JPH0836398A - Rhythm control system - Google Patents

Abstract

PURPOSE:To provide a rhythm control system in which high quality voice synthesis, that is close to a human voice, is conducted even though a small qualitative variation amount is used during a rule voice synthesis based on qualitative variation amount such as phoneme environment. CONSTITUTION:A pitch pattern of a mora is finely controlled by applying a quantification class I of multivariate analysis to the fundamental frequency control of the pitch pattern of a certain mora. When an analysis section 2 beforehand computes the number of categories for the quantized computational equations, item (qualitative variation amount) having higher partial correlation coefficients of accent patterns, etc., described by as mora, a proceeding mora and a suceeding mora are used. When a synthesis section 3 performs a rhythm control, the fundamental frequency of each pitch point at each mora of phoneme symbol column of voice synthesis object is successively computed. Thus, employing a smaller number of items and a smaller number of categories (classifications of each item), a fundamental frequency pattern of high quality voice close to a human voice is expressed.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a prosody control method relating to a fundamental frequency control method for phonological data input from a phonological database or the like to a speech synthesizing means in synthesizing arbitrary Japanese regular speech.

[0002]

2. Description of the Related Art In a conventional general speech synthesizer, for example, input text containing kanji and kana given from a host computer or the like is converted into a phoneme symbol string such as an accent phrase by Japanese processing, and this phoneme symbol string is converted into the phoneme symbol string. On the other hand, a prosodic control is performed by referring to a database that stores phoneme data such as mora fundamental frequency (pitch frequency), energy, and phoneme time length, and synthetic speech is obtained through an articulatory filter. This prosody control method was a modeling method.

As described above, the fundamental frequency, the phoneme duration and the energy are mainly used as the phoneme data used for the speech synthesis. Of these, particularly the fundamental frequency is a parameter relating to the accent of synthetic speech and is a portion that requires fine control in order to approach the speech generated by humans.

As one of the fundamental frequency control methods which enables such fine control, a method using quantification type I of multivariate analysis is known. This method calculates a quantitative variable (fundamental frequency) from a qualitative variable (phonological environment) and is formulated by the following equation. When j is the factor item of the qualitative variable of the i-th data, k is the category to which it belongs (classification of each item), and its category quantity (quantity (coefficient) given to the category) is ajk,

[0005]

[Equation 1]

Here, δ is a variable defined as follows.

[0007]

[Equation 2]

To predict the value {Yi} of the quantitative variate by the method of least squares

[0009]

(Equation 3)

The category quantity {ajk} is determined so as to satisfy the above.

The quantification type I is described in detail in the document "Handbook of PC statistical analysis II, multivariate analysis bias" (Yutaka Tanaka, Kyoyuki Tarumi, Kazumasa Wakimoto; Kyoritsu Shuppan).

[0012]

However, when the quantification type I of multivariate analysis is applied to the prosody control method of speech synthesis as described above, the influence of qualitative variables is very large. For example, even if the partial correlation coefficient (the degree to which the variable adapts to the quantitative variable) is low, the value of the partial correlation coefficient may change greatly depending on the data actually input.
Therefore, which item is effective for control cannot be uniquely determined. In addition, the coefficient (ajk) of the linear function (quantification calculation formula) that is finally output can be obtained well if input data that belongs to many categories is not given to all categories in one item to some extent. I had a problem such as not being. Therefore, the conventional prosody control method to which the quantification type I of the multivariate analysis is applied requires a large number of items and categories and has a problem that the partial correlation coefficient is unstable and high-quality speech synthesis is performed. There was room for improvement.

The present invention has been made in order to solve the above problems, and its object is to reduce a small number of qualitative variables when performing prosody control in regular speech synthesis based on a qualitative variable such as a phonological environment. However, it is to provide a prosody control method that enables high-quality speech synthesis close to that of humans.

