50 determines sub-cost weighing corresponding to the category. In addition, a phoneme rhythm extractor 60 extracts reading information and rhythm information of the input text, and selects a group of the reading information and the rhythm information belonging to the similar range of the group from a speech database memory 20 by regarding the group as a key. Next, a sub-cost calculator 80 uses the group of the reading information and the rhythm information of the input text and a group of selected reading information and rhythm information to calculate a sub-cost value. A total cost calculator 90 calculates a total cost value weighing and integrating it by sub-cost weighing. An elementary speech unit selector 100 selects elementary speech unit minimizing the total cost value, and an elementary speech unit connector 110 connects them to generate synthesis speech. <P>COPYRIGHT: (C)2006,JPO&amp;NCIPI

Description

  The present invention relates to a speech synthesizer that generates synthesized speech from text content, a speech synthesis method, and a speech synthesis program.

In recent years, with the reduction in the cost of using large-capacity storage devices, tens of minutes or more of large-capacity speech data is stored as it is in a large-capacity storage device, and speech segments are stored according to input text and prosodic information. A waveform-connected corpus-based speech synthesis method that synthesizes high-quality speech by appropriately selecting, connecting, and transforming has been proposed (see, for example, Patent Document 1 and Non-Patent Document 1).
In this method, first, a speech unit partially or completely matching a phoneme sequence corresponding to a character string to be synthesized is searched from a speech database using a speech unit dictionary composed of a binary tree or the like. Next, according to an evaluation scale based on a plurality of parameters for evaluating the similarity between speech units, the retrieved speech units are costed. Then, by using a method such as DP (Dynamic Programming), speech synthesis is performed by selecting an appropriate combination of speech units from these speech units and connecting the selected speech units in order. (For example, refer nonpatent literature 2).

In principle, it has become possible to generate high-quality synthesized speech equivalent to the real voice. However, high-quality synthesized speech can be generated with this method on the premise that an appropriate speech unit exists in the speech database, and in the first place, an appropriate speech unit is created in the speech database. If it does not exist, high-quality synthesized speech cannot be generated. Therefore, in order to synthesize various texts with high quality, it is essential to use a speech database including a variety of speech segment variations. Therefore, in recent years, in order to increase the variation of speech units and to improve the quality of synthesized speech, development has progressed in the direction of further increasing the capacity of the speech database.
Japanese Patent No. 2761552 M. Beutnagel, A. Conkie, J. Schoroeter, Y. Stylianou, and A. Sydral, "Chose the best to modify the least: A new generation concatenative synthesis system", in Proc. Eurospeech'99, 1999, pp. 2291 -2294 Waveform selection method in waveform editing type rule synthesis method, Hirokawa et al., IEICE Technical Report, SP89-114, pp.33-40 (1990)

However, the technique for improving the quality of synthesized speech by increasing the capacity of the speech database of speech units has a problem that, in practice, high-quality synthesized speech may not be generated.
In other words, although the quality of synthesized speech has been improved to some extent by increasing the volume of the speech database, the amount of speech information that can be recorded in the speech database is limited due to time and cost issues. . For this reason, it is virtually impossible to record all speech segments that can be assumed as Japanese in the speech database.

  In addition, in order to record the voice in the voice database, it is necessary to prepare a sentence (recording text) to be read out in advance, but here too, it is impossible to collect all the text that can be assumed as Japanese. For this reason, text recorded in the voice database is read to a certain extent, and as a result, the contents of the voice recorded in the voice database are biased. Specifically, for example, it is common for speech such as everyday conversation, but it is difficult to collect text in the field where the text that transcribes it usually does not exist, medical and legal fields etc. If there are a lot of domain-specific technical terms, such as the technical field of, it is difficult to collect technical terms of all fields and texts containing them. As a result, the speech recorded in the resulting speech database is mainly read out from the recording text generated based on the texts of fields that can be collected easily and in large quantities, such as newspapers and novels. For this reason, in a field where the basic text cannot be collected sufficiently, synthesized speech must be created using speech segments that are read out and recorded in text from other fields.

  Under the circumstances as described above, in Non-Patent Document 2, as a method of selecting an appropriate speech segment, a physical scale combining a single or a plurality of physical parameters, and a prosody such as pitch and power. An evaluation scale combining scales is set, and a speech unit is selected from the speech database based on those scales. Specifically, first, a sub-cost function for calculating the similarity between speeches from physical parameters is prepared for each physical parameter of speech. Next, the physical parameters calculated from the input text to be generated synthesized speech and the physical parameters of the speech unit based on the speech database are substituted into this sub-cost function, and the input text, speech unit, Find the subcost of. This sub cost is obtained for each physical parameter. Then, these sub-costs are weighted based on a priori knowledge, and they are combined to obtain the total cost of the input text and the speech unit, and whether or not the speech unit is appropriate based on this total cost. Judging.

