JP4787769B2 - F0 value time series generating apparatus, method thereof, program thereof, and recording medium thereof - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To generate an F0 value time series inexpensively in accordance with various tones at low cost. <P>SOLUTION: A text in which a boundary position is determined for every accent phrase, and starting time and ending time are determined for each multiple accent types and moras, is input (3, S2). By using a rhythm event table, a plurality of rhythm events are related to an indicated position of the accent phrase according to the accent type (12, S8). By using a rhythm event table classified by tones, a rhythm event classified by tones, which corresponds to generation condition of the accent phrase is added (13, S10). A rhythm event parameter is created for each rhythm event from a rhythm event parameter data base (22, S12), and a delta function is created for each rhythm event from a creation function table (16, S16). An initial F0 value is calculated for each accent phrase (18, S18), and the F0 value time series is created for each accent phrase from the delta function and the initial F0 value (20, S20). <P>COPYRIGHT: (C)2008,JPO&amp;INPIT

Description

  The present invention belongs to the field of text-to-speech synthesis that generates synthesized speech from text, and in particular, an F0 value time-series generation device that generates a prosodic pattern of speech in order to give appropriate inflection to speech, its method, its program, And a recording medium thereof.

In the following description, the F0 value indicates the fundamental frequency of speech at a certain point in time, and the F0 value time series indicates a sequence of F0 values over the duration of the synthesized speech.
As a prior art 1, a method for generating a conventional F0 value time series for generating a synthesized speech, and applying a multi-space probability distribution HMM (MSD-HMM) based on a multi-space probability distribution, There is a method for modeling pitch and spectrum in a unified manner using feature parameters obtained by combining pitch parameters and spectral parameters. This is a method of generating a plausible F0 value time series from a model obtained by learning a temporal change and duration of the F0 value for each phoneme using a statistical model such as an HMM. Details are described in Non-Patent Document 1.

As a prior art 2, a generation process model that expresses an F0 value time series by combining a phrase component that temporarily falls for each pause phrase composed of a plurality of accent phrases and an accent component that is specified for each accent phrase Using this model, there is a technique for obtaining an F0 value time series obtained by inputting the descending parameter of the phrase component, the amplitude parameter of the accent component, the position parameter, and the like. Details are described in Non-Patent Document 2.
Here, the mora described in Non-Patent Documents 1 and 2 is a syllable unit of a sound having a certain time length in phonological theory. For example, in the case of “chocolate”, “cho”, “co”, “le”, “-”, and “to” are respectively mora.

An accent phrase is a linguistic unit including zero or one accent core, and is usually formed from one or more phrases. An accent core is a mora with an accent. Japanese accent phrases are roughly classified into three types (1) to (3) according to the position of the accent nucleus.
(1) Type 0 accent phrase: F0 value of the first mora is relatively low, and F0 values of the mora after the second mora are relatively high, that is, an accent phrase that does not include an accent nucleus.
(2) Type 1 accent phrase: an accent phrase in which the first mora has a relatively high F0 value, and the second and subsequent mora have relatively low F0 values, that is, the first mora corresponds to the accent core.
(3) n-type accent phrase (n is an integer greater than or equal to 2): the F0 value of the first mora is relatively low, the F0 values from the second mora to the nth mora are relatively high, and the (n + 1) th mora An accent phrase that is relatively low after the eye, that is, the nth mora from the beginning corresponds to the accent nucleus.

Non-patent document 3 describes a technique for giving a boundary to the accent phrase and a technique for giving an accent type for each accent phrase.
Also, as the prior art 3, a large amount of F0 value time series extracted from real speech is collected, and an F0 value time series that is syntactically similar to the synthesized speech to be generated is searched for and used. There is also a method using a template based on a case. Details are described in Patent Document 1.
All of these techniques have succeeded in synthesizing a natural sound to some extent.
IEICE Transactions D-IIl.J38-D-II.7July, 2000, pp1600-1609 "Pitch pattern generation by multi-space probability distribution HMM" Journal of the Acoustical Society (E) Vol.5, No.4 (1984) ”Analysis of voice fundamental frequency contours for declarative sentences of Japanese” Asano, Matsuoka, Takagi, Ohara "Method of setting prosodic information for speech synthesis using morphological analysis by multistage analysis and its word dictionary structure", Natural Language Processing Vol6, No.2, Jan, 1999 Japanese Patent No. 3420964

  All of the conventional methods are based on the so-called speech tone that is used by an announcer to read aloud sentences. However, the technology of text-to-speech synthesis is not only used for reading tone. For example, by synthesizing the tone of the telephone reception operator's response, a part of the work of the telephone reception operator can be replaced with a machine, sports news can be introduced lively, and the live situation can be seen. By generating synthesized voices that resemble complex tone, information that is not usually introduced by professional announcers, such as the results of grass baseball team games, can be converted to voice and broadcast on local broadcasts closely related to the area. Is possible.

  In consideration of synthesizing speech in a manner similar to various tone in this way, all the conventional methods have problems and are difficult to use as they are.

First Problem In the conventional technique 1, in the method of synthesizing the F0 value time series from the HMM, the F0 value time series is learned and synthesized for each phoneme. In such a case, in order to generate a voice resembling a new tone, it is necessary to learn the average F0 value for each phoneme, its differential component, and in some cases, the second-order differential component as a model parameter. The number increases. For this reason, it is necessary to collect enormous amounts of learning data required for statistical learning, which increases the cost.