[0014]

In order to achieve the above object, in the prosody control method of the first example of the present invention, the analysis learning means to which the quantification method of multivariate analysis is applied is used, From the input data for learning that uses the mora, the accent pattern described by the following mora, the mora position from the beginning of the mora, the mora position from the end of the mora, the pitch point position within the mora, one or more of the long sound flag and the consonant flag. In order to quantify the fundamental frequency of each pitch point of the mora using the item and its category as variables, calculate the category quantity that is the coefficient of the quantification calculation formula for each category of the item, and calculate the phoneme symbol string of the speech synthesis target. Obtain the category of the item of each mora, and calculate the mora of the mora from the quantification formula using the category quantity calculated for the category. Using calculating means for sequentially calculating the fundamental frequency of the pitch point as the estimation means of the fundamental frequency, it is characterized by using the speech synthesis the fundamental frequencies the calculated.

Further, in the prosody control method of the second example of the present invention, the analysis learning means to which the quantification method of the multivariate analysis is applied is used to increase or decrease the accent, and to change the position of the accent from the mora position. , The order of the mora from before the first mora of the input is the first, the order of the mora after the last of the input is the first, the attributes and phonological symbols when only the vowel part of the phoneme signal is viewed. The quantification formula for each category of the item in order to quantify the fundamental frequency of each pitch point of the mora using the item and its category as variables from the input data for learning with the attribute when viewing the consonant part of The category quantity that is a coefficient of is calculated, the category of the item of each mora of the phoneme symbol string of the speech synthesis target is obtained, and the category calculated for the category is calculated. Using sequential calculation calculating means fundamental frequency of each pitch point of the mora from quantification calculation formula using the amounts as estimating means of the fundamental frequency, it is characterized by using the speech synthesis the fundamental frequencies the calculated.

Further, in the prosody control system of the third example of the present invention, a fuzzy neural network having an antecedent part and a consequent part is used as a learning means to learn coefficient data from learning input data, and A fuzzy neural network in which coefficient data obtained by learning is set is used as pitch frequency estimation means, and the estimated pitch frequency is used for speech synthesis.

In the prosody control system of the above third example,
The input data for learning in the learning means is reconstructed corresponding to the membership function of the antecedent part of the fuzzy neural network and processed by the quantification type I to calculate the coefficient of the input / output function, and the coefficient is the consequent. It is preferable to set the coefficient as the initial value of the linear function of the section in order to further reduce the estimation error.

[0018]

In the prosody control method of the first example of the present invention, in the case of determining the pitch pattern in a certain mora in the regular speech synthesis, by applying the quantification method of the multivariate analysis to the fundamental frequency control, An item with a high partial correlation coefficient (qualitative variate) such as the accent pattern described in the preceding mora and the subsequent mora in addition to the mora by finely controlling the pitch pattern of the mora. By sequentially quantifying the fundamental frequency of each pitch point in each mora of the phoneme symbol string for speech synthesis using, the high-quality fundamental frequency pattern close to that of a human can be obtained with a small number of items and categories (classification of each item). It can be expressed from numbers.

Further, in the prosody control system of the second example of the present invention, in the prosody control system to which the quantification method of the multivariate analysis is applied, the input format is changed to narrow down the items, and thereby the fundamental frequency in the prosody control is obtained. The accuracy of the estimation of is improved to enable high-quality speech synthesis closer to that of a human.

Further, in the prosody control method of the third example of the present invention, the fuzzy neural network is used for learning and estimation of the fundamental frequency in prosody control, so that the relationship between the combinations of all categories and the output in one data is considered. It is possible to estimate the fundamental frequency, improve the accuracy of the estimation, and enable high-quality speech synthesis closer to human beings.

[0021]

Embodiments of the present invention will now be described in detail with reference to the drawings.

FIG. 1 is a block diagram showing the configuration of the first embodiment of the present invention. In the figure, 1 is a phoneme database, 2 is an analysis unit, and 3 is a synthesis unit. The phoneme database 1 has (a) pitch values (learning data) and (b) accent patterns (1: LHL, 2: LHH, 3: LL).
H, 4: LLL, 5: HLH, 6: HLL, 7: HH
8 patterns of L, 8: HHH (H: High, L: Lo
w)), (c) Mora position (1, 2, ...) From the beginning of the word,
(D) Mora position from the ending (1, 2, ...), (e) Pitch point position in the mora (7 points per mora), (f) Mora name or phoneme type (unvoiced plosive, voiced fricative, etc.) ), (G) long sound flag (whether there is a long sound), and (h) consonant flag (whether there is a consonant sound) are accumulated. Of the above,
(B) to (h) indicate items, and the description in () following the item name indicates the type of category. In the accent pattern category, the preceding mora, the relevant mora,
Classify by the combination of high / low (H / L) of the accent of the subsequent mora. The feature of the present embodiment is that (b)
The input data is described by using the item and category of (h), and the quantification type I of the multivariate analysis is applied.

