JP4326251B2 - Text-to-speech synthesizer, text-to-speech synthesis method and program thereof - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To solve a problem that although trends of corrections are known by using a correction history before to manually improve a subsequent meter generation rule, it is very intricate and technical to automatically improve a meter generation rule itself generated from massive conditions by analyzing a very large correction history and it is difficult for an end user to correct the meter generation rule itself according to his or her preference. <P>SOLUTION: A text speech synthesizer is equipped with a dynamic meter generation part 102 which holds correction contents and correction conditions of past corrections of a meter correction part 103 as correction history information in a correction history holding part 106 and generates meter information thereafter by referring to the held correction history information and automatically reflecting correction contents made to correspond to a correction condition when there is information meeting the correction condition. <P>COPYRIGHT: (C)2005,JPO&amp;NCIPI

Description

[0001]
BACKGROUND OF THE INVENTION
The present invention relates to a text-to-speech synthesizer. More specifically, the present invention relates to a text-to-speech synthesizer having a function of automatically generating correction information and generating prosodic information. The present invention also relates to a text-to-speech synthesis method including a method for automatically generating a revision history and generating prosodic information, and a program for realizing the text-to-speech synthesis method.
[0002]
[Prior art]
In conventional speech synthesis, synthesized speech is generated with a predetermined pitch pattern by connecting phonemes constituting speech in accordance with a designated pitch pattern (prosodic information). Conventionally, a pitch pattern for text-to-speech synthesis has been generated according to a predetermined prosody generation rule based on an analysis result of an input text.
[0003]
The generated prosody may include errors at the time of text analysis or prosody generation. In this case, a synthesized speech different from the intention of the synthesized speech creator (hereinafter referred to as the author) is generated. . In order to correct such prosodic errors and to adjust to the prosody preferred by the author, the engineer rewrites the contents of the prosodic file in which the prosodic information is described, and performs specialized operations based on experience. It was necessary to directly change the parameters (pitch pattern, power, etc.) for determining the prosody.
[0004]
As a method for easily correcting the pitch pattern, the contents of the prosody file for controlling the prosody are graphically displayed on the display, and the displayed parameter (pitch pattern, power, etc.) pattern is changed with the mouse. Is known (see, for example, Patent Document 1). In this prosody modification method, the processed modification is stored in the storage device, and when performing the second speech synthesis using the same prosody file, the current pattern and the modification history pattern are displayed. As a result, the effort of correction is reduced, and the tendency of correction is grasped, so that information on subsequent prosody creation rules is obtained.
[0005]
The prosody modification procedure described in Patent Document 1 will be briefly described with reference to FIG. The language processing unit 401 extracts language-related information such as reading information, part-of-speech information, and dependency information from the input text. After that, the prosody generation unit 402 generates prosody files 409 serving as basic information for speech synthesis using the language related information. In the prosody modification unit 403, the content of the generated prosody file 409 is displayed on the screen by the graph file display generation unit 412, and the prosody file 409 is modified by modifying the graph file displayed on the screen. A portion for performing this correction work is a pattern correction unit 413.
[0006]
The value of each parameter corrected by the pattern correction unit 413 is recorded in the correction history DB (correction history database) 406 as correction history data. Therefore, when the same prosody file 409 is corrected again, the past correction history is also displayed on the screen by the graph file display generation unit 412, so that the equivalent prosody can be easily corrected. Further, since the correction tendency can be obtained by analyzing the correction history, it can be treated as one of information in the subsequent prosody generation.
[0007]
Based on the modified prosody file 419 created in this way, a segment selection unit 414 selects a segment for synthesis, and the speech synthesis unit 424 transforms and connects the segment according to the prosody file. Create synthesized speech.
[0008]
[Patent Document 1]
JP-A-5-232980 above
[0009]
[Problems to be solved by the invention]
In the conventional method shown in FIG. 18, every time speech synthesis is performed, the same correction is manually made to the prosody file 409 to be created, and the correction history is also held for each prosody file 409. Therefore, when synthesizing the same sentence, it is possible to refer to the past correction history, but when synthesizing different sentences, the same correction must be made without referring to the past correction history. It was.
[0010]
In addition, the prosody file 409 is panned to know the correction history correction tendency, and the tendency is used as one piece of information for the subsequent improvement of the prosody generation rules. Analyzing an enormous revision history for a rule and improving the prosody generation rule itself is a task with extremely high level of technical knowledge rather than individually modifying a prosody file. In other words, even if the general tendency can be grasped, it is difficult for the end user to easily modify the prosody generation rule itself according to his / her preference.
[0011]
The present invention has been made in view of the above. The purpose of the present invention is to make it possible for the end user to modify the prosody generation rules themselves directly without changing the prosodic generation rules themselves and without correcting a large amount of prosodic information every time speech synthesis is performed. Is to provide a method for easily performing speech synthesis suitable for the above, a text-to-speech synthesizer to which the method is applied, and a program for realizing the method.
[0012]
[Means for Solving the Problems]
In order to solve the above-mentioned problem, the present invention makes a pair of correction contents and correction conditions when prosodic information is corrected by an end user, and holds it as correction history data constituting correction history information. In the generation of prosodic information performed after the correction, if there is correction history data that matches the correction conditions, the correction details associated with the correction conditions are automatically reflected. This is a configuration having a function of generating the prosodic information.
[0013]
With the above configuration, it is not necessary to perform the same correction as each past correction registered in the correction history information, and the learning of the correction history information sequentially proceeds according to the correction made by the end user. Prosody information that suits the end user's preference can be easily generated.
[0014]
Specifically, the text-to-speech synthesizer according to the present invention performs language analysis on text data, extracts language information, a correction history holding unit that holds correction history information, and the correction history. A correction history management unit that manages information, a dynamic prosody generation unit that generates dynamic prosody information based on the language information with reference to the correction history information via the correction history management unit, and an external correction command Accordingly, the prosodic information is generated by correcting the dynamic prosodic information, and the correction history information is updated via the correction history management unit based on the correction according to the external correction command. A synthesis unit that generates a synthesized speech based on the correction unit, the language information, and the definite prosody information; The correction history management unit, a correction history extraction unit that extracts correction history information referred to by the dynamic prosody generation unit, a correction history update unit that updates the correction history information held in the correction history holding unit, The correction history holding unit includes a plurality of correction history databases having different prosodic styles, and the correction history management unit controls selection of the plurality of correction history databases according to a selection command. It further has a selection control means It is a configuration.
[0015]
The text-to-speech synthesis method according to the present invention includes a language analysis step for performing language analysis on text data and extracting language information, and correction history information. A revision history holding step for holding, a revision history management step for managing the revision history information, and the revision history information with reference to the revision history information via the revision history management unit, to generate dynamic prosodic information based on the language information A dynamic prosody generation step, generating the fixed prosody information by correcting the dynamic prosody information according to the external correction command, and the correction history management unit based on the correction according to the external correction command The prosody modification step for updating the modification history information via the above, the synthesized speech generation step for generating synthesized speech based on the language information and the definite prosody information, and the modification history management step, the dynamic prosody generation step The revision history extraction step for extracting revision history information referred to in the revision history update step for updating the revision history information held in the revision history holding step. The revision history holding step further holds a plurality of revision history databases having different prosodic styles, and the revision history management step stores the plurality of revision history databases in response to a selection command. A database selection control step for controlling the selection; .
[0016]
In addition, the text-to-speech synthesis program according to the present invention performs language analysis on text data. A language analysis step for extracting linguistic information, a correction history holding step for holding correction history information, a correction history management step for managing the correction history information, and the correction history information via the correction history management unit A dynamic prosody generation step for generating dynamic prosody information based on the language information, generating fixed prosody information by correcting the dynamic prosody information according to an external correction command, and A prosody modification step for updating the modification history information via the modification history management unit based on a modification according to the external modification command, and a synthesized speech for generating a synthesized speech based on the language information and the definitive prosody information A correction history extraction step in which the generation step, the correction history management step extracts correction history information referred to in the dynamic prosody generation step, and the correction history In the text-to-speech synthesis program for causing the computer to execute a correction history update step for updating the correction history information held in the holding step, the correction history holding step further holds a plurality of correction history databases having different prosodic styles. The correction history management step further includes a database selection control step for controlling selection of the plurality of correction history databases in response to a selection command. .
[0017]
DETAILED DESCRIPTION OF THE INVENTION
While describing the contents of the present invention, a preferred embodiment is described. Note that FIG. 1 and FIG. 2 are referred to as necessary. FIG. 1 is a block diagram conceptually showing the structure of a text-to-speech synthesizer according to the present invention. FIG. 2 is a block diagram conceptually showing the text-to-speech synthesis method according to the present invention.