  Next, problems of the prior art 3 will be described. In the method using a template based on the case, if the target tone when generating the synthesized speech changes, it is necessary to collect a large amount of the tone that matches the target, and to reconstruct the template. Similar to the method of synthesizing the F0 value time series, there is a problem that the cost increases. The first problem is a cost problem.

Second problem Next, the second problem will be described. In the case of using the generation process model as in the prior art 2, it is assumed that there is a component that gradually decreases. However, for example, in a tone that asks something to the other party, the voice F0 value rises toward the end of the word, and when speaking in a strong tone, it does not particularly drop and does not necessarily decrease gradually. Not exclusively. That is, the generation process model resembles a tone different from the reading speech phrasing tone, and there is a problem that when the synthesized speech is generated, the structure of the model mismatches with the features of the speech, and the correct expression may not be achieved. Therefore, as a second problem, there is a problem that when a synthesized speech is generated with a tone different from the reading speech phrasing tone, correct expression cannot be made.

  In the present invention, a boundary position and an accent type for each accent phrase are assigned to each accent phrase, and a text with a determined start time and end time for each mora is input to generate an F0 value time series of speech. The present invention relates to a sequence generation device. The F0 value time series generation apparatus of the present invention includes a prosodic event part and an F0 value time series part.

  The prosodic event unit generates a prosodic event from the accent type and the start time and end time for each mora using a prosodic event parameter table, and generates a prosodic event parameter for each prosodic event. The F0 value time series part generates an F0 value time series for each accent phrase using the prosodic event parameters and a predetermined generation function.

  The prosodic event unit may be composed of a prosody event generating unit, a tone-specific prosody event adding unit, and a prosodic event parameter generating unit. The prosodic event generation unit uses the prosodic event table to generate a plurality of prosodic events that are associated with the designated part of the accent phrase according to the accent type. The tone-specific prosodic event addition unit uses the tone-specific prosodic event table to add a tone-specific prosodic event corresponding to the occurrence condition to the specified location of the accent phrase if the accent phrase meets the occurrence condition. . The prosodic event parameter generation unit generates a prosodic event parameter for each prosodic event using the prosodic event parameter database and accent phrase information.

  Further, the F0 value time series unit may be configured by a delta function generation unit, an initial F0 value generation unit, and an F0 value time series generation unit. The delta function generation unit applies the prosodic event parameters to the generation function obtained from the generation function table for each prosodic event, and generates the sum of the generation functions corresponding to all the prosodic events as an accent phrase delta function. The initial F0 value generation unit obtains an initial F0 value for each accent phrase using the initial F0 value parameter database and accent phrase information. The F0 value time series generation unit generates an F0 value time series for each accent phrase from the delta function and the initial F0 value.

  Furthermore, the plurality of prosodic events may be ascending, descending, gently descending, and exciting.

  Furthermore, the prosodic event parameter database stores normalized prosodic event parameters (referred to as normalized prosodic event parameters), and converts the generated normalized prosodic event parameters according to mora information or accent phrase information. The prosodic event parameters may be output.

The fact that the first problem and the second problem are solved by the above configuration will be described.
First, the fact that the first problem has been solved will be described. The F0 value time series of the accent phrase is expressed only by the position parameter, the size parameter, the duration parameter for each of the plurality of prosodic events predetermined for each accent phrase, and the initial F0 value for each accent phrase. For example, in the accent phrase “in Kanagawa Prefecture”, six prosodic events are generated in the above configuration. Therefore, the F0 value time series of one accent phrase can be expressed by 3 × 6 + 1 = 19 parameters.

  In the F0 value time series generation method of the prior art 1, the average and variance for each F0 value, F0 value differential component, and second-order component differential of the F0 value are held for each phoneme, that is, 6 parameters are held. There is a need to. For example, if an F0 value time series of an accent phrase “in Kanagawa Prefecture” is to be generated, it is necessary to hold 6 parameters for each of 15 phonemes “KANAGAWAKEENDEWA”, so 90 parameters are used. There is a need.

  If the number of parameters to be used is small as in the configuration of the present invention, the number of learning data necessary to generate an appropriate parameter is reduced accordingly, and as a result, the cost of generating the F0 value time series Has the effect of lowering. Therefore, the problem of the prior art 1 can be solved.

  Further, in the configuration of the present invention as in the prior art 3, there is no concept of using a template, and naturally there is no need to reconfigure the template, and the problems of the prior art 3 can be solved.

  Next, the fact that the second problem has been solved will be described. As described above, in the related art 2, for example, an appropriate F0 value time series cannot be generated for a tone in which the F0 value increases at the end of the utterance and becomes a questionable tone. However, in the case of the configuration of the present invention, a prosodic event that generates a local movement of the F0 value time series is used, but a component that defines the movement of the entire utterance is not used. Therefore, if the F0 value at the end of the utterance is to be lowered, the “progress” type prosodic event may be used, and if the F0 value at the end of the utterance is to be increased, the “rising” type prosodic event may be used. Therefore, it is possible to generate an F0 value time series for synthesized speech resembling various tone. Therefore, the second problem can be solved by the configuration of the invention of the present application.

  The best mode for carrying out the invention will be described below.

  In this example, text is assumed as input. FIG. 1 is a diagram showing a functional configuration example of the first embodiment, and FIG. 2 is a flowchart showing a main processing flow of the first embodiment. In the following explanation, it is assumed that the input text is a question-like sentence “Are you sure?”.