In regular speech synthesis, when a pitch pattern in a certain mora is determined, its preceding mora,
The following mora, accent environment, and phoneme must be taken into consideration. Further, the pitch pattern of the mora requires fine control. When the quantified type I is used for this control, it suffices if the basic frequency pattern of the voice can be expressed with a small number of items and a small number of categories.

A method for determining a pitch pattern of a certain mora in response to the input of a phoneme symbol string will be described below.

First, in the analysis unit 2, the quantification type I of the multivariate analysis is executed by inputting the items (qualitative variables) of (b) to (h) of the phoneme database 1 in advance, and the pitch value of (a) is executed. (Quantitative variable) is learned to obtain each category quantity (coefficient ajk of input / output linear function). Analysis unit 2 for this
FIG. 2 shows an example of the block configuration of the above. Reference numeral 21 is an input data generation unit that generates input data from learning data, and 22 is a quantification type I calculation unit that calculates quantity data (category quantity) from the input data and stores it in a database or the like. FIG. 3 shows a control flow of the analysis unit 2. When estimating the pitch value (quantitative variable) of (a) from the qualitative variable with respect to the fundamental frequency at a certain position in the mora, it was shown by the conventional technique (1).
An expression is generated. Then, the pitch value at the next position is learned. When learning is completed for all data, the equation (2) shown in the prior art is also generated, and the category quantity can be obtained by the least square method.

Next, the voice synthesis in the synthesizer 3 will be described. FIG. 4 shows a block configuration example of the combining unit 3. Three
Reference numeral 1 is an input data generation unit that generates input data from test data, 32 is a pitch frequency estimation unit that obtains a pitch frequency from quantity data (category quantity) in a database,
Reference numeral 33 is a voice synthesizing section for performing voice synthesis with the obtained pitch frequency. FIG. 5 shows a control flow in the combining unit 3. At the time of speech synthesis, it is possible to estimate the fundamental frequency (pitch frequency) of the pitch point of each mora by calculating with a quantification formula of only the product-sum calculation generated by the category quantity. That is, the calculated category quantity is input, the data (2) to (8) of the phoneme database 1 is calculated from the target phonological phoneme symbol string of the input data, and the corresponding category quantity (coefficient ajk) is used (1). ) Is used to sequentially find the fundamental frequencies at all pitch point positions, and the pitch pattern is used to perform prosody control for speech synthesis.

FIG. 4 is a diagram showing the experimental results according to this embodiment. The learning value of the fundamental frequency of the pitch point performed by inputting 2741 points of the number of data from the real voice described in the six items shown in the table below and the estimated value of the fundamental frequency of the pitch point during speech synthesis It is a plot of the correlation of. The partial correlation coefficient of each item is as shown in the table below, and in particular, the partial correlation coefficient of each item of the accent pattern, the number of mora from the beginning and the number of mora from the end is high. Also,
The multiple correlation coefficient of the entire item is 0.525, which shows that it is effective to use this item and the category.

[0028]

[Table 1]

As described above, according to the method of this embodiment,
In regular speech synthesis, when determining a pitch pattern in a certain mora, its preceding mora, subsequent mora,
Accent environment, phoneme, etc. will be considered. Further, at that time, since the quantification type I of the multivariate analysis is applied to the fundamental frequency control, the pitch pattern possessed by the mora is finely controlled as compared with the conventional modeling method. . Further, in this embodiment, the number of items and the number of categories are smaller and the partial correlation coefficient is higher than in the conventional case where the quantified type I is used for the fundamental frequency control. Therefore, it is possible to express a high-quality fundamental frequency pattern of human voice.

Next, a second embodiment of the present invention will be described.