[0018]
The text-to-speech synthesizer shown in FIG. 1 performs language analysis on text data and extracts language information, a correction history holding unit 106 that holds correction history information, and correction history information. A correction history management unit 105 that manages the dynamic prosody information, a dynamic prosody generation unit 102 that generates dynamic prosody information based on linguistic information with reference to the correction history information via the correction history management unit 105, and dynamic prosody information A modified prosody modification unit 103 that performs modification according to an external modification command to generate definite prosody information, and updates modification history information via the modification history management unit 105 according to the modification, and linguistic information and definite prosody information And a synthesized speech generation unit 104 that generates synthesized speech based on the above.
[0019]
The text-to-speech synthesis method (text-to-speech program code) shown in FIG. 2 includes a language analysis step 201 (language analysis program code) for performing language analysis on text data and extracting language information, and correction history information. A dynamic prosody generation step 202 for generating dynamic prosodic information based on language information in cooperation with a correction history reference step 215 to be referred to (correction history reference program code) and a correction history reference step 215 (correction history reference program code). (Dynamic prosody generation program code), prosody modification step 203 (prosody modification program code) for generating definite prosody information by modifying dynamic prosody information according to an external modification command, and modification of dynamic prosody information Correction history update step 225 (correction history update program code) for updating the correction history information accordingly A configuration including a synthetic speech generation step 204 (speech synthesizer program code) that generates a synthesized speech based on the language information and defined prosodic information.
[0020]
As a general method for generating prosodic information, for example, a method of generating prosodic information (static prosodic information) according to a predetermined prosodic generation rule (static prosodic generation rule) using language information as an argument Prosody information (static prosodic information) by selecting one prosody pattern piece from a plurality of prosody pattern pieces prepared in advance according to a predetermined rule (static prosodic pattern piece selection rule) The method of producing | generating is mentioned. On the other hand, in the present invention, the essential feature is that the prosodic information (dynamic prosodic information) is dynamically generated by referring to the correction history information.
[0021]
In this specification, “static” and “dynamic” mean “fixed without depending on correction history information” and “variable depending on the correction history information, depending on the information”, respectively. . “Static prosodic information” means prosodic information generated without referring to correction history information as in the prior art. The “dynamic prosodic information” means prosodic information generated by referring to the correction history information. If the prosodic information matches the correction history information, the prosodic information is different from the basic prosodic information, and the correction history information If not, the prosodic information is the same as the basic prosodic information.
[0022]
First, the language analysis unit 101 of the text-to-speech synthesizer will be described. The language analysis unit 101 performs language analysis on text data to be subjected to speech synthesis. Various language information is extracted by this language analysis (language analysis step 201). Examples of language information include information for specifying reading (phoneme symbol string, etc.), information for specifying part of speech, and information for specifying dependency.
[0023]
Next, the dynamic prosody generation unit 102 of the text-to-speech synthesizer will be described. The dynamic prosody generation unit 102 refers to the correction history information (correction history reference step 215), and generates dynamic prosody information based on the language information (dynamic prosody generation step 202). The dynamic prosody generation unit 102 may refer to the correction history information by any method as long as the prosody information is dynamically generated by referring to the correction history information. For example, the following three reference methods may be used. Can be mentioned.
[0024]
In the first revision history reference method, the dynamic prosody generation unit 102 generates static prosody information according to a static prosody generation rule, and corrects the generated static prosody information according to the correction history information. In this way, dynamic prosodic information is generated.
[0025]
In the second modification history reference method, the dynamic prosody generation unit 102 modifies the setting of the prosody generation parameter of the prosody generation rule according to the correction history information, and the dynamic prosody generation rule changes depending on the setting of the prosody generation parameter. According to the method, dynamic prosodic information is generated.
[0026]
In the third revision history reference method, the dynamic prosody generation unit 102 selects one optimal prosodic pattern piece as a selected prosodic pattern piece from a plurality of prosodic pattern pieces based on linguistic information according to a static prosodic selection rule. In addition, the dynamic prosodic information is generated by correcting the selected prosodic pattern piece according to the correction history information.
[0027]
Next, the prosody modification unit 103 of the text-to-speech synthesizer will be described. The prosody modification unit 103 generates definite prosody information by modifying the dynamic prosody information generated by the dynamic prosody generation unit 102 according to the external modification command (prosody modification step 203).
[0028]
If an external correction command is not input to the prosody modification unit 103, the dynamic prosody information becomes definite prosody information without modification. On the other hand, if an external correction command is received, the correction according to the external correction command is applied to the dynamic prosodic information, and the corrected dynamic prosodic information becomes definite prosodic information.
[0029]
The correction contents in the correction according to the external correction command and the correction conditions at the time of correction (hereinafter, a pair of correction contents and correction conditions are referred to as correction history data) are the correction history information stored in the correction history holding unit 106. In order to update, it is delivered to the correction history management unit 105 (correction history update step 203). Here, updating the revision history information means adding revision history data to the revision history information or changing a part of the revision history data constituting the revision history information.
[0030]
As a correction element of the correction content in each correction history data constituting the correction history information, any prosodic parameter may be used as long as a change before and after the correction can be specified. As the correction element, it is preferable to use a prosodic parameter that can define the correction amount before and after the correction. The prosody parameters that can define the correction amount include, for example, the duration length, intensity pattern (power pattern), or fundamental frequency pattern (pitch pattern) for a phoneme symbol string unit, expiratory paragraph unit, or accent phrase unit composed of one or more phoneme symbols. ). The correction element of the correction content of each correction history data may include only one type of prosodic parameter, or may include a plurality of types of prosodic parameters.
[0031]
On the other hand, the correction conditions in the correction history data can be set using language information. Examples of the condition element in the correction condition include a phoneme symbol string including one or more phoneme symbols, a part of speech, a position in a sentence, and an accent type. When the dynamic prosody generation unit 102 has the first history reference method, at least one prosodic parameter included in the static prosodic information can be used as a conditional element. The correction condition of the correction history data may include only one type of condition element, or may include a plurality of types of condition elements.
[0032]
As a method of correcting dynamic prosody information in the prosody correcting unit 103, for example, a pattern of prosody parameters that can be corrected is displayed graphically, and the displayed pattern is corrected using a mouse or the like. There is a method in which the prosodic information is displayed as text and is corrected by editing the displayed text.
[0033]
Further, in the modification of the dynamic prosodic information, the modification may be adjusted interactively while sequentially listening to the synthesized speech generated using the modified prosodic information reflecting the modification. The modified prosody information in a state where the adjustment is finally completed is sent to the synthesized speech generation unit 104 as definite prosody information. In this case, the prosody modification unit 103 is configured to include sample speech synthesis means for generating synthesized speech based on the language information and the modified prosody information for the text data fragment to be modified.
[0034]
Next, the correction history holding unit 106 and the correction history management unit 105 of the text-to-speech synthesizer will be described. The correction history holding unit 106 holds the correction content corrected by the prosody correction unit 103 as correction history information together with the correction conditions. The correction history management unit 105 includes a correction history update unit that updates the correction history information according to the correction in the prosody correction unit 103 and a correction history extraction unit that extracts the correction history information referred to in the dynamic prosody generation unit 102. This is a configuration provided.
[0035]
The correction history extraction unit of the correction history management unit 105 extracts correction history data that matches the correction condition from the correction history holding unit 106 based on the correction condition sent from the dynamic prosody generation unit 102, and The associated correction content is sent to the dynamic prosody generation unit 102. If there are a plurality of correction history data that match the correction conditions in the correction history holding unit 106, the correction contents corresponding to all of them are extracted and sent to the dynamic prosody generation unit 102.
[0036]
When the correction history update unit of the correction history management unit 105 receives the correction history data including the correction contents and the correction conditions from the prosody correction unit 103, the correction history update unit updates the correction history information held in the correction history holding unit 106.
[0037]
Here, when the correction history data having the correction condition satisfying the received correction condition is not held in the correction history holding unit 106, the correction history information is updated by adding the correction history data to the correction history holding unit 106 To do.
[0038]
Further, when the existing correction history data having the correction condition satisfying the received correction condition and the received correction content and the correction content having a different correction element is held in the correction history holding unit 106, the received correction history The correction history information may be updated by adding data, or they may be integrated and replaced with one new correction history data having correction contents including a plurality of correction elements.
[0039]
When the correction history management unit 105 receives the correction history data from the prosody correction unit 103, the correction condition satisfying the received correction condition and the received correction content and the correction element are the same and the correction processing is different. Is stored in the correction history storage unit 106, the final correction content determined relative to the basic prosody information is reflected in the subsequent generation of dynamic prosody. Update so that you can. In the update in this case, the correction history data may be added in association with the existing correction history data, or may be replaced with one new correction history data in consideration of the difference from the past correction contents. .