  First, the text “Are you sure?” That is the target of generating the F0 value time series is input from the text input unit 3-1 (step S2). Further, a desired speed (hereinafter referred to as speech speed) of the F0 value time series to be generated is input from the speech speed input unit 3-2. In the following description, the speech speed is assumed to be 0.2 seconds / 1 mora.

  First, the accent phrase dividing / giving unit 2 assigns a boundary position to each accent phrase of the input text, and further assigns an accent type to each accent phrase (step S4). The contents of this processing are described in Non-Patent Document 3 above. For the text “Are you sure?”, A boundary line is added between the accent phrase “Now” and the accent phrase “Are you sure?”. Furthermore, an accent type is assigned to each of the accent phrase “Now” and the accent phrase “Are you sure?”, And a reading is also assigned. For the accent phrase “Now”, the third mora “de” becomes the accent core, and the accent type becomes type 3. As for “Are you sure?”, The third mora “shi” becomes the accent core, and the accent type becomes type 3. For example, as shown in FIG. 3, an accent type is given for each accent phrase. The accent phrase dividing / giving unit 2 outputs, for example, the format shown in FIG. The processing contents of the accent phrase dividing / giving unit 2 are described in Non-Patent Document 3.

  In the mora dividing / giving unit 4, the text is divided for each mora, and a start time and an end time are given to each mora (step S6). For simplification of description, the mora dividing / granting unit 4 divides all the lengths of one mora equally and assumes 0.2 seconds per mora which is the same as the speech speed. The method of mora division is not limited to this.

  As for “Is it OK?”, As shown in FIG. 4, “so” “re” “de” “ha” “yo” “ro” “shi” “−” “de” “su” “ka” Divided into mora. Furthermore, if the start time of the first mora “SO” is 0.11 seconds, the time per mora is 0.2 seconds, so the end time of the mora “SO” is 0.31 seconds. The next mora “re” has a start time of 0.31 seconds and an end time of 0.51 seconds. In this way, the start time and end time are assigned to all remaining mora as shown in FIG. For example, the mora dividing / giving unit 4 outputs the data in the format shown in FIG.

  In addition, if the input text is “Today is sunny and a pleasant day”, the mora to be divided, the start and end times given to each mora, the accent phrase, and the accent phrase The accent type is given as shown in FIG.

  The input text in which the boundary position for each accent phrase and the accent type for each accent phrase are given and the start time and end time for each mora are determined is input to the prosodic event generation unit 12. The prosodic event generation unit 12 uses the prosodic event table to generate a plurality of prosodic events corresponding to the accent type at the location where the accent phrase is specified (step S8).

  Here, the prosodic event is, for example, a command for generating four types of local movements such as a sudden rise, a sudden fall, a gentle fall, and a rise in the F0 value time series. The prosodic event table is stored in the prosodic event table storage unit 28. An example of the prosodic event table is shown in FIG. For example, if the accent phrase is type 0, prosodic events corresponding to prosodic event IDs 0 to 2, that is, a descending event, an ascending event, and a gently descending event are assigned to this accent phrase. A descending event is given at the start time of the first mora of the accent phrase, a rising event is given at the end time of the first mora, and a gentle descending event is given at the end time of the accent phrase, that is, the end time of the last mora. Is done. If the accent phrase is type 1 or type n, a prosodic event is similarly assigned to the generation location shown in FIG.

  Specifically, since the accent type of the accent phrase “Now” is type 3, prosodic events corresponding to prosodic event IDs 6 to 11 are given. Specifically, a descending event is given at the start time 0.11 second of the first mora “so”, and a rising event is given at the end time 0.31 second of the first mora “so”. In this way, the prosodic event generation unit 12 generates a plurality of prosodic events at designated locations for one accent phrase. FIG. 7A shows the prosodic event assigned to the accent phrase “Now”.

  Similarly, a plurality of prosodic events are given to the accent phrase “Are you sure?” As shown in FIG. 7B. The prosody event generation unit 12 outputs, for example, in the format of FIGS. 7A and 7B and inputs to the tone-specific prosody event addition unit 13.

  The tone-specific prosodic event adding unit 13 uses the tone-specific prosodic event table to add a tone-specific prosodic event corresponding to the occurrence condition to the specified location of the accent phrase if the accent phrase meets the occurrence condition. (Step S10). The tone-specific prosodic event table is stored in the tone-specific prosodic event table storage unit 30.

  FIG. 8 is an example of a tone-specific prosodic event table. For example, if the accent phrase corresponds to the occurrence condition “the particle“ ka ”exists at the end of the utterance”, that is, if the last mora of the accent phrase is the particle “ka”, the prosodic event ID is at the start time of “ka”. A rising event that is 100 is added.

  When the tone-specific prosodic event table is FIG. 8, the accent phrase “So” does not meet the occurrence condition, but the accent phrase “Are you sure?” Corresponds to the occurrence condition “The particle“ ka ”exists at the end of the utterance” ” . Therefore, a rising event (tone-based prosody event) is added to 2.11 seconds, which is the start time of “ka”. An example of the added result is shown in FIG. The prosodic event IDs shown in FIG. 6 and FIG. 8 are symbols used for convenience of explanation of this embodiment, and are not necessarily required when the invention is carried out.