However, it is usually difficult to weight this sub-cost with an appropriate balance. This is because the frequency of words used differs depending on the type of text, and the sub-cost weighting for obtaining the best synthesized speech also differs. For this reason, conventionally, voices are synthesized using various texts, and the sub-cost weighting is often tuned so that the quality of the synthesized speech is improved on average.
As a result, the weighting determined thereby is suitable for synthesizing texts in general fields such as those recorded in the average text or speech database. Therefore, when synthesizing text in a specialized field or a field that is not recorded in the speech database, it is not possible to select the optimal speech segment from the speech database, resulting in higher quality synthesized speech. However, there is a problem that the quality of the synthesized speech is low, although there is a possibility that it can be generated.

  The present invention has been made in view of these points, and an object of the present invention is to provide a technology capable of realistically improving the quality of synthesized speech generated from input text.

  In the present invention, in order to solve the above problem, the text is assigned to at least one of predetermined categories based on the contents of the input text, and the sub-cost corresponding to the category to which the text is assigned. Determine the weight. Further, the reading information and prosodic information of the input text are extracted, and from the speech database associating the speech unit, the reading information and the prosodic information with the set of the input text reading information and the prosodic information as a key, A set of reading information and prosodic information belonging to the similar range of these sets is selected. Next, using a set of reading information and prosodic information of the input text and a reading information and prosodic information set selected by the search means, a sub-cost value indicating the similarity for each phoneme and prosodic element is calculated. Then, the total cost value obtained by integrating the sub cost values by weighting the sub cost weights is calculated. Then, a speech unit that minimizes the total cost value is selected, and the selected speech unit is connected to generate a synthesized speech.

  Here, the sub cost weight corresponding to the input text category is determined, and the total cost value is calculated based on the sub cost weight. Therefore, the total cost can be calculated using the optimum sub cost weight for each category. As a result, it is possible to select a speech unit that is more suitable than the case where the total cost is calculated using sub-cost weights that are averagely suitable for all categories.

  As described above, according to the present invention, the total cost is calculated using the sub-cost weight corresponding to the input text category, and the speech segment is selected, so that the quality of the synthesized speech can be improved practically. Is possible.

Hereinafter, embodiments of the present invention will be described with reference to the drawings.
<Configuration>
FIG. 1 is an illustration of a conceptual configuration diagram of a speech synthesizer 1 in the present embodiment.
As illustrated in this figure, the speech synthesizer 1 includes an input unit 5, a text memory 10, a speech database memory 20, a memory 30, a text classification unit 40, a sub-cost weight determination unit 50, a phonological prosody extraction unit 60, and a search unit 70. , A sub cost calculation unit 80, a total cost calculation unit 90, a speech unit selection unit 100, a speech unit connection unit 110, and a control unit 120.

Here, the memory 30 includes a category information storage area 31, a sub cost correspondence table storage area 32, a sub cost weight information storage area 33, a phoneme prosody storage area 34, a search result storage area 35, a sub cost storage area 36, an overall cost storage area 37, and A selected speech segment storage area 38 is provided. The phonological prosody extraction unit 60 includes a text analysis unit 61 and a prosodic physical parameter extraction unit 62.
Note that the speech synthesizer 1 according to the present embodiment causes a known computer including a CPU (Central Processing Unit), a RAM (Random Access Memory), a hard disk device, and the like to execute a predetermined program (speech synthesis program). It is comprised by.

<Pretreatment>
Next, preprocessing in the speech synthesis method of this embodiment will be described.
[Audio database]
First, the speech database memory 20 stores a speech database having speech segments necessary for speech synthesis.
FIG. 2 shows an example of the data structure of the voice database 200 in this embodiment.
As illustrated in FIG. 2, the speech database 200 of this example includes a phoneme string 221 (corresponding to “reading information”), prosodic information 222 and speech segment data (digital signal data, corresponding to “speech segment”). Are associated with each other.

The phoneme string 221 in this example is a string of Roman characters (phonemes) when the text is written in Roman letters. In this example, “A”, “Ai”, “Au”, “I”, and the like are stored as phoneme strings.
The prosodic information 222 includes a prosodic index 222a and a prosodic physical parameter 222b. Here, the prosody index 222a is data indicating a rough index of the prosody of the corresponding speech segment. In this example, the front phoneme environment 222aa, the rear phoneme environment 222ab, and the accent 222ac correspond to the prosody index 222a. Further, the previous phoneme environment 222aa in this example is data indicating the previous phoneme in time series with respect to the corresponding phoneme string 221, and in this example, “#” indicating a pause (silent state) is set. . Further, the postphoneme environment 222ab in this example is data indicating a phoneme subsequent to the corresponding phoneme string 221 in time series. In this example, “#” indicating a pause (silent state) or “ S, “G”, “N”, “T”, “R” and the like are set. Further, the accent 222ac in this example is data indicating whether the corresponding phoneme string 221 has an accent, and if so, where the accent is. In this example, “0” indicates that the corresponding phoneme string 221 has no accent, “1” has an accent on the first note of the corresponding phoneme string 221, and “2” has an accent on the second note. It is shown that.

