JP2011191528A - Rhythm creation device and rhythm creation method - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To provide a rhythm creation device and a rhythm creation method for creating rhythm information which is a stable and natural rhythm. <P>SOLUTION: By referring to representative rhythm information 102 created based on input language information 101, and detailed rhythm information 103, from a detailed rhythm information storage section 2 which stores the detailed rhythm information 103 for indicating each of a plurality of rhythm feature patterns beforehand, the detailed rhythm information 103 similar to the representative rhythm information 102 is selected. <P>COPYRIGHT: (C)2011,JPO&INPIT

Description

  The present invention relates to a prosody creation device and a prosody creation method for creating prosody information of a synthesized speech in the case of creating a synthesized speech from a sentence input to a computer.

A text-to-speech synthesis system that mechanically creates a speech waveform from an arbitrary sentence has been developed. In general, a text-to-speech synthesis system includes a language processing unit, a prosody generation unit, and a speech waveform generation unit. The prosody generation unit performs processing related to the naturalness of speech inflection, rhythm, and volume (power) when creating a speech waveform in the system.
In order to generate natural inflection, a method has been proposed in which a large number of pitch patterns extracted from natural speech are used as they are (see, for example, Patent Document 1). In this method, a pitch pattern extracted from natural speech is stored in the prosodic database, and one optimal pitch pattern is selected from the prosodic database according to the linguistic information corresponding to the input text to generate a pitch pattern.
On the other hand, Patent Document 2 discloses a plurality of prosody databases that store a large number of pitch patterns extracted from natural speech based on linguistic information obtained by analyzing the text for each prosodic control unit of the text to be synthesized. A method of selecting a pitch pattern and generating a new pitch pattern therefrom is disclosed.

JP 2002-297175 A JP 2006-309162 A

  The conventional technique has a problem that when there are a plurality of pitch patterns having the same language information as the input text, an appropriate pitch pattern cannot be selected, resulting in an unnatural inflection. For example, when creating a pitch pattern by analyzing human-generated speech, the pitch of the voice varies depending on the recording time or the state of the speaker, even if the language information is the same, and the pitch of the voice is different. A pattern can be a selection candidate. In this case, an appropriate pitch pattern cannot be selected successfully based on language information.

  In addition, when selecting a pitch pattern based on the distance between the linguistic information of the input text and the linguistic information in the prosodic database, if there is no linguistic information similar to the linguistic information of the input text, the distance between the linguistic information is There is also a problem that an unnatural inflection occurs when an unexpected pitch pattern is selected.

  The present invention has been made to solve the above-described problems, and an object of the present invention is to obtain a prosody creation device and a prosody creation method for creating prosody information that is stable and natural.

  The prosody creation device according to the present invention represents a representative prosody information creation unit that creates representative prosody information indicating a pattern of prosodic features corresponding to the linguistic information from input language information, and shows a plurality of prosodic feature patterns, respectively. With reference to the detailed prosodic information storage unit that stores the detailed prosodic information in advance, the representative prosodic information created in the representative prosodic information creation unit and the detailed prosodic information stored in the detailed prosodic information storage unit, A detailed prosodic information selection unit that selects similar detailed prosodic information from the detailed prosodic information storage unit.

According to the present invention, the detailed prosodic information storage unit that preliminarily stores detailed prosodic information indicating each of a plurality of prosodic feature patterns refers to representative prosodic information and detailed prosodic information created based on the input language information. Then, detailed prosodic information similar to the representative prosodic information is selected.
In this way, since the detailed prosodic information similar to the representative prosodic information is selected, even if there are a plurality of patterns that become the same language information, the detailed prosodic information greatly deviated from the prosodic feature pattern indicated by the representative prosodic information is There is an effect that it becomes difficult to be selected and a stable prosody (no discontinuity between phonemes, words, and sentences for speech synthesis) can be obtained.

It is a block diagram which shows the structure of the prosody creation apparatus by Embodiment 1 of this invention. It is a figure which shows an example of the memory content of a detailed prosodic information storage part. It is a figure which shows an example of input language information. It is a figure which shows the representative prosody information created from the input language information in FIG. It is a figure which shows an example of output prosodic information. It is a figure explaining the process which weights the representative prosodic information and detailed prosodic information by a predetermined weighting ratio, and produces output prosodic information.