  As described above, the prosodic event generation unit 12 and the tone-specific prosody event addition unit 13 do not use a global event such as the generation process model of the prior art 2 that presupposes an influence over the entire utterance. . In addition, the presence of the particle “ka” at the end of the utterance in the tone-specific prosodic event adding unit 13 means that this accent phrase is a questionable tone, and “ka” increases. Therefore, an accurate F0 value time series is generated even in such a questionable tone. In addition, even when a synthesized speech is generated in various tone different from the phonetic tone, for example, “natural tone”, the correct expression can be made depending on the setting of the tone-specific prosody event table. The point can be solved. The tone-specific prosody event adding unit 13 outputs, for example, the format shown in FIG. 9 and the prosody event parameter generating unit 14.

  The prosodic event parameter generation unit 14 generates prosodic event parameters for each prosodic event using the speech / linguistic situation at the location where the prosodic event parameter database and the prosodic event are associated with each other (step S14). The prosodic event parameter database is stored in the prosodic event parameter database storage unit 24. First, the prosodic event parameters will be described.

Each of the prosodic events generated by the prosodic event generating unit 12 and the tone-specific prosody event adding unit 13 is associated with a generation function corresponding to its type. As shown in a generation function table shown in FIG. 10 to be described later, for example, a generation function of the following equation (1) is associated with a rising event.

If it is a descending event, the following equation (2) is associated.

If it is a gentle descent event, the following equation (3) is associated.
-A (mt) σ ² exp (-σ (mt)) (3)

If it is a climax event, the following equation (4) is associated.

  In Expressions (1) to (4), t represents time, A, m, and σ represent prosodic event parameters, and A, m, and σ need to be generated. Specifically, A is an amplitude parameter that represents the amplitude of the function, and corresponds to the amount of intonation caused by the prosodic event. m is a position parameter indicating the position of the generation function, and represents how much the F0 value is actually changed by the prosodic event at a position shifted from the generation position associated with the prosodic event. σ is a duration parameter indicating how much time the change of the F0 value due to the prosodic event occurs.

  Different prosodic event parameters may be generated for the same type of prosodic event. For example, as shown in FIG. 5, when the input text is “Today is a sunny and pleasant day”, the accent type “Today is” and the accent type “Sunny” Are the same type. Accordingly, the prosodic event generation unit 13 is assigned events of prosodic event IDs 3, 4, and 5 shown in FIG. However, the accent phrase “Today” at the beginning of a sentence and the accent phrase “Sunny” in the middle of a sentence generally do not occur with exactly the same inflection. For example, the accent phrase “sunny” in the middle of a sentence is small, that is, it may be appropriate to reduce the value of the amplitude parameter A.

  Further, it is observed that the timing at which the F0 value starts to fall is different when comparing the case where there is a sound repellent “n” next to the accent nucleus and the case where there is no sound repellent. In such a case, it is necessary to appropriately generate the position parameter m of the prosodic event according to the situation. In this way, even for the same type of prosodic event, since the situation such as the beginning of a sentence, the middle of a sentence, or the presence of a repelling sound after an accent nucleus, the generated prosodic event parameters are different. May be different.

  While it is necessary to appropriately generate prosodic event parameters for each prosodic event according to these various situations, in the field of application of text-to-speech synthesis, synthesized speech can be used for any text. In consideration of the necessity to generate, the prosodic event parameter generation unit 14 must be able to generate appropriate prosodic event parameters for every situation.

  Therefore, as a method for generating appropriate prosodic event parameters corresponding to such various situations, the prosodic event parameter database is configured by, for example, a context clustering method using a decision tree for each prosodic event. It is possible to do.

  On the other hand, the speech / linguistic situation at the location associated with the prosodic event may be the situation of an accent phrase, for example. Therefore, as a method for generating prosodic event parameters, the following will be described using the prosodic event parameter database and the situation of the accent phrase.

FIG. 11 is an example of a decision tree that is a configuration of a prosodic event parameter database of rising events. As is clear from FIG. 11, the decision tree is, for example, a binary tree, and a question that can be answered with YES / NO is given to the node. If the answer to the question about the status of the generated prosodic event is YES, if the answer is NO to the right child node, if NO, the tree is traced to the left child node. Well, finally reach one of the leaves. Prosodic event parameters A, p, and q are designated for the leaves (final nodes). The position parameter p and the duration parameter q are values obtained by normalizing the position parameter m and the duration parameter σ, respectively (hereinafter, referred to as a normalized position parameter p and a normalized duration parameter q). This normalization will be described later.
If the structure of the prosodic event parameter database is such a decision tree, accurate prosodic event parameters can be generated for prosodic events in any situation.

  Next, a specific flow of prosodic event parameter generation processing will be described. The prosody event parameter generation processing of the rising event (hereinafter referred to as prosodic event 8201) having the reference number 8201 described in FIG. 9B will be described with reference to FIG. The accent phrase to which this rising event is added is "Are you sure?"

  First, regarding the accent phrase “Are you sure?”, It is examined whether or not the question “is it the phrase at the beginning of the sentence” of the node 601 that is the root node. The accent phrase “Are you sure?” Is not the phrase at the beginning of the sentence but the second phrase, so the answer is NO. The bus moves to the node 602 through the bus with the symbol “NO”.

  Next, the question of the node 602 is examined whether or not the current accent type is type 1. Since the accent type of the current accent phrase “Are you sure?” Is type 3, the answer is NO. The bus moves to the node 603 through a bus assigned with a code of NO.

  The node 603 question “whether the accent type of the immediately preceding phrase is type 0” or not is examined. The immediately preceding accent phrase is “Now,” and the accent phrase is type 3, so the answer is NO. The bus moves to the node 604 through a bus assigned with a code of NO. The node 604 is a leaf node, no question is given, and the values of the amplitude parameter A, the normalized position parameter p, and the normalized duration parameter q are described. Therefore, the prosodic event parameters of the prosodic event 8201 are generated as A = 2.2, p = −0.2, and q = 0.1.