[0040]
The revision history holding unit 106 may be constituted by a single revision history DB or may be constituted by a plurality of revision history DBs having different prosodic styles. The prosodic style means a tone style according to dialects such as Osaka dialect and Kyoto dialect, and a tone style according to emotions such as a sad tone, a pleasant tone, an intense tone, and a gentle tone. When the revision history holding unit 106 has a plurality of revision history DBs, the revision history information is the entire prosody modification data included in all revision history DBs, that is, the prosody revisions held in the revision history holding unit. Note that it means the entire data.
[0041]
In the following, a case where the correction history holding unit 106 has a plurality of correction history DBs will be described. Refer to FIGS. 3 and 4 as necessary. FIG. 3 is a block diagram conceptually showing the structure of the text-to-speech synthesizer provided with a correction history holding unit composed of a plurality of correction history DBs. FIGS. 4A to 4C are block diagrams conceptually showing a configuration example of DB selection control means for controlling selection of a plurality of correction history DBs.
[0042]
As shown in FIG. 3, when the revision history holding unit 106 has a plurality of revision history DBs 116, the revision history management unit 105 together with the revision history extraction unit 115 and the revision history update unit 125 reads from the revision history DB. It is configured to include DB selection control means 135 for controlling which correction history DB is extracted in the correction history data extraction or correction history information update to the correction history DB.
[0043]
The DB selection control unit 135 includes a correction history DB (hereinafter also referred to as a reference correction history DB) referred to by the dynamic prosody generation unit 102 and a correction history DB (hereinafter referred to as a correction history DB updated based on the correction in the prosody correction unit 103). As the update correction history DB, the same correction history DB may be selected, or the reference correction history DB and the update correction history DB may be selected independently of each other. . Hereinafter, a specific configuration of the DB selection control means 135 will be described.
[0044]
As shown in FIG. 4A, the correction history management unit 105 includes a common DB selection control unit 145 that selects at least one of the plurality of correction history DBs 116 as a common correction history DB in response to the common selection command. It can be set as the structure which it has (1st structure). In the case of the first configuration, the dynamic prosody generation unit 102 selectively refers to the correction history information included in the common correction history DB, and the prosody correction unit 103 selects the correction history information included in the common correction history DB. Will be updated. A common selection instruction is used as the selection instruction.
[0045]
With the above configuration, the correction history information of one or a plurality of correction history DBs among the plurality of correction history DBs 116 can be selectively reflected in dynamic prosody generation according to the purpose, and one time Correction history information of one or a plurality of correction history DBs can be selectively updated by prosody correction. In addition, the dynamic prosody generation unit 102 can easily generate dynamic prosody information reflecting a desired prosodic style.
[0046]
As shown in FIG. 4B, the correction history management unit 105 selects a reference DB selection control that selects at least one of the plurality of correction history DBs 116 as a reference correction history DB in response to a reference selection command. A configuration (second configuration) including means 155 and update DB selection control means 165 that selects at least one of the plurality of correction history DBs 116 as an update correction history DB in response to an update selection command. it can. In the case of the second configuration, the dynamic prosody generation unit 102 selectively refers to the correction history information included in the reference correction history DB, and the prosody correction unit 103 includes the correction history information included in the update correction history DB. Will be updated selectively. Note that a reference selection instruction and an update selection instruction are used as selection instructions.
[0047]
When the correction history management unit 105 is in the first configuration, common control is performed for the reference correction history DB and the update correction history DB. In the second configuration, the reference correction history is used. The DB and the update correction history DB can be controlled independently. Thereby, the correction history information can be updated flexibly and effectively. That is, the correction history can be learned efficiently.
[0048]
The correction history management unit 105 includes a common DB selection control unit 145 that selects at least one of the plurality of correction history DBs 116 as a common correction history DB in response to the common selection command, and responds to the selection change command. The selection DB change for canceling the selection for any of the correction history DBs selected by the common DB selection control unit 145 and / or for additional selection of a correction history DB other than the correction history DB selected by the common DB selection unit 145 A configuration having the means 175 can be adopted.
[0049]
The selection DB changing means 175 is a change common to both the correction history DB referred to by the dynamic prosody generation unit 102 and the correction history DB updated based on the correction in the prosody correction unit 103 or a change independent of both. May be added. Further, only one of the correction history DB referred to by the dynamic prosody generation unit 102 and the correction history DB updated based on the correction in the prosody correction unit 103 may be changed.
[0050]
FIG. 4C shows a configuration in which the correction history management unit 105 includes selection DB changing means 175 for changing the correction history DB updated based on the correction in the prosody correction unit 103 (third configuration). It is shown. In the case of the third configuration, the dynamic prosody generation unit 102 selectively refers to correction history information included in at least one correction history DB (reference correction history DB) selected by the common DB selection control unit 145. The prosody modification unit 103 selectively updates the modification history information included in at least one modification history DB (update modification history DB) determined by the common database selection control unit 145 and the selection DB modification unit 175. Become. A common selection instruction and a selection change instruction are used as selection instructions.
[0051]
With the third configuration, the update correction history DB can be made independent of the reference correction history DB and can be arbitrarily selected from the plurality of prosody history DBs 116 in the correction history holding means 106. When the prosodic correction unit 103 corrects the dynamic prosody information, the correction is usually reflected on the prosodic style (correction history DB) referred to in the generation of the dynamic prosodic information. Compared to the above configuration, the correction history information can be updated easily, flexibly and effectively.
[0052]
The correction history management unit 105 includes a common DB selection unit 145 that selects at least one of the plurality of correction history databases 116 as a common correction history DB in response to the common selection command, and responds to the selection change command. Thus, a configuration (fourth configuration) having selection DB changing means 175 that only adds a new correction history DB to the update correction history DB configured by the common correction history DB may be adopted. In the case of the fourth configuration, the dynamic prosody generation unit 102 selectively refers to the correction history information included in the reference correction history DB configured by the common correction history DB selected by the common DB selection control unit 145. The prosody modification unit 103 selectively updates the modification history information included in the modification modification history DB configured by the common modification history DB and at least one modification history DB added by the selection DB modification unit; Become. A common selection instruction and a selection change instruction are used as selection instructions.
[0053]
When the prosody modification unit 103 corrects the dynamic prosody information, the correction history for updating is usually reflected in the prosodic style (correction history DB) referred to in the generation of the dynamic prosody information. More preferably, the DB includes all the correction history DBs constituting the reference correction history DB. Therefore, if it is the 4th composition, although the composition is simple compared with the 3rd composition, the same effect as the 3rd composition will be expressed.
[0054]
Which of the plurality of correction history DBs 116 is selected may be determined every time the apparatus or application is started, or may be determined for each text data. Further, on the application, the correction history DB may be appropriately determined every time the prosodic correction unit 103 corrects the dynamic prosodic information. Further, when the DB selection control unit 135 of the correction history management unit 105 is the second configuration, the third configuration, the fourth configuration, or the like (in the case of a configuration having at least two types of units), A plurality of determination methods can be used in combination.
[0055]
When selecting the correction history DB for each text data, the author may select on the application, or may be selected according to the control code (style selection information) included in the text file together with the text data. . In the former case, the selection command input unit for inputting a selection command is used. In the latter case, the text-to-speech synthesizer further includes a selection command generation unit for analyzing the text file and generating a selection command.
[0056]
Finally, the synthesized speech generation unit 104 of the text speech synthesizer will be described. The speech synthesis generation unit 104 generates synthesized speech by selecting a segment, deforming the segment, and connecting the segment based on the linguistic information and the definite prosodic information (synthesized speech creation step). It should be noted that any conventional known technique may be used for generating a synthesized speech using linguistic information and definite prosodic information.
[0057]
(Embodiment 1)
In the first embodiment, a text-to-speech synthesizer having a dynamic prosody generation unit to which the first correction history reference method is applied will be described. Refer to FIG. 5 and FIG. 6 as necessary. FIG. 5 is a block diagram conceptually showing the structure of the text-to-speech synthesizer according to the first embodiment. FIG. 6 is an explanatory diagram for explaining in detail the characteristic part in the text-to-speech synthesizer according to the first embodiment.
[0058]
The text-to-speech synthesizer shown in FIG. 5 includes a text holding unit 107, a language analysis unit 101, a dynamic prosody generation unit 102 having a static prosody generation unit 112, a filtering unit 122, and a filter control unit 132, and a prosody. The correction history management unit 105 including the correction unit 103, the display unit 117, the correction command input unit 127, the correction history holding unit 106 having a single correction history DB, the correction history extraction unit 115, and the correction history update unit 125. A synthesized speech creating unit 104 having a segment holding unit 134, a segment selecting unit 114, and a speech synthesizing unit 124, and a speech output unit 137.
[0059]
The operation of the text-to-speech synthesizer shown in FIG. 5 will be described. The language analysis unit 101 performs language analysis on the text data held in the text holding unit 107 in a predetermined unit, and as a result, language information is extracted (language processing step). The extracted language information is sent to the static prosody generation means 112 of the dynamic prosody generation unit 102.
