JPH0990970A - Speech synthesis device - Google Patents

Abstract

PROBLEM TO BE SOLVED: To synthesize speech of one designated speaker by controlling a size of an accent phrase based on an accent type of an accent phrase of a speech synthesis object and the position of an accent phrase in sentences. SOLUTION: A parameter system generation part 1 by which a feature parameter system for a speech synthesis is generated based on the input character string, and a speech synthesis part 2 outputting to speaker 3 which generates a speech signal based on the generated feature parameter system are provided. And an inputted character string is converted into the voice of a prescribed speaker by using F0 control rules 31-33 controlling the pitch frequency of the voice made for each speaker. The F0 Control rules 31-33 are rules which control the size of the phrase based on the number of the morae of the phrases of the speech synthesis object and the numbers of the morae of the phrase preceding to this phrase, and which control the size of the accent phrase and the pitch frequency of the speech based on the accent type of the accent phrase of the voice synthesis object and the position of the accent phrase in the sentence.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a voice synthesizing device for synthesizing voice based on an input character string.

[0002]

2. Description of the Related Art In order to model a pitch frequency (hereinafter referred to as F ₀ frequency) which is a fundamental frequency of speech, a time series pattern of the F ₀ frequency (hereinafter referred to as F ₀ pattern) is efficiently used with a small number of parameters. A superposed model that can be parameterized is used (see, for example, the conventional literature 1 “Fujisaki et al.,“ Fundamental frequency pattern of Japanese word accents and model of its generation mechanism ”), Acoustical Society of Japan. Magazine, Vol.27, No.9, pp.445.
-453, September 1971 ". In this superposed model, the F ₀ pattern is regarded as the sum of the phrase component (also called speech component) that gently falls from the beginning of the phrase to the end of the phrase and the accent component corresponding to the accent phrase.
Parameterization of the F ₀ pattern by the superposition model includes:
It has 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 F ₀ pattern is separated into two components, that is, the phrase component and the accent component, the F ₀ pattern is optimized. Therefore, the control rule obtained as a result of the optimization is relatively easy to interpret.

A conversion rule (hereinafter, referred to as a conventional example) between three utterance patterns of a normal tone, a commercial tone, and a reading tone for diversifying a rule-synthesized voice (hereinafter referred to as a conventional example) is disclosed in, for example, the conventional document 2 "Abe. In addition, "Change of utterance style and its evaluation",
Proceedings of ASJ, 3-P-18, 1993 1
Proposed in "October". In this conventional example, by converting the parameters of the formant frequency, the duration, the fundamental frequency, and the power, each voice of normal tone, commercial tone, and reading tone, and voice converted from normal tone to commercial tone, and from normal tone A total of 5 voices converted to reading
We prepare two speaking styles and evaluate their similarity.

[0004]

However, in the above-mentioned conventional example, although the conversion of the utterance style is targeted, the speech synthesis is performed without considering the speaker characteristics. That is, if the accent type is different, the height of the accent will be different for each individual,
In the conventional example, the voice of one designated speaker cannot be synthesized.

An object of the present invention is to solve the above problems and to provide a voice synthesizer capable of synthesizing the voice of one designated speaker.

[0006]

According to a first aspect of the present invention, there is provided a voice synthesizing apparatus for synthesizing a voice based on an input character string, the pitch of voices produced for each speaker. It is characterized in that it is provided with a conversion means for converting a character string input using a control rule for controlling a frequency into a voice of a speaker designated in advance.

A speech synthesizer according to a second aspect is the speech synthesizer according to the first aspect, wherein the control rule is
The size of the phrase is controlled based on the number of moras of the phrase to be voice-synthesized and the number of mora of the preceding phrase preceding the phrase, and the accent type of the accent phrase to be voice-synthesized and the sentence of the above-mentioned accent phrase are included. It is a rule to control the pitch frequency of the voice by controlling the size of the accent phrase based on the position of.