  FIG. 11 shows a decision tree corresponding to a rising event. Similar decision trees are prepared for a falling event, a gentle falling event, and a rising event. Then, for all types of prosodic events, prosodic event parameters are generated for each prosodic event by the above processing using the decision tree corresponding to the prosodic event type.

  As shown in FIG. 12, the accent phrase “Are you sure?” Is the generation location, amplitude parameter A, normalized position parameter p, and normalized duration parameter q, which are associated with each prosodic event. Expressed as a set of two values. The prosody event parameter generation unit 14 outputs, for example, in the form of a table shown in FIG. 12 and inputs to the prosody event parameter conversion unit 22.

  Further, FIG. 11 shows a decision tree for collectively determining the amplitude parameter A, the normalized position parameter p, and the normalized duration parameter q, but it is also conceivable to construct a different decision tree for each parameter type. It is also conceivable to use a different decision tree for each prosodic event ID shown in FIGS. Further, in the example of FIG. 11, the question is related to the accent phrase accent type and the dependency relationship, but in addition to this, the position of the accent phrase in the input text or the prosodic event is generated. Questions can be considered from various points of view, such as morphological information and phonological information of words before and after a given location, or the sum of amplitudes of prosodic events generated before the target prosody event for which parameters are generated.

  In the prosodic event parameter conversion unit 22, the normalized prosodic event parameters (normalized position parameter p and normalized duration parameter q) generated by the prosodic event parameter generating unit 14 are prosodic according to mora information or accent phrase information. It is converted into an event parameter (step S14). Specifically, the normalized position parameter p and the normalized duration parameter q are converted into a position parameter m and a duration parameter σ, respectively. In the following description, a case where conversion is performed in accordance with mora information will be described.

  The normalization of the position parameter m and the duration parameter σ will be described. The unit of the position parameter p and the duration parameter q described above is a value normalized by the average mora length of the accent phrase including the corresponding prosodic event. For example, an accent phrase “Are you sure?” Includes seven mora in the accent phrase. Also, referring to FIG. 4 and the like, the accent phrase “Are you sure?” Has a start time of 0.91 seconds for the first mora “yo” and an end time of 2.31 seconds for the last mora “ka”. is there. Therefore, the duration of the accent phrase is 1.4 seconds, and the average mora length of the entire accent phrase is 0.2 seconds / mora. Moreover, the generation location of the rising event 902 illustrated in FIG. 12 is 1.11 seconds, and the normalized position parameter p is −0.2. This is from the end time of 1.11 seconds, which is the end time of the first mora “Yo”, to −0.2 mora, that is, −0.2 (normalized positional parameter of the rising event 902) × 0.2 (average mora length) = −0.04. That is, 1.07, which is 0.04 seconds before, is the value of the position parameter m.

  Similarly, the normalization duration parameter q is 0.1. In this case, 0.01 obtained by multiplying the average mora length of 0.1 is the value of the duration parameter σ. For other prosodic events, the normalized position parameter p and the normalized duration parameter q are converted, for example, a table as shown in FIG. 13 is generated and output from the prosodic event parameter converting unit 22, and the delta function generating unit 16 Is input.

  The unit of the normalized position parameter p and the normalized duration parameter q is not limited to the average mora length, and units such as seconds and milliseconds can be directly used. However, if absolute units such as seconds or milliseconds are used, the value of the prosodic event parameter is strongly influenced by the speech speed. For this reason, when generating synthesized speech corresponding to a faster or slower speech speed than usual, it is necessary to use a decision tree corresponding to the position parameter and duration parameter corresponding to the desired speech speed. In the event parameter database, it is necessary to prepare a large number of decision trees corresponding to a desired speech speed. Therefore, the cost for constructing the prosodic event parameter database increases, and enormous data must be stored in the prosodic event parameter database storage unit 24. Therefore, in order to express a stable position and duration regardless of the speech speed, the example of FIG. 12 uses the average mora length as a unit.

  The delta function generation unit 16 applies the prosodic event parameters A, m, and σ to a predetermined generation function for each prosodic event, and calculates the sum of the generation functions corresponding to all the prosodic events, thereby obtaining F0 in the accent phrase. A value time series delta function FD (t) is generated (step S16). Examples of the predetermined function include the above formulas (1) to (4), but are not limited thereto. In the following description, the predetermined generation function is described as the above formulas (1) to (4). The generation function table is stored in the generation function table storage unit 32, and FIG. 10 described above is an example of the generation function table.

  As shown in FIG. 10, the generation function and the general type of the generation function are associated with the type of prosodic event. For example, in the case of a descending event, the generation function is the above equation (2). The delta function generation unit 16 includes a generation function generation unit 162 and an addition unit 164.

  First, the generation function generation unit 162 generates generation functions corresponding to prosodic events for all prosodic events included in the input. The generated generation function is shown in FIG. Reference numbers 1001 to 1007 of the generation function shown in FIG. 14 correspond to the prosodic event reference numbers 901 to 907 of FIG. All the generated generation functions 1001 to 1007 are input to the adding unit 164.

  The addition unit 164 obtains the delta function FD (t) for the generation functions 1001 to 1007 by adding the sum of the times from the start time to the end time of the input accent phrase. An example of the delta function FD (t) is shown in FIG. The delta function FD (t) is a function obtained by differentiating the F0 value time series, that is, a function indicating increase / decrease of the F0 value time series. In this way, the delta function FD (t) is generated for each accent phrase.

