JPH0519780A - Device and method for voice rule synthesis - Google Patents

Abstract

PURPOSE:To improve the ability to locally emphasize a voice which is called prominence by giving rhythm control parameters according to classifications of the prominence of Japanese. CONSTITUTION:When a sentence unit to which the prominence is added is the whole phrase, a variation quantity DELTADAa which is the difference between 'accent command increment' defined as an increment for the size of an accent component when the prominence is not added and 'accent command increment' for an adjacent accent component is set to a positive value. When the sentence unit to which the prominence is added is part of a word and the accent type is not an accent modified type, the DELTADAa is set to the positive value and a pause is inserted right before or behind the sentence unit; when the accent type is the accent modified type, the DELTADAa is set to the positive value. Further, when the sentence unit to which the prominence is added is one beam of a subject, the DELTADAa is set to the positive value and the pause is inserted right before and behind the sentence unit.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a rule synthesizing apparatus and method for sentence speech, and more particularly to quality improvement of rule synthesizing speech.

[0002]

The following documents are known as techniques related to the present invention.

1. Satoshi Ichikawa, et al .; Experimental study on the naturalness of synthetic speech, Proceedings of the Acoustical Society of Japan 1-3-8 (Sho 42) 2. Tsuyoshi Nakayama, et al. Expression of emphasis, IEICE Conference 64 (Sho 43) 3. Japanese Patent Laid-Open No. 59-081697 (using Fujisaki model for word rule synthesis) 4. Japanese Laid-Open Patent No. 60-074224 (Voice tone of each paragraph 5) Japanese Patent Laid-Open No. 62-138898 (Intonation of question sentence, imperative sentence, desire sentence, etc. is generated by Fujisaki model) 6. H. Fujisaki et. Al., "Analysis of voice fundame
ntal frequency contours for declarative sentences o
f Japanese, "J. Acoust. Soc. Jpn. (E) 5,4 (1984). 7. Sato Toshio; Differences in time elements of voiced and unvoiced plosives, Vol. 14, No. 2 of the Acoustical Society of Japan ( 1958) 8. Kazuo Ochiai; Auditory examination of pitch frequency change in unvoiced plosives, Proceedings of the Acoustical Society of Japan 2-3-12 (Sho 43-1)
1) 9. JP-A-63-174100 (Model in which a phoneme control mechanism, a sentence pattern control mechanism, and an emphasis control mechanism are further added to the Fujisaki model) 10. Hirokichi Hirose, Hiroya Fujisaki, et al. 2; Fundamental frequency Text-to-speech synthesis based on pattern generation process model, IEICE Transactions A, J72-A, 1, pp.32-40 (1989-1) 11. Hisashi Kawai, Keikichi Hirose, Hiroya Fujisaki; Japanese Speech Rules for synthesis of prosodic features in the synthesis of speech, IEICE Technical Report Speech, SP88-129 (1989-1) 12. Japanese Patent Application No. 1-214799 (application for basic patent for prominence production rules) 13. Hiroya Fujisaki, Hirose Keikichi, et al. 2; Realization of Accent Components in Continuous Speech, Voice Study Group Material, S84-36 (1984-
7) 14. Shoichi Takeda, Satoshi Ichikawa; Estimation of model parameters for pitch control mechanism for 4-mora words, Proceedings of the Acoustical Society of Japan 1-5-13 (Sho57-3) A conventional technique will be described.

From the text of any sentence or word,
The method of synthesizing the corresponding speech is called “speech synthesis by rule” or simply “rule synthesis”. Rule-synthesized speech is different from natural speech because it generally has features such as phoneme connections, durations, and changes in pitch (voice pitch) from the outside. Therefore, the speech by rule synthesis is
It is worse than the sound quality of so-called "analysis and synthesis," which preserves the characteristics of natural speech. The factors of the sound quality deterioration of the rule-synthesized speech include the ones caused by the deterioration of the phoneme clarity and the ones caused by the unnaturalness of sentence intonation.

Regarding rules governing the intonation of sentences, that is, prosodic rules, there are already known Japanese rules for generating intonation of sentences including ordinary sentences, interrogative sentences, imperative sentences, and various expressions (( (See References 1 and 2 above). However, since the model used here only gives point pitch information in syllable units, it is not sufficient to express the difference between a question sentence, an imperative sentence and a desire sentence. Therefore, the intonation of speech synthesized by giving such a pitch pattern (fundamental frequency change pattern) sounds unnatural.

In order to sufficiently express the difference in intonation of various sentences, it is necessary to clarify the relationship between the fundamental frequency (pitch frequency) in the syllable and time. The "pitch control mechanism model" described by the critical damping quadratic linear system has been used as a model that can describe the pitch pattern in such a syllable and can clearly define the time structure.

