JP2880433B2 - Speech synthesizer - Google Patents

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 speech based on an input character string.

[0002]

BACKGROUND ART pitch frequency is a fundamental frequency of the speech to the modeling (hereinafter. Referred to F ₀ frequency), the pattern of the time series of efficiently F ₀ frequency with a small number of parameters conventionally (hereinafter, F ₀ pattern (For example, Fujisaki et al., “Model of Fundamental Frequency Pattern of Japanese Word Accent and Its Generation Mechanism”, Papers of the Acoustical Society of Japan. 27, No. 9, pp. 445.
-453, September 1971 ". ) This superposition model, taken as the sum of an accent component corresponding to the F ₀ pattern is also called a phrase component (talking tone component that gradually drops toward phrase end from Kuatama.) And accent phrase.
For parameterization of the F ₀ pattern by the superposition model,
There are the following advantages. (1) The number of free parameters used in the model is small,
It is easy to optimize F ₀ control by statistical analysis. (2) Since the separation of F ₀ pattern into two components phrase component and accent component, since the optimization, it is relatively easy interpretation of the control rules obtained as a result of the optimization.

Further, a conversion rule (hereinafter, referred to as a conventional example) between three utterance styles of a normal tone, a commercial tone, and a reading tone in order to diversify the rule-synthesized speech is disclosed in, for example, a conventional document 2 “Abe”. In addition, “Changes in speech style and its evaluation”,
Proceedings of the Acoustical Society of Japan, 3-P-18, 1993 1
October ". In this conventional example, by converting the parameters of the formant frequency, the duration, the fundamental frequency, and the power, the normal tone, the commercial tone, the reading tone, the tone converted from the normal tone to the commercial tone, and the normal tone, Total of 5 voices converted to reading style
We prepare two utterance styles and evaluate their similarity.

[0004]

However, in the above-mentioned conventional example, the conversion of the utterance style is targeted, but the speech is synthesized without considering the speaker characteristics. In other words, if the accent type is different, the height of the accent is different for each individual,
In the conventional example, it is impossible to synthesize the voice of one specified speaker.

It is an object of the present invention to solve the above problems and to provide a speech synthesizer capable of synthesizing speech of one specified speaker.

[0006]

Speech synthesis apparatus according to the present invention SUMMARY OF THE INVENTION, the voice data memory means (11-13), F ₀ control rule storage means (31-33), learning means (20, 21),
A voice synthesizing device comprising a parameter sequence generating means (1) and a voice synthesizing means (2), wherein the voice data storage means (1
1-13) accumulates audio data, F ₀ control rule storage means (31-33) accumulates F ₀ control rule learning means (20, 21), the extraction means, generating means, rule generation Means, and the extraction means is a voice data storage means (11-
13) A pitch pattern is extracted from the voice data accumulated in 13), and the generating means analyzes the pitch pattern extracted from the extracting means using an analysis method based on a critical control model, and outputs an accent command, a phrase command, an accent phrase boundary. The rule generating means controls the pattern of the pitch frequency of the voice by paying attention to a predetermined control factor based on the pitch pattern extracted by the extracting means and the model parameter generated by the generating means. ₀ control rules are generated and stored in the F ₀ control rule storage means (31-33). The control factors include the number of mora of the phrase, the number of mora of the preceding phrase preceding the phrase, the accent type of the accent phrase, wherein a position within the sentence accent phrase, the parameter sequence generating means (1), the string and F ₀ control rule memory means input Generating a pattern of pitch frequency based on the F ₀ control rules stored in the 31-33), and generate an acoustic parameter sequence including the pitch frequency,
The voice synthesizing means (2) includes a parameter sequence generating means (1)
Synthesizes a speech based on the acoustic parameter sequence output by.