[0008] Furthermore, the voice synthesizing device according to claim 3 is
The voice synthesizer according to claim 1 or 2, further comprising:
The learning means includes a learning means for generating the control rule, wherein the learning means extracts the pitch frequency pattern of the voice based on the voice data, and the pitch frequency pattern of the voice extracted by the extracting means. Generating means for generating a model parameter of the critical control model by using an analysis method by the critical control model, a pitch frequency pattern of the voice extracted by the extracting means, and the critical control model of the critical control model generated by the generating means. Generating means for generating a control rule for controlling the pitch frequency of the voice based on the model parameter.

[0009]

DETAILED DESCRIPTION OF THE INVENTION Embodiments of the present invention will be described below with reference to the drawings. FIG. 1 shows the F of this embodiment.
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 exemplary embodiment has an F ₀ control rule (one of 31, 32, and 33) of one speaker selectively connected based on an input character string. And a voice quality control rule 41 and a phoneme duration control rule 42 to generate a feature parameter sequence for voice synthesis, and a voice signal based on the generated feature parameter sequence. And a voice synthesizer 2 for outputting to the speaker 3. In the present embodiment, in particular, the character string input using the F ₀ control rules 31, 32 and 33 for controlling the pitch frequency of the voice created for each speaker is converted into the voice of the speaker specified in advance. The F ₀ control rules 31, 32, 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. Then, the pitch frequency of the voice is controlled by controlling the size of the accent phrase based on the accent type of the accent phrase of the voice synthesis target 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 a F ₀ control rule learning process which will be described later in detail. 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 SW 1, while the switch SW 2 is connected.
Via the F ₀ control rules 31, 32,
One of 33 is selectively connected. The switching of these switches SW1 and SW2 is performed by the F ₀ control rule learning unit 2
0, the voice data of the same speaker and the F ₀ control rule are controlled so as to be simultaneously connected. further,
One of the F ₀ control rules 31, 32, 33 of the speakers A, B, C is selectively connected to the parameter sequence generation unit 1 via the switch SW3. This switch SW3
Is switched so that the operator selects the F ₀ control rule of the speaker whose voice is synthesized. The parameter sequence generation unit 1 is also connected to a conventional voice quality conversion control rule 41 and a conventional phoneme duration control rule 42, which will be described in detail later.

In this embodiment, the voice data 11, 1
2, 13, working memory 21, 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. In addition, the F ₀ control rule learning unit 20,
The parameter sequence generation unit 1 is composed of, for example, a digital computer.

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

First, the processing of step S1 will be described. The voice data 11, 12, and 13 each include voice signal data of a read-aloud sentence (also referred to as a vocalized voice sentence) of one speaker. In this step S1, the voice signal data is subjected to A / D conversion and LPC analysis to extract characteristic parameter data, and then based on the extracted characteristic parameter data, for example, a known critical braking model analysis method is used. (For example, in conventional document 3, “Fujisaki et al.,“ Analys
is of voice fundamental frequency contours for dec
larative sentences of Japanese (Analysis of fundamental frequency patterns of Japanese syllabary sentences) ”, The Acoustical Society of Japan,
(E), Vol. 5, No. 4, pp. 233-24
4, April 1984 ”. ), Detects the F ₀ pattern and the model parameter, and labels them in phoneme units, accent phrase units, and phrase units. Here, the characteristic parameter data includes a logarithmic power, a 16th-order cepstrum coefficient, a Δlogarithmic power, and a 16th-order Δcepstrum coefficient, and the model parameter includes 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 above critical braking model, the phrase component and the accent component are regarded as the response of the critical braking quadratic linear system to the phrase command and the accent command, respectively. The precise timing and magnitude of each command is analyzed by automatic synthesis based on the approximate timing of the phoneme labeling information, accent phrase information, phrase command obtained from phrase boundary information, and accent command (Analysis-by-S).
It can be obtained by using