Embodiment 1 FIG.
FIG. 1 is a block diagram showing a configuration of a prosody creation apparatus according to Embodiment 1 of the present invention. 1, the prosody creation device according to the first exemplary embodiment includes a representative prosody information creation unit 1, a detailed prosody information storage unit 2, a detailed prosody information selection unit 3, and a mixed prosody information creation unit 4. The representative prosody information creation unit 1 is a component that inputs input language information 101 to the prosody creation device according to the first embodiment, and creates representative prosody information 102 with reference to the input language information 101. The detailed prosodic information storage unit 2 is a storage unit that stores a plurality of detailed prosodic information 103.

  The detailed prosodic information selection unit 3 is a component that selects the detailed prosodic information 103, selects the detailed prosodic information 103 similar to the representative prosodic information 102 from the detailed prosodic information storage unit 2 with reference to the representative prosodic information 102, The detailed prosodic information 104 is output. The mixed prosodic information creation unit 4 is a component that creates prosodic information in which the representative prosodic information 102 and the detailed prosodic information 104 are mixed. The prosodic feature pattern similarity between the representative prosodic information 102 and the detailed prosodic information 104 is determined. According to the degree of similarity obtained in the time series, either the representative prosodic information 102 or the detailed prosodic information 104 is set as the output prosodic information 105 to be output in the time series.

The detailed prosodic information storage unit 2 stores, as detailed prosodic information 103, a plurality of sound information and patterns of prosodic features respectively corresponding thereto. Here, the pattern of prosodic features is, for example, a pitch analysis (pitch of voice) that is a prosodic feature of speech uttered by humans, and four pitches per mora arranged in time order.
FIG. 2 is a diagram showing an example of the contents stored in the detailed prosodic information storage unit, showing sound information 201 and pitch pattern 202 of two utterances “stop” and one utterance “sold”. Yes. The sound information 201 of the utterance “I will quit” is “Chussy”, and the pattern 202 is a pitch of 24 points (6 mora × 4) arranged in time order. As shown in FIG. 2, even if the language information is the same, human speech is not completely the same, and thus the pitch pattern 203 and the pitch pattern 204 are different in shape.

The representative prosody information creation unit 1, the detailed prosody information selection unit 3, and the mixed prosody information creation unit 4 execute a prosody creation program according to the gist of the present invention in a computer so that the hardware and software cooperate. As a practical means, it can be realized by the computer.
The detailed prosodic information storage unit 2 is constructed in a storage device mounted on the computer, such as a hard disk device or an external storage medium.

Next, the operation will be described.
The principle of operation of the prosody generation apparatus according to Embodiment 1 will be described in detail. For the configuration of the prosody creation apparatus, refer to FIG.
First, the input language information 101 includes sound information, accent information, position information, and the like. Sound information, accent information, and position information can be obtained by analyzing kanji-kana mixed sentences using conventional known morphological analysis techniques.
FIG. 3 is a diagram showing an example of input language information, and shows language information obtained by morphological analysis of a kanji-kana mixed sentence “display a map”. In FIG. 3, the sound information 301 of “Map” is “Chizuo”, the accent type 302 is “1 type”, and the position information 303 is “first”. The sound information 301 of “display” is “Hyoojishi”, the accent type 302 is “5 type”, and the position information 303 is “second”.

The representative prosodic information creation unit 1 creates representative prosodic information 102 from the input language information 101. The representative prosodic information 102 is a pattern of prosodic features (pitch, power, rhythm).
For example, the pitch pattern can be created using a generally well-known point pitch model. The point pitch pattern is disclosed on pages 168-169 of Reference 1.
(Reference 1) Yukinori Takubo, Kikuo Maekawa, Haruo Kubozono, Kiyoshi Honda, Katsuhiko Shirai and Seiichi Nakagawa, “Iwanami Lecture Language Science 2 Speech”, published by Iwanami Shoten

  In the point pitch model, the pitch tends to decrease in the whole sentence, so this basic inclination pattern is represented by a straight line, and the accent component added on it is represented by a trapezoid, and each mora (almost equivalent to a syllable unit) The pitch of the center point is determined, and a pitch pattern is obtained by interpolating between the pitches of the center points.