[0060]
In the static prosody generation means 112, static prosody information is generated according to a static prosody generation rule based on the transmitted language information (static prosody generation step). The generated static prosodic information is sent to the filtering unit 122 together with the language information.
[0061]
In the filtering means 122, a correction condition is generated from the transmitted language information and static prosodic information. The generated correction condition is sent to the filter control means 132 in order to determine the processing content of the filtering process.
[0062]
In the filter control means 132, in order to check whether or not the prosody modification data that matches the sent modification condition is included in the prosody modification DB of the modification history holding unit 106, the modification condition is stored in the modification history management unit 105. Is sent to the correction history extraction means 115. Here, since the correction history holding unit 106 has only one type of prosody correction DB, the entire information included in the correction history DB is the correction history information.
[0063]
The correction history extraction unit 115 searches the correction history DB of the correction history holding unit 106. As a result, if there is correction history data that matches the sent correction conditions, the correction history data is extracted. The extracted prosody modification data is sent to the filter control means 132. If there are a plurality of correction history data matching the correction conditions, all the correction history data are sent to the filter control means 132. If there is no correction history data that matches the correction condition, the filter control means 132 is notified to that effect.
[0064]
The filter control unit 132 determines the correction amount of each prosodic parameter in the static prosody information from the received correction history data, and notifies the filtering unit 122 of the correction amount. Here, when a plurality of correction history data is sent from the correction history extraction means 115, the correction amount of each prosodic parameter is determined so as to reflect all of them. If there is a notification that the correction history data does not exist, the filtering unit 122 is notified that the correction is not performed or that the correction amount of each prosodic parameter is zero.
[0065]
The filtering means 122 corrects the static prosodic information based on the correction amount of each prosodic parameter determined by the filter control means 132 to generate dynamic prosodic information. The generated dynamic prosody information is sent to the prosody modification unit 103 together with the language information. Here, it should be noted that the correction in the filtering unit 122 is automatically performed according to the correction history information, and that the dynamic prosody information is prosodic information in which the correction history information is reflected.
[0066]
The prosody modification unit 103 causes the display unit 117 to display the transmitted dynamic prosody information as a graphical image. The desired additional correction is performed on the dynamic prosody information displayed on the display unit 117 according to the external correction command from the correction command input unit 127. Here, it should be noted that the correction in the prosody correction unit 103 is performed manually as in the prior art.
[0067]
For the additional correction in the prosody correction unit 103, correction history data in which a correction condition and correction content are set is generated. The generated correction history data is sent to the correction history update unit 125 of the correction history management unit 105 in order to update the correction history DB in the correction history holding unit 106. The dynamic prosody information with the desired correction is sent to the segment selection means 114 of the synthesized speech generation unit 104 together with the language information as definite prosody information.
[0068]
The correction history update unit 125 of the correction history management unit 105 updates the correction history DB of the correction history holding unit 106 based on the received correction history data. Here, it should be noted that the updated correction history DB is referred to in the next filtering process by the filtering unit 122. Also, note that modifications to the first file, not limited to the same file, are also reflected in the filtering process when text-to-speech text data contained in a second file different from the first file. Cost.
[0069]
The unit selection unit 114 of the synthesized speech generation unit 104 selects an optimal unit from the unit group held in the unit holding unit 134 based on the language information sent from the prosody modification unit 103. The selected segment is sent to the speech synthesizer 124.
[0070]
The unit held by the unit holding unit 134 may be a single unit or a composite unit. Examples of the composite segment include a CV unit (C: consonant, V: vowel) segment, a VC unit segment, a CVC unit segment, and a VCV unit segment. The element group is composed of only a single element, composed of only one kind of composition element, composed of a plurality of kinds of composition elements, and composed of a element element and one or more composition elements. It may be.
[0071]
In the speech synthesizer 124 of the synthesized speech generation unit 104, the transmitted speech is deformed and connected based on the definite prosodic information, thereby generating synthesized speech. The generated synthesized speech is output from the speech output unit 137.
[0072]
After the above processing, since the author's past correction is automatically reflected and dynamic prosodic information that generates synthesized speech close to preference is manually added, basic prosodic information (static Compared with conventional text-to-speech synthesis that manually corrects prosodic information) or manual text-to-speech synthesis that manually corrects prosody generation rules, any text data is easily output as synthesized speech that suits your preferences. Can be made.
[0073]
Here, based on the specific example shown in FIG. 6, the filtering process in the filtering means 122 will be described in detail. The prosodic information that is automatically corrected may be a duration, power pattern, pitch pattern, or the like. In this specific example, a case where the prosodic parameter included in the generated prosodic information is a pitch pattern will be described. In the following description, the reference numerals shown in FIG. 5 are assigned to the members and the means.
[0074]
First, prosody correction performed in the past will be described. In the prosody modification unit 103, modification 305a using the mouse is performed on the graphically displayed pitch pattern, or modification 305b is performed on the pitch pattern displayed in text by text editing. Here, the pitch pattern before correction means that the pitch of the first phoneme symbol / a / is 400 Hz, and the pitch of the last phoneme symbol / a / is 300 Hz. Since no pitch is defined for consonants, no numerical value is given to / k /. In past prosody correction, the pitch of the last phoneme symbol / a / is changed from 300 Hz to 200 Hz.
[0075]
Even if the prosody correction is performed by any of the above methods, the same correction history data 306 is generated. The generated correction history data 306 is stored in the correction history DB 116.
[0076]
In the correction history data 306 registered in the correction history DB 116, the position in the sentence is the end of the sentence and the phoneme symbol string (target phoneme symbol, preceding phoneme symbol, subsequent phoneme symbol) is (/ a /, / k /, / − /), The number of mora is 2, and the accent type is 0 type, and the correction content means that the pitch of the target phoneme phoneme / a / is lowered by 100 Hz. Includes correction processing.
[0077]
Next, generation of dynamic prosodic information after past prosodic correction will be described. As the static prosody information 301 generated by the static prosody generation unit 112, a pitch pattern (400 Hz, −, 300 Hz) corresponding to the phoneme symbol string (/ a /, / k /, / a /) is sent to the filtering unit 122. If given, a modification condition 302 is generated.
[0078]
The filter control means 132 refers to the correction history DB 116 based on the correction condition 302 and obtains correction history data 306 that matches the correction condition. Since there is a correction content (−100 Hz: lower by 100 Hz) in the correction history data 306, the correction content 303 that “lowers the pitch of the last phoneme symbol / a / by 100 Hz” is sent to the filtering means 122.
[0079]
The filtering unit 122 lowers the pitch pattern for the last phoneme symbol / a / by 100 Hz based on the correction content 303 sent. That is, the pitch pattern is corrected from (400 Hz, −, 300 Hz) to (400 Hz, −, 200 Hz). The corrected pitch pattern is sent to the prosody correcting unit 103 as dynamic prosody information 304.
[0080]
Thereby, when the correction history data matching the correction condition generated in the static prosody generation means 112 is already included in the correction history information, the same correction as the correction performed in the past is automatically applied. Dynamic prosodic information can be generated.
[0081]
(Embodiment 2)
In the second embodiment, the text-to-speech synthesizer includes a dynamic prosody generation unit to which the first correction history reference method is applied and a correction history holding unit having a plurality of prosody DBs having different prosodic styles. Will be described. Note that FIG. 7 and FIG. 8 are referred to as necessary. FIG. 7 is a block diagram conceptually showing the structure of the text-to-speech synthesizer according to the second embodiment. FIG. 8 is an explanatory diagram for explaining in detail the characteristic part of the text-to-speech synthesizer according to the second embodiment.
[0082]
The text-to-speech synthesizer shown in FIG. 7 includes a text holding unit 107, a language analysis unit 101, a dynamic prosody generation unit 102 having a static prosody generation unit 112, a filtering unit 122, and a filter control unit 132, and a prosody. A correction unit 103, a display unit 117, a correction command input unit 127, a correction history holding unit 106 having a plurality of types of correction history DBs 116, a correction history extraction unit 115, a correction history update unit 125, and a DB selection control unit 135 are provided. The composition includes a modification history management unit 105, a selection command input unit 127, a unit holding unit 134, a unit selection unit 114, and a voice synthesis unit 124, and a voice output unit 137. is there. With this configuration, the correction history holding unit 106 includes a plurality of correction history DBs 116, so that prosody correction that selectively reflects a desired prosody style from a plurality of prosody styles can be automatically performed.
[0083]
The operation of the text-to-speech synthesizer shown in FIG. 7 will be described. Note that the operation of the text-to-speech synthesizer shown in FIG. 7 is basically the same as that of the text-to-speech synthesizer of the first embodiment, except that the revision history information is updated and referred to by the revision history management unit. Since there is, explanation about a common part is omitted.