The prosodic physical parameter 222b is data indicating the physical quantity of the prosody of the corresponding speech unit. In this example, the average F0 (reference frequency) (Hz) 222ba, the slope of F0 (Hz / ms) 222bb, time The length (ms) 222bc and the power (dB) 222bd correspond to the prosodic physical parameter 222b.
[Sub cost correspondence table]
In this embodiment, the sub cost correspondence table is stored in the sub cost correspondence table storage area 32 of the memory 30.

FIG. 3 is an example of the data configuration of the sub-cost correspondence table 300 in the present embodiment.
The sub-cost correspondence table 300 in this example is a table in which text categories and sub-costs are associated with each other. The sub-cost correspondence table 300 in the example of FIG. 3 is a table in which sub-cost weights are associated with N categories 1 to N and M sub-costs A to M.
In this example, “Category 1” is associated with sub-cost weights “W _1A ”, “W _1B ”,... “W _1M ” of “Sub Cost A”, “Sub Cost B”,. Are associated with sub-cost weights “W _2A ”, “W _2B ”, “W _2M ” of “sub-cost C”, and “sub-cost A” “sub-cost” for “category N” Sub-cost weights “W _NA ”, “W _NB ”,... “W _NM ” of “B”.

  Here, “category” is information indicating the field and type of text. This category is determined in advance, but the category may be a category or category classified based on a single classification criterion, or a category or category based on multiple classification criteria. May be classified. Specifically, for example, categories such as “dictionary”, “magazine”, and “book” may be set based on a single classification standard “medium type”. For example, “medium type” “field” The categories “dictionary”, “magazine”, “book” and “natural science”, “literature”, and “law” may be set multidimensionally based on the two classification criteria.

  In addition, the “sub cost” in this example is a value indicating the similarity of each element of reading information and phonological information between two sounds. That is, a sub cost indicating the similarity of reading information, a sub cost indicating a similarity of average F0 that is phonemic information, a sub cost indicating a similarity of accent, and the like can be exemplified. Note that in the speech synthesis processing of this embodiment, the sub-cost for only reading information may be used, the sub-cost for only phonemic information may be used, or only the sub-cost related to a part of phonemic information may be used. However, higher quality synthesized speech can be generated using many types of sub-costs.

Furthermore, the “sub cost weight” means a weighting constant for each sub cost used when calculating the total cost (cost indicating the similarity between two voices) from a plurality of sub costs. For example, sub-cost A, B, sub-cost weight for C is W _A, W _B, if a W _C, W _A · A and W _B · B and total cost by the W _C · C is determined.
[Method for generating sub-cost correspondence table]
The categories of the sub-cost correspondence table 300 are, for example, a predetermined text is converted into a preliminary category set based on a single classification criterion or a plurality of classification criteria by a procedure (described later) executed by the text classification unit 40. When the assignment and the text are assigned to a plurality of the preliminary categories, the determination is made by a process of integrating the plurality of preliminary categories into one category.

  That is, based on a priori knowledge about the spoken language, a category in which text can be classified is preliminarily determined based on a single classification standard. Next, the recording list (text) read out when recording the speech segment data 223 included in the speech database 200 is classified into these preliminary categories. As a result, when the same text is classified into a single preliminary category, this preliminary category is determined as a category of the sub-cost correspondence table 300, and the same text is duplicated into a plurality of preliminary categories. When classified, a plurality of preliminary categories into which the text is classified are newly generated as one category. For example, when 10 categories from A to J are considered as preliminary categories, the recording list (text) is a single category B, D, G, category A and category C, and category B and category D. It is classified as a multiple force category. In this case, categories B, D, and G are newly defined as single categories, multiple categories of categories A and C are defined as category K, and multiple power categories of categories B and D are newly defined as category L. In practice, this includes text with category A and category C features (or category A and category C were originally one category), and category B and category D features. Since the prepared texts exist, categories K and L of these texts are newly set. Further, for example, categories C, E, F, H, T, and J in which the recording list (text) is not classified may be integrated to newly provide category M. In reality, there is no text that exists in categories C, E, F, H, T, and J, and these categories are integrated into category M. In this case, as a result, there are seven categories of A, B, D, G, K, L, and M. In this way, it is possible to set a category suitable for the speech database actually used in speech synthesis, rather than simply categorizing categories based on the type of foresight text.