As an application of this pitch control mechanism model, there are an example of application to word voice synthesis (reference 3), an example of application to sentence voice synthesis of interrogative sentences, command sentences, desire sentences, etc. (reference 5). , A considerable sound quality improvement effect is recognized.

[0008] Document 9 further adds a component expressing a local fluctuation of phoneme level (Documents 7 and 8) effective for improving phonological clarity. In addition, a component (reference 5) that expresses a subtle change in fundamental frequency unique to various emotions and expressions, such as a rising tone appearing in an interrogative sentence, a command sentence, and a desire sentence, is added. Reference 9 provides a method of synthesizing a voice having a natural human-like intonation using a modified pitch control mechanism model that generates these components.

[0009]

Among the various pitch control mechanism models described above, the introduction of the phoneme control mechanism has improved the phoneme clarity of synthesized speech. However, ordinary sentences without emotions or special expressions do not remove the monotonous and mechanical feeling of speech. Such monotonousness and mechanical feeling impose a heavy burden on a person who uses the synthetic speech system for a long time and causes fatigue. Unless these monotonous and mechanical feelings are removed, it cannot be applied to a system that is used for a long time, such as a reading operation in newspaper editing.

[0010] On the other hand, one of the reasons why the feeling of fatigue is small even when listening to a natural voice uttered by a human for a long time is that the utterance is changed locally by strengthening it or weakening it conversely. Because it is attached. In other words, human beings, when they want to strengthen, relatively raise their voice, make their voice louder, and speak slowly. On the contrary, when it is not important, he tries to speak with a low and low voice, and swiftly and vaguely. That is, the spoken language is also used to emphasize expressions corresponding to the "brackets" and "bold characters" in the written language. This strengthening or weakening relieves the listener of the need to constantly pay attention to the utterance, thus reducing the burden.

The present invention provides an apparatus and method for realizing the strengthening and weakening of natural speech in rule-synthesized speech.

[0012]

Means for Solving the Problems The above-mentioned strengthening and weakening in the sentence voice is performed by the relative strength with respect to other parts in the sentence. In this way, the strength that makes the other parts stand out relative to other parts (is made prominent) is called “prominence” or “contrast emphasis”.

According to the classification of prominences from a linguistic standpoint, the present invention introduces a scale for quantitatively expressing the prosodic features of these prominences. That is, a prominence generation rule that stores prosody control parameters obtained based on the result of natural speech analysis is stored in accordance with the prominence classification, and the prosody control parameters when prominences are added are controlled according to the prominence generation rules.

In terms of speech information processing, these prominences have (1) fundamental frequency, (2) speech waveform amplitude (power), and (3) time length (phoneme or "pause" duration). It is realized by increasing and decreasing. Particularly, in the present invention, prominence is realized by controlling the size of the accent command. If necessary, the control of the time length by inserting the pause or the control of the power level is performed.

The power has a strong correlation with the fundamental frequency, and as the fundamental frequency increases due to prominence, the magnitude of the power also increases accordingly.

[0016]

Since the prosody control by the puminence generation rule of the present invention is obtained based on the quantitative analysis of natural speech, it is possible to give human-like natural strengthening and weakening to the speech synthesized from the input document (text). it can. According to the present invention, it is possible to realize strengthening and weakening in almost all cases in which a real sentence voice can occur. Therefore, the user can easily understand the utterance content without paying special attention, and the burden on the user can be significantly reduced. In particular, the effect of reducing fatigue when working for a long time such as newspaper editing is remarkable, and improvement in work efficiency can be expected.

[0017]

First, the "pitch control mechanism model" used in the embodiments of the present invention will be described. Here, the pitch control mechanism model is a model as described below.

It is the pitch control mechanism model that the fundamental frequency which gives the information of the voice pitch is considered to be generated in the following process. The frequency of vocal cord vibration, that is, the fundamental frequency, is controlled by an impulse command issued each time the brain switches the phrase and a step command issued each time the accent is raised or lowered. At that time, due to the delay characteristic of the physiological mechanism, the impulse command of becomes a gentle downward curve (phrase component) from the beginning of the sentence to the end of the sentence, and the step command becomes a curve of local undulation (accent component). These two components are modeled as the response of the critical damping quadratic linear system of each command, and the time change pattern of the logarithmic fundamental frequency is expressed as the sum of these two components. FIG. 2 shows a pitch control mechanism model.
The model fundamental frequency F ₀ (t) (t is time) is formulated as the following equation.