[0007]

[0008]

[0009]

Embodiments of the present invention will be described below with reference to the drawings. FIG. 1 illustrates the F
₀ is a block diagram of a speech synthesis apparatus having a F ₀ control rule learning unit 20 to generate a F ₀ control rules for controlling the frequency.
In FIG. 1, the speech synthesizer according to the present embodiment is configured such that the F ₀ control rule (one of 31, 32, and 33) of one speaker selectively connected based on an input character string. And a parameter sequence generator 1 for generating a feature parameter sequence for speech synthesis using a voice quality control rule 41 and a phoneme duration control rule 42, and generating a speech signal based on the generated feature parameter sequence. And a voice synthesizing unit 2 for outputting the voice to a speaker 3. In the present embodiment, in particular, a character string input using F ₀ control rules 31, 32, and 33 for controlling a pitch frequency of a speech created for each speaker is converted into a speech of a speaker specified in advance. The F ₀ control rules 31, 32, and 33 control the size of the phrase based on the number of mora of the phrase to be synthesized and the number of mora of the preceding phrase preceding the phrase. The rule is to control the pitch frequency of the voice by controlling the size of the accent phrase based on the accent type of the accent phrase to be synthesized and the position of the accent phrase in the sentence.

The F ₀ control rule learning unit 20 is connected to a working memory 21 used as a work area when executing an F ₀ control rule learning process described in detail later. Further, one of the voice data 11, 12, and 13 of the speakers A, B, and C is selectively connected to the F ₀ control rule learning unit 20 via the switch SW1, while the switch SW2 is connected.
Via the F ₀ control rules 31, 32,
One of the 33 is selectively connected. Switching of these switches SW1 and SW2 is performed by the F ₀ control rule learning unit 2
By 0, the voice data of the same speaker and the F ₀ control rule are linked and controlled simultaneously. further,
One of the F ₀ control rules 31, 32, and 33 of speakers A, B, and C is selectively connected to the parameter sequence generation unit 1 via a switch SW3. This switch SW3
Is switched by the operator so as to select the _F0 control rule of the speaker who synthesized the voice. Further, a conventional voice quality conversion control rule 41 and a conventional phoneme duration control rule 42, which will be described in detail later, are connected to the parameter sequence generation unit 1.

In the present embodiment, the audio data 11, 1
2, 13, working memory 21, and F ₀ control rule 3
1, 32, 33, the voice quality control rule 41, and the phoneme duration control rule 42 are configured by a memory such as a hard disk, for example. Further, the F ₀ control rule learning unit 20 includes:
The parameter sequence generation unit 1 is constituted by, for example, a digital computer.

FIG. 2 is a flowchart showing the F ₀ control rule learning process executed by the F ₀ control rule learning section 20 of FIG. First, in step S1, audio data 1
After extracting the F ₀ pattern based on the audio data in 1,12,13, in step S2, the extracted F ₀
The model parameters of the critical control model are generated using an analysis method based on the critical control model based on the pattern.
Further, in step S3, the extracted F ₀ pattern,
Based on the model parameters of the criticality control model, a control rule for controlling the pitch frequency of the voice is generated by focusing on a predetermined control factor. Here, the control factors are the number of mora of the phrase to be synthesized, the number of mora of the preceding phrase preceding the phrase, the accent type of the accent phrase to be synthesized, and the position of the accent phrase in the text. The F ₀ control rule controls the size of the phrase based on the number of mora of the phrase to be speech-synthesized and the number of mora of the preceding phrase preceding the phrase, and The pitch frequency of the voice is controlled by controlling the size of the accent phrase based on the accent type and the position of the accent phrase in the text. Next, details of the processing of each of the above steps will be described.