In this embodiment, the input frequency and the learning data are changed in the first embodiment, and the pitch frequency control method in the first embodiment using the quantification type I in the arbitrary Japanese rule speech synthesis is changed. It is intended to improve the error between the estimated value and the actual measured value in. That is, basically, this embodiment is also a method that uses the quantification type I as the pitch frequency control method in arbitrary Japanese rule speech synthesis,
The coefficient (quantity) of the input / output function is calculated using a statistical method in advance based on the phonological environment data, and the pitch frequency for an arbitrary mora of arbitrary Japanese is calculated using the coefficient at the time of synthesis. Is.

FIG. 7 shows data of 36965 points contained in 490 words uttered by one female speaker as learning data for obtaining the category quantity (coefficient of input / output function) in the first embodiment. Use as a test word 10
This is the result of estimation for the test word using the data of 9214 points in 0 word, and the vertical axis is the estimated value and the horizontal axis is the measured value. At this time, the multiple correlation coefficient was 0.76 and the average relative error was 15.40%.

In the above, the multiple correlation function is a value with which it is possible to evaluate how well it is predicted when a certain category quantity is used, and it is estimated as an observed value {y _i : i = 1, ..., N}. Equal to the correlation coefficient with the value {Y _i : i = 1, ..., N},
It is a value calculated by the following formula.

[0034]

[Equation 4]

The average relative error J is a value calculated by the following equation.

[0036]

(Equation 5)

As can be seen from FIG. 7, there is an estimation in which the error between the estimated value and the actually measured value is separated by 200 Hz or more. The possible causes are as follows.

First, as for the accent type of (b), the front mora, the relevant mora, and the rear mora only consider three mora. Also, information about whether the mora is high or low is expected to be very important, but there is no direct input for it. The method of considering the pitch point position in the mora of (e) as one of the input data cannot express the complicated pitch frequency for each position of the mora point in correspondence with the coefficient of the input / output function. (F),
Regarding the phoneme and long sound of (g), if the phonemes are classified into at most 6 patterns, the number of data belonging to one pattern increases, and the variation (dispersion) of the pitch frequency increases accordingly.

Therefore, in the second embodiment, the input format (item and category) is as shown below.

(1) Indicates whether the accent is rising or falling. Up is 2, down is 1.

(2) The number of moras after the mora position where the accent has changed is shown. Change point of accent 1
Count as the second. (1 to 11) (3) The first mora of the input is the first, and the number of mora from the front is shown. (1 to 13) (4) The last mora of the input is the first, and the number of mora from the back is shown. (1 to 13) (5) Looking only at the vowel part of the phoneme signal / A / I / U / E / O / A- / I- / U- / E- / O-
Indicates one of /. Each numerical value takes a value of 1 to 11 from the front.

(6) Looking at the consonant part of the phonological symbol, / B, D, G / / BY, DY, GY / / Z, J / / N, M / / NY, MY / / R / / RY / / W / / WY / / P, T, K, TS, CH / / PY, KY / / F, H, S, SH / / FY, HY / / no consonant / It belongs to //. Numerical values take values from 1 to 14 from the top.

In the above input format, (1) directly gives rise and fall of accents and is a very important input.

(2) is the position from the accent change point,
Shows how many rising (or falling) accents have continued.

(3) and (4) indicate which mora position in the entire input as a set.

In (5) and (6), phonological symbols are finely classified into vowel parts and consonant parts, which reduces the number of data belonging to one input pattern and reduces the variation in pitch frequency to some extent.

Further, it is ideal to learn one pitch frequency for one input pattern. Therefore, in order to further reduce the variation of the pitch frequency for one input pattern of the learning data, the average value and the standard deviation of the pitch frequency for all the input patterns are obtained, and the data of the pitch frequency deviated from the average value by the amount of the standard deviation is obtained. Was deleted and used as learning data. By doing so, a pitch frequency that is determined to some extent is learned for one input pattern.

Further, in order to express the pitch frequency with respect to the intra-mora point position as versatilely as possible in correspondence with the coefficient of the input / output function, a data set corresponding to each intra-mora pitch point position is prepared and learned. , Will be estimated.

