JP2000221989A - Sound synthesizing device, regular sound synthesizing method, and memory medium - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide a sound synthesizing device, a regular sound synthesizing method, and a memory medium, capable of appropriately controlling the pause lengths by a simple technique without using a connection information etc., used in text analysis at all, and capable of obtaining a synthesized sound of natural tempo feeling. SOLUTION: This sound synthesizing device determines the pause lengths for regular sound synthesis by being equipped with a just-before-pause breath group length computing part 210 for counting the length of a breath group just before a breath group boundary that is not a non-pause, a just-after-pause breath group length computing part 211 for counting the length of a breath group just after a breath group boundary that is not a non-pause, an after-two-pause breath group length computing part 212 for counting the length of a breath group two pauses after a breath group boundary that is not a non-pause, a categorizing means 203 for categorizing the length of the respective breath groups, a pause-length learning/ predicting part 204 for predicting the length of pause time from the categorized length of the breath groups by using a pause time length weighted table obtained by the quantification class 1, a learning/prediction control part 207 for controlling the pause-length learning/ predicting part 204, and a pause length modifying part 206 for modifying a predicted value of pause length according to the speed of sound production.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a speech synthesizer for synthesizing an arbitrary speech according to rules, a rule speech synthesis method and a storage medium.
The present invention relates to a speech synthesis device, a rule speech synthesis method, and a storage medium in which a pause length determination method for rule speech synthesis capable of outputting synthetic speech with high naturalness is improved.

[0002]

2. Description of the Related Art In a speech synthesis system, a speech waveform itself is recorded, a speech waveform is created by combining the speech waveforms, and a parameter representing speech characteristics is analyzed and recorded. Some use. further,
There are various voice synthesis methods such as recording and editing, voice unit editing, parameter editing, and rule synthesis, depending on the combination of the methods of the control unit and the waveform forming unit.

[0003] Among them, the rule synthesis of speech converts a sequence represented by discrete symbols such as characters and speech symbols into continuous speech. In the process of conversion, universal properties and artificial properties of speech generation are applied as synthesis rules.

[0004] Text-to-speech synthesis converts text (text data) into speech, and usually does not correspond one-to-one with speech notation. Therefore, language processing such as morphological analysis and syntax analysis is required to convert the input text into a sequence of phonetic symbols and to automatically generate prosodic features.

Conventionally, a text-to-speech conversion apparatus for converting a text sentence into a voice and outputting the text includes a text analysis section and a rule speech synthesis section (a parameter generation section and a speech synthesis section).

The text analysis unit inputs a sentence mixed with kanji and kana, performs morphological analysis with reference to a word dictionary, determines reading, accent, intonation, and outputs phonetic symbols with prosodic symbols (intermediate language).

[0007] The parameter generation unit sets the pitch frequency pattern, the phoneme duration, and the like.
Performs speech synthesis processing.

The speech synthesis unit selects a speech synthesis unit from a target phoneme sequence (intermediate language) from speech data stored in advance and combines / deforms the speech according to the parameters determined by the parameter generation unit. Perform synthesis processing.

The speech synthesis units are phonemes, syllables (CV), V
There are units such as CV and CVC (C: consonant, V: vowel), and units obtained by expanding phoneme chains.

As a voice synthesis method, a pitch mark (reference point) is previously attached to a voice waveform, and the voice waveform is cut out at the center thereof, and at the time of synthesis, the pitch mark position is shifted according to the synthesis pitch cycle while shifting the synthesis pitch cycle. A superposition combining method is known.

In order to output a synthesized voice with higher naturalness by the text-to-speech conversion having the above-described configuration, the parameter (pitch) in the parameter generation unit is determined along with the manner of holding the unit of the voice unit, the unit quality, and the synthesis method. It is extremely important how to appropriately control the frequency pattern, phoneme duration, pause, and amplitude) so as to be close to natural speech. A pause is a short pause before and after a phrase.

As a pause setting method, there is a method described in Reference 1: Japanese Patent Laid-Open No. 6-59695 (ATR Automatic Translation and Telephone Research Institute).

[0013]

However, such a conventional pause length control method has the following problems.

That is, in the conventional pause length control method, since the pause length is controlled by using a local dependency relationship, a text analysis unit is required to realize this method. In the text analysis unit, syntactic analysis is required to obtain the dependency relationship, and there is a problem that the storage capacity of the word dictionary and the processing of the entire text analysis unit increase. In addition, there is a drawback that an incorrect syntax analysis has a direct adverse effect.

As described above, in the above-mentioned conventional speech synthesis method, it was difficult to determine the pause length by a simple method.

The present invention provides a speech synthesizing apparatus which can appropriately control a pause length by a simple method without using any dependency information or the like in text analysis and can obtain a synthesized speech with a natural tempo feeling. It is an object to provide a synthesis method and a storage medium.

[0017]

According to the present invention, there is provided a voice synthesizing apparatus for obtaining an arbitrary synthesized voice according to a rule, comprising: means for detecting a non-pause exhaled paragraph boundary; First counting means for counting the immediately preceding expiration paragraph length; second counting means for counting the expiration paragraph length immediately after the non-pause expiration paragraph boundary; and the expiration paragraph two times after the non-pause expiration paragraph boundary. Third counting means for counting the length, means for categorizing the first, second and third expiration paragraph lengths, and a pause time length weight table obtained by multivariate analysis from the categorized expiration paragraph lengths And a pause length determining means for predicting a pause time length.

The speech synthesis apparatus according to the present invention is a speech synthesis apparatus for obtaining an arbitrary synthesized speech according to a rule, wherein a means for detecting a non-pause exhalation paragraph boundary and an exhalation paragraph length immediately before the non-pause exhalation paragraph boundary are used. A first counting means for counting, a second counting means for counting the exhalation paragraph length immediately after a non-pause exhalation paragraph boundary, a means for categorizing the first and second expiration paragraph lengths, And a pause length determining means for predicting a pause time length using a pause time length weight table obtained by the multivariate analysis from the expired paragraph length.

The speech synthesizer according to the present invention is a speech synthesizer for obtaining an arbitrary synthesized speech according to a rule, wherein the first counting means for counting an exhalation paragraph length immediately before a pause symbol;
A second counting means for counting the length of the expiration paragraph immediately after the pause symbol; and a third means for counting the length of the expiration paragraph immediately after the pause symbol.
Counting means, means for categorizing the first, second and third exhalation paragraph lengths, and the categorized expiration paragraph lengths,
A pause length determining means for predicting a pause time length by using a pause time length weight table obtained by the multivariate analysis is provided.

The speech synthesizer according to the present invention is a speech synthesizer for obtaining an arbitrary synthesized speech according to a rule, wherein: a first counting means for counting an exhalation paragraph length immediately before a pause symbol;
Second counting means for counting the exhalation paragraph length immediately after the pause symbol, means for categorizing the first and second expiration paragraph lengths, and a pause obtained by multivariate analysis from the categorized expiration paragraph lengths A pause length determining means for predicting a pause time length using a time length weight table is provided.

According to the speech synthesizing apparatus of the present invention, in a speech synthesizing apparatus for obtaining an arbitrary synthesized speech according to a rule, a means for detecting a non-pause exhalation paragraph boundary and an exhalation paragraph length immediately before the non-pause exhalation paragraph boundary are used. First counting means for counting, second counting means for counting the length of the expiration paragraph immediately after the non-pause exhalation paragraph boundary, and third means for counting the exhalation paragraph length two times after the non-pause expiration paragraph boundary. And a pause length determining means for predicting a pause time length from each expiration paragraph length using a pause time length weight table obtained by multivariate analysis.