First, the processing in step S1 will be described. Each of the voice data 11, 12, and 13 includes voice signal data of a sentence read by one speaker (also referred to as an uttered voice sentence). In this step S1, A / D conversion and LPC analysis are performed on the audio signal data to extract characteristic parameter data, and based on the extracted characteristic parameter data, for example, an analysis method using a known critical braking model is performed. (For example, see Reference 3 “Fujisaki et al.,“ Analys
is of voice fundamental frequency contours for dec
larative sentences of Japanese (analysis of fundamental frequency patterns of Japanese declarative sentences) ”, Transactions of the Acoustical Society of Japan,
(E), Vol. 5, No. 4, pp. 233-24
4, April 1984 ". ) And detects the F ₀ pattern and the model parameters, and labels them by phoneme unit, accent phrase unit and phrase unit to generate them. Here, the feature parameter data includes log power, 16th order cepstrum coefficient, Δlogarithmic power, and 16th order cepstrum coefficient, and the model parameters include an accent command and a phrase command. Contains boundary information. In the above analysis, F
₀ is decomposed into an accent component indicating a local relief phrases component and F ₀ frequency is gentle downward component of the frequency. In the critical braking model, the phrase component and the accent component are regarded as the response of the critical braking secondary linear system to the phrase command and the accent command, respectively. The precise timing and size of each command are analyzed by automatic synthesis based on the approximate timing of the phrase command obtained from phoneme labeling information, accent phrase information, phrase boundary information, and accent command (Analysis-by-S).
(synthesis).

Next, the processing in step S2 will be described. As one of the above-described superposition control models, a Fujisaki model, which has been researched and proposed by Fujisaki, is known (for example, see Conventional Document 1). Conventionally, a hill-climbing method has been used for parameterization using the Fujisaki model (for example, see Conventional Document 3). That is, by changing all the free parameters of Fujisaki model, a set of parameters that minimize the mean estimation squared error F ₀ pattern, it is an analysis result of the F ₀ pattern. This can be regarded as a search problem in which the total number of parameters is the number of search space dimensions. Conventionally, the angular frequency, the input time, and the size of the phrase command, and the angular frequency, the rising time, the falling time, and the size of the accent command are treated as free parameters. Therefore, the search space is (3I + 4J) -dimensional ( Here, I is the number of phrase commands, and J is the number of accent commands.) Of these parameters, the magnitude of the accent command can be uniquely obtained by using the least squares method if an actual measured value of the F ₀ frequency and other parameters are given.
The calculation time can be reduced by reducing the number of dimensions of the search space by J. In this method, F ₀ at each time point
Reliability of the frequency value (here, the maximum value of the autocorrelation function obtained in calculating the F ₀ frequency from the voice) formulation that may be used in the weighting of the estimated error evaluation of F ₀ frequency at each time point Is becoming This is for the reliability of the value of F ₀ frequency obtained from the voice data corresponds to it is not uniform at each time point. In the present embodiment, it was parameterization of F ₀ pattern by the hill-climbing method using this method.

Further, generation of the F ₀ control rule in step S3 will be described. A rule for estimating the attribute of each command from the above-mentioned control factors which are considered to affect the phrase command and the accent command is defined by a well-known spatial multiple division quantification method (Multiple Split Regressi).
on (for example, Conventional Document 4, “Iwahashi et al.,“ Statistical Modeling of Voice Control by Space Division Type Quantification Method ”, Proceedings of the Acoustical Society of Japan, 1-5-11, pp. 237-238, October 1994) ); Hereinafter, referred to as the MSR method. ). In the MSR method, similar to the analysis procedure in the regression tree, a binary tree is grown by a classification method that minimizes the sum of square errors between the model estimation value and the actually measured value, and a model is generated. Further, the MSR method allows nodes other than the leaf node of the binary tree to share the branching condition over the entire subtree below it, and modeling can be performed efficiently with a small number of parameters. Control factor used for the growth of the binary tree at a node closer to the root node, so affects the estimate of the number of samples, it can be determined that they are important control factor deeply involved in the control of F ₀ frequency.

The target of the command estimation model includes timing information such as the size of the command and the time of rising. The timing information can be estimated by a small number of rules, but the estimation of the size of the command is complicated. In the present embodiment, the size of each command is set as an estimation target because a simple rule is required. Here, the command estimation model of each of the phrase command and the accent command is collectively called an _F0 control rule.