  Note that the generation function of the descending event (the above formula (2)) and the generation function of the climax event (the above formula (4)) of the generation function of the generation function table shown in FIG. Can be expressed by the generation function (the above formula (1)). In the following description, the descending event generation function is obtained by adding “-” to the ascending event generation function.

  As for the generation function of the rising event, first, the duration of the generation function of the rising event and the falling event is halved, that is, σ is replaced with σ / 2. Then, the generation function of the rising event whose duration is ½ is moved in the negative direction by σ / 2, that is, m is replaced with m−σ / 2. Further, the generation function of the falling event whose duration is ½ is moved in the positive direction by σ / 2, that is, m is replaced with m + σ / 2. By adding the replaced rising event generating function and falling event generating function, a rising event generating function can be obtained. From the above, the generation function can be expressed by the generation function of the rising event (the above formula (1)) and the gentle generation function of the falling event (the above formula (3)).

  The initial F0 value generation unit 18 obtains an initial F0 value for each accent phrase using the initial F0 value parameter database and information on the accent phrase (step S18).

  The initial F0 value parameter database is stored in the initial F0 value parameter database storage unit 26. As the initial F0 value parameter database, for example, a binary tree configuration is conceivable as in the above-mentioned prosodic event parameter database. A configuration example of the initial F0 value parameter database is shown in FIG.

  The accent phrase “Are you sure?” Will be specifically described as an example. First, the question of the node 1201 as the root node “whether the current accent phrase is accent type 1” is examined. The accent type of the accent phrase “Are you sure?” Is type 3, and the answer is NO. Therefore, the vehicle moves to the node 1202 through the bus to which the symbol “NO” is assigned. Next, it is examined whether or not the question of node 1202 is “the accent type of the current phrase is type 0” or not. Since the answer is NO, the vehicle moves to the node 1204 through the bus to which the code of NO is assigned. Next, it is examined whether or not the question of the node 1204 is “beginning of sentence”. Since the accent phrase “Are you sure?” Is not the phrase at the beginning of the sentence, the answer is NO, and the bus moves to the node 1207 through the bus assigned the sign of NO. The node 1207 is a leaf node, no question is given, and an initial F0 value is described. Therefore, the initial F0 value of “Are you sure?” Is determined to be 5.2. The configuration example of the initial F0 value parameter database is not limited to the binary tree, and various configurations can be considered. In this way, the initial F0 value generation unit 18 determines the initial F0 value for each accent phrase and inputs it to the F0 value time series generation unit 20. Further, as shown in FIG. 17, the initial F0 value generation unit 18 may output the table shown in FIG. 13 and a combination of the delta function and the initial F0 value shown in FIG.

  In the F0 value time series generation unit 20, the F0 value time series for each accent phrase is calculated from the delta function for each accent phrase from the delta function generation unit 16 and the initial F0 value for each accent phrase from the initial F0 value generation unit 18. Is generated (step S20).

Specifically, for example, the initial F0 value is added to the integral value of the delta function FD (t) to generate the F0 value time series F (t) for each accent phrase. t is an arbitrary time between the start time and the end time. That is, the F0 value time series F (t) is generated by the following equation (5).

  Here, t1 indicates the start time of the accent phrase. The meaning of the integral operation on the right side is that, as described above, the delta function FD (t) is obtained by differentiating the F0 value time series, so that the F0 value time series is obtained by integrating the delta function FD (t). I can do it. FIG. 18 shows the calculation result of the above equation (5) that is the processing result in the F0 value time series generation unit 20, that is, the F0 value time series of the generated accent phrase “Are you sure?”.

  As described above, for example, in the case of the accent phrase “Are you sure?”, Three prosodic event parameters are generated for each of the initial F0 value and the seven prosodic events. Therefore, the F0 value time series can be expressed with only a total of 22 prosodic parameters. On the other hand, in the prior art 1, 6 parameters are required for every 13 phonemes “YOROSIDE ESUKA”, that is, 78 parameters are required. Therefore, in this embodiment, it becomes possible to generate the F0 value time series with a small number of parameters, and as a result, the cost can be reduced, and the first problem is solved.

  Further, the accent type of the accent phrase “Are you sure?” Is type 3, and in addition to the movement of the corresponding F0 value, the increase of the F0 value corresponding to the last “ka” is realized. Therefore, since the F0 value time series corresponding to various tone can be generated without being limited to the question tone, the second problem can be solved.

  In the second embodiment, in order to simplify the processing, the configuration of the prosody event generation unit 12, the tone-specific prosody event addition unit 13, the prosody event parameter generation unit 14, and the prosody event parameter conversion unit 22 described in the first embodiment is used. In combination, the F0 value time-series generating device 52 is operated as the prosodic event unit 54.

  FIG. 19 is a diagram illustrating a functional configuration example of the second embodiment. The prosodic event unit 54 generates a prosodic event from the start time and end time of the accent type and the mora type using the prosodic event parameter table, and further adds a tone-specific prosody event. Prosodic event parameters are generated.

  The prosodic event parameter table is stored in the prosodic event parameter table storage unit 29. FIG. 20 shows a prosodic event parameter table. For example, the prosodic event parameter table is obtained by integrating the prosodic event table shown in FIG. 6 and the tone-specific prosodic event table shown in FIG. 8 and adding prosodic event parameters corresponding to each prosodic event and tone-specific prosodic event. . In the prosodic event section, the prosodic event table shown in FIG. 6, the tone-specific prosodic event table shown in FIG. 8, and the prosodic event parameter database shown in FIG. 11 are not used.