FIG. 4 is a diagram showing representative prosody information created from the input language information in FIG. In FIG. 4, the horizontal axis represents time (Time), and the vertical axis represents voice pitch (pitch). In addition, sound information 401 is shown along the horizontal axis, and a basic inclination pattern 402 in the point pitch model is indicated by a broken line. A black circle plot is the pitch 403 of the center point of the mora. By interpolating between the pitches 403, a pitch pattern 404 is obtained.
In the representative prosody information 102 shown in FIG. 3, for example, the start point and end point of the basic gradient pattern 402 are determined by the maximum value of the position information 303 (“2” (th) in the case of FIG. 3). The accent component (the shape of the trapezoidal part) is determined by the number (“3” in the case of “Chizue”) and the accent information (type 1 in the case of “Chizue”).

The detailed prosodic information selection unit 3 compares the representative prosodic information 102 obtained as described above with the pattern stored in the detailed prosodic information storage unit 2 (detailed prosodic information 103), and a pattern similar to the representative prosodic information 102 This is selected as the detailed prosody information 104.
For example, the similarity between the two of the representative prosodic information 102 and the pattern stored in the detailed prosodic information storage unit 2 is calculated, and the detailed prosodic information 103 having a large similarity is used as the detailed prosodic information 104.
The similarity can be calculated using, for example, the distance d obtained by the following equation (1) by referring to the input language information 101, the representative prosody information 102, the detailed prosody information 103, and the corresponding language information. .
In the following formula (1), i represents time and w (i) represents weighting for each time i. F0t (i) indicates the pitch at time i of the representative prosodic information 102, and F0s (i) indicates the pitch at time i of the pitch pattern of the detailed prosodic information 103.

  In the above formula (1), the weight w (i) is the sound information corresponding to the time i of the input language information 101 and the sound information corresponding to the time i of the detailed prosodic information 103 stored in the detailed prosodic information storage unit 2. If (Language information) is the same, the value is small (0 or more and less than 1), and if it is different, the value is large (1 or more). That is, when sound information is similar, the distance d is reduced (similarity is increased) so that it can be easily selected. On the other hand, when the sound information is not similar, the distance d is increased (similarity is decreased) so that the selection is difficult.

The mixed prosody information creation unit 4 creates output prosody information 105 with reference to the representative prosody information 102 created by the representative prosody information creation unit 1 and the detailed prosody information 104 selected by the detailed prosody information selection unit 3. . Specifically, the similarity between the representative prosodic information 102 and the detailed prosodic information 104 is calculated, and either the representative prosodic information 102 or the detailed prosodic information 104 is set as output prosodic information 105 according to the similarity.
The similarity can be calculated using, for example, the distance d obtained by the above equation (1).
In this case, in the above formula (1), F0s (i) is the pitch at time i of the pitch pattern of the detailed prosodic information 104 selected by the detailed prosodic information selection unit 3.

Even in this case, the weight w (i) is a small value (0 to 1) if the sound information corresponding to the time i of the input language information 101 and the sound information corresponding to the time i of the detailed prosodic information 104 are the same. If it is different, a larger value is set (1 or more). That is, when the sound information is the same, the distance d is decreased, and conversely, when the sound information is different, the distance d is increased.
If the distance d is a section with a certain value or more, the representative prosody information 102 is the output prosody information 105, and if the distance d is less than a certain value, the detailed prosody information 104 is the output prosody information 105.

FIG. 5 is a diagram illustrating an example of output prosody information. In FIG. 5, when the distance d corresponding to the sound information “chizu wo” is a certain value or more (similarity is small), the portion corresponding to the sound information “chizu wo” in the output prosodic information 105 is representative prosodic information. 102 (a pitch pattern corresponding to “chizu wo” in FIG. 4). Further, the distance d of the portion corresponding to the sound information “Hyojoshi-ma” in the output prosodic information 105 is less than a certain value (high similarity), and the detailed prosodic information 104 is output.
FIG. 5 shows the output prosodic information 105 when the detailed prosodic information 104 is the pitch pattern 404.
As described above, the mixed prosodic information creation unit 4 converts the mixed prosodic information 102 and the detailed prosodic information 104 into the output prosodic information according to the similarity between the representative prosodic information 102 and the detailed prosodic information 104. 105 is output.