A representative one of the statistical modeling methods for generating an estimation model is quantification class I (for example, see Reference 5).
"Hayashi et al.," Quantification Theory and Data Processing ", Asakura Shoten, 1
982 ". ) Or a regression tree (for example, see
rieman et al. , "Classifica
Tion And Regression Tree
s ", Wadsworth Statistics / P
roboticity Series, U.S.A. S. A. ,
1984 ". )and so on. Here, the quantification class I is a linear multiple regression model using the control factors as an explanatory variable space, and cannot assume the independence between the control factors. In addition, in a regression tree in which the explanatory variable space is sequentially divided, the independence of the divided explanatory variable space is assumed, so that the dependency between the divided spaces cannot be expressed. On the other hand, the MSR method solves the problems of both the quantification class I and the regression tree by considering parameters shared by a plurality of divisions in the process of analyzing the regression tree. Note that the result obtained by quantification class I is a special solution of the MSR method that allows division only at the root node, and the result obtained by the regression tree is a special solution of the MSR method that prohibits simultaneous division at multiple nodes. Each can be considered as a solution.

Similar to the regression tree, a tree structure model is generated in the analysis of the MSR method. FIG. 3 shows an example of a model based on the MSR method. In this example, the observed value is represented by two types of control factors C
Can be estimated by _1, C _2. Observed estimated value is the quantity of the quantity at the terminal node when the tree branch satisfying the condition is sequentially traced from the top root node based on the control factor and the quantity a _i written in the node is added. Obtained as a sum. The value of the quantity a _i is calculated using control factors obtained from a large amount of data, observed values and normal equations,
As in the case of the quantification class I and the regression tree, parameter values cannot be uniquely obtained, so that it is necessary to restrict some parameter values. In the present embodiment, the quantity on the node side that does not satisfy the condition (for example, the node that branches to No when the condition C ₁ ≦ 5) is set to 0, and another quantity is obtained. In the example of FIG. 3, a ₃ , a ₇ , a ₉ , and a ₁₁ are set to 0. Under this condition, the quantity a of the root node
_{1, a 3, a 7, a} 9, a 11 from that both are 0, the data group to reach the bottom right edge of the node (i.e., the data for selecting the side of the node is not the case condition in any branch) Is the average of the observed values. The quantity a ₁ can be regarded as an initial value when obtaining an estimated value.

In FIG. 3, the tree structure surrounded by the dotted line is MS
It is an example of the analysis result peculiar to R method. This subtree is a result of the division of a node of a ₃ is performed, is generated node a _6, a _7, then again, a ₃ node split in the performed by node a _6, a ₇ is divided It is generated by doing this. a ₁₀ and a ₁₁ can be regarded as quantities as shared parameters. In the case of quantification class I, the division is allowed only at the root node, and in the case of the regression tree, the division is allowed only at the terminal node. .

As described above, the MSR method used in the present embodiment is an expanded version that includes the concepts of quantification type I and the regression tree. Furthermore, the increase in the number of model parameters can be suppressed due to the presence of shared parameters, and modeling can be efficiently performed with a small number of parameters. From this point of view, the present embodiment uses the MSR method as a statistical modeling method.

By analyzing the relationship between the phrase command and the accent command obtained from the large amount of voice data and the control factors affecting each command by the MSR method, the phrase command and the accent command are obtained from the control factors. The model to be estimated is obtained. Each model has a binary tree structure and model parameters. The binary tree is used as a rule for classifying each command according to a control factor. The model parameters are used for calculating an estimated value. By examining the structure of the binary tree obtained by the analysis, it becomes possible to analyze what control factors influence each command. The size of the model parameter is also a criterion for determining whether the classification involving the parameter has a large effect on each command.

FIGS. 4 and 5 show four speakers M1, M2,
F1, F2 (where M1 and M2 are male speakers and F1
1 and F2 are female speakers. ) Shows each F ₀ control rule. Here, FIG. 4 shows the control amount for the number of mora of the phrase and the control amount for the number of mora of the preceding phrase. FIG. 5 shows the control amount for the accent type of the accent phrase and the control amount for the accent phrase in the sentence of the accent phrase. And a control amount for the position. Here, the mora is a beat substantially corresponding to one kana character. The accent type is referred to as type 1 when the accent phrase is on the first beat, and is referred to as type 2 when the accent phrase is on the second beat.
Table 1 shows the _F0 control rules for the speaker F2 in FIGS.