Moreover, the sub cost weight of the sub cost correspondence table 300 is determined as follows, for example.
That is, after determining a category as described above, a large number of text corpora are classified according to a newly set category (described later), a plurality of texts are extracted for each category, and first, those sub-cost weights are used as initial values. Create and listen to synthesized speech corresponding to the text to check the quality. After that, an appropriate subcost is determined by repeating the creation and listening of the synthesized sound while appropriately adjusting the subcost value for each category, and the subcost weight of each subcost corresponding to each category is determined. Thus, the sub cost correspondence table 300 is created.

[Enter text]
A text for generating a synthesized speech is input to the input unit 5, and the input text is stored in the text memory 10.
<Speech synthesis processing>
Next, speech synthesis processing in this embodiment will be described. Note that the following processing is performed under the control of the control unit 120.
[Process overview]
First, in the text classification unit 40 (FIG. 1), the input text is classified into at least one category. Next, the sub cost weight determination unit 50 determines a sub cost weight coefficient based on the classified category. Further, the text analysis unit 61 of the phonological prosody extraction unit 60 acquires reading information and prosodic indices from the input text, and the prosodic physical parameter extraction unit 62 calculates prosodic physical parameters from the prosodic indices. Next, after the speech unit corresponding to the reading information and the prosodic index is searched by the search unit 70, the reading information and the prosodic physics extracted by the phoneme prosody extraction unit 60 by the sub cost calculation unit 80 and the total cost calculation unit 90 are searched. The total cost is calculated using the sub-cost function and the sub-cost weight coefficient from the parameters and the reading information and prosodic physical parameters corresponding to the speech segment. Finally, the speech unit selection unit 100 and the speech unit connection unit 110 synthesize speech by selecting and connecting speech units based on the calculated total cost value.

[Details of processing]
FIG. 4 is a flowchart for explaining speech synthesis processing in this embodiment.
Hereinafter, the details of the speech synthesis processing of this embodiment will be described with reference to FIG.
First, the text classification unit 40 reads text (corresponding to “input text”) from the text memory 10 (step S1), and based on this content, the text is classified into at least one of predetermined categories. Assigned to one category (step S2). The “predetermined category” is a category of the sub-cost correspondence table 300 described above. Various methods have been proposed to classify this text into predetermined categories. For example, Joachims, T. "Text Categorization with Support Vector Machines: Learning with Many Relevant Features", Proc. Of 10th European Conference on Machine Learning (ECML-98), pp.137-142 (1998), Japanese Patent Application No. 11-191064 “Text Classification Learning Method and Device, and Storage Medium Stored Text Classification Learning Program” ”And Japanese Patent Application No. 2002-204434“ Method and apparatus for extracting multiple topics of text, program for extracting a large amount of topics of text, and recording medium on which the program is recorded ”, and the like. .

When this category is classified based on a single classification standard (see <Preprocessing>), the text classification unit 40 assigns the input text to any one category. On the other hand, when this category is classified based on a plurality of classification criteria, the text classification unit 40 may assign the input text to a plurality of categories. Then, the category information specifying the assigned category as described above is sent to the memory 30 and stored in the category information storage area 31.
Next, the sub-cost weight determination unit 50 reads the category information from the category information storage area 31, refers to the sub-cost correspondence table 300 (FIG. 3) of the sub-cost correspondence table storage area 32 based on this category information, and assigns text. A sub-cost weight corresponding to the given category is determined (step S3).

The method of determining the sub-cost weight differs depending on whether the input text is assigned to only one category or a plurality of categories.
[When the entered text is assigned to only one category]
In this case, the sub cost weight corresponding to one assigned category is used as it is, and the sub cost weight information indicating the sub cost weight is stored in the sub cost weight information storage area 33. For example, when “category 1” in FIG. 3 is assigned, sub-cost weight “W” of “sub-cost A” “sub-cost B”... “Sub-cost C” associated with “category 1” in the sub-cost correspondence table 300 is assigned. “ _1A ”, “W _1B ”... “W _1M ” are specified, and these pieces of information are stored in the sub-cost weight information storage area 33.

[When the entered text is assigned to multiple categories]
In this case, the preliminary sub cost weights corresponding to the plurality of categories are respectively determined, and the preliminary sub cost weights are weighted based on the similarity (including likelihood) between the input text and each category, and totaled. (Weighted sum of sub-cost weights), and this total value is determined as a sub-cost weight corresponding to the category.

以上のように決定されたサブコスト重みを示すサブコスト重み情報は、メモリ３０に送られ、そのサブコスト重み情報格納領域３３に格納される。 For example, when the input text is duplicated and classified into N categories, the sub cost weight W is determined as follows. Here, sub-cost weights _{W ci} = category i is _{(w1 ci, w2 ci, w3} ci, ..., wm ci), m the sub-cost number, sub-cost weights in subcost j to Wj, likelihood corresponding to Pi to category i Degree or similarity.