FIG. 8 shows the flow of learning processing in this embodiment.
Shown in the block. In the figure, 23 is an input data generation unit that generates input data from learning data, 24 is a data reduction unit that deletes from the input data data of pitch frequencies that are apart from the average value by the amount of standard deviation, and 25 is reduced. It is a quantification type I calculation unit that calculates quantity data (category quantity) from input data and stores it in a database or the like. The loop 26 from the quantification type I calculation unit 25 to the input data generation unit 23 side is for performing the calculation of the quantitative data for all the pitch points in the mora. The voice synthesis is performed in the same manner as in the first embodiment. However,
The different input formats are as described above.

FIG. 9 shows the estimation result of the test data in the second embodiment. The learning data is the data included in the 490 words uttered by one female speaker, which is the same as the one shown in the case of FIG. 7, and is extracted for one mora inner point position and deleted in the width of the standard deviation of the pitch frequency. 424
Two points of data. The test data is the data of 1476 points extracted for one mora inner point position in 100 words. At this time, the multiple correlation coefficient was improved to 0.90 and the average relative error was improved to 11.6%.

Next, a third embodiment of the present invention will be described.

FIG. 10 shows another estimation result according to the input format of the second embodiment. This result is obtained by learning with the data of 4242 points included in 490 words uttered by one female speaker to obtain the quantity (coefficient of the input / output function), and using the quantity, 100 by the utterance of the same speaker. The estimation was performed using 1048 points of test data included in the word. However, at this time, the test data used was the one that was subjected to the deletion processing in the same manner as the learning data in order to investigate the performance when estimating one value for one input pattern. The vertical axis of the figure is the measured value, and the horizontal axis is the estimated value. At this time, the multiple correlation coefficient is 0.93 and the average relative error is 9.8. %Met.

In the second embodiment, one for each category
Since the quantity is assigned to one, the number of moras of the input at the time of estimation is limited. Quantification type I is a type of multiple regression analysis that determines one quantity for one category in factor items by the least squares method, but when calculating category quantity, the relationship between the combination of categories and the output is Not considered. What is actually considered is the output for a combination of one category with another category,
If the output largely depends only on the combination of two or more categories, the estimation accuracy becomes poor. It is difficult to obtain a priori knowledge of human beings, that is, an input / output function corresponding to the relationship between the combination depending on the magnitude of the input and the output from the result of the quantification type I, and it is difficult to interpret the result in this sense.

Therefore, the pitch frequency control method proposed in this embodiment is based on the phoneme environment data in advance.
The fuzzy neural network (hereinafter,
It is calculated by Bop propagation (BP) learning as a coupling coefficient of FNN), and the pitch frequency for an arbitrary mora of arbitrary Japanese is calculated by using the coefficient at the time of synthesis.

FIG. 1 shows the flow of learning processing in this embodiment.
1 shows the flow of the estimation process in the block diagram of FIG. Regarding the construction and learning method of FNN, the paper "Construction method and learning method of fuzzy neural network" (Shinichi Horikawa, Takeshi Furuhashi, Yoshiki Naito; Journal of Japan Fuzzy Society, Vol.4, No.5, pp906-928)
(1982)) and the like.

In FIG. 11, 27 is an input data generator for generating input data from learning data, and 28 is an FN for obtaining coefficient data from the input data and storing it in a database.
N learning unit.

Further, in FIG. 12, reference numeral 34 is an input data generating section for generating input data from test data, 35 is an FNN pitch frequency estimating section for obtaining a pitch frequency from the coefficient data of the database, and 36 is the obtained pitch frequency. Is a voice synthesis unit for performing voice synthesis by.

FIG. 13 shows the configuration of the FNN used in this embodiment. The FNN of this embodiment is composed of the antecedent part of the fuzzy rule composed of layers (A) to (E) and the consequent part of the fuzzy rule composed of layers (F) to (L). This corresponds to the case where the antecedent variable and the antecedent variable are different in the type II FNN in the above-mentioned reference. Type here
II means that the consequent part of the fuzzy rule is represented by a linear function of the input variable. Σ in the (B), (J), and (L) layers outputs the sum of the inputs, and Π in the (K) layer outputs the product of the inputs.
Π ^ of the (E) layer outputs a value obtained by dividing the product of the inputs by the total sum of the inputs to the (E) layer. The f of the (C) layer outputs from the sigmoid function. □ 1 is a neuron that outputs 1. The function realized by this FNN is fuzzy inference in which the consequent part of the fuzzy rule is represented by a linear function. Learning is performed by changing the coupling coefficients W _c , W _g , and W _a using a back propagation (hereinafter, BP) method. Where W _c and W _g are coefficients of the antecedent membership function,
W _a corresponds to the coefficient of the consequent linear function.