[0023]

[Table 1] Specific example of F ₀ control rule for each phoneme string <Case of speaker F2> (corresponding to (d) in FIG. 4 and (d) in FIG. 5) ────────────────────────── (1) The size of the relevant phrase, preceding phrase, and accent command are each set to a predetermined initial value of the relevant speaker. Initialize to (0.6). ─────────────────────────────────── (2) Judgment control regarding the number of mora of the phrase (2-1) ) If the length of the phrase is not less than 1 mora and not more than 3 mora, the size of the phrase is reduced by 0.15 from the initial value. (2-2) If the length of the phrase is not less than 4 mora and not more than 6 mora, the size of the phrase is reduced by 0.05 from the initial value. (2-3) If the length of the phrase is 7 to 12 mora, the size of the phrase is reduced by 0.025 from the initial value. (2-4) If the length of the phrase is 13 mora or more, reduce the size of the phrase by 0.025 from the initial value. ─────────────────────────────────── (3) Judgment control on the number of mora in the preceding phrase (3-1) ) If the length of the preceding phrase is 1 mora or more, reduce the size of the preceding phrase by 0.0125 from the initial value. (3-2) If there is no preceding phrase, the size of the preceding phrase does not change from the initial value. ─────────────────────────────────── (4) Judgment control regarding accent type of accent phrase (4-1 ) If the accent type is type 1 or type 2, increase the size of the accent phrase by 0.05 from the initial value. (4-2) If the accent type is 3 or more, the size of the accent phrase does not change from the initial value. (4-3) If there is no accent phrase, reduce the size of the accent phrase by 0.2 from the initial value. ─────────────────────────────────── (5) Judgment control regarding the position of the accent phrase in the sentence (4) -1) If the accent phrase is at the beginning of the sentence, the size of the accent phrase does not change from the initial value. (4-2) If the accent phrase is in the sentence, the size of the accent phrase does not change from the initial value. (4-3) If the accent phrase is at the end of the sentence, reduce the size of the accent phrase by 0.25 from the initial value. ─────────────────────────────────── (Note) The size of the phrase command is controlled in Table 1. The control amount of (2) and (3) is the sum of the control amounts, and the control of the accent phrase size is the sum of the control amounts of (4) and (5) in Table 1.

Next, the synthesized speech "Today is fine weather"
, The F ₀ control rules 31, 32, 32 used to control each parameter for each phoneme or phoneme sequence
33, voice quality control rule 41 and phoneme duration control rule 4
Table 2, Table 3, and Table 4 respectively show an example of No. 2. In Table 3, the acoustic feature parameters are log power, 16th order cepstrum coefficient, Δlog power, and 1
This is a 34-dimensional parameter including a sixth-order ΔCepstrum coefficient.

[0025]

[Table 2] Example of F ₀ control rule ─────────────────────────────────── Phoneme sequence F ₀ control Rule ─────────────────────────────────── kyo'uwa F ₀ control rule of phrase size Accent F ₀ control rule of size 'yo'ite ＮNkidusu Phrase the size of the F ₀ control rules Accents size of one of the F ₀ control rules Accents 2 magnitude of F ₀ control rules ─────────────────────── ────────────

[0026]

[Table 3] Example of voice quality control rule ───────────────── phoneme acoustic feature parameter ｋ ky ( ..., o (0.45, 0.38, ...) u (0.25, 0.42, ...) w (0.32, 0.30, ...) a (0. 12, 0.45,…)… ───────────────── ─────────────────

[0027]

[Table 4] Example of phoneme duration control rule ───────────── phoneme phoneme duration ───────────── ky 0.054 seconds o 0 120 seconds u 0.095 seconds w 0.080 seconds a 0.110 seconds ─────────────