Next, the processing of step S2 will be described. The Fujisaki model, which has been researched and proposed by Fujisaki, is known as one of the above-described superimposed control models (see, for example, conventional document 1). The hill climbing method has been conventionally used for parameterization using the Fujisaki model (see, for example, 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, since the angular frequency, the input time point and the size of the phrase command and the accent frequency, the rising time point, the falling time point and the size of the accent command are handled as free parameters, the search space is (3I + 4J) -dimensional ( Here, I is the number of phrase commands and J is the number of accent commands.). Among these parameters, the size of the accent command can be uniquely obtained by using the least squares method if the measured value of the F ₀ frequency and other parameters are given.
The calculation time can be shortened by decreasing the number of dimensions of the search space by J. In this method, F ₀ at each time point is
A formula that allows the reliability of the frequency value (here, the maximum value of the autocorrelation function obtained when the F ₀ frequency is calculated from the speech) to be used for weighting the estimation error evaluation of the F ₀ frequency at each time point. It has become. This is because the reliability of the value of the F ₀ frequency obtained from the audio data is not uniform at each time point. In the present embodiment, the F ₀ pattern was parameterized 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 considered to affect the phrase command and the accent command is a known spatial multiple division quantification method (Multiple Split Regressi).
on (for example, the conventional literature 4 "Iwahashi et al.," Statistical modeling of voice control by space division quantification method ", Proceedings of the Acoustical Society of Japan, 1-5-11, pp. 237-238, October 1992). Hereinafter)); hereinafter referred to as MSR method. ). In the MSR method, similar to the analysis procedure using the regression tree, the binary tree is grown by the classification method that minimizes the sum of squared errors between the model estimated value and the measured value, and the model is generated. Further, in the MSR method, nodes other than the leaf node of the binary tree are allowed to share branching conditions over the entire subtrees smaller than that, and modeling can be performed efficiently with a small number of parameters. Since the control factors used for growing the binary tree at the node close to the root node affect the estimated values of many samples, it can be determined that they are important control factors deeply involved in the control of the F ₀ frequency.

By the way, although the target of estimation of the command estimation model includes timing information such as the command size and the rising time point, the timing information can be estimated by a small number of rules, but the command size estimation is complicated. In this embodiment, the size of each command is used as an estimation target because various rules are required. Here, the command estimation models of the phrase command and the accent command are collectively referred to as the F ₀ control rule.

A typical statistical modeling method for generating an estimation model is a quantification type I (see, for example, conventional document 5).
"Hayashi et al.," Quantification theory and data processing ", Asakura Shoten, 1
982 ". ) And a regression tree (for example, in conventional document 6 “B
rieman et al. , "Classifica
tion And Repression Tree
s ", Wadsworth Statistics / P
robbability series, U.S.A. S. A. ,
1984 ". )and so on. Here, the quantification type I is a linear multiple regression model in which control factors are used as explanatory variable spaces, and since independence between control factors is assumed, it is not possible to express a dependency relationship between factors. In addition, in a regression tree in which the explanatory variable space is sequentially divided, independence of the divided explanatory variable space is assumed, so that a dependent relationship between the divided spaces cannot be expressed. On the other hand, the MSR method solves the problems of both the quantification type I and the regression tree by considering the parameters shared by a plurality of divisions in the regression tree analysis process. The result obtained by the quantification I is a special solution of the MSR method that allows only the root node to split, and the result obtained by the regression tree is a special solution of the MSR method that prohibits simultaneous splits at multiple nodes. Each can be thought of as a solution.

Similar to the regression tree, the MSR method analysis produces a tree structure model. FIG. 3 shows an example of a model by the MSR method. In this example, the observed value is set to two types of control factors C
It is possible to estimate from ₁ and C ₂ . The observed estimated value is the quantity at the terminal node when the branches of the tree satisfying the conditions are sequentially traced from the top root node based on the control factor and the quantity a _i written in the node is added at the same time. Obtained as the sum. The value of the quantity a _i is calculated using the control factor obtained from a large amount of data, the observed value and the normal equation,
Similar to the quantification type I and the regression tree, the parameter value cannot be uniquely obtained, so that it is necessary to place restrictions on the values of some parameters. In this embodiment, the quantity on the node side that does not meet the condition (for example, the node on the side 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 is a group of data that reaches the node at the right end of the bottom row (that is, data that selects the node that does not meet the condition in any branch) because a ₃ , a ₇ , a ₉ , and a ₁₁ are all 0 Is the average of the observed values of. The quantity a ₁ can be regarded as an initial value when obtaining an estimated value.