Fuzzy inference has a feature that an algorithm including ambiguity such as human judgment can be linguistically described by if-then type fuzzy rules.
However, as seen in the fuzzy controller design example, a great deal of effort is generally required to identify or adjust fuzzy rules or membership functions in fuzzy reasoning. On the other hand, there is a neural network (hereinafter referred to as NN) as another method suitable for handling human knowledge. The feature of NN is that any input / output relation can be identified by the learning function. However, since the knowledge about the input / output relation is distributed and stored in the connection weights in the network, it is necessary to qualitatively grasp this knowledge. Is difficult
Therefore, the FNN is proposed as a combination of fuzzy reasoning and NN and having both features. F
The NN has a structure suitable for fuzzy reasoning, and while applying the Pack Propagation (BP) method that is a learning method of the NN, the feature of the fuzzy reasoning that the knowledge can be described linguistically and the knowledge are automatically described. It also has the characteristic of NN that you can learn.

In such an FNN configuration, the process of calculating an inference value by fuzzy inference is realized by the structure of NN (BP model), and the parameter to be identified or adjusted in fuzzy inference is associated with the NN coupling weight. Is. By updating these connection weights using the BP method, the fuzzy rules can be identified and the membership function can be automatically adjusted, and the learning result can be grasped as the fuzzy rules.

Generally, the inference value obtained by fuzzy inference is calculated from the given input values as follows: (a) Membership function value of the antecedent part (b) Antecedent suitability of each fuzzy rule (c) Consequent part of each fuzzy rule It can be obtained by sequentially obtaining the estimated value and integrating the results of (a) and (c).

In this embodiment, the backpropagation (BP) of the antecedent part membership function coefficient and the consequent part linear function coefficient of the fuzzy inference is performed in the FNN in which the antecedent part for realizing the input / output of the fuzzy inference is a linear function. This is a method that is determined by learning. In this method, when BP learning is performed, the initial value of the consequent linear function is set to 0 and learning is performed. Input here (1) ~
The input (6) corresponds to the input forms (1) to (6) of the quantification type I in the second embodiment described above, except that the input (5) and the input (6) of the phonological symbols are as follows. 11 + 14
= 25 input values of 0 or 1. There are two types of membership functions for the antecedent part, and their labels are Small and Big.

FIG. 14 shows the estimation result of the test data by FNN. The learning rates W _c , W _g , W _a (constant term) and W _a (variable term) are 0.000005 and 0.0000, respectively.
The learning was performed by setting it to 1, 0.1, 0.05. The data used are the same as the data used in the estimation in the quantification type I for both the training data and the test data. That is,
4242 included in 490 words spoken by one female speaker
Learning is performed with point data, each coefficient of FNN is obtained, and the results are used to estimate with 1048 points of test data included in 100 words uttered by the same speaker. The learning was performed by sequentially inputting the learning data into the network, and the number of times of learning was set to be once for one cycle of the data. As a result, when the learning was performed 8650 times, the multiple correlation coefficient was improved to 0.97 and the average relative error was improved to 8.41%.

This is because BP learning was performed in order to realize the relationship between the combinations of all the categories in one input data and the outputs as the coefficients of FNN.
That is, the nonlinear input / output learning is performed by the BP learning.

As shown in Table 2, the consequent linear functions formed by learning are classified according to the size of the input and are input / output functions corresponding to human a priori knowledge. Here, x ₁ ~x ₄ corresponds to the input, respectively (1) to the input (4), the input x ₁ for the consequent part of Table 2
The values of up to x ₄ are each multiplied by W _s in FIG.
The W _s values corresponding to the two inputs are 0.50, 0.
It is 09, 0.08, 0.08.