  First, a boundary position and an accent type for each accent phrase are assigned to each accent phrase, and a text in which a start time and an end time for each mora are determined is input to the prosodic event unit 54. In the prosodic event unit 54, a prosodic event is generated according to the accent type of the accent phrase using the prosodic event parameter table, and at the same time, an amplitude parameter A, a position parameter m, and a duration parameter σ corresponding to the prosodic event are obtained. It is done. Subsequent processing by the delta function generation unit 16 and the like is the same as that in the first embodiment, and thus will be omitted.

  The second embodiment can be implemented at a lower cost than the first embodiment.

  In the third embodiment, the delta function generation unit 16, the initial F0 value generation unit 18, and the F0 value time series generation unit 20 described in the first embodiment are integrated into a F0 value time series unit 58 as a prosodic event parameter generation device 56. Is to be processed. FIG. 21 is a functional configuration example of the third embodiment.

  The F0 value time series unit 58 generates an F0 value time series for each accent phrase using the prosodic event parameters and a predetermined generation function. The predetermined generation function includes, for example, a rising event generation function (the above formula (1)), a gentle downward generation function (the above formula (3)), and the like. As described above, the generation function of the descending event and the generation function of the rising event can be obtained from the generation function of the rising event.

  First, the generation function corresponding to the prosodic event is obtained by the F0 value time series unit 58. In addition, for example, the initial F0 value is obtained by the above method, and the prosodic event parameter obtained by the prosodic event parameter converting unit 22 is applied to the generation function. Further, for example, the initial F0 value is obtained by the method described in the first embodiment, and the F0 value time series is obtained from these values.

  In the first embodiment, the F0 value time series is obtained by integration calculation after adding the generation function. In this third embodiment, the F0 value time series is obtained by adding the generation function subjected to the integral calculation. Etc. The third embodiment is effective in that the object is achieved even if the order of processing described in the first embodiment is not used.

  In the fourth embodiment, the F0 value time series generation device 60 is configured by the prosodic event unit 54 described in the second embodiment and the F0 value time series unit 58 described in the third embodiment. FIG. 22 is a functional configuration example of the fourth embodiment. Since the processing content is as described in the second and third embodiments, the description is omitted.

  In the process of F0 value time series generation processing described above, a time series of logarithmic values of F0 values is generated, and then an F0 value time series is synthesized using an exponential function. Therefore, as for the prosodic event parameter and the initial F0 value of the generation function, numerical values obtained by taking the logarithm of the F0 value are shown as examples. This is a process reflecting the knowledge that the change in the F0 value in the logarithmic region corresponds well to the change in audibility. Of course, even when a linear F0 value is used without using a logarithmic F0 value, if the numerical values included in the prosodic event parameter database and the initial F0 value parameter database are linear F0 values, the F0 value time series can be directly converted by the same processing. It is possible to generate.

  In addition to the above embodiments, the F0 value time-series generation apparatus according to the present invention is not limited to the above-described embodiments, and can be appropriately changed without departing from the spirit of the present invention. In addition, the processing described in the F0 value time-series generation apparatus is not only executed in time series in the order described, but also executed in parallel or individually as required by the processing capability of the apparatus that executes the processing. It is good.

  Further, when the processing in the F0 value time series generation device of the present invention is realized by a computer, the processing contents of the functions that the F0 value time series generation device should have are described by a program. Then, by executing this program on a computer, the processing function in the F0 value time-series generation device is realized on the computer.

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

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

  A computer that executes such a program first stores, for example, a program recorded on a portable recording medium or a program transferred from a server computer in its own storage device. When executing the process, the computer reads a program stored in its own recording medium and executes a process according to the read program.

  As another execution form of the program, the computer may directly read the program from a portable recording medium and execute processing according to the program, and the program is transferred from the server computer to the computer. Each time, the processing according to the received program may be executed sequentially. Further, the above-described processing may be executed by a so-called ASP (Application Service Provider) type service that realizes a processing function only by an execution instruction and result acquisition without transferring a program from the server computer to the computer. Good. Note that the program in this embodiment includes information that is used for processing by an electronic computer and that conforms to the program (data that is not a direct command to the computer but has a property that defines the processing of the computer).

  In this embodiment, the F0 value time series generation device is configured by executing a predetermined program on the computer. However, at least a part of these processing contents may be realized by hardware. Good.

The block diagram which shows the function structural example of Example 1 of this invention. The flowchart which shows the flow of the main processes of Example 1 of this invention. The figure which shows the output example of the accent phrase division | segmentation and provision part 2. FIG. The figure which shows the example of an output of the mora division | segmentation and provision part 4. FIG. The figure which shows the other output example of the mora division | segmentation and provision part 4. FIG. The figure which shows the example of a prosodic event table. The figure which shows the example of an output of the prosodic event generation part 12. FIG. The figure which shows the example of the prosodic event table classified by tone. The figure which shows the output example of the prosodic event addition part 13 according to a tone. The figure which shows the example of a production | generation function table. The figure which shows the structural example of a prosodic event parameter database. The figure which shows the example of an output of the prosodic event parameter production | generation part 14. FIG. The figure which shows the example of an output of the prosodic event parameter conversion part 22. FIG. The figure which shows the output example of the production | generation function production | generation part 162. FIG. The figure which shows the output example of the addition part 164. FIG. The figure which shows the structural example of an initial F0 value parameter database. The figure which shows the example of an output of the initial F0 value production | generation part 18. The figure which shows the output example of F0 value time series production | generation part 20. FIG. The figure which shows the function structural example of Example 2 of this invention. The figure which shows the example of a prosodic event parameter table. The figure which shows the function structural example of Example 3 of this invention. The figure which shows the function structural example of Example 4 of this invention.