In FIG. 3, the tree structure surrounded by a dotted line is MS.
It is an example of the analysis result peculiar to R method. This subtree is split at the node a ₃ to generate nodes a ₆ and a ₇ , and then split again at the node a ₃ to split the nodes a ₆ and a _7. It is generated by doing. It is possible to regard a ₁₀ and a ₁₁ as quantities as shared parameters. In the case of quantification type I, division is allowed only at the root node, and in the case of regression tree, division is allowed only at the end nodes, so division in a subtree as in the example cannot be expressed. .

As described above, the MSR method used in this embodiment is an extension of the concept of quantification I and the concept of regression tree. Further, the presence of the shared parameter can suppress an increase in the number of parameters of the model, and the modeling can be efficiently performed with a small number of parameters. From this point of view, the MSR method is used as the statistical modeling method in this embodiment.

By using the MSR method to analyze the relationship between the phrase command and accent command obtained from a large amount of voice data by the above-mentioned processing and the control factor affecting each command, the phrase command and the accent command are extracted from the control factor. The model to estimate is obtained. Each model is composed of 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 parameter is also used to calculate the estimated value. By examining the structure of the binary tree obtained by the analysis, it is possible to analyze what kind of control factors influence each command. The size of the model parameter also serves as a criterion for judging whether or not the classification related to the parameter has a large influence on each command.

4 and 5, four speakers M1, M2,
F1, F2 (where M1 and M2 are male speakers, and F
1, F2 are female speakers. ) Shows each F ₀ control rule. Here, FIG. 4 shows the control amount for the mora number of the phrase and the control amount for the mora number of the preceding phrase, and FIG. 5 shows the control amount for the accent type of the accent phrase and the sentence in the sentence of the accent phrase. The control amount with respect to the position is shown. Here, the mora is a beat substantially corresponding to one kana character. Further, 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 F ₀ control rules for the speaker F2 in FIGS. 4 and 5.

[0023]

[Table 1] Specific example of F ₀ control rule for each phoneme sequence <In the case of speaker F2> (corresponding to (d) of FIG. 4 and (d) of FIG. 5) ────────── ────────────────────────── (1) The size of the phrase, the preceding phrase, and the accent command are respectively set to the specified initial values of the 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 1 mora or more and 3 mora or less, reduce the size of the phrase by 0.15 from the initial value. (2-2) If the length of the phrase is 4 mora or more and 6 mora or less, the size of the phrase is reduced by 0.05 from the initial value. (2-3) If the length of the phrase is 7 mora or more and 12 mora or less, 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, the size of the phrase is reduced by 0.025 from the initial value. ─────────────────────────────────── (3) Judgment control regarding the number of mora of the preceding phrase (3-1 ) If the length of the preceding phrase is 1 mora or more, the size of the preceding phrase is reduced by 0.0125 from the initial value. (3-2) If there is no preceding phrase, the size of the preceding phrase is not changed 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, the size of the accent phrase is increased by 0.05 from the initial value. (4-2) If the accent type is 3 or more, the size of the accent phrase is not changed 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 accent phrase in the sentence (4 -1) If the accent phrase is at the beginning of a sentence, the size of the accent phrase does not change from the initial value. (4-2) If the accent phrase is included in the sentence, the size of the accent phrase is not changed 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) Phrase command size control is shown in Table 1. The control amounts of (2) and (3) are to be summed, and the size of the accent phrase is to be summed of the control amounts of (4) and (5) in Table 1.

Next, a synthetic voice "Today is a nice day"
F ₀ control rules 31, 32, used to control each parameter for each phoneme or phoneme sequence when
33, voice quality control rule 41 and phoneme duration control rule 4
Tables 2, 3 and 4 show examples of No. 2 respectively. In Table 3, acoustic characteristic parameters are logarithmic power, 16th-order cepstrum coefficient, Δlogarithmic power, and 1
It is a 34-dimensional parameter including a 6th-order Δ cepstrum coefficient.