The input number of moras is an input of the antecedent part membership function, and its value is not limited unlike the quantification type I.

[0067]

[Table 2]

Next, a fourth embodiment of the present invention will be described.

In the third embodiment, since learning is performed by setting the initial value of the linear function coefficient of the FNN consequent part to 0, even if the learning data converges to a value with a small error, the unlearned data ( It may be a local minimum that has nothing to do with test data). That is, a coefficient that should greatly contribute to unlearned data may not be learned at all, or may be dragged by the learning of another coefficient, resulting in a random value. Therefore, in FIG. 14, there are some points where the measured value and the estimated value are greatly separated.

Therefore, in this embodiment, in order to improve the accuracy of estimation even for unlearned data, a method of determining the initial value of the FNN consequent linear function coefficient by quantification type I is proposed. . In this method, (1) the input data for learning is reconstructed in correspondence with the FNN membership function, and (2) the reconstructed data is processed by quantification type I to calculate the coefficient of the input / output function. , (3) The coefficient is F
NN Set the initial value of the linear function coefficient of the consequent part.

That is,

As a simple example, the input of FNN is the antecedent part,
The case where the consequent part has two inputs and there are two membership functions for one input will be described. Assuming that the two inputs A and B take values of, for example, 1 to 10, the values of each input are divided by a certain threshold value (for example, 5) to correspond to the two membership functions, If it is less than 5, set it to 1, and if it is 5 or more, set it to 2. The output uses the value as it is. Then, the recreated data is processed by the quantification type I to obtain the coefficient (quantity) of the input / output function. The obtained coefficient is the coefficient of Class I for the value of 1 of input A, which is the initial value of the coefficient of A when the value of Input A is Small in FNN, and the coefficient of quantification I for the value of 2 of input A is The coefficient of the class is the initial value of the coefficient of A when the value of the input A is FNN and is Big. The same applies to the input B.

FIG. 15 shows an estimation result when the above processing is applied to the above pitch frequency control. However, in this case, since the inputs (5) and (6) for the phonological symbols are not input to the antecedent part, it is not necessary to binarize them according to the membership function. Therefore, the quantification type I processing may be performed with the value as it is, and the quantification type I coefficient regarding each phoneme may be used as the coefficient of the corresponding FNN consequent linear function. The data used here is the same as that used above for both the learning data and the test data. As can be seen by comparing the figures of the estimation results, it can be seen that the points with large errors are eliminated and the estimation accuracy is improved for unlearned data. At this time, the average relative error with respect to the test data was improved to 6.84%.

[0074]

As is clear from the above description, according to the prosody control system of the first example to which the quantification type I of the present invention is applied, (1) the position of one position in one mora is determined by the sequential method. The fundamental frequency can be obtained, and the pitch pattern of the mora can be finely controlled.

(2) Quantification of multivariate analysis with few items (phonological environment) and categories (classification of each item) I
Class can be executed, and stable high-quality speech synthesis prosody control with a high partial correlation coefficient can be realized.

(3) The input data is the phoneme, the preceding phoneme,
Since it can be described by a subsequent phoneme, it is possible to realize a human-like control that considers the influence of a phoneme that cannot be realized by a conventional phoneme symbol string such as an accent phrase.

Further, according to the prosody control method of the second example in which the input format of the quantification type I of the present invention is changed, the following effects can be obtained.

(1) Prediction accuracy can be improved by determining and learning the input format.

(2) The prediction accuracy can be improved by creating and learning one set of data at one mora inner point position.

(3) The prediction accuracy can be improved by learning the pitch frequency of a certain value for one input pattern of the learning data.

According to the prosody control system of the third example using the fuzzy neural network (FNN) of the present invention, the following effects can be obtained.

(1) The number of moras at the time of estimation is an input to the FNN membership function and is not limited.

(2) Since the nonlinear input / output learning can be performed by the BP method, an input / output system can be constructed in consideration of the relationship between the combinations of all the categories in one input data and the output. Therefore, the estimation accuracy is improved.

(3) It is possible to obtain an input / output function corresponding to human a priori knowledge.