[0019]

[Equation 1]

Where Fmin is the minimum frequency, I is the number of phrase commands, Ap (i) is the size of the i-th phrase command, and T is the number of phrase commands.
₀ (i) is the time of the i-th phrase command, J is the number of accent commands, Aa (j) is the size of the j-th accent command, T
₁ (j) and T ₂ (j) are the start time and end time of the jth accent command, respectively. Further, Gp (i, t) and Ga (j, t) are the impulse response function of the phrase control mechanism and the step response function of the accent control mechanism, which are given by the following equations.

[0021]

[Equation 2]     Gp (i, t) = α (i) t exp (-α (i) t) u (t)… (Equation 2)

[0022]

[Equation 3] Ga (j, t) = Min [1- (1 + β (j) t) exp (-β (j) t) u (t), θ (j)] (Equation 3) where , Α (i) is the natural angular frequency of the phrase control mechanism for the i-th phrase command, β (j) is the natural angular frequency of the accent control mechanism for the j-th accent command, and u (t) is the unit step function. Further, θ (j) is the upper limit value of the accent component, and is selected as 0.9, for example.

Here, the fundamental frequency (pitch frequency)
And pitch control parameters (Ap (i), Aa (j), T ₀ (i), T
The unit of the value of ₁ (j), T ₂ (j), α (i), β (j), Fmin) is defined as follows. That is, the units of F ₀ (t) and Fmin are
The unit of [Hz], T ₀ (i), T ₁ (j) and T ₂ (j) is [s], and the unit of α (i) and β (j) is [1 / s]. Also Ap (i) and Aa (j)
The value of is used when the units of the values of the fundamental frequency and the pitch control parameter are determined as described above.

As an analysis method, an optimization method is used. That is, the best approximate estimation of the pitch pattern is performed by obtaining a pitch control parameter that minimizes the error between the pitch pattern generated by the pitch control mechanism model and the actual measurement value obtained by the analysis / extraction of the original speech (references). 6).

Next, the "corrected pitch control mechanism model" will be described. FIG. 3 (a) shows a modified pitch control mechanism model.

The characteristic of this modified model is that a phoneme control mechanism, a sentence pattern control mechanism, and an emphasis control mechanism are added to the model composed of the phrase control mechanism and the accent control mechanism. is there. By introducing these three control mechanisms, it is possible to add various fluctuation components to the pitch pattern.

That is, the above-mentioned phoneme control mechanism is a mechanism for generating a local fluctuation component of the fundamental frequency for each phoneme, for example, voiced consonants / d /, / m /, / n /, / r /, / w It is possible to represent the local fundamental frequency drop of / etc., and the descent characteristic from the high fundamental frequency often found in the transition part to the following vowels of unvoiced plosives / t /, / k / etc. In addition, the sentence pattern designation control mechanism is a mechanism for generating a component expressing the rising of the fundamental frequency at the end of the question sentence. The emphasis control mechanism is a mechanism aiming to generate components expressing various emotions and facial expressions such as imperative sentences and desire sentences.

As a formula for simply describing the modified pitch control mechanism model, for example, the above-mentioned formulas 17 to 24 may be used. Here, the unit of each parameter of the equations 17 to 24 is determined according to the conventional pitch control mechanism. Of course, the formulas to be specifically realized are not limited to the above formulas 17 to 24. Further, the pitch pattern can be generated by a combination of arbitrary control mechanisms of Expressions 17 to 22 depending on the nature of the text voice and the selection of the control method. For example,
If the enhancement is expressed using the emphasis component, the relationship between the accent command and the emphasis command becomes a superposition type as shown in (1) of FIG. 3B. However, the pitch pattern identical to the pitch pattern obtained by these commands can be obtained only by the accent command as shown in (2) of FIG. Such a stepwise change to another command value at the end of one accent command is called "accent transformation".
"Emphasis component superimposed on accent component" and "accent transformation"

[0029]

[Equation 4] Aa ₂ = Aa ₁ + As (Equation 4)

[0030]

[Equation 5] T ₁₂ = T ₇₁ ... (Equation 5)

[0031]

## EQU6 ## Mutual conversion is possible due to the relationship of T ₂₂ = T ₈₁ (Equation 6).

The estimation (analysis) of the model parameters can be performed by the optimization method as in the case of the conventional pitch control mechanism model (Reference 6).