The operation of the speech synthesizer shown in FIG. 1 will be described below. As shown in FIG. 1, a character string to be speech-synthesized is input to a parameter sequence generation unit 1. Parameter sequence generating unit 1, based on the character string input, F ₀ control rules for controlling the F ₀ frequency (31,32,3
3 one of) the voice control rules 41 for controlling the acoustic feature parameter, by using the phoneme duration control rules 42 for controlling the phoneme duration, F ₀ frequency and the acoustic feature parameter The control parameter data including the phoneme duration is selected, and based on the selected parameter data, a process such as a time alignment process and a voice spectrum interpolation process is executed by, for example, the DTW method, and a 16th-order cepstrum is processed. The time series data of the coefficients is generated and output to the speech synthesis unit 2. The voice synthesizer 2 includes a pulse generator, a noise generator, a variable gain amplifier, and a filter.
By generating an audio signal based on the input time-series data and outputting the generated audio signal to the speaker 3, a synthetic voice corresponding to the input character string is generated.

[0029] In the above embodiments, to speak the small number of the audio data to the speaker of the conversion target, the F ₀ control rules generated based on this, one of the F ₀ control rules generated from a large amount of speech data The F ₀ control rule may be generated by replacing

[0030]

DESCRIPTION OF THE PREFERRED EMBODIMENTS The inventor uses the speech synthesizer shown in FIG.
Subjecting the speech database to F ₀ control rule learning processing, generating F ₀ control rules for estimating the magnitude of phrase commands and accent commands, generating and analyzing F ₀ control rules for a plurality of speakers, The commonality of important control factors among speakers was investigated.

As the audio data, the generation of the F ₀ control rule requires 500 sentences uttered by two male and female speakers, for a total of 2,
000 sentences (for example, conventional literature 7 "Abe et al.,
Proceedings of the Acoustical Society of Japan, pp. 267-268,19
October 1989 ". ). The utterance contents are sentences selected from newspapers and magazines. Table 5 shows the number of phrase commands and accent commands for each voice data. Using the above-described processing method, F ₀ control rules for each voice database were generated, and individual control rules were analyzed.

[0032]

[Table 5] Number of commands included in each voice database ─────────────────────────────────── Speaker M1 M2 F1 F2 フ レ ー ズ Phrase command 1903 1684 1425 1532 Accent command 3200 3176 3306 3119 ───────────────────────────────────

The analysis of the control factors and control rules used for generating the F ₀ control rule will be described. The parameters used in the critical braking model include the input time and size for the phrase command, the rising time for the accent command,
At the time of falling, there is a size. Among them, it has been reported that the time information such as the input time can be controlled by a small number of simple rules (for example, the conventional art 8 "Miki et al., IEICE technical report, SP92-
6, March 1992 ". ). On the other hand, appropriate control of the magnitude of the command is important for improving the naturalness and intelligibility of the synthesized sound. Therefore, in generating the F ₀ control rule, the magnitudes of the phrase command and the accent command were set as targets of estimation.

First, the control factors used for estimating the magnitude of the phrase command and the effects thereof will be described. In order to estimate the magnitude of the phrase command, the following four control factors were considered. (A1) The phrase length (specifically, the number of mora of the phrase) (divided into 5 categories) (A2) The preceding phrase length (specifically, the number of mora of the preceding phrase) (divided into 6 categories (A3) Position in the sentence of the phrase (divided into two categories at the end of the sentence or non-sentence.) (A4) Accent type of the first accent phrase of the phrase (divided into four categories.)

If the phrase is short, it is not necessary to keep the phrase component at a high value for a long time. Therefore, it is conceivable that the phrase command becomes smaller as the phrase becomes shorter. When the preceding phrase is short, the phrase starts before the phrase component of the preceding phrase is sufficiently attenuated. In this case, it is expected that the phrase command will also be reduced. For these reasons, the lengths of the phrase and the preceding phrase were used as control factors for estimating the size of the phrase command. In addition, in voice, the F ₀ frequency is significantly reduced at the end of a sentence, and the phrase command at the end of the sentence is considered to be smaller than that at other positions. Therefore, the position of the phrase in the sentence is determined by the size of the phrase command. It was used to estimate the length. Furthermore, in order to suppress that the value of F ₀ frequency phrase top portion becomes too large, with the accent type is a factor having a magnitude strong correlation accent component the size of a phrase command as a factor inhibiting.