A speech synthesis apparatus according to the present invention is a speech synthesis apparatus for obtaining an arbitrary synthesized speech according to a rule, wherein a means for detecting a non-pause expiration paragraph boundary and an expiration paragraph length immediately before the non-pause expiration paragraph boundary are set. A first counting means for counting, a second counting means for counting the exhalation paragraph length immediately after the non-pause exhalation paragraph boundary, and:
A pause length determining means for predicting a pause time length by using a pause time length weight table obtained by the multivariate analysis is provided.

The speech synthesizer according to the present invention is a speech synthesizer for obtaining an arbitrary synthesized speech according to a rule, wherein: a first counting means for counting an exhalation paragraph length immediately before a pause symbol;
A second counting means for counting the length of the expiration paragraph immediately after the pause symbol; and a third means for counting the length of the expiration paragraph immediately after the pause symbol.
And a pause length determining means for predicting a pause time length from each expiration paragraph length using a pause time length weight table obtained by multivariate analysis.

The speech synthesizer according to the present invention is a speech synthesizer for obtaining an arbitrary synthesized speech according to a rule, wherein: a first counting means for counting an exhalation paragraph length immediately before a pause symbol;
Second counting means for counting the expiration paragraph length immediately after the pause symbol, and pause length determination means for predicting a pause time length from each expiration paragraph length using a pause time length weight table obtained by multivariate analysis. It comprises.

The pause length determining means may include a pause length correcting means for correcting and setting a predicted pause length value in accordance with the utterance speed.

Further, the rule speech synthesis method according to the present invention comprises:
A method for detecting a non-paused exhalation paragraph boundary, a step for counting the exhalation paragraph length immediately before the non-pause exhalation paragraph boundary, and a non-pause exhalation paragraph boundary. Counting the exhalation paragraph length immediately after the step, counting the exhalation paragraph length two times after the non-pause exhalation paragraph boundary, categorizing each exhalation paragraph length, and categorizing the exhalation paragraph length. And a step of predicting the pause time length using the pause time length weight table obtained by the multivariate analysis to sequentially determine the pause length for the ruled speech synthesis.

According to the rule speech synthesis method of the present invention, in a rule speech synthesis method for obtaining an arbitrary synthesized speech according to a rule, a step of detecting an exhalation paragraph boundary which is not non-paused, and an exhalation paragraph immediately before the non-pause exhalation paragraph boundary. Counting the length of the exhalation paragraph immediately after the non-pause exhalation paragraph boundary; categorizing each exhalation paragraph length; and multivariate analysis from the categorized exhalation paragraph length. And a step of predicting a pause time length by using the pause time length weight table to determine a pause length for regular speech synthesis.

According to a rule speech synthesis method according to the present invention, in a rule speech synthesis method for obtaining an arbitrary synthesized speech according to a rule, a step of counting an exhalation paragraph length immediately before a pause symbol, and Counting, counting the exhalation paragraph length after the pause symbol, categorizing each exhalation paragraph length, and a pause time length obtained by multivariate analysis from the categorized expiration paragraph length. And a step of predicting a pause time length using a weight table are sequentially performed to determine a pause length for rule-based speech synthesis.

According to a rule speech synthesis method according to the present invention, in a rule speech synthesis method for obtaining an arbitrary synthesized speech according to a rule, a step of counting an exhalation paragraph length immediately before a pause symbol, and A counting step, a step of categorizing each expiration paragraph length, and a step of predicting a pause time length from the categorized expiration paragraph length using a pause time length weight table obtained by multivariate analysis. The method is characterized in that it is executed to determine a pause length for rule speech synthesis.

According to the rule speech synthesis method of the present invention, in a rule speech synthesis method for obtaining an arbitrary synthesized speech by a rule, a step of detecting an exhalation paragraph boundary which is not non-paused, and an exhalation paragraph immediately before the non-pause exhalation paragraph boundary. Counting the length of the exhalation paragraph immediately after the non-pause exhalation paragraph boundary; counting the exhalation paragraph length two times after the non-pause exhalation paragraph boundary; and And a step of predicting the pause time length using a pause time length weight table obtained by the multivariate analysis are sequentially executed to determine a pause length for rule-based speech synthesis.

According to the rule speech synthesis method of the present invention, in a rule speech synthesis method for obtaining an arbitrary synthesized speech by a rule, a step of detecting an exhalation paragraph boundary which is not non-paused, and an exhalation paragraph immediately before the non-pause exhalation paragraph boundary. Counting the length of the exhalation paragraph immediately after the non-pause exhalation paragraph boundary; and, from each exhalation paragraph length, using the pause time length weight table obtained by the multivariate analysis to determine the pause time length. And a predicting step are sequentially performed to determine a pause length for rule-based speech synthesis.

According to a rule speech synthesis method according to the present invention, in a rule speech synthesis method for obtaining an arbitrary synthesized speech according to a rule, a step of counting an exhalation paragraph length immediately before a pause symbol, and Counting, counting the exhalation paragraph length after the pause symbol, and predicting the pause duration from each expiration paragraph length using a pause duration weight table obtained by multivariate analysis. Are sequentially executed to determine a pause length for rule-based speech synthesis.

In the rule speech synthesis method according to the present invention, in the rule speech synthesis method for obtaining an arbitrary synthesized speech according to a rule, the step of counting the length of the breath paragraph immediately before the pause symbol and the step of counting the length of the breath paragraph immediately after the pause symbol are performed. A counting step and a step of predicting a pause time length from each expiration paragraph length using a pause time length weight table obtained by multivariate analysis are sequentially performed to determine a pause length for rule-based speech synthesis. It is characterized by the following.

In the rule speech synthesis method according to the present invention, a step of predicting a pause time length and then correcting and setting a predicted pause length value according to the utterance speed may be executed.

The above-mentioned multivariate analysis may be a quantification type 1 for calculating an objective external criterion based on a qualitative factor, and also calculate an objective variable based on a plurality of explanatory variables. Multiple regression analysis may be used.

A storage medium according to the present invention is a storage medium that stores a program that is read by a computer as needed and executed on the computer.
A program for sequentially executing the rule speech synthesis method according to any one of 2 to 19 is stored.

[0037]

Embodiments of the present invention will be described below with reference to the drawings. First Embodiment FIG. 1 is a block diagram showing a configuration of a speech synthesis device and a rule speech synthesis method according to a first embodiment of the present invention. The speech synthesis method according to the present embodiment is an example applied to a text-to-speech conversion device that inputs text data.

In FIG. 1, reference numeral 101 denotes a text analysis unit;
102 is a word dictionary, 103 is a parameter generator, 104
Is a speech synthesis unit, 105 is a speech unit dictionary, 106 is a speech input unit, and 107 is a speech unit creation unit.

The text analysis unit 101 inputs a sentence mixed with kanji and kana, performs morphological analysis with reference to the word dictionary 102, determines reading, accent, intonation, and outputs phonetic symbols with prosodic symbols (intermediate language). . Accent and intonation are most closely related to the temporal pattern of pitch frequency.Pitch frequency patterns not only give natural and easy-to-hear tones, but also show the unity of words and phrases to understand sentence speech. It plays a role in facilitating.

The word dictionary 102 is, for example, a ROM or a RAM.
And stores a word dictionary and a word search table that defines the types of subsequent words that can be grammatically linked.

The parameter generation unit 103 receives the phonetic symbols with prosodic symbols (intermediate language) directly or from the text analysis unit 101 and uses them from the intermediate language itself.
A segment address in the segment dictionary 105 is selected, and a pitch frequency pattern, a phoneme duration time, an amplitude, a pause length, and the like are set. The parameter generation unit 103 is a main part of the present speech synthesis method, and will be described later in detail with reference to FIGS.