[0084]
First, in the update of the correction history information, the correction history update means 125 according to the new correction history data from the prosody correction unit 103 is at least one correction selected according to the correction command from the selection command input unit 147. The correction history information is updated with respect to the history DB.
[0085]
Next, in referring to the revision history information, the revision history extraction unit 115 is selected according to the selection command from the selection command input unit 147 in response to a request from the dynamic prosody generation unit 102 (filter control unit 132). Then, a search is performed on at least one correction history DB, and correction history data matching the correction conditions is extracted.
[0086]
Here, based on the specific example shown in FIG. 8, a method of updating and referring to the revision history information in the revision history management unit 105 will be described in detail. The prosodic information that is automatically corrected may be a duration, power pattern, pitch pattern, or the like. In this specific example, a case where the prosodic parameter included in the generated prosodic information is a pitch pattern will be described. In the following description, the reference numerals shown in FIG. 7 are assigned to the members and the means.
[0087]
First, prosody correction performed in the past will be described. In the prosody modification unit 103, the prosody modification 314 is performed, and the modification history data 315 is generated. In the generated correction history data 315, the pitch pattern (400 Hz,-, 300 Hz) corresponding to the phoneme symbol string (/ a /, / k /, / a /) is corrected to the pitch pattern (400 Hz,-, 200 Hz). The correction contents to be included are included. The correction history data 315 is stored in the correction history DB-A 126 by a selection command. The correction conditions are not shown.
[0088]
Further, the prosody modification unit 103 performs the prosody modification 316 and the modification history data 317 is generated. In the generated correction history data 317, the pitch pattern (400 Hz,-, 300 Hz) corresponding to the phoneme symbol string (/ a /, / k /, / a /) is corrected to the pitch pattern (400 Hz,-, 350 Hz). The correction contents to be included are included. The correction history data 317 is stored in the correction history DB-B 136 by a selection command. Although the illustration of the correction condition in the correction history data 317 is omitted, it is assumed that it is the same as the correction condition in the correction history data 315.
[0089]
Next, generation of dynamic prosodic information after past prosodic correction will be described. As the static prosody information 311 generated by the static prosody generation means 112, the pitch pattern 311 (400 Hz,-, 300 Hz) corresponding to the phoneme symbol string (/ a /, / k /, / a /) is the filtering means 122. The correction condition is generated, and the generated correction condition is sent to the correction history extraction unit 115 of the correction history management unit 105 via the filter control unit 132. It is assumed that the generated correction conditions are the same as the correction conditions in the correction history data 315 and the correction conditions in the correction history data 317.
[0090]
In the DB selection control means 135, the correction history DB-B 136 is selected 312 by a selection command inputted in advance, so that the correction history extraction means 125 only reads the correction history data held in the correction history DB-B 136. Search is performed, and correction history data 317 that matches the correction condition is extracted. Here, it should be noted that although the correction history data matching the correction condition exists in the correction history DB-A 126, the correction history data 315 of the correction history DB-A 126 is not extracted. The extracted correction history data 317 is sent to the filter control means 132 of the dynamic prosody generation unit.
[0091]
In the filter control means 132, the correction history data 317 includes a correction process (+50 Hz: increase by 50 Hz) associated with the correction condition, so that “the pitch of the last phoneme symbol / a / is increased by 50 Hz”. Determine the correction contents. The determined correction content is sent to the filtering means 122.
[0092]
In the filtering means 122, the pitch pattern corresponding to the last phoneme symbol / a / is raised by 50 Hz based on the sent correction contents. That is, the pitch pattern (static prosodic information) 311 is corrected from (400 Hz, −, 300 Hz) to (400 Hz, −, 350 Hz). The corrected pitch pattern (dynamic prosody information) 313 is sent to the prosody correcting unit 103.
[0093]
When the prosody modification unit 103 performs further modification on the pitch pattern 313 sent thereto, the modification history DB-B 136 is selected unless the modification history DB selection is changed again by a selection command.
[0094]
As a result, dynamic prosodic information reflecting a desired prosodic style among a plurality of prosodic styles can be generated. In addition, it is possible to update only correction history information corresponding to a desired prosodic style among a plurality of prosodic styles.
[0095]
Here, a method for selecting the correction history DB will be described. In the selection command input unit 147, the correction history DB 116 is switched by a selection command input from the outside by pressing a prosodic style selection button (correction history DB selection button) on the application.
[0096]
The text-to-speech synthesizer shown in FIG. 7 is configured to include a selection command input unit 147 that inputs a selection command for selecting the correction history DB 116. Can also be entered. FIG. 9 is a block diagram conceptually showing the structure of a text-to-speech synthesizer including a selection command generation unit instead of the selection command input unit 147 in FIG.
[0097]
As shown in FIG. 9, the text-to-speech synthesizer extracts the prosodic style information by analyzing the control code from the text data held in the text holding means 107, and selects a command based on the extracted prosodic style information. The selection command generation unit 108 for generating
[0098]
The selection command generation unit 108 includes only one prosodic style information in the text data, and even if it is a means for determining the prosodic style for each text data, the text data includes a plurality of prosodic style information for each piece of text data. It may be a means for determining the prosodic style. For example, when the analyzed control code includes contents describing that the prosodic style A and the prosodic style B are applied to the sentence 1 and the sentence 2 in the same file, respectively, the dynamic prosody generation unit 102 The pitch pattern can be generated with the prosody style A for the sentence 1 and the pitch pattern can be generated with the prosody style B for the sentence 2.
[0099]
(Embodiment 3)
In the third embodiment, a text-to-speech synthesizer having a dynamic prosody generation unit to which the second correction history reference method is applied will be described. Refer to FIGS. 10 and 11 as necessary. FIG. 10 is a block diagram conceptually showing the structure of the text-to-speech synthesizer according to the third embodiment. FIG. 11 is an explanatory diagram for explaining in detail the characteristic part in the text-to-speech synthesizer according to the third embodiment.
[0100]
The text-to-speech synthesizer shown in FIG. 10 includes a text holding unit 107, a language analysis unit 101, a dynamic prosody generation unit 102 including a prosody generation parameter control unit 142 and a prosody generation rule control unit 152, and a prosody modification unit. 103, a display unit 117, a correction command input unit 127, a correction history holding unit 106 having a correction history DB, a correction history management unit 105 having a correction history extracting unit 115 and a correction history updating unit 125, and a unit holding A synthesized speech generation unit 104 having a unit 134, a segment selection unit 114 and a speech synthesis unit 124, and a speech output unit 137 are included.
[0101]
The operation of the text-to-speech synthesizer shown in FIG. 10 will be described. Since the configuration other than the dynamic prosody generation unit is the same as that of the first embodiment, the description thereof is omitted.
[0102]
In the dynamic prosody generation unit 142 of the dynamic prosody generation unit 102, basic prosody generation parameters necessary for generating prosody information are determined based on the linguistic information sent from the language analysis unit 101, and correction conditions are set. Is generated. The generated correction condition is sent to the prosodic generation rule control means 152 in order to determine the prosodic generation parameters used for generating the dynamic prosodic information.
[0103]
In the prosody generation rule control means 152, the sent correction condition confirms whether or not the prosody correction data that matches the correction condition is included in the correction history DB (correction history information) of the correction history holding unit 106. Therefore, it is sent to the correction history extraction means 115 of the correction history management unit 105.
[0104]
The correction history extraction unit 115 searches the correction history DB of the correction history holding unit 106 based on the received correction conditions, and as a result, if there is correction history data that matches the correction conditions, the correction history data To extract. The extracted correction history data is sent to the prosody generation rule control means 152. If there are a plurality of correction history data matching the correction conditions, all the correction history data are sent to the prosody generation rule control means 152. If there is no correction history data that matches the correction condition, the prosody generation rule control means 152 is notified of this fact.
[0105]
The prosody generation rule control unit 152 determines the correction of each prosody generation parameter based on the received correction history data and notifies the dynamic prosody generation unit 142 of the correction. If there is a notification that the correction history data does not exist, the dynamic prosody generation unit 142 is notified that the correction is not performed, or that the correction of each prosody parameter is zero. Here, when a plurality of correction history data is sent from the correction history extraction means 115, the correction of each prosody generation parameter is determined so as to reflect the correction contents included in all the correction history data.
[0106]
The dynamic prosody generation unit 142 generates dynamic prosody information according to the prosodic generation rules using the prosodic generation parameters determined by the prosody generation rule control unit 152 based on the language information generated by the language analysis unit 101. The generated dynamic prosody information is sent to the prosody modification unit 103 together with the language information. Here, it should be noted that a dynamic prosody generation rule is determined based on each prosody generation parameter determined by the prosody generation rule control means 152. Further, the change of the prosody generation rule in the dynamic prosody generation unit 142 is automatically performed according to the correction history information, and the dynamic prosody information is the prosody information in which the correction history information is reflected. Need attention.