When a model for estimating the magnitude of the phrase command from these control factors was generated and analyzed, it was confirmed that the length of the phrase and the preceding phrase were important control factors in all voice databases. . Also,
Regarding the above factor (A4), it was found that the phrase becomes smaller when an accent phrase having an accent nucleus (hereinafter, referred to as an undulating accent phrase) is present at the beginning of the phrase in three of the four speakers.

Next, the control factors used for estimating the magnitude of the accent command and the effects thereof will be described. In order to estimate the magnitude of the accent command, the following four control factors were considered. (B1) The length of the accent phrase (specifically, the number of moras of the accent phrase) (divided into four categories) (B2) Accent type of the accent phrase (divided into four categories) (B3) Preceding accent Accent type of phrase (divided into five categories) (B4) Position of the accent phrase in the sentence (divided into three categories: beginning, middle, and end of sentence)

As is well known, it is known that when the accent phrase is short, and when the accent type is a flat type, the accent component becomes small, and these are considered as control factors. According to the experimental results of the present inventor, the smaller the number indicating the accent type, that is, the smaller the number of beats pronounced at “high”, the greater the tendency for the accent command to be. Classification (type 1, type 2, type 3 to type 5, type 6 or more) and analysis were performed. If the leading accent phrase is undulating,
Since it is conceivable that the energy for raising the _F0 frequency in the preceding accent phrase is consumed and the accent phrase becomes smaller, the accent type of the preceding accent phrase is added as a control factor. Furthermore, as described above, the position in the sentence is described as a control factor for estimating the size of the phrase command. However, the size of the accent command may be different at the beginning, in the middle, and at the end of the sentence. This was also taken into account.

An accent command estimation model was generated by using these control factors and the measured value of the size of the accent command and analyzed. As a result, the accent command of the accent phrase located at the end of the sentence in the above factor (B4) was analyzed. The reduction in size was confirmed in any speaker estimation model. In this experiment using a larger amount of voice data, there was no particular difference in the effect of the phrase command and the accent command on the size.

As described above, according to the embodiment of the present invention, the character string input using the _F0 control rule for controlling the pitch frequency of the speech created for each speaker is specified in advance. The F ₀ control rule controls the size of the phrase on the basis of the number of mora of the phrase to be synthesized and the number of mora of the preceding phrase preceding the phrase. By controlling the size of the accent phrase based on the accent type of the accent phrase to be synthesized and the position of the accent phrase in the sentence,
It is configured to control the pitch frequency of the voice. Therefore, it is possible to provide a speech synthesizer capable of synthesizing the voice of one specified speaker. Also, F
_{0 The} control rule learning unit 20 extracts the pitch frequency pattern of the voice based on the voice data, and uses the critical control model analysis method based on the extracted pitch frequency pattern of the voice to model parameters of the critical control model. And a control rule for controlling the pitch frequency of the voice can be generated. Therefore, it is possible to automatically and easily create an optimal and faithful F ₀ control rule for synthesizing the voice of one specified speaker.

[0041]

As described above in detail, according to the speech synthesizing apparatus according to the present invention, the speech data storage means (11-13),
F ₀ control rule storage means (31-33), learning means (2
0, 21), a parameter sequence generating means (1), and a voice synthesizing means (2), wherein a voice data storage means (11-13) stores voice data and stores F _0.
Control rule memory means (31-33) accumulates F ₀ control rule learning means (20, 21), the extraction means, generating means consists rule generating means, extracting means, speech data storage means (11 -13) The pitch pattern is extracted from the voice data accumulated in (13), and the generation means analyzes the pitch pattern extracted from the extraction means using an analysis method based on a critical control model, and outputs an accent command, a phrase command, an accent phrase. Generating a model parameter including a boundary, the rule generating means controls the pattern of the pitch frequency of the voice by focusing on a predetermined control factor based on the pitch pattern extracted by the extracting means and the model parameter generated by the generating means. An F ₀ control rule is generated, and the F ₀ control rule storage means (31-
33), and the control factors include the number of mora of the phrase, the number of mora of the preceding phrase preceding the phrase, the accent type of the accent phrase, and the position of the accent phrase in the sentence. ) generates a pitch frequency pattern based on the stored F ₀ control rules to a string and F ₀ control rule memory means input (31-33), to generate an acoustic parameter sequence including the pitch frequency, sound The synthesizing unit (2) synthesizes speech based on the acoustic parameter sequence output from the parameter sequence generating unit (1). Therefore, according to the present invention, the optimal and faithful F ₀ for synthesizing the voice of one specified speaker is described.
Control rules can be created automatically and easily.
Further, it is possible to provide a speech synthesizer capable of easily synthesizing the voice of one specified speaker.