The unit dictionary 105 receives a natural speech signal and is created in advance by the unit creating unit 107.

The audio signal input unit 106 is an RS-232C
And the like, and an internal buffer for storing data, and natural speech data serving as a basis for a unit used for speech synthesis is input through the communication port and the FDD.

The segment creating section 107 labels a natural speech data in advance and creates a segment which is a base of a synthetic sound.

The speech synthesis unit 104 selects a segment in the segment dictionary 105 and synthesizes speech. Various conventional methods can be applied, for example, a waveform superposition method can be used.

The pitch period determined by the parameter generation unit 103 is the period of unit connection, and the duration of the phoneme is
This is realized by compression, expansion and contraction of the segments in the segment dictionary 105.

The word dictionary 102 and the speech synthesizer 1
04, the unit dictionary 105, and the unit creating unit 107 may be the same as those of the prior art.

FIG. 2 is a block diagram for explaining a method of determining a pause length for the synthesis of ruled speech by the parameter generation unit 103.

In FIG. 2, 201 is an intermediate language, 210
Is a pause immediately before expiration paragraph length calculation unit (first counting means) for counting the expiration paragraph length immediately before the non-pause expiration paragraph boundary, and 211 is immediately after the pause for counting the expiration paragraph length immediately after the non-pause expiration paragraph boundary. Expiration paragraph length calculation unit (second
Counting means), 212 is an expiration paragraph length calculating section (third counting means) for counting the expiration paragraph length two cycles after the non-pause expiration paragraph boundary, and 203 is each expiration paragraph length calculating section 210. A categorizing means 204 for categorizing the respective exhalation paragraph lengths 213 to 212, and a pause length for estimating a pause time length from the categorized expiration paragraph lengths using a pause time length weight table obtained by quantification class 1. Learning / prediction unit (pause length determining means), 205 is learning data, 20
7, a learning / prediction control unit for controlling the pause length learning / prediction unit 204; 208, utterance speed data; 206, a pause length correction unit (pause length correction means) for correcting a pause length based on the utterance speed data 208; Is the target pose length data. Quantification 1 will be described later.

FIG. 3 is a block diagram showing a detailed configuration of the pause length learning / prediction unit 204 and the pause length correction unit 206.

In FIG. 3, the pause length learning / prediction unit 20
4 includes a quantified first-class prediction model 301, a category quantity memory 302, and a quantified first-class learning model 303, and a pause length correction unit 206 includes a utterance speed ratio calculation unit 305.
And a pose length extension / contraction section 307. Also, 20
A learning / prediction control unit 7 controls the pause length learning / prediction unit 204.

The categorized first-class learning model 303 receives the learning categorized respective exhalation paragraph lengths 205, and the quantified first-class prediction model 301 receives the predicted categorized respective exhalation paragraph lengths 213. The estimated value data 304 is output to the pause length expansion / contraction unit 307 from the first type prediction model 301. On the other hand, the pronunciation speed data 208 is input to the utterance speed ratio calculation unit 305, and the utterance speed ratio 306 is output from the utterance speed ratio calculation unit 305 to the pause length expansion / contraction unit 307.
From the pose length expansion / contraction unit 307, target pose length data 209
Is output.

The operation of the pause length determination method in the speech synthesis method configured as described above will be described below.

First, text data such as a text file in a personal computer communication or a text file in a floppy disk (FD) is input via a communication port such as RS232C or an FDD, temporarily stored in an internal buffer, and exceeds a certain amount. Thus, the text is sent to the text analysis unit 101 for each unit (for example, for each sentence).

In the text analysis unit 101, a ROM or RA
M is read while collating the word dictionary of the word dictionary 102 composed of M with its text data, and generates a phonological prosodic symbol in which information such as accent, intonation, and pause is described as a character string, and this is output to the parameter generation unit 103. send.

The parameter generation unit 103 receives a phonetic symbol with a prosodic symbol (intermediate language) directly from the text analysis unit 101 and selects a segment address in the segment dictionary 105 to be used from the intermediate language itself. ,Also,
A pitch frequency pattern, a phoneme duration time, an amplitude, a pause length, and the like are set, and a synthesis parameter including these pieces of information is generated and sent to the synthesis voice unit 104.

The synthesized voice section 104 generates voice waveform data based on the generated synthesized parameters, and
/ A converter (not shown).

Next, the parameter generator 103 shown in FIG.
Will be described with reference to FIG.

The pause may be explicitly indicated by a symbol such as a comma (,) or the expiration paragraph boundary may be set as a pause because the expiration paragraph boundary is a part of breathing without explicitly indicating the pause. When the exhalation paragraph boundary is paused, a non-pause symbol indicating that no pause occurs at a specific exhalation paragraph boundary may be employed. In the following, a pose refers to a pose that is explicitly indicated by a pose symbol, or a part that is not non-paused when not explicitly indicated.

As shown in FIG. 2, an intermediate language 201, an expiration paragraph length calculation section 210 immediately before a pause, an expiration paragraph length calculation section 211 immediately after a pause, an expiration paragraph length calculation section 212 after 2 pauses,
Categorizing means 203, pose length learning / prediction unit 204,
It comprises learning data 205, learning / prediction control unit 207, utterance speed data 208, pause length correction unit 206, and target pause length data 209.

The operation is divided into (1) the case of model learning and (2) the case of prediction according to the instruction of the learning / prediction control unit 207. (1) In the case of model learning When model learning is instructed by the learning / prediction control unit 207, the expiration paragraph length calculation unit 210 immediately before the pause, the expiration paragraph length calculation unit 211 immediately after the pause, and the expiration paragraph length calculation unit 212 after two pauses. Thus, the exhalation paragraph length immediately before, immediately after, and two times after is counted. In counting the length, the number of mora is counted, and in consideration of the characteristics of the Japanese language, the sound repellent (N), the prompting sound (tsu), and the long sound are also counted in one beat. After that, these exhalation paragraph lengths are categorized by the categorizing means 203 as factors. The learning data 205 is created using these values and the measured pose length as external criteria.

In the pose length learning unit 204, the learning data 205 is input, and the measured pose length value is used as an external criterion.
Learn the quantification type 1 model. Learning this model of quantification type 1 is to determine the category quantity.

Quantification type 1 is one of multivariate analyses,
A target external criterion is calculated based on qualitative factors, and is formulated by the following equations (1) to (3).

When the factor item of the i-th data is j, the category to which it belongs is k, and the category quantity (coefficient given to the category) is x (jk), the predicted value y (i)
Is represented by Expressions (1) and (2).

[0065]

(Equation 1)

The above x (jk) is obtained by the least square method. That is, as shown in the equation (3), the square error between the predicted value y (i) and the actually measured value Y (i) is obtained so as to be minimized.

[0067]

(Equation 2)

In the solution, a normal equation can be obtained by partially differentiating the above equation (3) with x (jk) and setting it to zero.

As the actual calculation by the computer,
It can be reduced to a numerical analysis problem that solves simultaneous equations. (2) In the case of prediction On the other hand, in the case of prediction in accordance with an instruction from the learning / prediction control unit 207, the exhalation paragraph length calculation unit 210 immediately before pause and the exhalation paragraph length calculation unit immediately after pause from the input phoneme symbol string of the intermediate language 201. 211 and two pauses after expiration paragraph length calculation unit 212
By the above operation, each exhalation paragraph length 213 is counted. In counting the length, the number of moras is counted as in the case of the learning data, and the sound repellency (N), the prompting sound (tsu), and the long sound are also counted in one beat in consideration of the characteristics of Japanese.