In the above description, the case where the input language information 101 is obtained by analyzing the kanji mixed sentence is described, but the input language information 101 may be created manually instead.
Further, although the case where the representative prosodic information 102 is created using the input language information 101 as sound information, accent information, position information, and the like has been shown, the present invention is not limited to this.
For example, the representative prosody information 102 may be created only from sound information. That is, the start point and end point of the basic gradient pattern are determined from the total number of sounds in the sound information, and the accent component (which is an average trapezoidal shape independent of the accent type) is determined from the number of sounds in each sound information. Good.

In addition, although the pitch pattern is used as an example of the representative prosody information 102, a power pattern representing the loudness of a voice or a rhythm pattern representing the length of each sound may be used instead.
In particular, when the representative prosodic information 102 is a pitch pattern, it is shown that it is created by a point pitch model. Instead, a pitch pattern is created by manually giving a pitch for each mora and interpolating these. May be.
Furthermore, feature quantities (pitch, power, length of one note) estimated from linguistic information using statistical learning methods (quantification type I model and regression tree model) are set for each mora, and these are interpolated. The pattern obtained in this way may be used as the representative prosody information 102.

In the above description, the case where the sound information corresponding to the pattern is stored in the detailed prosodic information storage unit 2 is shown, but the sound information may not be stored. That is, the detailed prosodic information storage unit 2 may store only prosodic feature patterns (pitch pattern, power pattern, rhythm pattern).
Further, although the patterns stored in the detailed prosodic information storage unit 2 are four feature amounts per 1 mora, they may be more or less. The number per mora may not be a fixed number, but may be a variable number depending on the length of sound.

  In the above description, the length of a phrase is used as the pattern length. However, the pattern length may be longer (for example, the length of a phrase or sentence), or conversely, may be shorter (for example, a phoneme). And the length of the morpheme).

  In addition, although the description has been made using Japanese (and mora) as the input language information 101, other languages such as English and Chinese are also applicable. For languages that do not have a mora, it may be controlled in units equivalent to mora or in syllable units.

  Furthermore, although the case where the pitch pattern (detailed prosody information 103) memorize | stored in the detailed prosodic information storage part 2 was obtained by pitch-analyzing the voice uttered from the person was described, it obtained by pitch-analyzing the voice uttered by the person. The pitch pattern may be created manually with reference to the pitch and stored in the detailed prosodic information storage unit 2.

最大ノルム

In the above description, the detailed prosodic information selection unit 3 calculates the similarity between the representative prosodic information 102 and the detailed prosodic information 103 with the weighted Euclidean distance expressed by the above formula (1). Alternatively, any known distance measure may be used instead.
For example, the calculation may be performed using the distance of the p-order average norm obtained by the following equation (2) or the distance of the maximum norm obtained by the following equation (3).
p-order average norm (eg, p = 1, 3, 4, etc.)

Maximum norm

  Further, as the similarity, a correlation coefficient between the representative prosodic information 102 and the detailed prosodic information 103 may be calculated, and the detailed prosodic information 103 having the maximum correlation coefficient may be used as the selection result (detailed prosodic information 104). .

In the above description, the mixed prosody information creation unit 4 calculates the similarity between the representative prosodic information 102 and the detailed prosodic information 104 using the weighted Euclidean distance expressed by the above formula (1). Alternatively, any known distance measure may be used instead.
For example, the calculation may be performed using the distance of the p-order average norm obtained by the above equation (2) or the distance of the maximum norm obtained by the above equation (3).

  Further, as the similarity, a correlation coefficient between the representative prosodic information 102 and the detailed prosodic information 104 may be calculated, and the detailed prosodic information 104 having the maximum correlation coefficient may be used as the selection result (output prosodic information 105).

  In the above description, the weight w (i) is set to a variable value by the detailed prosodic information selection unit 3, but may be a constant value (for example, 1) when there is no sound information in the detailed prosodic information storage unit 2.