The sub cost weight information indicating the sub cost weight determined as described above is sent to the memory 30 and stored in the sub cost weight information storage area 33.

  Next, the phonological prosody extraction unit 60 reads the same text (corresponding to “input text”) read from the text memory 10 in step S1 (step S4), and extracts the reading information and prosodic information. Stored in the phoneme prosody storage area 34 of the memory 30. In the case of this example, first, the text analysis unit 61 of the phonological prosody extraction unit 60 performs text analysis processing of this text, extracts reading information and prosodic indices, and stores them in the phonological prosody storage area 34 (steps). S5). The text analysis process in this example mainly includes a morphological analysis process and a reading / accenting process. However, there are various conventional methods for these processing methods. For example, Japanese Patent No. 3337943 “Morphological Analysis Method and Morphological Analysis”. The processing can also be performed based on a method such as “a recording medium on which an analysis program is recorded” or Japanese Patent No. 3518340 “a recording medium on which a reading prosodic information setting method and apparatus and a reading prosodic information setting program are stored”.

  Next, the prosodic physical parameter extraction unit 62 reads out the prosodic indices extracted in step S5 from the phonological prosody storage area 34 of the memory 30, obtains the prosodic physical parameters based on the prosodic indices, and stores them in the phonological prosody storage area 34. Store (step S6). Here, the prosodic physical parameters include pitch (fundamental frequency), phoneme duration, and the like. Conventionally, there are methods for obtaining them. For example, the pitch (fundamental frequency) can be obtained by the method of Japanese Patent No. 3240691 “Pitch Pattern Generation Method, Apparatus and Program Recording Medium” and Japanese Patent No. 3344487 “Fundamental Audio Frequency Pattern Generation Device”. Also, for example, Kaiki et al., “Control of vowel duration using linguistic information” vol. 75, No. 3 pp. 467-473, Shinsei theory, MD Riley. “Tree-based modeling for speech synthesis. . "In G. Bailly, C. Benoit, and TR Sawallis, editors, Talking Machine: Theories, Models, and Designs, pages 265-273. Elsevier, 1992.

  Next, the search unit 70 reads the reading information and the prosodic index (stored in step S5 and corresponds to “a set of input text reading information and prosodic information”) from the phonological prosody storage area 34 of the memory 30. The speech database 200 (FIG. 2) in the speech database memory 20 is searched using this set as a key, and the phoneme string 221 (reading information) and prosody information 222 (prosodic index) belonging to the similar range of these sets are searched from the speech database. 222a + prosodic physical parameters 222b) and speech segment data 223 associated therewith are selected / extracted. The search results of the extracted speech segment data 223, phoneme string 221 (reading information), and prosody information 222 (prosodic index 222a + prosodic physical parameter 222b) are stored in the search result storage area 35 of the memory 30 (step). S7).

  Note that the “similarity range” here is a concept including, for example, a case where reading information and prosodic information are completely matched, a portion of which is matched, and a high similarity specified by cost. For example, if the reading information is “A”, the speech unit pronounced as “A”, that is, the speech unit data 223 in which the phoneme string 221 of the speech database 200 is “A”, and the prosodic index are also used. Is searched for as a similar range, that is, the speech unit data 223 in which the phoneme sequence 221 of the speech database 200 is “A” and the accent 222ac is “1”. Become. In addition, the search result extracted in this step is not limited to one. Extraction of all speech segment data 223, phoneme string 221 (reading information) and prosodic information 222 (prosodic index 222a + prosodic physical parameter 222b) satisfying the conditions is performed. • Storage is performed.

  Next, the sub-cost calculator 80 reads the input text reading information (extracted in step S5), the prosodic physical parameters (extracted in step S6), and ("input text reading information) from the phoneme prosody storage area 34 of the memory 30. And the prosody information set ”), and from the search result storage area 35, the reading information and the prosodic physical parameter 222b (searched in step S7 and“ the set of the reading information and the prosody information selected by the search unit 70 ”). Is equivalent). Then, the sub cost calculation unit 80 calculates a sub cost indicating the similarity for each phoneme and prosodic element using these, and stores the calculated sub cost in association with the corresponding speech element data in the sub cost storage area 36 ( Step S8).

This sub-cost can be calculated using, for example, a sub-cost function as follows ("Waveform Selection Method Considering Spectral Continuity in Waveform Editing Type Synthesis Method", Proceedings of the Acoustical Society of Japan, 2 -6-10, pp. 239-240, 1990/9).
[Sub-cost function]
The sub-cost function of this example is shown below.
(1) sub-cost function C ₁ corresponding to the reading information (n) = 1 / ^{e n}
However, the number of phonemes that coincides between the phoneme sequence as the reading information of the input text and the phoneme sequence as the reading information of the speech unit (phoneme sequence associated with the speech unit data) is n.