After that, by the operation of the categorizing means 203, the exhalation paragraph length is classified into the same category as that at the time of learning, and is used as factor data (category exhalation paragraph lengths) as qualitative data. To enter. The pose length prediction unit 204 predicts a pose length, which is an external criterion, using the category quantity determined at the time of learning.

After that, the pose length correction unit 206
The pause length is corrected based on the utterance speed data 208 to obtain target pause length data 209.

Here, as the factor data, in the first embodiment, the expiration paragraph length immediately before the pause (number of mora), the expiration paragraph length immediately after the pause (number of mora), and the expiration paragraph length after two pauses (the number of mora). (Mora number).

Next, the operations of the pause length learning / prediction unit 204 and the pause length correction unit 206 will be described in detail with reference to FIG.

This operation is also divided into (1) the case of learning and (2) the case of prediction according to the instruction of the learning / prediction control unit 207. (1) In the case of learning According to the instruction of the learning / prediction control unit 207, in the case of learning, the quantification type 1 learning model 303 is activated from the categorization of each expiration paragraph (factor) for learning and the actually measured pause value data 205. Then, the equation (3) is solved, and the obtained category quantity is stored in the category quantity memory 302. (2) In the case of prediction In the case of prediction according to the instruction of the learning / prediction control unit 207, the categorization for each exhalation paragraph length (factor) data 21
3 is calculated from an input phoneme symbol string which is an intermediate language. For each of the predicted categorized expiration paragraph length (factor) data 213, the category quantity in the category quantity memory 302 is used as a weighting factor, and the quantified first-class prediction model 301
Is activated to calculate the equation (1), and obtain the estimated value data 304 of the pause length.

The pause correction unit 206 is composed of a utterance speed ratio calculation unit 305 and a pause length expansion / contraction unit 307. The operation is performed by inputting the current utterance speed and inputting the current utterance speed.
In step 05, the ratio 306 to the reference utterance speed at the time of learning is obtained and output to the pause length expansion / contraction unit 307. The pause length expansion / contraction unit 307 multiplies the pause length estimated value data 304 by a constant determined from the utterance speed ratio 306 to obtain the target pause length data 2.
09.

As described above, the speech synthesizing device according to the first embodiment calculates the expiration paragraph length immediately before a non-pause expiration paragraph boundary.
10. Exhaled paragraph length immediately after pause counting the exhalation paragraph length immediately after the non-pause expiration paragraph boundary 211, Pause two after expiration paragraph length to count the exhalation paragraph length two after the non-pause expiration paragraph boundary A calculating unit 212, a categorizing unit 203 for categorizing each exhalation paragraph length, and a pause length for estimating a pause time length from the categorized expiration paragraph length using a pause time length weight table obtained by quantification class 1. Learning / prediction unit 204, pose length learning / prediction unit 204
A learning / prediction control unit 207 for controlling the pause length and a pause length correction unit 206 for correcting the predicted pause length value according to the utterance speed are provided to determine the pause length for rule-based speech synthesis. An appropriate pause length can be obtained even at a more compact rule speech synthesis level that is not always necessary, improving the overall rhythmic feeling,
A synthesized voice with a natural tempo can be obtained.

Hereinafter, the effects of the present embodiment will be described specifically.

As a result of learning one female speaker using the ATR503 sentence database, the number of effective reading data 671 was obtained.
For each individual, there is a correlation between the exhalation paragraph length and the pause length immediately before, immediately after and two pauses later, and the estimation accuracy of the multiple correlation coefficient 0.562 and the mean square error 99.7 msec from the actually measured value is improved. Obtained.

As a result, the pause length can be appropriately controlled by a simple method without using any dependency information or the like in the text analysis, and there is an effect that a synthesized voice with a natural tempo feeling can be obtained.

Second Embodiment FIG. 4 is a diagram for explaining a method of determining a pause length for regular speech synthesis by a speech synthesis apparatus and a parameter generation unit of a regular speech synthesis method according to a second embodiment of the present invention. It is a block diagram. In the description of the speech synthesis method according to the present embodiment, the same components as those in FIG. 2 are denoted by the same reference numerals.

The present pause length determination method can be applied to the pause length determination method by the parameter generation unit 103 shown in FIG.

In FIG. 4, reference numeral 201 denotes an intermediate language;
Is a pause immediately before expiration paragraph length calculation section (first counting means) for counting the expiration paragraph length immediately before the non-pause expiration paragraph boundary, and 402 is immediately after the pause for counting the expiration paragraph length immediately after the non-pause expiration paragraph boundary. Expiration paragraph length calculation unit (second
Counting means), 203 is each exhalation paragraph length calculation unit 401, 4
A categorizing means 204 for categorizing each expiration paragraph length 213 of 02, and a pause length learning for estimating a pause time length from the categorized expiration paragraph length using a pause time length weight table obtained by quantification class 1. / Prediction unit, 205
Denotes learning data, 207 denotes a learning / prediction control unit, 208 denotes utterance speed data, 206 denotes a pause length correction unit, and 209 denotes target pause length data.

The detailed configurations of the pause length learning / prediction unit 204 and the pause length correction unit 206 are the same as those in the block diagram shown in FIG.

The operation of the pause length determination method in the speech synthesis method configured as described above will be described below.

The pause may be explicitly indicated by a symbol such as a comma (,), or the exhalation paragraph boundary may be set as a pause because the exhalation paragraph boundary is a point of breathing without explicitly indicating the pause. When the exhalation paragraph boundary is paused, a non-pause symbol indicating that no pause occurs at a specific exhalation paragraph boundary may be employed. In the following, a pose refers to a pose that is explicitly indicated by a pose symbol, or a part that is not non-paused when not explicitly indicated.

As shown in FIG. 4, an intermediate language 201, an expiration paragraph length immediately before a pause calculation unit 401, an immediately after expiration paragraph length calculation unit 402, a categorizing unit 203, a pause length learning /
Prediction unit 204, learning data 205, learning / prediction control unit 2
07, utterance speed data 208, pause length correction unit 206,
And target pose length data 209.

The operation is divided into (1) the case of model learning and (2) the case of prediction according to the instruction of the learning / prediction control unit 207. (1) In the case of model learning When model learning is instructed by the learning / prediction control unit 207, the expiration paragraph length immediately before and immediately after the pause is counted by the expiration paragraph length calculation unit 401 immediately before the pause and the expiration paragraph length calculation unit 402 immediately after the pause. Keep it. When counting the length, the number of mora is counted, and in consideration of the characteristics of Japanese, the sound repellency (N),
Prompting sounds and long sounds are counted as one beat. After that, these exhalation paragraph lengths are categorized by the categorizing means 203 as factors. The learning data 205 is created using these values and the measured pose length as external criteria.

In the pose length learning unit 204, the learning data 205 is input, and the measured pose length value is used as an external criterion.
The model of the quantification type 1 is learned (category quantity is determined). (2) In the case of prediction On the other hand, in the case of prediction in accordance with an instruction from the learning / prediction control unit 207, an expiration paragraph length calculation unit 401 immediately before a pause and an expiration paragraph length calculation unit immediately after a pause from an input phoneme symbol string of the intermediate language 201. By the operation of 402, each exhalation paragraph length 213 is counted. In counting the length, the number of moras is counted as in the case of the learning data, and the sound repellency (N), the prompting sound (tsu), and the long sound are also counted in one beat in consideration of the characteristics of Japanese.