[0042]

[Brief description of the drawings]

FIG. 1 is a block diagram of a speech synthesizer according to an embodiment of the present invention.

2 is a flowchart showing the F ₀ control rule learning process executed by the F ₀ control rule learning unit of FIG.

FIG. 3 is a diagram showing an example of modeling by a spatial multiplexing type quantification method (MSR) used in the F ₀ control rule learning unit of FIG. 1;

FIG. 4 is a graph showing an example of an F ₀ control rule relating to a phrase command created by an F ₀ control rule learning unit in FIG. 1;

FIG. 5 is a graph showing an example of an F ₀ control rule for an accent phrase created by an F ₀ control rule learning unit in FIG. 1;

[Explanation of symbols]

1 ... parameter sequence generating unit, 2 ... speech synthesis unit, 3 ... speaker, 11 ... audio data of the speaker A, 12 ... speaker B in the speech data, 13 ... speaker C sound data, 20 ... F ₀ Control Rule learning unit, 21 ... working memory, 31 ... F ₀ control rules of the speaker a, 32 ... F ₀ control rules of the speaker B, 33 ... speaker C of F ₀ control rule, 41 ... sound quality control rule, 42 ... phoneme Duration data.

────────────────────────────────────────────────── ─── Continuation of the front page (72) Inventor Norio Higuchi Kyoto, Soraku-gun, Seika-cho, 5th, Inani, 5th, Sanpira-ya ATI Ron Co., Ltd. Voice Translation and Communication Research Laboratories (56) References JP-A-3 JP-A-249800 (JP, A) JP-A-6-43891 (JP, A) JP-A-64-61796 (JP, A) JP-A-2-113299 (JP, A) JP-A-6-27984 (JP, A) JP-A-64-28695 (JP, A) (58) Fields investigated (Int. Cl. ⁶ , DB name) G10L 3/00 G10L 5/02

Claims (1)

(57) [Claims] An audio data storage means (11-13);
₀ control rule storage means (31-33), learning means (20,
21) A voice synthesizing device including a parameter sequence generating means (1) and a voice synthesizing means (2), wherein the voice data storage means (11-13) stores voice data, and stores the F ₀ control rule storage means ( 31-33) accumulates the F ₀ control rules, the learning means (20, 21) comprises an extraction means, a generation means, and a rule generation means, and the extraction means accumulates in the voice data storage means (11-13). A pitch pattern is extracted from the extracted voice data. The generating means analyzes the pitch pattern extracted from the extracting means using an analysis method based on a critical control model, and outputs model parameters including an accent command, a phrase command, and an accent phrase boundary. The rule generation unit generates a voice signal based on the pitch pattern extracted by the extraction unit and the model parameter generated by the generation unit, and pays attention to a predetermined control factor. Generate F ₀ control rules for controlling the pattern of the pitch frequency, F ₀ is stored in the control rule memory means (31-33), the control factor, the number of moras phrase, and number of moras preceding phrase preceding the phrase includes accent type of accent phrase, a position within the sentence accent phrase, the parameter sequence generating means (1) is, F _0, which is stored in the string and F ₀ control rule memory means input (31-33) A pattern of pitch frequency is generated based on the control rule, and an acoustic parameter sequence including the pitch frequency is generated. The speech synthesis unit (2) includes a parameter sequence generation unit (1)
A speech synthesizer that synthesizes speech based on the acoustic parameter sequence output by the device.