Further, although the detailed prosodic information selection unit 3 sets the weight w (i) to a variable value according to the difference in sound information, it may be set to a variable value according to the difference in position information. In this case, if the position information is the same, the weight w (i) (for all times i) is set to a small value (0 to less than 1), and if the position information is different, the weight w (i) (all times i). Is set to a large value (1 or more).
Furthermore, the weight w (i) may be set by taking into account the difference between the sound information and the position information at the same time.

In the above description, the detailed prosodic information selection unit 3 calculates the similarity between the representative prosody information 102 and the pattern (the detailed prosodic information 103) stored in the detailed prosodic information storage unit 2, and the similarity Although the detailed prosodic information 104 is set to be small, it is not limited to this.
For example, a representative value (mode value, median value, average value) may be obtained from a plurality of detailed prosodic information 103 in which the degree of similarity between two parties is a certain value or less, and this may be used as the detailed prosodic information 104.

The mixed prosodic information creation unit 4 calculates the distance between the representative prosodic information 102 and the detailed prosodic information 104, and either the representative prosodic information 102 or the detailed prosodic information 104 is set as the output prosodic information 105 based on the similarity. However, the present invention is not limited to this.
For example, a weighted ratio is set according to the difference in sound information between the input language information 101 and the detailed prosodic information 104, and the weighted average of the representative prosodic information 102 and the detailed prosodic information 104 with this weighted ratio is used as the output prosodic information 105. Good.
FIG. 6 is a diagram for explaining a process of creating output prosodic information by weighted averaging representative prosodic information and detailed prosodic information at a predetermined weighting ratio. In FIG. 6, the output prosodic information 604 is created by performing weighted averaging of the representative prosodic information 601 and the detailed prosodic information 602 at a weighted ratio 603.
The weighted ratio 603 in FIG. 6 is set so that the ratio to the representative prosodic information 601 is large in the first half time section and the ratio to the detailed prosodic information 602 is large in the second half time section.
Comparing the sound information "Hyojishimasu" and "Chushishushi" one sound, there is a difference in the first half, and in the second half "Sushi", the sound is the same.
Thus, by setting a large weight ratio to the latter half pattern of the detailed prosodic information 602, the pattern of the detailed prosodic information 602 can be reproduced in the second half of the pattern of the output prosodic information 604.

As described above, according to the first embodiment, the representative created based on the input language information 101 from the detailed prosodic information storage unit 2 that stores in advance the detailed prosodic information 103 each indicating a plurality of prosodic feature patterns. The detailed prosodic information 103 similar to the representative prosodic information 102 is selected with reference to the prosodic information 102 and the detailed prosodic information 103.
By configuring in this way, the detailed prosodic information 103 similar to the prosodic feature pattern indicated by the representative prosodic information 102 is selected as the detailed prosodic information 104 constituting the output prosodic information 105. Even if there are a plurality of pieces, it becomes difficult to select the detailed prosodic information 104 greatly deviating from the prosodic feature pattern of the representative prosodic information 102. For this reason, a stable prosody (there is no discontinuity across phonemes, words, and sentences to be synthesized) can be obtained.
In addition, since the output prosodic information 105 becomes detailed prosodic information that is a prosodic feature pattern obtained by analyzing human speech, it is possible to generate a prosodic pattern that is close to the prosodic feature of speech produced by humans. As a result, natural prosody is obtained.
For example, when generating a pattern (detailed prosody information 103) to be stored in the detailed prosody information storage unit 2 by analyzing a voice uttered by a human, the voice feature fluctuates depending on the recording time and the state of the speaker, and the voice feature differs. Even if the detailed prosodic information becomes a selection candidate (in the case of pitch, for example, the height is different), by selecting with reference to the representative prosodic information 102, the prosodic feature pattern indicated by the representative prosodic information 102 deviates greatly. The detailed prosodic information 103 is less likely to be selected, and a stable prosody can be obtained.

  Further, even when there is no similar language information and the distance between the language information cannot be defined well, it is difficult to select the detailed prosodic information that is greatly deviated based on the representative prosodic information 102, and the output prosodic information 105 Since it becomes a pattern (detailed prosodic information) obtained by analyzing the emitted speech, there is an effect that a stable and natural prosody can be obtained.