After that, the operation of the categorizing means 203 classifies the exhalation paragraph length into the same category as that at the time of learning, and sets the data as qualitative data (categorical exhalation paragraph lengths) as qualitative data. To enter. The pose length prediction unit 204 predicts a pose length, which is an external criterion, using the category quantity determined at the time of learning.

The pause correction unit 206 is composed of a utterance speed ratio calculation unit 305 and a pause length expansion / contraction unit 307 as shown in the above 3, and its operation is to input the current utterance speed,
The utterance speed ratio calculation unit 305 obtains a ratio 306 from the reference utterance speed at the time of learning, and outputs the ratio 306 to the pause length expansion / contraction unit 307. The pause length expansion / contraction unit 307 multiplies the estimated pause length data 304 by a constant determined from the utterance speed ratio 306 to obtain target pause length data 209.

Here, in the second embodiment, the exhalation paragraph length immediately before the pause (number of mora) and the exhalation paragraph length immediately after the pause (number of mora) are used as the factor data.

As described above, the speech synthesizing apparatus and the rule speech synthesizing method according to the second embodiment include a pause-pre-expiration-paragraph length calculating unit 401 that counts the expiration paragraph length immediately before a non-pause exhalation paragraph boundary. A breathing paragraph length immediately after pause calculating section 402 for counting the breathing paragraph length immediately after the non-pause paragraph boundary, a categorizing means 203 for categorizing each breathing paragraph length, and a quantification class 1 from the categorized breathing paragraph length. Length learning / prediction unit 20 that predicts the pause time length using the pause time length weight table obtained by
4. A learning / prediction control unit 207 for controlling the pause length learning / prediction unit 204, and a pause length correction unit 206 for correcting the pause length prediction value according to the utterance speed, and determining a pause length for rule speech synthesis. Therefore, similarly to the first embodiment, the pause length can be appropriately controlled with a simple configuration, and a synthesized voice with a natural sense of tempo can be obtained.

Hereinafter, the effects of the present embodiment will be specifically described.

[0094] Female speaker 1 of ATR503 sentence database
As a result of learning by name, the number of valid reading data 671
An estimation accuracy of a multiple correlation coefficient of 0.44 and a mean square error of 106 msec from an actually measured value was obtained.

Therefore, similarly to the first embodiment, there is an effect that the pause length can be appropriately controlled with a simple configuration, and a synthesized voice with a natural sense of tempo can be obtained.

In addition to these effects, in particular, in the second embodiment, the category quantity memory 302, which is a weighting factor, is used without significantly lowering the pose length estimation accuracy compared to the first embodiment. Can be reduced to 2/3, and the amount of calculation such as addition at the time of prediction can be reduced to 2/3.

Third Embodiment FIG. 5 is a diagram for explaining a pause length determination method for rule-based speech synthesis by a speech synthesis apparatus and a parameter generation unit of a rule-based speech synthesis method according to a third embodiment of the present invention. It is a block diagram. In the description of the speech synthesis method according to the present embodiment, the same components as those in FIG. 2 are denoted by the same reference numerals.

This pose length determination method can be applied to the pause length determination method by the parameter generation unit 103 in FIG.

In FIG. 5, reference numeral 201 denotes an intermediate language;
Is a pause immediately before expiration paragraph length calculation unit (first counting means) for counting the expiration paragraph length immediately before the non-pause expiration paragraph boundary, and 211 is immediately after the pause for counting the expiration paragraph length immediately after the non-pause expiration paragraph boundary. Expiration paragraph length calculation unit (second
Counting means), 212 is a pause-two-expiration-paragraph length calculation unit (third counting means) for counting the expiration paragraph length two times after the non-pause expiration paragraph boundary, and 501 is a weight from each expiration paragraph length. Pose length learning / prediction unit (pause length determining means) for predicting a pause time length using a pause time length weight table obtained by regression analysis, 505 is learning data, 207 is a learning / prediction control unit, and 208 is utterance speed data. , 206 is a pose length correction unit, and 209 is target pose length data.

FIG. 6 is a block diagram showing a detailed configuration of the pause length learning / prediction unit 501 and the pause length correction unit 206. The same components as those in FIG. 3 are denoted by the same reference numerals.

In FIG. 6, a pose length learning / prediction unit 50 is used.
1 is a multiple regression prediction model 601, a regression coefficient memory 602
The pause length correction unit 206 includes a utterance speed ratio calculation unit 305 and a pause length expansion / contraction unit 307. Reference numeral 207 denotes a learning / prediction control unit that controls the pause length learning / prediction unit 204.

The multiple regression learning model 603 receives the learning breath length 505, the multiple regression prediction model 601 receives the prediction breath length 600, and the estimated value from the multiple regression prediction model 601. The data 304 is output to the pause length expansion / contraction unit 307. On the other hand, the utterance speed ratio calculation unit 305
, The utterance speed data 308 is input from the utterance speed ratio calculator 305 to the utterance speed ratio calculator 305.
07. The pose length expansion / contraction unit 307 outputs target pose length data 209.

The operation of the pause length determination method in the speech synthesis method configured as described above will be described below.

The pause may be explicitly indicated by a symbol such as a comma (,), or the exhalation paragraph boundary may be set as a pause because the exhalation paragraph boundary is a breathing point without explicitly indicating the pause. When the exhalation paragraph boundary is paused, a non-pause symbol indicating that no pause occurs at a specific exhalation paragraph boundary may be employed. In the following, a pose refers to a pose that is explicitly indicated by a pose symbol, or a part that is not non-paused when not explicitly indicated.

As shown in FIG. 5, the intermediate language 201, the expiration paragraph length calculation section 210 immediately before a pause, the expiration paragraph length calculation section 211 immediately after a pause, the expiration paragraph length calculation section 212 after two pauses,
It comprises a pause length learning / prediction unit 501, learning data 205, learning / prediction control unit 207, utterance speed data 208, pause length correction unit 206, and target pause length data 209.

The operation is divided into (1) the case of model learning and (2) the case of prediction according to the instruction of the learning / prediction control unit 207. (1) In the case of model learning When model learning is instructed by the learning / prediction control unit 207, the exhalation paragraph length immediately before, immediately after, and immediately after the pause is counted from the database. When counting the length, the number of mora is counted, and in consideration of the characteristics of Japanese, the sound repellency (N),
Prompting sounds and long sounds are counted as one beat. In addition, learning data 505 is created using the exhalation paragraph length as an explanatory variable and the measured pause length as an objective variable.

The learning data 505 is transferred to the pause length learning unit 5
01, the model of the multiple regression analysis is learned using the measured pose length as the objective variable, and the regression coefficient is determined.

The multiple regression analysis is one of the multivariate analyses,
The purpose variable is calculated based on multiple explanatory variables.
Formulated by the following equations (4) to (6).

When the objective variable y is modeled from n pieces of learning data using the explanatory variables x1, x2,..., Xp, the i-th data predicted value is expressed by the following equation (4).

[0110]

(Equation 3)

Here, a0, a1,..., Ap are selected so as to minimize the sum of squares of the prediction error ei.

In the solution, the equation (5) is converted to a0, a1,.
By partially differentiating each with p and setting it to zero, it can be reduced to a numerical analysis problem for solving simultaneous equations.

[0113]

(Equation 4)

A0, which minimizes the sum of squares of the prediction error ei,
The solutions (regression coefficients) of a1,..., ap are α0, α1,.
When p is set, the prediction model is represented by Expression (5).