(2) Sub-cost function for average pitch C ₂ (Vp, Vs) = | Vp−Vs | ²
However, let Vp be the average pitch of the phoneme physical parameters extracted from the input text, and let Vs be the average pitch of the speech segments (average F0 associated with the speech segment data).
(3) Sub-cost function for pitch inclination C ₃ (Fp, Fs) = | Fp−Fs | ²
However, the pitch inclination of the phoneme physical parameter extracted from the input text is Fp, and the pitch inclination of the speech segment (the slope of F0 associated with the speech segment data) is Fs.

(4) Sub cost function for time length C ₄ (Tp, Ts) = | Tp−Ts | ²
However, the time length of the phoneme physical parameter extracted from the input text is Tp, and the time length of the speech unit (the time length associated with the speech unit data) is Ts.
(5) Sub cost function for amplitude C ₅ (Ap, As) = | Ap−As | ²
However, the amplitude of the phoneme physical parameter extracted from the input text is Ap, and the time length of the speech unit (power associated with the speech unit data) is As.

Each sub cost value can be obtained by substituting each information read in step S8 into the above sub cost function. Note that the processing in step S8 is performed for each data set extracted in step S7, and the sub cost calculated for each is stored in association with the corresponding speech segment data.
Next, the total cost calculation unit 90 reads the sub cost weight information calculated in step S3 from the sub cost weight information storage area 33 of the memory 30, reads each sub cost calculated in step S8 from the sub cost storage area 36, The total cost value obtained by integrating the weighted sub cost values with the sub cost weight is calculated and stored in the total cost storage area 37 of the memory 30 (step S9).

The calculation of the total cost in this example is performed for each speech unit. For example, when the sub cost function is the above (1) to (5), the total cost is calculated as follows.
[Total cost]
(6) Ω = ω ₂ · C ₂ + ω ₃ · C ₃ + ω ₄ · C ₄ + ω ₅ · C ₅ is calculated.
(7) Calculate ω ₁ · C ₁ + (1−ω ₁ ) · Ω.
(8) The total cost Pnew = (1 + G) · P is calculated.
C ₁ , C ₂ , C ₃ , C ₄ , and C ₅ are sub-cost functions C ₁ (n), C ₂ (Vp, Vs), C ₃ (Fp, Fs), and C ₄ (Tp) in step S8. , Ts) and C ₅ (Ap, As) are sub-costs calculated by substituting n, Vp, Vs, Fp, Fs, Tp, Ts, Ap, As, respectively. Further, ω ₁ , ω ₂ , ω ₃ , ω ₄ , and ω ₅ indicate sub-cost weights of the sub-costs C ₁ , C ₂ , C ₃ , C ₄ , and C ₅ . Further, G represents an acoustic constant.

This step is performed for all data sets calculated in step S8, and the calculated total costs are stored in the total cost storage area 37 of the memory 30 in association with the corresponding speech segment data. Is done.
Next, the speech element selection unit 100 extracts the total cost stored in step S9 from the total cost storage area 37 of the memory 30, and obtains the minimum value thereof. The calculation of the minimum value can be easily realized by using a general DP method. Then, the speech unit selection unit 100 extracts speech unit data (corresponding to “speech unit that minimizes the total cost value”) associated with the determined minimum total cost from the total cost storage area 37. Read (corresponding to “select”) and store it in the selected speech segment storage area 38 of the memory 30 (step S10).

Next, the control unit 120 determines whether or not all speech segment data for one text stored in the text memory 10 has been selected (step S11). Here, if all the speech unit data has not been selected, the process returns to step S5, and if all the speech unit data has been selected, the process proceeds to step S12.
In step S12, the speech unit connection unit 110 reads each speech unit data (corresponding to “speech unit selected by the speech unit selection unit”) from the selected speech unit storage area 38 of the memory 30, and these are read out. By connecting in order, a synthesized speech is generated and output (step S12).

Note that the selected speech unit data may be simply connected in temporal order, but it is also easy to interpolate between different speech unit data in terms of time or frequency (Japanese Patent Application No. 5-217337). "Speech synthesis method and apparatus"). In addition, after processing speech unit data selected based on prosodic physical parameters, these may be connected (Y. Stylianou, “Concatenative Speech Synthesis using a Harmonic plus Noise Model.” In : The 3rd ESCA / COCOSDA Workshop on Speech Synthesis, Jenolan Caves, NSW, Australia, NOV. 1998, Paper H.1.)
<Features of this embodiment>
As described above, in this embodiment, the input text is classified into at least one of the text categories determined in advance according to the contents, and the sub-cost weight which is a weight for the sub-cost function based on the category. As a result, the calculation method of the total cost, which is the basis for selecting speech segments, was changed. Thereby, it is possible to generate an optimal synthesized speech corresponding to the content of the text.