Further, in the prosody control method using the FNN, when the initial value of the FNN consequent part linear function is determined using the method of quantification type I, the estimation accuracy is further improved. To be

[Brief description of drawings]

FIG. 1 is a block configuration diagram for realizing a first embodiment of the present invention.

FIG. 2 is a block configuration diagram of an analysis unit in the first embodiment.

FIG. 3 is a diagram showing a control flow of an analysis unit showing the first embodiment of the present invention.

FIG. 4 is a block diagram of a combining unit in the first embodiment.

FIG. 5 is a diagram showing a control flow of a combining unit in the first embodiment.

FIG. 6 is a diagram showing an experimental result according to the first embodiment.

FIG. 7 is a diagram showing another experiment result according to the first embodiment.

FIG. 8 is a block diagram showing the flow of processing according to the second embodiment of the present invention.

FIG. 9 is a diagram showing experimental results according to the second embodiment.

FIG. 10 is a diagram showing another experimental result according to the second embodiment.

FIG. 11 is a block diagram showing the flow of learning processing according to the third embodiment of this invention.

FIG. 12 is a block diagram showing the flow of an estimation process according to the third embodiment.

FIG. 13 is a block diagram of an FNN used in the third embodiment.

FIG. 14 is a diagram showing an experimental result according to the third embodiment.

FIG. 15 is a diagram showing experimental results according to the fourth embodiment of the present invention.

[Explanation of symbols]

 DESCRIPTION OF SYMBOLS 1 ... Phonological database 2 ... Analysis part 21, 23, 27 ... Input data generation part 22, 25 ... Quantification type I calculation part 24 ... Data reduction part 28 ... FNN learning part 3 ... Synthesis part 31, 34 ... Input data generation part 32 ... Pitch frequency estimation unit 33, 36 ... Speech synthesis unit 35 ... FNN pitch frequency estimation unit

Front page continuation (72) Inventor Shigeru Kashiwagi 2-1-117 Osaki, Shinagawa-ku, Tokyo Stock Company, Meidensha

Claims (4)

[Claims] 1. An accent pattern described by the preceding mora, the relevant mora, and the subsequent mora, the mora position from the beginning of the mora, the mora position from the end of the mora, and the mora concerned by using the analysis learning means to which the quantification method of the multivariate analysis is applied. In order to quantify the fundamental frequency of each pitch point of the mora using the item and its category as variables from the input data for learning in which the pitch point position, the long sound flag, and one or more of the consonant flags are items. For each category, calculate the category quantity that is the coefficient of the quantification formula, find the category of the item of each mora of the phoneme symbol string of the speech synthesis target, and calculate the quantification formula using the calculated category quantity for the category. Using the calculation means for sequentially calculating the fundamental frequency of each pitch point of the mora from Prosody control method which comprises using the calculated fundamental frequency in the speech synthesis. 2. An analysis learning means to which a quantification method of multivariate analysis is applied is used, and the position from the mora position where the accent rises or falls and the accent changes, and the leading mora of the input is before the first. The items are the order of the mora, the order of the mora after the last one of the input is the first, the attribute when only the vowel part of the phoneme signal is viewed, and the attribute when the consonant part of the phonological symbol is viewed. In order to quantify the fundamental frequency of each pitch point of the mora from the input data for learning using the item and its category as variables, calculate the category quantity which is the coefficient of the quantification formula for each category of the item, and the speech synthesis target The category of the item of each mora in the phonological symbol sequence of the Using calculating means for sequentially calculating the fundamental frequency of the switch point as the estimation means of the fundamental frequency, the prosody control method is characterized by using the speech synthesis the fundamental frequencies the calculated. 3. A fuzzy neural network obtained by learning coefficient data from learning input data by using a fuzzy neural network having an antecedent part and a consequent part as learning means, and setting the coefficient data obtained by the learning. A prosody control method characterized in that a network is used as a pitch frequency estimating means and the estimated pitch frequency is used for speech synthesis. 4. The coefficient of the input / output function is calculated by reconstructing the learning input data in the learning means in correspondence with the membership function of the antecedent part of the fuzzy neural network and processing it by quantification I, The prosody control method according to claim 3, wherein the coefficient is set as an initial value of a coefficient of a linear function of the consequent part.