[0115]

(Equation 5)

(2) In the case of prediction On the other hand, in the case of prediction according to the instruction of the learning / prediction control unit 207, the exhalation paragraph length calculation unit 210 immediately before the pause, the exhalation paragraph immediately after the pause, from the input phoneme symbol string of the intermediate language 201. Length calculation unit 211 and expiration paragraph length calculation unit 212 after two pauses
, Each exhalation paragraph length 600 is counted. In counting the length, the number of mora is counted in the same manner as in the learning, and at that time, in consideration of the characteristics of Japanese, the repellent sound (N), the prompting sound (tsu), and the long sound are also counted in one beat.

The pause length learning / prediction unit 501 inputs each expiration paragraph length 600, reads out the regression coefficient determined at the time of learning, calculates the linear sum of the above equation (6), and calculates the pause length as the objective variable. Predict.

The pause correcting section 206 receives the current utterance speed and corrects the pause length.

Here, as the factor data, in the third embodiment, the expiration paragraph length immediately before the pause (number of mora), the expiration paragraph length immediately after the pause (number of mora), and the expiration paragraph length after two pauses (the number of mora). (Mora number).

Next, the operations of the pause length learning / prediction unit 501 and the pause length correction unit 206 will be described in detail with reference to FIG.

This operation is also divided into (1) the case of learning and (2) the case of prediction according to the instruction of the learning / prediction control unit 207. (1) In the case of learning According to the instruction of the learning / prediction control unit 207, in the case of learning, in the case of learning, the multiple regression learning model 603 is obtained from the learning data 505 composed of each expiration paragraph length (explanatory variable) for learning and the actual measured value of the pose. Is started, the equation (5) is solved, and the obtained regression coefficient is stored in the regression coefficient memory 602. (2) In the case of prediction In the case of prediction according to the instruction of the learning / prediction control unit 207, each of the expiratory paragraph length (explanatory variable) data 600 for prediction is
First, it is calculated from an input phoneme symbol string which is an intermediate language. In the multiple regression prediction model 601, a regression coefficient memory 602
Are read out, the multiple regression prediction model 601 is started, and the above equation (6) is calculated, and the estimated value data 304 of the pause length is obtained.

The pause correction unit 206 is composed of a utterance speed ratio calculation unit 305 and a pause length expansion / contraction unit 307. First, the current utterance speed is input, and the utterance speed ratio calculation unit 305 performs the learning operation. Ratio to the standard utterance speed of 30
Then, in the pause length expansion / contraction unit 307, the pause length estimated value data 304 is multiplied by a constant determined from the utterance speed ratio 306.

As described above, the speech synthesizing device according to the third embodiment calculates the expiration paragraph length immediately before the non-pause expiration paragraph boundary.
10. Exhaled paragraph length immediately after pause counting the exhalation paragraph length immediately after the non-pause expiration paragraph boundary 211, Pause two after expiration paragraph length to count the exhalation paragraph length two after the non-pause expiration paragraph boundary The calculation unit 212, a pause length learning / prediction unit 501 that predicts a pause time length from each expiration paragraph length using a pause time length weight table obtained by multiple regression analysis, and learning that controls the pause length learning / prediction unit 501. The prediction control unit 207 includes a pause length correction unit 206 that corrects the pause length prediction value according to the utterance speed, and determines the pause length for ruled speech synthesis, as in the first embodiment. In addition, the pause length can be appropriately controlled by a simple method without using dependency information or the like in text analysis, and a synthesized voice with a natural tempo feeling can be obtained.

In particular, in the third embodiment, since the multiple regression model is used, each exhalation paragraph length is not categorized.
Variations in pose length due to variations in explanatory variables can be controlled more precisely, and highly accurate pose length prediction can be performed.

Fourth Embodiment FIG. 7 is a diagram for explaining a method of determining a pause length for regular speech synthesis by a speech synthesis apparatus and a parameter generation unit of a regular speech synthesis method according to a fourth embodiment of the present invention. It is a block diagram. In the description of the speech synthesis method according to the present embodiment, the same components as those in FIG. 4 are denoted by the same reference numerals.

This pose length determination method can be applied to the pause length determination method by the parameter generation unit 103 in FIG.

In FIG. 7, reference numeral 201 denotes an intermediate language;
Is a pause immediately before expiration paragraph length calculation section (first counting means) for counting the expiration paragraph length immediately before the non-pause expiration paragraph boundary, and 402 is immediately after the pause for counting the expiration paragraph length immediately after the non-pause expiration paragraph boundary. Expiration paragraph length calculation unit (second
A pause length learning / prediction unit 501 for predicting a pause time length from each expiration paragraph length using a pause time length weight table obtained by a multiple regression analysis; 505, learning data; The control unit 208 is utterance speed data, 206 is a pause length correction unit, and 209 is target pause length data.

The detailed configurations of the pause length learning / prediction unit 501 and the pause length correction unit 206 are the same as those in the block diagram shown in FIG.

Hereinafter, the operation of the pause length determination method in the speech synthesis method configured as described above will be described.

The pause may be explicitly indicated by a symbol such as a comma (,), or the expiration paragraph boundary may be set as a pause because the expiration paragraph boundary is a point of breathing without explicitly indicating the pause. When the exhalation paragraph boundary is paused, a non-pause symbol indicating that no pause occurs at a specific exhalation paragraph boundary may be employed. In the following, a pose refers to a pose that is explicitly indicated by a pose symbol, or a part that is not non-paused when not explicitly indicated.

As shown in FIG. 6, an intermediate language 201, an expiration paragraph length immediately before a pause calculation section 401, an expiration paragraph length immediately after a pause calculation section 402, a pause length learning / prediction section 501, learning data 505, a learning / prediction control section 207. Utterance rate data 2
08, a pose length correction unit 206, and target pose length data 209.

The operation is divided into (1) the case of model learning and (2) the case of prediction according to the instruction of the learning / prediction control unit 207. (1) In the case of model learning In the case of learning in accordance with an instruction from the learning / prediction control unit 207, the exhalation paragraph length immediately before and immediately after a pause is counted. In counting the length, the number of mora is counted, and at that time, in consideration of the characteristics of Japanese, the sound repellency (N), the prompting sound (tsu), and the long sound are also counted in one beat. The breath paragraph length immediately before and immediately after the pause and the measured pause length value are used as learning data, input to the pause length learning unit 501, and the measured pause length values are used as explanatory variables.
The equation (5) is solved, and a model for multiple regression analysis is learned (regression coefficients are determined). (2) In the case of prediction On the other hand, in the case of prediction in accordance with an instruction from the learning / prediction control unit 207, an expiration paragraph length calculation unit 401 immediately before a pause and an expiration paragraph length calculation unit immediately after a pause from an input phoneme symbol string of the intermediate language 201 By the operation of 402, each exhalation paragraph length is counted. In counting the length, the number of mora is counted in the same manner as in the learning, and at that time, in consideration of the characteristics of Japanese, the repellent sound (N), the prompting sound (tsu), and the long sound are also counted in one beat. The pause length prediction unit 501 inputs the expiration paragraph length, reads the regression coefficient in the regression coefficient memory 602 determined at the time of learning, activates the multiple regression prediction model 601, and executes the equation (6).
Is calculated, and a pause length, which is an objective variable, is predicted.

The pause correcting section 206 inputs the current utterance speed and corrects the pause length.

Here, in the fourth embodiment, the exhalation paragraph length immediately before the pause (number of mora) and the exhalation paragraph length immediately after the pause (number of mora) are used as the factor data.