[0025]

[Table 2] Example of F ₀ control rule ─────────────────────────────────── Phoneme string F ₀ control Rule ─────────────────────────────────── kyo'uwa phrase size F ₀ control rule accent F ₀ control rule of size ─────────────────────────────────── yo'i te'Nkidese 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] Examples of voice quality control rules ───────────────── Phoneme Acoustic feature parameters ───────────────── ky ( 0.05, 0.03, ...) 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 voice-synthesized is input to the 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), a voice quality control rule 41 that controls the acoustic feature parameter, and a phoneme duration control rule 42 that controls the phoneme duration, using the 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 time matching process and a voice spectrum interpolation process are executed by, for example, the DTW method, and the 16th-order cepstrum, for example. The time series data of the coefficient is generated and output to the voice synthesis unit 2. The voice synthesizer 2 is configured to include a pulse generator, a noise generator, a variable gain amplifier and a filter,
A voice signal is generated based on the input time-series data and is output to the speaker 3, thereby generating a synthetic voice corresponding to the input character string.

In the above embodiment, a small number of voice data is uttered by the conversion target speaker, and the F ₀ control rule generated based on this is the one of the F ₀ control rule generated from a large amount of voice data. The F ₀ control rule may be generated by replacing

[0030]

DESCRIPTION OF THE PREFERRED EMBODIMENTS The present inventor uses the speech synthesizer shown in FIG.
Subjected to F ₀ control rule learning process on speech database, phrase command, and generates a F ₀ control rules to estimate the size of an accent command, generating and analyzing the F ₀ control rules a plurality of speakers, We investigated the commonality of important control factors among speakers.

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

[0032]

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

The analysis of the control factors and control rules used to generate the F ₀ control rule will be described. The parameters used in the critical braking model are input time and size for phrase commands, rising time for accent commands,
There is a magnitude at the time of the fall. Of these, it has been reported that time information such as the input time can be controlled by a small number of simple rules (for example, conventional document 8 "Kaiki et al., Technical Report of IEICE, SP92-").
6, March 1992 ”. ). On the other hand, proper control of the command size is important for improving the naturalness and intelligibility of the synthetic speech. Therefore, in the generation of the F ₀ control rule, the sizes of the phrase command and the accent command are used as targets for estimation.

First, the control factors used to estimate the magnitude of the phrase command and their effects 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 of the phrase in the sentence (divided into two categories, sentence end or non-sentence.) (A4) Accent type of the leading accent phrase of the phrase (divided into four categories.)

When the phrase is short, it is not necessary to keep the phrase component at a high value for a long time, so it is conceivable that the shorter the phrase, the smaller the phrase command. Further, when the preceding phrase is short, the phrase starts before the phrase component of the preceding phrase is sufficiently attenuated, and in this case also, it is expected that the phrase command will become small. Therefore, the lengths of the phrase and the preceding phrase were used as control factors for estimating the size of the phrase command. In addition to this, in speech, 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. It was used to estimate the height. Further, in order to prevent the value of the F ₀ frequency from becoming too large at the beginning of the phrase, the accent type, which is a factor having a strong correlation with the magnitude of the accent component, is used as a factor for suppressing the size of the phrase command.

When a model for estimating the size of the phrase command was generated from these control factors and analyzed, it was confirmed that the lengths of the phrase and the preceding phrase are 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 a relief accent phrase) is present at the beginning of the phrase in 3 out of 4 speakers.

Next, the control factors used to estimate the magnitude of the accent command and their effects will be described. The following four control factors were considered in order to estimate the magnitude of the accent command. (B1) the accent phrase length (specifically, the number of mora of the accent phrase) (divided into 4 categories) (B2) the accent type of the accent phrase (divided into 4 categories) (B3) preceding accent Phrase accent type (divided into 5 categories.) (B4) Position of the accent phrase in the sentence (divided into 3 categories: sentence beginning, sentence, and sentence end).

As is well known, it is known that the accent component becomes small when the accent phrase is short or when the accent type is a flat type, so these factors were considered as control factors. In the results of experiments conducted by the present inventor, the smaller the number indicating the accent type, that is, the smaller the number of beats pronounced "high", the larger the accent command tends to be. Classification was performed (Type 1, Type 2, Type 3 to Type 5, Type 6 and above) for analysis. If the preceding accent phrase is undulating,
Since it is considered that the preceding accent phrase consumes energy for increasing the F ₀ frequency and the accent phrase becomes small, the accent type of the preceding accent phrase is added as a control factor. Furthermore, as described above, the position in the sentence is treated as a control factor for estimating the size of the phrase command, but the size of the accent command may differ at the beginning, in the sentence, and at the end of the sentence. , This was also considered as a factor.