[0107]
Here, based on the specific example shown in FIG. 11, the method of controlling the prosody generation rule in the dynamic prosody modification means 142 will be described in detail. The generated dynamic prosodic information may be time length, power, pitch pattern, etc. In this specific example, the case where the prosodic parameter included in the generated dynamic prosodic information is a pitch pattern will be described. . In the following description, the reference numerals shown in FIG. 10 are assigned to the members and the means.
[0108]
As a model for generating a pitch pattern in the dynamic prosody generation means 142, a Fujisaki model is used. The Fujisaki model expresses a pitch pattern by superimposing a phrase component that expresses attenuation of an exhalation paragraph and an accent component that expresses pitch fluctuation for each accent. In the Fujisaki model, various parameters such as the slope of phrase attenuation can be adjusted. In this example, the pitch pattern is determined by two parameters: a phrase command indicating the size of the phrase and an accent command indicating the size of the accent. Is generated.
[0109]
First, prosody correction performed in the past will be described. In the prosody modification unit 103, modification 325 using the mouse is performed on the pitch pattern displayed graphically. Here, the pitch pattern before correction generated by using the phrase command “h” and the accent command “a” is indicated by a dotted line, and the pitch pattern after correction is indicated by a solid line.
[0110]
Along with this correction, calculation 326 of prosody generation parameters necessary for generating a corrected pitch pattern is performed. In this specific example, “h” remains unchanged for the phrase command, and “a” changes to “a ′” for the accent command.
[0111]
A publicly known method can be used as a method for calculating the prosody generation parameter necessary for generating the corrected pitch pattern. For example, the prosody generation parameter for generating the pitch pattern is obtained by solving a linear equation with the prosody generation parameter to be estimated as an unknown using a least square method targeting the pitch pattern after correction.
[0112]
When the prosody generation parameters necessary for generating the corrected pitch pattern are calculated, the correction history data 327 is subsequently generated. The generated correction history data 327 is stored in the correction history DB 116.
[0113]
The generated correction history data 327 has a correction condition that the position in the sentence is the end of the sentence, and the phoneme symbol string (corrected phoneme, preceding phoneme, subsequent phoneme) is (/ a /, / k /, //-/), The number of mora is 3 and the accent type is 0 type, and the correction content includes the processing content which means that the accent command is corrected from “a” to “a ′”. Yes.
[0114]
Next, generation of dynamic prosodic information after past prosodic correction will be described. In the dynamic prosody generation unit 142, the phrase designation “h” and the accent command “a” are generated as the basic prosody generation parameter 321, and the correction condition 322 is generated. The generated correction condition 322 is sent to the prosody normal rule control means 152. Here, in the generated correction condition, the position in the sentence is the end of the sentence, and the phoneme symbol string (target phoneme symbol, preceding phoneme symbol, subsequent phoneme symbol) is (/ a /, / k /, / − /). , The number of mora is 3 and the accent type is 0 type.
[0115]
The prosody generation rule control means 152 searches the correction history DB 116 based on the received correction conditions, and obtains correction history data 327 that matches the correction conditions as a result of the search. The correction history data 327 includes correction contents 323 (a is changed to a ′: the pitch pattern is changed from the dotted line in the upper left diagram to the solid line pattern) associated with the received correction condition. The accent command “a” is corrected to “a ′”, and the corrected prosody generation parameter 324 is determined. The modified prosody generation parameter 324 is sent to the dynamic prosody generation unit 142.
[0116]
The dynamic prosody generation unit 142 determines and determines the prosody generation rule using the phrase designation “h” and the accent command “a ′” as the modified prosody generation parameter based on the received modified prosody generation parameter 324. In accordance with the prosodic generation rules, dynamic prosodic information is generated based on the linguistic information. The generated dynamic prosody information is sent to the prosody modification unit 103.
[0117]
As a result, when the correction history data matching the correction condition generated by the dynamic prosody generation unit 142 is already included in the correction history information, the same correction as the correction performed in the past is automatically applied. Dynamic prosodic information can be generated.
[0118]
In the above description, the Fujisaki model has been described as a model for generating a pitch pattern. However, corrections made in the past can be automatically reflected using other models as well. When performing control other than the pitch pattern, for example, when controlling the duration or power pattern in the prosodic information, it is necessary to use a model suitable for each prosodic parameter to be controlled.
[0119]
In the above description, the configuration in which the correction history holding unit has a single correction history DB has been described. However, as described in the second embodiment, a configuration having a plurality of correction history DBs may be employed. 12 and 13 show a conceptual configuration of a text-to-speech synthesizer including a dynamic prosody generation unit of the second history reference method and a correction history holding unit having a plurality of prosody modification DBs having different prosodic styles. FIG. The configuration is substantially the same as that of the text-to-speech synthesizer shown in FIG. 10 except that the correction history management unit has DB selection control means and the correction history holding means has a plurality of correction history DBs.
[0120]
Also, the text-to-speech synthesizer shown in FIGS. 12 and 13 includes static prosody generation means 112, filtering means 122, and filter control means 132 in the text-to-speech synthesizer shown in FIGS. 7 and 9, respectively. The configuration is the same except that the dynamic prosody generation unit is changed to a dynamic prosody generation unit having dynamic prosody generation unit 142 and prosody generation rule control unit 152.
[0121]
A method for referring to and updating a plurality of correction history DBs in the text-to-speech synthesizer shown in FIGS. 12 and 13 is the same as that in the second embodiment, and a description thereof will be omitted.
[0122]
(Embodiment 4)
In the fourth embodiment, a text-to-speech synthesizer having a dynamic prosody generation unit to which the third modification history reference method is applied will be described. Reference is made to FIGS. 14 and 15 as necessary. FIG. 14 is a block diagram conceptually showing the structure of the text-to-speech synthesizer according to the fourth embodiment. FIG. 15 is an explanatory diagram for explaining in detail the characteristic part in the text-to-speech synthesizer according to the fourth embodiment.
[0123]
The text-to-speech synthesizer shown in FIG. 14 includes a text holding unit 107, a language analysis unit 101, a prosody pattern selection unit 162, a prosody pattern correction unit 172, a pattern piece correction control unit 182 and a prosody pattern piece holding unit 192. The dynamic prosody generation unit 102, the prosody modification unit 103, the display unit 117, the modification command input unit 127, the modification history holding unit 106 including the modification history DB, the modification history extraction unit 115, and the modification history update unit 125. The modified history management unit 105 including the unit, the unit holding unit 134, the unit selection unit 114, and the voice synthesis unit 124, and the voice output unit 137.
[0124]
The operation of the text-to-speech synthesizer shown in FIG. 14 will be described. Since the configuration other than the dynamic prosody generation unit is the same as that of the first embodiment, the description thereof is omitted.
[0125]
In the prosodic pattern piece selection means 162, one optimal prosodic pattern piece is selected as the selected prosodic pattern piece from the prosodic pattern piece holding means 192 based on the transmitted language information, and a correction condition is generated. . The selected selected prosodic pattern piece and the generated correction condition are sent to the prosodic pattern piece correcting means 172.
[0126]
In the prosody pattern piece correcting means 172, the sent correction condition is sent to the pattern piece correction control means 182 in order to determine the correction for the selected prosodic pattern piece.
[0127]
In the pattern piece correction control means 182, the received correction condition is corrected in order to check whether the correction history data matching the correction condition is included in the prosody correction DB of the correction history holding unit 106. It is sent to the correction history extraction means 115 of the history management unit 105.
[0128]
The correction history extraction unit 115 searches the correction history DB of the correction history holding unit 106 based on the sent correction conditions. As a result, if there is correction history data that matches the sent correction conditions, The correction history data is extracted. The extracted prosody correction data is sent to the pattern piece correction control means 182. If there are a plurality of correction history data matching the correction conditions, all the correction history data are sent to the pattern piece correction control means 182. If there is no correction history data that matches the correction condition, the pattern piece correction control means 182 is notified of this fact.
[0129]
The pattern piece correction control means 182 determines the correction contents for the selected prosodic pattern piece based on the sent correction history data. The determined correction content is sent to the prosodic pattern piece correcting means 172. Here, when a plurality of correction history data is sent from the correction history extraction means 115, the correction contents of the prosodic pattern pieces are determined so as to reflect the correction contents of all the correction history data. If there is a notification that the correction history data does not exist, the prosody pattern piece correcting means 172 is notified that the correction is not performed or that each prosody pattern piece is not corrected.
[0130]
The prosodic pattern piece correcting means 172 corrects the selected prosodic pattern piece based on the correction contents determined by the pattern piece correction control means 182 to generate dynamic prosodic information. The generated dynamic prosody information is sent to the prosody modification unit 103 together with the language information. Here, it should be noted that the correction in the prosodic pattern piece correcting means 172 is automatically performed according to the correction history information, and that the dynamic prosody information is prosodic information in which the correction history information is reflected. .