  That is, as described above, in this embodiment, various categories are assumed in advance, a plurality of categories are set according to the contents, and several texts representing the categories are collected for each classified category. Then, by synthesizing those texts and adjusting the sub-cost weight, which is a weight for the sub-cost function, so that the quality of the synthesized speech is best, the optimum sub-cost weight is found for each classified category. . Then, the total cost is calculated using the optimum sub-cost weight for the text category, and a speech segment is selected based on the total cost. As a result, speech unit selection based on the same total cost calculation method is conventionally performed for all input texts, and as a result, it is high in fields other than general texts that exist in many speech databases. It is possible to solve the problem that it is not possible to generate synthesized speech of sound quality.

<Hardware configuration>
FIG. 5 is a block diagram when the speech synthesizer 400 according to the present embodiment is realized by a Neumann computer.
As illustrated in this figure, the speech synthesizer 400 of this example includes an input unit 410, an output unit 420, a CPU 430, a RAM 440, a ROM 450, a hard disk device 460, and a bus 470 that connects these components so as to communicate with each other.
The input unit 410 is, for example, an interface such as a USB (Universal Serial Bus) that accepts input of text data or the like, or an input device such as a keyboard, a mouse, and a touch panel. The output unit 420 is an audio output device such as an interface such as a USB or a speaker that outputs the generated synthesized audio data.

The hard disk device 460 stores an OS (Operating System) program 461 such as Microsoft Windows (registered trademark), a speech synthesis program 462, a speech database 463, a sub-cost correspondence table 464, and text information 465.
Here, the speech synthesis program 462 is application software for executing the processing of FIG. 4 described above. The speech database 463 is a database similar to the speech database 200 of FIG. 3 is the same data as the sub-cost correspondence table 300 in FIG. Further, the text information 465 is text in which synthesized speech input from the input unit 410 is generated.

  When the speech synthesizer 400 executes speech synthesis processing, the CPU 430 first loads the OS program 461, speech synthesis program 462, speech database 463, sub cost correspondence table 464, and text information 465 of the hard disk device 460 into the RAM 440 once. The CPU 430 executes the speech synthesis program 462 after executing the OS program 461 read from the RAM 440, and performs each process of FIG. In this case, the RAM 440 corresponds to the text memory 10, the voice database memory 20, and the memory 30. In addition, the CPU 430 may directly read at least a part of the OS program 461, the speech synthesis program 462, the speech database 463, the sub cost correspondence table 464, and the text information 465 of the hard disk device 460 from the hard disk device 460 and perform processing. Needless to say.

<Speech synthesis program>
Also, the speech synthesis program 462 describing the above-described processing contents can be recorded on a computer-readable recording medium. The computer-readable recording medium may be any medium such as a magnetic recording device, an optical disk, a magneto-optical recording medium, or a semiconductor memory. Specifically, for example, the magnetic recording device may be a hard disk device or a flexible Discs, magnetic tapes, etc. as optical disks, DVD (Digital Versatile Disc), DVD-RAM (Random Access Memory), CD-ROM (Compact Disc Read Only Memory), CD-R (Recordable) / RW (ReWritable), etc. As the magneto-optical recording medium, MO (Magneto-Optical disc) or the like can be used, and as the semiconductor memory, EEP-ROM (Electronically Erasable and Programmable-Read Only Memory) or the like 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.
For example, a computer (speech synthesizer 400) that executes such a program first stores a program recorded on a portable recording medium or a program transferred from a server computer in the hard disk device 460, and performs processing in accordance with the program. As described above, as another execution form of the program, the computer may directly read the program from the portable recording medium and execute processing according to the program. Each time a program is transferred from the server computer, processing according to the received program may be executed sequentially. Also, the program is not transferred from the server computer to the computer, and the above-described processing is executed by a so-called ASP (Application Service Provider) type service that realizes the processing function only by the execution instruction and result acquisition. 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).

<Modifications>
The present invention is not limited to the embodiment described above. For example, in this embodiment, the phoneme index is used for searching the speech unit data in step S7, and the phoneme physical parameter is used for calculating the sub-cost in step S8. However, the phoneme physics is used for searching the speech unit data in step S7. A parameter (corresponding to “prosodic information”) may be used, and a phoneme index (corresponding to “prosodic information”) may be used for calculating the sub-cost in step S8.
In addition, the various processes described above are not only executed in time series according to the description, but may be executed in parallel or individually according to the processing capability of the apparatus that executes the processes or as necessary. Further, the configuration of the processing loop is not limited to that described.

In this embodiment, the present apparatus is configured by executing a predetermined program on a computer. However, at least a part of these processing contents may be realized by hardware.
Needless to say, other modifications are possible without departing from the spirit of the present invention.

  The field of application of the present invention includes the field of speech synthesis.