Further, according to the first embodiment, the detailed prosodic information storage unit 2 stores the detailed prosodic information 103 and the language information corresponding thereto, and the detailed prosodic information selecting unit 3 includes the input language information 101, the representative prosodic information. 102, in calculating the similarity between the representative prosodic information 102 and the detailed prosodic information 103 with reference to the detailed prosodic information 103 and the language information corresponding thereto, the language corresponding to the input language information 101 and the detailed prosodic information 103 When the information matches, the similarity is calculated to be higher than when the information does not match.
Thus, by calculating the similarity with the input language information 101 taken into account, the detailed prosodic information 104 similar to the input language information 101 and similar to the representative prosodic information 102 is selected, and the input language information Prosody information that includes 101-specific prosodic features (pattern changes due to phonemes, differences in pattern shape at the beginning, middle, and end of sentences, etc.) can be easily obtained, and stable and natural prosodic information can be obtained. There is an effect that it is.

  Further, according to the first embodiment, the prosodic feature patterns of the representative prosodic information 102 and the detailed prosodic information 104 are sections in which both are not similar to the prosodic feature pattern of the representative prosodic information 102, and both are similar. Is provided with the mixed prosody information creation unit 4 that creates mixed prosody information mixed as a prosodic feature pattern of the detailed prosodic information 104 as output prosodic information 105, so that the selected detailed prosodic information 104 and representative prosodic information are included. When the prosodic information 102 is not similar, the representative prosodic information 102 is used as the output prosodic information 105, so that the detailed prosodic information 104 greatly deviating from the prosodic feature pattern of the representative prosodic information 102 is not selected, and stable prosodic information is obtained. There is an effect that it is.

In the first embodiment, the output prosodic information 105 is obtained by concatenating two or more detailed prosodic information 104, and the selected detailed prosodic information 104 and the representative prosodic information 102 are similar. However, when the output prosodic information 105 in this section is the representative prosodic information 102, the other detailed prosodic information 104 is selected as being similar to the representative prosodic information 102 (with a short distance).
For this reason, the continuity of connection between the representative prosodic information 102 and the detailed prosodic information 104 becomes high, and there is an effect that stable prosodic information can be obtained. In other words, the output prosodic information 105 obtained by connecting the detailed prosodic information 104 and the representative prosodic information 102 becomes prosody information with high continuity, and stable prosodic information is obtained.

Also, according to the first embodiment, the mixed prosody information is created as the output prosodic information 105 by mixing the prosodic feature patterns of the representative prosodic information 102 and the detailed prosodic information 104 by weighted averaging at a predetermined weight ratio. A mixed prosody information creation unit 4 is provided.
By having this configuration, the selection of the detailed prosodic information 104 greatly deviating from the prosodic feature pattern of the representative prosodic information 102 is suppressed, and the output prosodic information 105 is a pattern obtained by analyzing human-generated speech Since it is close to (detailed prosodic information), there is an effect that stable and natural prosodic information can be obtained.
In particular, when there are few patterns stored in the detailed prosodic information storage unit 2, there is a case where detailed prosodic information greatly deviated is selected. In this case, the representative prosodic information and the detailed prosodic information are weighted and averaged to obtain output prosodic information, so that selection of the detailed prosodic information that deviates greatly is suppressed, and stable prosodic information can be obtained.

  Furthermore, according to the first embodiment, the weight ratio is set according to the difference in sound information between the input language information 101 and the detailed prosodic information 104, and the representative prosodic information 102 and the detailed prosodic information 104 are weighted with this weight ratio. The mixed prosody information created on average is output as output prosody information 105. By doing so, prosodic information similar to sound information is reflected in the output prosodic information 105, and natural prosodic information can be obtained.

  DESCRIPTION OF SYMBOLS 1 Representative prosodic information creation part, 2 Detailed prosodic information storage part, 3 Detailed prosodic information selection part, 4 Mixed prosodic information creation part, 101 Input language information, 102,601 Representative prosodic information, 103,104,602 Detailed prosodic information, 105 , 604 Output prosodic information, 201, 301, 401 Sound information, 202 pattern, 203, 204, 404 Pitch pattern, 302 Accent type, 303 Position information, 402 Basic gradient pattern, 403 pitch, 603 Weighted ratio.