[0131]
Here, a method of correcting the selected prosodic pattern in the prosodic pattern piece correcting means 172 will be described in detail based on the specific example shown in FIG. Note that the prosodic information to be corrected may be duration time, power, pitch pattern, and the like. In this specific example, a case where the prosodic parameter included in the generated prosodic information is a pitch pattern will be described. In the following description, the reference numerals shown in FIG. 14 are assigned to the members and the means.
[0132]
First, prosody correction performed in the past will be described. In the prosody modification unit 103, modification 337 using a mouse is performed on the graphically displayed pitch pattern. Here, the pitch pattern 334a (selected prosodic pattern piece) before correction is indicated by a dotted line, and the corrected pitch pattern 334b (corrected prosodic pattern piece) is indicated by a solid line.
[0133]
Along with this prosody correction, correction history data 338 is generated, and the generated correction history data 338 is stored in the correction history DB 116.
[0134]
The generated correction history data 338 includes, as correction conditions, that the position in the sentence is the end of the sentence, the number of mora is 3, and the accent type is 2 type. This includes correcting the pattern 334a to the corrected pitch pattern 334b. Here, in this specific example, the corrected pitch pattern itself is held as correction contents.
[0135]
Next, generation of dynamic prosodic information after past prosodic correction will be described. In the prosodic pattern piece selection means 162, the optimal prosodic pattern piece is selected as the selected prosodic pattern piece from the prosodic pattern piece DB 334 based on the language information, and the correction condition 331 is generated. The selected selected prosodic pattern piece and the generated correction condition are sent to the prosodic pattern piece correcting means 172. The generated correction condition 331 includes that the position in the sentence is the end of the sentence, the number of mora is 3, and the accent type is 2.
[0136]
The prosody pattern piece correcting means 172 sends the sent correction condition to the pattern piece correction control means 182.
[0137]
The pattern piece correction control means 182 searches the correction history DB 116 based on the correction condition, and as a result, the correction history data 338 that matches the correction condition is obtained. Since there is a correction content that satisfies the correction conditions (the selected pitch pattern 334a is changed to the held pitch pattern 334b: the pitch pattern is changed from the dotted line in the upper left diagram), the correction content is prosodic. This is sent to the pattern correction means 172.
[0138]
The prosodic pattern piece correcting means 172 corrects the prosodic pattern 334a to the prosodic pattern 334b based on the received correction contents, and generates dynamic prosodic information 336. The generated dynamic prosody information is sent to the prosody modification unit 103.
[0139]
Thereby, when the correction history data matching the correction condition generated by the prosodic pattern piece selection means 162 is already included in the correction history information, the same correction as the correction made in the past is automatically applied. Dynamic prosodic information can be generated.
[0140]
In the above, the changed pitch pattern is retained, but any method can be used to retain the modified pitch pattern as long as it can be reproduced finally. May be. For example, the changed pitch pattern can also be reproduced by storing the difference value for each time with respect to the pitch pattern selected from the prosodic pattern piece DB 334.
[0141]
In the above description, the configuration in which the correction history holding unit 106 has a single correction history DB 116 has been described. However, as described in the second embodiment, a configuration having a plurality of correction history DBs may be employed. FIGS. 16 and 17 are conceptual diagrams of a text-to-speech synthesizer including a dynamic prosody generation unit of the third correction history reference method and a correction history holding unit having a plurality of prosody modification DBs having different prosodic styles. FIG. The configuration is substantially the same as that of the text-to-speech synthesizer shown in FIG. 14 except that the correction history management unit has DB selection control means and the correction history holding means has a plurality of correction history DBs.
[0142]
Also, the text-to-speech synthesizer shown in FIGS. 16 and 17 has static prosody generation means 112, filtering means 122, and filter control means 132 in the text-to-speech synthesizer shown in FIGS. 7 and 9, respectively. The dynamic prosody generation unit has the same configuration except that the dynamic prosody generation unit is changed to a dynamic prosody generation unit having prosody pattern piece selection means 162, prosody pattern correction means 172, pattern piece correction control means 182 and prosody pattern piece holding means 192.
[0143]
The reference method and the update method of the plurality of correction history DBs in the text-to-speech synthesizer shown in FIGS. 16 and 17 are the same as those in the second embodiment, and a description thereof will be omitted.
[0144]
【The invention's effect】
As described above, in the text-to-speech synthesizer of the present invention, the correction history information is stored via the correction history management unit that manages the correction history information, the correction history management unit that manages the correction history information, and the correction history information. With the configuration including the dynamic prosody generation unit that generates dynamic prosody information based on language information, it is not necessary to perform the same correction as the past correction registered in the correction history information again And since learning of correction history information progresses sequentially according to past corrections, it becomes a device that easily generates prosodic information that suits the end user's preference.
[0145]
The dynamic prosody generation unit generates static prosody information according to static prosody generation rules, and modifies the generated static prosody information according to the correction history information, thereby generating dynamic prosody information. The first correction history reference method is adopted.
[0146]
The dynamic prosody generation unit corrects the prosody generation parameter setting of the prosody generation rule according to the correction history information, and generates dynamic prosody information according to the dynamic prosody generation rule that changes according to the setting of the prosody generation parameter. The second correction history reference method is adopted.
[0147]
The dynamic prosody generation unit selects one optimal prosody pattern piece as a selected prosodic pattern piece from a plurality of prosodic pattern pieces based on linguistic information according to a static prosodic selection rule, and the selected prosodic pattern piece. The third correction history reference method for generating dynamic prosodic information by correcting the information according to the correction history information is adopted.
[0148]
A text-to-speech synthesizer equipped with a dynamic prosody generation unit adopting any one of the first to third correction history reference methods assures easy generation of prosodic information that matches the end user's preference. it can.
[0149]
Further, the revision history holding unit is configured by a plurality of revision history DBs having different prosodic styles, and the revision history management unit performs selective reference to the revision history DB and update of the revision history DB selectively. The prosody modification that selectively reflects the desired prosody style can be automatically applied from the prosody styles, and the entire modification history information can be updated (learned) flexibly and effectively.
[0150]
In the text-to-speech synthesis method of the present invention, a correction history reference step for referring to correction history information, a dynamic prosody generation step for generating dynamic prosody information based on language information in cooperation with the correction history reference step, By including a revision history update step for updating revision history information according to the revision, it is not necessary to perform the same revision as the past revision registered in the revision history information, and the past revision Accordingly, the learning of the correction history information proceeds sequentially, so that prosodic information suitable for the end user's preference can be easily generated.
[0151]
The text-to-speech synthesis program according to the present invention includes a correction history reference program code that references correction history information, and a dynamic prosody generation program that generates dynamic prosody information based on language information in cooperation with the correction history reference program code. By including the code and the correction history update program code for updating the correction history information according to the correction, it becomes unnecessary to perform the same correction as the past correction registered in the correction history information, and Since learning of correction history information sequentially proceeds according to past prosodic corrections, it is possible to realize a text-to-speech synthesis method that easily generates prosodic information that matches the end user's preference.
[Brief description of the drawings]
FIG. 1 is a block diagram conceptually showing the structure of a text-to-speech synthesizer according to the present invention.
FIG. 2 is a block diagram conceptually showing the structure of the text-to-speech synthesis method according to the present invention.
FIG. 3 is a block diagram conceptually showing the structure of a text-to-speech synthesizer having a plurality of correction history DBs in a correction history holding unit.
FIG. 4 is a block diagram conceptually showing a configuration of a correction history management unit in a text-to-speech synthesizer having a plurality of correction history DBs in a correction history holding unit, and FIG. FIG. 4B is a block diagram illustrating a configuration of a correction history management unit including a DB selection control unit, and FIG. 4B illustrates a configuration of a correction history management unit including a reference DB selection control unit and an update DB selection control unit. FIG. 4C is a block diagram illustrating a configuration of a correction history management unit having a common DB selection control unit and a selection DB change unit.
FIG. 5 is a block diagram conceptually showing the structure of a text-to-speech synthesizer provided with a dynamic prosody generation unit of the first modification history reference method.
FIG. 6 is an explanatory diagram for explaining a first correction history reference method;
FIG. 7 conceptually illustrates a first configuration example of a text-to-speech synthesizer including a correction history holding unit including a plurality of correction history DBs and a dynamic prosody generation unit using a first correction history reference method. FIG.
FIG. 8 is an explanatory diagram for explaining a selective update method of correction history information and a selective extraction method of correction history information in a correction history management unit;
FIG. 9 conceptually illustrates a second configuration example of a text-to-speech synthesizer including a correction history holding unit including a plurality of correction history DBs and a dynamic prosody generation unit of the first correction history reference method. FIG.