It is an illustration of the conceptual block diagram of the speech synthesizer in this form. It is an example of the data structure of the audio | voice database in this form. It is an example of the data structure of the sub cost correspondence table | surface in this form. It is a flowchart for demonstrating the speech synthesis process in this form. It is a block diagram in the case of implement | achieving the speech synthesizer 400 in this form with a Neumann computer.

Explanation of symbols

DESCRIPTION OF SYMBOLS 1,400 Speech synthesis apparatus 20 Speech database memory 40 Text classification part 50 Sub cost weight determination part 60 Phonological prosody extraction part 61 Text analysis part 62 Phonological physical parameter extraction part 70 Search part 80 Sub cost calculation part 90 Total cost calculation part 100 Speech unit Select part

Claims (7)

In a speech synthesizer that selects and synthesizes an appropriate speech segment from a speech database corresponding to the input text,
A speech database memory storing a speech database in which speech segments, their reading information and prosodic information are associated;
Text classification means for assigning the text to at least one of predetermined categories based on the contents of the input text;
Sub-cost weight determining means for determining a sub-cost weight corresponding to the category to which the text is assigned;
Phonological prosody extraction means for extracting the input text reading information and prosodic information;
Search means for searching the speech database using a set of input text reading information and prosodic information as a key, and selecting from the speech database a set of reading information and prosodic information belonging to a similar range of these sets;
Using the set of reading information and prosodic information of the input text and the reading information and prosodic information set selected by the search means, a sub-cost value indicating the similarity for each phoneme and prosodic element is calculated. Sub-cost calculation means;
A total cost calculating means for calculating a total cost value obtained by weighting and integrating the sub cost value with the sub cost weight;
Speech unit selection means for selecting the speech unit that minimizes the total cost value;
Speech unit connection means for connecting the speech units selected by the speech unit selection means;
A speech synthesizer characterized by comprising: The speech synthesizer according to claim 1,
The prosodic information of the speech database is
A prosodic index that is a prosodic index of the corresponding speech segment, and a prosodic physical parameter that is a physical quantity of the prosody,
The prosodic information extracted from the input text by the phonological prosody extracting means is:
Prosodic indices and prosodic physical parameters,
The search means includes
Searching the speech database using the input text reading information and prosodic index as a key, and selecting from the speech database a set of reading information and prosodic information belonging to a similar range of these sets,
The prosodic information used by the sub-cost calculating means for calculating the sub-cost value is:
Prosodic physics parameters,
A speech synthesizer characterized by the above. The speech synthesizer according to claim 1,
The predetermined category is:
A category classified based on a single classification criterion,
The text classification means includes:
Means for assigning the inputted text to any one of the categories;
The sub-cost weight determining means includes
Means for determining a sub-cost weight corresponding to one category to which the text is assigned;
A speech synthesizer characterized by the above. The speech synthesizer according to claim 1,
The predetermined category is:
A category classified based on multiple classification criteria,
The sub-cost weight determining means includes
If the entered text is assigned to only one category, determine a sub-cost weight corresponding to this one category;
When the input text is assigned to a plurality of categories, spare sub-cost weights corresponding to the plurality of categories are respectively determined, and the spare sub-cost weight is determined based on the similarity between the input text and each category. A means for weighting and summing the original values and determining the total value as a sub-cost weight corresponding to the category.
A speech synthesizer characterized by the above. The speech synthesizer according to claim 1,
The predetermined category is:
When a predetermined text is assigned to a preliminary category according to a procedure executed by the text classification means, and the text is assigned to a plurality of the preliminary categories, the plurality of preliminary categories are assigned to one Have categories generated by integrating into categories,
A speech synthesizer characterized by the above. In a speech synthesis method for selecting and synthesizing appropriate speech segments from a speech database in accordance with input text,
In the speech database memory, a speech database in which speech segments, their reading information and prosodic information are associated is stored,
A text classification means assigns the text to at least one of predetermined categories based on the content of the input text,
A sub cost weight determining means determines a sub cost weight corresponding to the category to which the text is assigned;
The phonological prosody extraction means extracts the input text reading information and prosody information,
A search means searches the speech database using a set of input reading information and prosodic information of the text as a key, and selects a set of reading information and prosodic information belonging to a similar range of these sets from the speech database. ,
The sub-cost calculating means uses the set of reading information and prosodic information of the input text and the set of reading information and prosodic information selected by the searching means, and calculates the similarity for each of these phonemes and prosodic elements. Calculate the sub-cost value shown,
Comprehensive cost calculation means calculates the integrated cost value obtained by weighting and integrating the sub cost value with the sub cost weight,
A speech segment selection means selects the speech segment that minimizes the total cost value;
A speech unit connection means connects the speech units selected by the speech unit selection means;
A speech synthesis method characterized by the above.   A speech synthesis program for causing a computer to function as the speech synthesizer according to claim 1.

Images