As described above, the voice synthesizing apparatus according to the fourth embodiment calculates the expiration paragraph length immediately before a non-pause expiration paragraph boundary.
01, a pause immediately after expiration paragraph length calculation unit 402 that counts the expiration paragraph length immediately after the non-pause expiration paragraph boundary. Length learning / prediction unit 501 for predicting the length, a learning / prediction control unit 207 for controlling the pause length learning / prediction unit 501, and a pause length correction unit 206 for correcting the pause length prediction value according to the utterance speed. Since the pause length for speech synthesis is determined, similarly to the above embodiments, the pause length can be appropriately controlled with a simple configuration, and a synthesized voice with a natural sense of tempo can be obtained. effective. In addition, since multiple regression analysis is used, there is no quantization error when the factors are categorized, and the estimation accuracy is improved.

Further, in addition to these effects, in particular, in the fourth embodiment, the accuracy of estimating the pause length is greatly reduced as compared with the third embodiment. Without having to do this, the number-of-times coefficient memory 602, which is a weight quantity memory, can be reduced to 2/3, and the amount of calculations for multiplication and addition at the time of prediction can be reduced to 2/3, respectively. There is an effect.

In the speech synthesizing method and apparatus according to each of the above embodiments, a non-pause exhalation paragraph boundary is detected.
Although each exhalation paragraph length at the non-pause exhalation paragraph boundary is counted, each exhalation paragraph length of the pause symbol is counted by adopting a non-pause symbol indicating that a specific exhalation paragraph boundary does not enter a pause. The same effect can be obtained.

In the first and second embodiments, the length of each exhalation paragraph is calculated and then categorized. However, the calculation and categorization of the length of each exhalation paragraph are performed. It may be configured to execute at the same time.

The method of determining the pause length for the rule speech synthesis in each of the above embodiments may be realized by a dedicated hardware device (for example, a text speech synthesis LS)
A configuration that realizes the device in I) may be adopted. Also, floppy disks containing such software,
There may be no problem if a configuration is adopted in which a recording medium such as a CD-ROM is used to read data as needed and execute the program on a general-purpose computer.

The speech synthesis method and apparatus according to each of the above embodiments can be applied to any speech synthesis method using text data as an input. Any device may be used as long as it is a part of a circuit incorporated in various terminals.

Further, the number of dictionaries and various circuit units constituting the speech synthesizing method and apparatus according to each of the above embodiments, the form of the model, and the like are not limited to the above embodiments.

[0142]

According to the speech synthesizing apparatus and the rule speech synthesizing method of the present invention, a means for detecting a non-paused exhalation paragraph boundary and a first counting for counting the exhalation paragraph length immediately before the non-paused exhalation paragraph boundary. Means, a second counting means for counting the exhalation paragraph length immediately after the non-pause exhalation paragraph boundary, a means for categorizing the first and second expiration paragraph lengths, and a categorized expiration paragraph length, A pause length determining means for predicting a pause time length using a pause time length weight table obtained by quantification class 1 is provided, so that a simple method can be used without using any dependency information in text analysis. , The pause length can be appropriately controlled, and a synthesized voice with a natural sense of tempo can be obtained.

In the voice synthesizing apparatus and the rule voice synthesizing method according to the present invention, a means for detecting a non-paused exhalation paragraph boundary, and a first counting means for counting the exhalation paragraph length immediately before the non-paused exhalation paragraph boundary. Second counting means for counting the breath paragraph length immediately after the non-pause breath paragraph boundary;
A pause length determining means for predicting a pause time length using a pause time length weight table obtained by multiple regression analysis from each exhalation paragraph length is provided, so that the variation of the pause length due to the variation of the explanatory variable is further reduced. Fine control can be performed, and the pause length can be predicted with high accuracy, so that a synthesized voice with a more natural tempo can be obtained.

[Brief description of the drawings]

FIG. 1 is a block diagram showing a configuration of a speech synthesis device and a rule speech synthesis method according to a first embodiment to which the present invention is applied.

FIG. 2 is a block diagram for explaining a method of determining a pause length for regular speech synthesis by the speech synthesis apparatus and the parameter generation unit of the regular speech synthesis method.

FIG. 3 is a block diagram showing a detailed configuration of a pause length learning / prediction unit and a pause length correction unit of the voice synthesis device and the rule voice synthesis method.

FIG. 4 is a block diagram for explaining a pause length determination method for regular speech synthesis by a speech synthesis device and a parameter generation unit of a regular speech synthesis method according to a second embodiment of the present invention.

FIG. 5 is a block diagram for explaining a pause length determination method for regular speech synthesis by a speech synthesis device and a parameter generation unit of the regular speech synthesis method according to a third embodiment to which the present invention is applied.

FIG. 6 is a block diagram showing a detailed configuration of a pause length learning / prediction unit and a pause length correction unit of the voice synthesis device and the rule voice synthesis method.

FIG. 7 is a block diagram for explaining a pause length determination method for regular speech synthesis by a speech synthesis apparatus and a parameter generation unit of the regular speech synthesis method according to a fourth embodiment to which the present invention is applied.

[Explanation of symbols]

101 text analysis unit, 102 word dictionary, 103
Parameter generation unit, 104 speech synthesis unit, 105 unit dictionary, 106 speech input unit, 107 unit creation unit, 20
1 intermediate language, 203 categorization means, 204, 50
1 Pose length learning / prediction unit (pause length determination means), 20
5,505 learning data, 206 pose length correction unit (pose length correction means), 207 learning / prediction control unit, 208
Utterance speed data, 209 target pause length data, 21
0,401 Expiration paragraph length calculation section immediately before pause (first counting means), 211,402 Expiration paragraph length calculation section immediately after pause (second counting means), 212 Expiration paragraph length calculation section after 2 pauses (third counting section) Counting means), 301 quantification type 1 prediction model, 302 category quantity memory, 303 quantification 1
Class learning model, 305 utterance speed ratio calculation unit, 307 pause length expansion / contraction unit, 601 multiple regression prediction model, 602 regression coefficient memory, 603 multiple regression learning model

Claims (23)