FIG. 10 is a block diagram conceptually showing the structure of a text-to-speech synthesizer including a second modification history reference method dynamic prosody generation unit.
FIG. 11 is an explanatory diagram for explaining a second correction history reference method;
FIG. 12 is a conceptual diagram illustrating a first configuration example of a text-to-speech synthesizer including a correction history holding unit including a plurality of correction history DBs and a dynamic prosody generation unit of a second correction history reference method; It is a block diagram shown in FIG.
FIG. 13 is a conceptual diagram illustrating a second configuration example of a text-to-speech synthesizer including a correction history holding unit including a plurality of correction history DBs and a dynamic prosody generation unit of a second correction history reference method; It is a block diagram shown in FIG.
FIG. 14 is a block diagram conceptually showing the structure of a text-to-speech synthesizer including a third modification history reference method dynamic prosody generation unit.
FIG. 15 is an explanatory diagram for explaining a third correction history reference method;
FIG. 16 is a conceptual diagram illustrating a first configuration example of a text-to-speech synthesizer including a correction history holding unit including a plurality of correction history DBs and a dynamic prosody generation unit of a third correction history reference method; It is a block diagram shown in FIG.
FIG. 17 is a conceptual diagram illustrating a second configuration example of a text-to-speech synthesizer including a correction history holding unit including a plurality of correction history DBs and a dynamic prosody generation unit of a third correction history reference method; It is a block diagram shown in FIG.
FIG. 18 is a block diagram conceptually showing the structure of a conventional text-to-speech synthesizer.
[Explanation of symbols]
101 Language Analysis Department
102 Dynamic prosody generation part
112 Static prosody generation means
122 Filtering means
132 Filter control means
142 Dynamic Prosody Generation Means
152 Prosody Generation Rule Control Means
162 Prosodic pattern segment selection means
172 Prosody pattern fragment correction means
182 Prosodic pattern fragment correction control means
192 Prosody pattern fragment holding means
103 Prosody modification part
104 synthesized speech generator
114 Segment selection means
124 Speech synthesis means
134 Element holding means
105 revision history management
115 Correction history extraction means
125 Correction history update means
135 DB selection control means
145 Common DB selection control means
155 DB selection control means for reference
165 Update DB selection control means
175 Selection DB changing means
106 Correction history holding unit
116 revision history DB
107 Text holding part
117 Display
127 Correction command input part
137 Audio output unit
147 Selection command input part
108 Selection instruction generator
201 Language analysis step (language analysis program code)
202 dynamic prosody generation step (dynamic prosody generation program code)
203 Prosody modification step (prosody modification program code)
204 synthesized speech generation step (synthesized speech generation program code)
215 Correction history extraction step (correction history extraction program code)
225 revision history update step (modification history update program code)

Claims (13)

A language analysis unit that performs language analysis on text data and extracts language information;
A revision history holding unit for holding revision history information;
A correction history management unit for managing the correction history information;
A dynamic prosody generation unit that generates dynamic prosody information based on the language information with reference to the correction history information via the correction history management unit;
The fixed prosody information is generated by correcting the dynamic prosodic information according to an external correction command, and the correction history information is generated via the correction history management unit based on the correction according to the external correction command. A prosody modification section for updating
A synthesized speech generation unit that generates a synthesized speech based on the language information and the definite prosodic information;
The correction history management unit extracts correction history information for extracting correction history information referred to by the dynamic prosody generation unit; a correction history update unit for updating correction history information held in the correction history holding unit; Have
The correction history holding unit has a plurality of correction history databases having different prosodic styles,
The text-to-speech synthesizer further comprising a database selection control means for controlling the selection of the plurality of correction history databases in accordance with a selection command . The text-to-speech synthesizer according to claim 1.
The dynamic prosody generation unit generates static prosody information according to a static prosody generation rule, and modifies the static prosody information generated by the static prosody generation unit A text-to-speech synthesizer comprising: filtering means for generating the dynamic prosody information by means of; and filter control means for controlling correction by the filtering means in accordance with the correction history information. The text-to-speech synthesizer according to claim 1.
The dynamic prosody generation unit generates the dynamic prosody information according to a dynamic prosody generation rule that changes according to the setting of the prosody generation parameter, and the prosody generation parameter according to the correction history information A text-to-speech synthesizer comprising: prosody generation rule control means for controlling The text-to-speech synthesizer according to claim 1.
The dynamic prosody generation unit selects one of the plurality of prosodic pattern pieces as a selected prosodic pattern piece according to a prosodic pattern piece holding unit that holds a plurality of prosodic pattern pieces, and a static prosodic selection rule. Prosodic pattern fragment selecting means, prosodic pattern fragment correcting means for generating the dynamic prosodic information by correcting the selected prosodic pattern fragment, and controlling correction to the selected prosodic pattern fragment according to the correction history information A text-to-speech synthesizer. The text-to-speech synthesizer according to claim 4 .
A text-to-speech synthesizer, further comprising a selection command input unit for inputting the selection command. The text-to-speech synthesizer according to claim 4 .
A text-to-speech synthesizer, further comprising: a selection command generation unit that detects style selection information included in the text data and generates the selection command based on the style selection information. The text-to-speech synthesizer according to claim 4 .
The correction history management unit has a common database selection control means for selecting at least one of the plurality of correction history databases as a common correction history database in response to the selection command,
The dynamic prosody generation unit selectively refers to correction history information included in the common correction history database;
The text-to-speech synthesizer, wherein the prosody modification unit selectively updates modification history information included in the common modification history database. The text-to-speech synthesizer according to claim 4 .
The selection instruction includes a reference selection instruction and an update selection instruction;
The correction history management unit is configured to select at least one of the plurality of correction history databases as a reference correction history database according to the reference selection command, and according to the update selection command. An update database selection control means for selecting at least one of the plurality of correction history databases as an update correction history database;
The dynamic prosody generation unit selectively refers to correction history information included in the reference correction history database,
The text-to-speech synthesizer, wherein the prosody modification unit selectively updates modification history information included in the update modification history database. A language analysis step for performing language analysis on text data and extracting language information;
A revision history holding step for holding revision history information;
A correction history management step for managing the correction history information;
A dynamic prosody generation step of generating dynamic prosody information based on the language information with reference to the correction history information via the correction history management unit;
The fixed prosody information is generated by correcting the dynamic prosodic information according to the external correction command, and the correction history information is generated via the correction history management unit based on the correction according to the external correction command. A prosody modification step for updating
A synthesized speech generating step for generating a synthesized speech based on the language information and the definite prosodic information;
The correction history management step includes a correction history extraction step for extracting correction history information referred to in the dynamic prosody generation step; a correction history update step for updating the correction history information held in the correction history holding step; Have
The revision history holding step further holds a plurality of revision history databases having different prosodic styles,
The text-to-speech synthesis method, wherein the modification history management step further includes a database selection control step for controlling selection of the plurality of modification history databases in response to a selection command . The text-to-speech synthesis method according to claim 9 .
The dynamic prosody generation step generates static prosody information according to a static prosody generation rule, and corrects the static prosody information according to the correction history information to generate the dynamic prosody information. A method for synthesizing text-to-speech. The text-to-speech synthesis method according to claim 9 .
The dynamic prosody generation step generates a modified prosody generation parameter by correcting a prosody generation parameter for determining a prosody generation rule according to the correction history information, and the prosody generation parameter according to the prosodic generation rule using the modified prosody generation parameter. A text-to-speech synthesis method characterized by generating dynamic prosodic information. The text-to-speech synthesis method according to claim 9 .
The dynamic prosody generation step selects any one of a plurality of prosodic pattern pieces as a selected prosodic pattern piece based on a static prosodic selection rule, and the selected prosodic pattern piece corresponds to the correction history information The text-to-speech synthesis method is characterized in that the dynamic prosody information is generated by correcting the information. A language analysis step for performing language analysis on text data and extracting language information;
A revision history holding step for holding revision history information;
A correction history management step for managing the correction history information;
A dynamic prosody generation step of generating dynamic prosody information based on the language information with reference to the correction history information via the correction history management unit;
The fixed prosody information is generated by correcting the dynamic prosodic information according to the external correction command, and the correction history information is generated via the correction history management unit based on the correction according to the external correction command. A prosody modification step for updating
A synthesized speech generating step for generating a synthesized speech based on the language information and the definite prosodic information;
The correction history management step includes a correction history extraction step for extracting correction history information referred to in the dynamic prosody generation step; a correction history update step for updating correction history information held in the correction history holding step; In a text-to-speech program that causes a computer to execute
The revision history holding step further holds a plurality of revision history databases having different prosodic styles,
The text-to-speech synthesis program characterized in that the revision history management step further comprises a database selection control step for controlling selection of the plurality of revision history databases in response to a selection command .

Images