Claims (10)

A representative prosody information creating unit that creates representative prosody information indicating a pattern of prosodic features corresponding to the language information from the input language information;
A detailed prosodic information storage unit that preliminarily stores detailed prosodic information respectively indicating a plurality of prosodic feature patterns;
With reference to the representative prosody information created in the representative prosody information creation unit and the detailed prosody information stored in the detailed prosody information storage unit, detailed prosody information similar to the representative prosody information is stored in the detailed prosody information storage unit A prosody creation device comprising a detailed prosody information selection unit to select from. The detailed prosodic information storage unit stores the detailed prosodic information and linguistic information corresponding thereto,
The detailed prosodic information selection unit refers to the input language information, the representative prosodic information, the detailed prosodic information, and the language information corresponding thereto, and the similarity between the representative prosodic information and the detailed prosodic information 2, when the input language information and the language information corresponding to the detailed prosodic information match, the similarity is calculated to be larger than when they do not match. The prosody creation device described.   The prosodic feature patterns of the representative prosodic information created by the representative prosodic information creating unit and the detailed prosodic information selected by the detailed prosodic information selecting unit, and the sections where they are not similar are the prosodic feature patterns of the representative prosodic information And a mixed prosody information creating unit that creates mixed prosody information as a prosody information to be output in a section in which both are similar to each other so as to be a pattern of the prosodic features of the detailed prosodic information. The prosody creation device according to claim 1 or 2.   Mixed prosody information obtained by weighted averaging of prosodic feature patterns of the representative prosodic information created in the representative prosodic information creating unit and the detailed prosodic information selected by the detailed prosodic information selecting unit at a predetermined weight ratio, The prosody creation device according to claim 1, further comprising a mixed prosody information creation unit that creates the prosody information to be output.   5. The mixed prosody information creation unit sets a weighted ratio according to a difference in sound information between the input language information and the detailed prosody information selected by the detailed prosody information selection unit. Prosody creation device. In the prosody creation method for creating prosody information by the prosody creation device,
Representative prosodic information creation step of creating representative prosodic information indicating a pattern of prosodic features corresponding to the language information from the input language information;
From the detailed prosodic information storage unit that stores in advance detailed prosodic information each indicating a plurality of prosodic feature patterns, the representative prosodic information and the detailed prosodic information created by the representative prosodic information creating step are referred to. A prosody creation method comprising: a detailed prosodic information selection step for selecting detailed prosodic information similar to information. Storing the detailed prosodic information and the corresponding language information in the detailed prosodic information storage unit;
In the detailed prosodic information selection step, referring to the input language information, the representative prosodic information, the detailed prosodic information and the language information corresponding thereto, the similarity between the representative prosodic information and the detailed prosodic information 2. When calculating the linguistic information, the linguistic information is calculated so that the degree of similarity is greater when the linguistic information corresponding to the detailed prosodic information matches the linguistic information than when the linguistic information does not match. 6. The prosody creation method according to 6.   The prosodic feature patterns of the representative prosodic information created in the representative prosodic information creating step and the detailed prosodic information selected in the detailed prosodic information selecting step are the prosodic feature patterns of the representative prosodic information in sections where they are not similar A mixed prosody information creating step for creating mixed prosodic information as a prosodic information to be output in a section in which both are similar to each other so as to be a pattern of prosodic features of the detailed prosodic information. The prosody creation method according to claim 6 or 7.   Mixed prosodic information obtained by weighted averaging of prosodic feature patterns of the representative prosodic information created in the representative prosodic information creating step and the detailed prosodic information selected in the detailed prosodic information selecting step at a predetermined weighting ratio, 8. The prosody creation method according to claim 6 or 7, further comprising a mixed prosody information creation step for creating the prosody information to be output.   The weight ratio is set according to the difference in sound information between the input linguistic information and the detailed prosodic information selected in the detailed prosodic information selecting step in the mixed prosodic information creating step. Prosody creation method.

Images