An accent command estimation model was generated using these control factors and the measured value of the size of the accent command, and its analysis was performed. As a result, the accent command of the accent phrase located at the end of the sentence in the factor (B4) was detected. It was confirmed that the size was small in all speaker estimation models. In addition, in this experiment dealing with a larger amount of voice data, no particular individual difference in the influence of the phrase command and the accent command on the size was found.

As described above, according to this embodiment of the present invention, the input character string is designated in advance by using the F ₀ control rule for controlling the pitch frequency of the voice created for each speaker. The F ₀ control rule controls the size of the phrase based on the number of mora of the phrase to be voice-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 voice synthesizer capable of synthesizing the voice of one designated speaker. Also, F
_{0 The} control rule learning unit 20 extracts a pattern of the pitch frequency of the voice based on the voice data, and based on the extracted pattern of the pitch frequency of the voice, a model parameter of the critical control model is used by the analysis method by the critical control model. And a control rule for controlling the pitch frequency of the voice can be generated. Therefore, the optimum and faithful F ₀ control rule for synthesizing the voice of one designated speaker can be automatically and easily created.

[0041]

As described in detail above, according to the voice synthesizer of the present invention, the character string input is designated in advance by using the control rule for controlling the pitch frequency of the voice created for each speaker. And a conversion means for converting into a speaker's voice, wherein the control rule is that the size of the phrase is based on the number of mora of the phrase to be voice-synthesized and the number of mora of the preceding phrase preceding the phrase. This is a rule for controlling the pitch frequency of a voice by controlling the pitch frequency and 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. Therefore, there is a peculiar effect that it is possible to provide a voice synthesizing device capable of synthesizing the voice of one designated speaker.

Further, the learning means for generating the control rule is provided, and the learning means extracts the pattern of the pitch frequency of the voice based on the voice data, and the pitch frequency of the voice extracted by the extracting means. Generating means for generating a model parameter of the critical control model using an analysis method based on the pattern of the critical control model, a pattern of the pitch frequency of the voice extracted by the extracting means, and the pattern generated by the generating means. Generating means for generating a control rule for controlling the pitch frequency of the voice based on the model parameter of the critical control model. This makes it possible to automatically and easily create an F ₀ control rule that is optimal and faithful for synthesizing the voice of one designated speaker.

[Brief description of 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 multiple division 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 the F ₀ control rule learning unit of FIG. 1.

5 is a graph showing an example of F ₀ control rules relating to accent phrases created by the F ₀ control rule learning unit of FIG. 1. FIG.

[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) Yoshinori Kozaka, Yoshinori Kozaka, No. 5, Mihiraya, Seiji-cho, Seika-cho, Soraku-gun, Kyoto Prefecture, Ltd. Shiraka-gun Seika-cho, Osamu Osamu, Osamu Osamu, No. 5, Mihiraya, ATR Co., Ltd.

Claims (3)

[Claims] 1. A voice synthesizer for synthesizing a voice based on an input character string, wherein the input character string is designated in advance using a control rule for controlling the pitch frequency of the voice created for each speaker. A speech synthesizer comprising a conversion means for converting the speech of the speaker. 2. The accent rule to be voice-synthesized is controlled by controlling 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. The speech synthesis apparatus according to claim 1, wherein the rule is a rule for controlling the pitch frequency of the voice by controlling the size of the accent phrase based on the accent type of the phrase and the position of the accent phrase in the sentence. . 3. The speech synthesizer further comprises learning means for generating the control rule, the learning means comprising: an extracting means for extracting a pitch frequency pattern of a voice based on voice data; and the extracting means. Generating means for generating model parameters of the critical control model by using an analysis method by a critical control model based on the extracted pitch frequency pattern of the speech, and a pattern of the pitch frequency of the speech extracted by the extracting means, 3. The speech generating apparatus according to claim 1, further comprising: generating means for generating a control rule for controlling the pitch frequency of the speech based on the model parameter of the critical control model generated by the generating means. Synthesizer.