[Claims] 1. A speech synthesizer for obtaining an arbitrary synthesized speech according to a rule, means for detecting a non-paused exhalation paragraph boundary, and a first counting unit for counting an exhalation paragraph length immediately before the non-pause exhalation paragraph boundary. Means, a second counting means for counting an exhalation paragraph length immediately after the non-pause exhalation paragraph boundary, and a third counting means for counting the exhalation paragraph length two times after the non-pause exhalation paragraph boundary. Means for categorizing the first, second, and third exhalation paragraph lengths, and a pause time length weight table obtained by multivariate analysis from the categorized expiration paragraph lengths to determine a pause time length. A speech synthesizer comprising: a pause length determining means for predicting the pause length. 2. A speech synthesizer for obtaining an arbitrary synthesized speech according to a rule, means for detecting a non-paused exhalation paragraph boundary, and a first counting unit for counting an exhalation paragraph length immediately before the non-pause exhalation paragraph boundary. Means, second counting means for counting the exhalation paragraph length immediately after the non-pause exhalation paragraph boundary, means for categorizing the first and second expiration paragraph lengths, and the categorized exhalation paragraph. A pause length determining means for predicting a pause time length using a pause time length weight table obtained by multivariate analysis from the length. 3. A speech synthesizer for obtaining an arbitrary synthesized speech according to a rule, wherein: a first counting means for counting a breath paragraph length immediately before a pause symbol; and a second counting means for counting a breath paragraph length immediately after a pause symbol. Counting means; third counting means for counting the exhalation paragraph length two times after the pause symbol; means for categorizing the first, second and third exhalation paragraph lengths; and the categorized exhalation A speech synthesis apparatus comprising: a pause length determining unit that predicts a pause time length using a pause time length weight table obtained by multivariate analysis from a paragraph length. 4. A voice synthesizing apparatus for obtaining an arbitrary synthesized voice according to a rule, wherein: a first counting means for counting a breath paragraph length immediately before a pause symbol; and a second counting means for counting a breath paragraph length immediately after a pause symbol. Counting means; means for categorizing the first and second expiration paragraph lengths; and, from the categorized expiration paragraph lengths, the pause time length is calculated using a pause time length weight table obtained by multivariate analysis. A speech synthesizer comprising: a pause length determining means for predicting the pause length. 5. A speech synthesizer for obtaining an arbitrary synthesized speech according to a rule, means for detecting a non-paused exhalation paragraph boundary, and a first counting unit for counting an exhalation paragraph length immediately before the non-pause exhalation paragraph boundary. Means, a second counting means for counting an exhalation paragraph length immediately after the non-pause exhalation paragraph boundary, and a third counting means for counting the exhalation paragraph length two times after the non-pause exhalation paragraph boundary. And a pause length determining means for predicting a pause time length from each of the expiration paragraph lengths using a pause time length weight table obtained by multivariate analysis. 6. A speech synthesizer for obtaining an arbitrary synthesized speech according to a rule, means for detecting a non-paused exhalation paragraph boundary, and a first counting unit for counting an exhalation paragraph length immediately before the non-pause exhalation paragraph boundary. Means, a second counting means for counting the expiration paragraph length immediately after the non-pause expiration paragraph boundary, and a pause time length weight table obtained by multivariate analysis from each expiration paragraph length. And a pause length determining means for predicting the length. 7. A speech synthesizer for obtaining an arbitrary synthesized speech according to a rule, wherein: a first counting means for counting a breath paragraph length immediately before a pause symbol; and a second counting means for counting a breath paragraph length immediately after a pause symbol. Counting means, third counting means for counting the length of the expiration paragraph after the pause symbol, and a pause time length from each of the expiration paragraph lengths using a pause time length weight table obtained by multivariate analysis. A speech synthesizer comprising: a pause length determining means for predicting the pause length. 8. A speech synthesizer for obtaining an arbitrary synthesized speech according to a rule, wherein: a first counting means for counting a breath paragraph length immediately before a pause symbol; and a second counting means for counting a breath paragraph length immediately after a pause symbol. A speech synthesizer comprising: counting means; and pause length determining means for predicting a pause time length from each of the expiration paragraph lengths using a pause time length weight table obtained by multivariate analysis. 9. The method according to claim 1, wherein the multivariate analysis is a quantification class for calculating a target external criterion based on qualitative factors. Speech synthesizer. 10. The speech synthesizer according to claim 5, wherein the multivariate analysis is a multiple regression analysis that calculates an objective variable based on a plurality of explanatory variables. 11. The apparatus according to claim 1, wherein said pause length determining means includes a pause length correcting means for correcting and setting a pause length predicted value in accordance with a utterance speed. Voice synthesizer. 12. A rule speech synthesis method for obtaining an arbitrary synthesized speech according to a rule, comprising: detecting a non-pause exhalation paragraph boundary; and counting an exhalation paragraph length immediately before the non-pause exhalation paragraph boundary. Counting the exhalation paragraph length immediately after the non-pause exhalation paragraph boundary; counting the exhalation paragraph length two times after the non-pause exhalation paragraph boundary; and categorizing each of the exhalation paragraph lengths. And a step of predicting a pause time length from the categorized expiration paragraph lengths using a pause time length weight table obtained by multivariate analysis to determine a pause length for rule speech synthesis. A rule speech synthesis method. 13. A rule speech synthesis method for obtaining an arbitrary synthesized speech according to a rule, comprising: detecting a non-paused exhalation paragraph boundary; and counting an exhalation paragraph length immediately before the non-paused exhalation paragraph boundary. Counting the exhalation paragraph length immediately after the non-pause exhalation paragraph boundary; categorizing each of the exhalation paragraph lengths; and, from the categorized exhalation paragraph length, a pause time length obtained by a multivariate analysis. And a step of predicting a pause time length using a weight table to sequentially determine a pause length for regular voice synthesis. 14. A rule speech synthesis method for obtaining an arbitrary synthesized speech according to a rule, comprising: a step of counting an exhalation paragraph length immediately before a pause symbol; a step of counting an expiration paragraph length immediately after a pause symbol; Counting the breath length after two, and categorizing each breath length, from the categorized breath length, using a pause time length weight table obtained by multivariate analysis, And a step of predicting a pause time length is sequentially performed to determine a pause length for the synthesis of a rule speech. 15. A rule speech synthesis method for obtaining an arbitrary synthesized speech according to a rule, wherein: a step of counting an exhalation paragraph length immediately before a pause symbol; a step of counting an expiration paragraph length immediately after a pause symbol; Categorizing paragraph lengths; and predicting pause time lengths from the categorized expiration paragraph lengths using a pause time length weight table obtained by multivariate analysis, sequentially executing rule speech synthesis. A ruled speech synthesis method characterized by determining a pause length for a speech. 16. A rule speech synthesis method for obtaining an arbitrary synthesized speech according to a rule, comprising: detecting a non-paused exhalation paragraph boundary; and counting an exhalation paragraph length immediately before the non-paused exhalation paragraph boundary. Counting the exhalation paragraph length immediately after the non-pause exhalation paragraph boundary; counting the exhalation paragraph length two times after the non-pause exhalation paragraph boundary; and multivariate analysis from each of the exhalation paragraph lengths. And a step of predicting the pause time length using the pause time length weight table obtained in step (a), to sequentially determine a pause length for the rule voice synthesis. 17. A rule speech synthesis method for obtaining an arbitrary synthesized speech according to a rule, comprising: detecting a non-paused exhalation paragraph boundary; and counting an exhalation paragraph length immediately before the non-pause exhalation paragraph boundary. Counting the exhalation paragraph length immediately after the non-pause exhalation paragraph boundary; and, from the respective exhalation paragraph lengths, estimating the pause time length using a pause time length weight table obtained by multivariate analysis. A rule speech synthesis method characterized by sequentially determining a pause length for rule speech synthesis. 18. A rule speech synthesis method for obtaining an arbitrary synthesized speech according to a rule, comprising: counting a breath paragraph length immediately before a pause symbol; counting a breath paragraph length immediately after a pause symbol; A step of counting the breath length after two, and a step of predicting the pause time length from each of the breath paragraph lengths using a pause time length weight table obtained by multivariate analysis. A method of synthesizing a rule speech, comprising determining a pause length for synthesis. 19. A rule speech synthesis method for obtaining an arbitrary synthesized speech according to a rule, wherein: a step of counting an exhalation paragraph length immediately before a pause symbol; a step of counting an exhalation paragraph length immediately after a pause symbol; And a step of predicting a pause time length using a pause time length weight table obtained by multivariate analysis from the paragraph length to sequentially determine a pause length for rule speech synthesis. Synthesis method. 20. The multivariate analysis is a quantification class for calculating a target external criterion based on qualitative factors.
A method for synthesizing a rule speech according to any one of the above. 21. The rule speech synthesis method according to claim 16, wherein the multivariate analysis is a multiple regression analysis for calculating an objective variable based on a plurality of explanatory variables. 22. The method according to claim 12, wherein after predicting a pause time length, a step of correcting and setting the pause length prediction value according to a utterance speed is executed. Rule speech synthesis method. 23. A storage medium storing a program read out by a computer as needed and executed on the computer, wherein the rule-based speech synthesis method according to claim 12 is sequentially executed. A storage medium storing a program.