Claims (10)

An F0 value time series generating device that generates a F0 value time series of speech by inputting a boundary position and an accent type for each accent phrase for each accent phrase, and inputting text with a determined start time and end time for each mora. There,
A prosodic event section that generates a prosodic event using the prosody event parameter table from the accent type and the start time and end time for each mora, and generates a prosodic event parameter for each prosodic event;
An F0 value time series part for generating an F0 value time series for each accent phrase using prosodic event parameters and a predetermined generation function;
Have
The prosodic events are ascending, descending, gentle descent, and excitement.
The generation function has the prosodic event

And
The F0 value time-series generation apparatus, wherein the prosodic event parameters are A, σ, and m. The F0 value time-series generation device according to claim 1,
The F0 value time series part is
For each prosodic event, a prosody event parameter is applied to the generating function obtained from the generating function table, and a sum of generating functions corresponding to all prosodic events is generated as a delta function of the F0 value time series in the accent phrase. When,
Using an initial F0 value parameter database and accent phrase information, an initial F0 value generation unit for obtaining an initial F0 value for each accent phrase;
An F0 value time series generation unit for generating an F0 value time series for each accent phrase from the delta function and the initial F0 value;
An F0 value time-series generation apparatus characterized by comprising: The F0 value time series generation device according to claim 1 or 2,
The prosodic event part is
In place of the prosodic event parameter table, using a prosodic event table, a prosodic event generating unit that generates a plurality of prosodic events that are associated with a designated part of an accent phrase according to an accent type;
If the accent phrase matches the occurrence condition using the tone-specific prosody event table, the tone-specific prosody event addition unit that adds the tone-specific prosody event corresponding to the occurrence condition to the specified location of the accent phrase,
A prosodic event parameter generating unit that generates a prosodic event parameter for each prosodic event, using a speech / linguistic situation at a location where the prosodic event parameter database and the prosodic event are associated,
Have
The F0 value time-series generating device , wherein the tone-specific prosodic events are ascending, descending, gentle descending, and rising . The F0 value time-series generation device according to claim 3 ,
The prosodic event parameter database stores normalized prosodic event parameters (hereinafter referred to as normalized prosodic event parameters),
The F0 value has a prosodic event parameter conversion unit that converts the normalized prosodic event parameter generated by the prosodic event parameter generation unit according to mora information or accent phrase information and outputs the prosodic event parameter Sequence generation device. An F0 value time series generation method in which a boundary position and an accent type for each accent phrase are assigned to each accent phrase, text having a determined start time and end time for each mora is input, and an F0 value time series of speech is generated. There,
The prosodic event means generates a prosodic event using the prosodic event parameter table from the accent type and the start time and end time for each mora, and generates a prosodic event parameter for each prosodic event;
An F0 value time series means for generating an F0 value time series for each accent phrase using a prosodic event parameter and a predetermined generation function;
Have
The prosodic events are ascending, descending, gentle descent, and excitement.
The generation function has the prosodic event

And
The F0 value time-series generation method, wherein the prosodic event parameters are A, σ, and m. The F0 value time series generation method according to claim 5,
The F0 time series process is
The delta function generating means applies the prosodic event parameters to the generating function obtained from the generating function table for each prosodic event, and the sum of the generating functions corresponding to all prosodic events is used as the F0 value time-series delta function in the accent phrase. The delta function generation process to generate,
An initial F0 value generating means for generating an initial F0 value for each accent phrase using the initial F0 value parameter database and accent phrase information;
An F0 value time series generating means for generating an F0 value time series for each accent phrase from the delta function and the initial F0 value;
An F0 value time series generation method characterized by comprising: The F0 value time series generation method according to claim 5 or 6,
The prosodic event process is
A prosodic event generating means for generating a plurality of prosodic events that are associated with a designated portion of an accent phrase according to an accent type using a prosodic event table instead of the prosodic event parameter table; ,
The tone-specific prosodic event addition means uses the tone-specific prosodic event table to add a tone-specific prosodic event corresponding to the occurrence condition to the specified location of the accent phrase if the accent phrase meets the occurrence condition. The process of adding prosodic events by tone,
The prosodic event parameter generating means generates a prosodic event parameter for each prosodic event using a speech / linguistic situation at a location where the prosodic event parameter database and the prosodic event are associated, and
Have
The F0 value time series generation method, wherein the tone-specific prosodic events are ascending, descending, gently descending, and exciting . The F0 value time series generation method according to claim 7 ,
The prosodic event parameter database stores normalized prosodic event parameters (hereinafter referred to as normalized prosodic event parameters),
The prosodic event parameter converting means has a prosodic event parameter converting process of converting the normalized prosodic event parameter generated in the prosodic event parameter generating process according to mora information or accent phrase information and outputting the prosodic event parameter The F0 value time series generation method characterized by the above.   An F0 value time series generation program for causing a computer to execute each process of the F0 value time series generation device according to claim 1.   A computer-readable recording medium on which the F0 value time series generation program according to claim 9 is recorded.

Images