Here, the Japanese Patent Application No.
1-214799 (Reference 12), Japanese Patent Application No. 2-183947, Japanese Patent Application No. 2-
According to No. 250172, (1) the size of the accent command, (2) the power, or (3) the phoneme duration of the prominent part, in the case of strong sentences, as compared with the case of no strong sentences. It can be seen that there is an increase and in some cases poses occur. On the contrary, (1) the size of the accent command, or (2) the power may decrease, as in the case of weakening the end of a plain text. Therefore, strengthening or weakening by prominence increases or decreases each value of these (1) to (3) (these (1) to (3) are collectively called “prosody”). It is realized by reducing. Each element of prosody (1) ~
(3) may increase or decrease independently, or may increase or decrease depending on the combination. As a matter of course, the effect of excellence becomes greater when the amount is increased or decreased by the combination.

The present invention introduces a scale for quantitatively expressing the prosodic features of prominence. That is, the following various quantities are defined as a scale showing the strong position and degree of a prominence-containing sentence (target voice) with reference to a non-strong sentence (reference voice). (1) F ₀ ratio (F ₀ R): Fundamental frequency F ₀ x of target speech with respect to fundamental frequency F ₀ r of reference speech
Is defined by the following equation. However, the fundamental frequency is
The value estimated by the Fujisaki model was used.

[0035]

[Equation 7] F0R = 20log (F ₀ x / F ₀ r) (dB) (Equation 7) (2) Accent command increment (DAa): The accent command of the target voice with respect to the reference command accent command magnitude Aar It is an increment of size Aax and is defined by the following formula.

[0036]

[Equation 8] DAa = Aax-Aar (Equation 8) (3) Power ratio (POWR): The ratio of the power Px of the target voice to the power Pr of the reference voice, which is defined by the following equation.

[0037]

[Equation 9] POWR = 10log (Px / Pr) (dB) (Equation 9) (4) Time change rate (TIME WARP): Indicates the degree of time expansion / contraction of the target voice with respect to the reference voice. Now, let Tr be the duration of the corresponding phonemes of the reference and target voices.
(i) and Tx (i) (i is the meaning of the i-th phoneme), the time change rate TW (i) of the i-th phoneme is defined by the following equation.

[0038]

[Equation 10] TW (i) = (Tx (i) -Tr (i)) / Tr (i) × 100 (%) (Equation 10) Prominence quantitative analysis results are obtained using the above prosodic feature scale. The summary is as follows (Japanese Patent Application No. 2-183947)
No., see Japanese Patent Application No. 2-250172).

[1] Fundamental frequency The characteristics of the fundamental frequency of the prominence are examined with respect to the accent command size Aa, the starting time point T ₁ , the ending time point T ₂ , and the accent deformation starting time point T ₁₂ . The accent start / end and accent transformation start time are based on the syllable boundary where the accent rises from low to high, the syllable boundary where the accent falls from high to low, and the syllable boundary time when the accent changes from high to other high. The values are obtained as ΔT ₁ , ΔT ₂ , and ΔT ₁₂ . However, in the case of head-height accent or prominence of the _first syllable, the start time of the _first syllable is used as the reference time for ΔT ₁ measurement, and in the case of flat-type or tail-high accent or prominence of the last syllable, the end syllable end time is set. Let ΔT ₂ be the reference time for measurement.

In the case of intentional prominence, the accent command increment DAa is defined as a scale showing the degree of enhancement by the fundamental frequency, as described above. However, prominence may be realized not by increasing the target accent command itself, but by relatively reducing the size of accent commands before and after the accent command. In this case, DAa does not take a large value. Therefore, the difference between the accent command increments defined by the following equation is evaluated as a measure of the prominence effect by the value of the fundamental frequency.

[0041]

[Equation 11] ΔDAa = DAap-DAan (Equation 11) Here, DAap represents the command increment of the accent component in which prominence is placed, and DAan represents the smaller value of the command increments of the accent components adjacent to this accent component. .

Regarding the fundamental frequency, the following tendencies were observed.

(1) The magnitude of the accent command tends to depend on the accent type (accent modified type or not) in the accent components other than the head and end sentence clauses. However, the effect of the presence or absence of a pose was not observed (Fig. 5 (a)). On the other hand, in the case of the accent component of the end sentence,
There is a large variation in the size of the accent command. Furthermore, the accent component of the head bunsetsu does not depend on the speaker or the accent type, and the data shows almost the same distribution as the data of the accent-modified accent component other than the head and end bunsetsu. This is thought to be due to the influence of the default prominence described below, in which the degree of intentional prominence strengthening becomes relatively small (Fig. 6 (a)).

(2) Regarding the interrogative sentence, there was a tendency that the size of the accent command strengthened at the end of the sentence (FIG. 5 (b)).

(3) Regarding the size of the accent command, the accent of the first bunsetsu of the plain text has a default prominence (FIG. 6 (b)). Here, the value of the difference between the accent command size Aa ₁ of the first phrase (prosodic word) and the accent command size Aa ₂ of the second phrase (prosodic word) was used as a measure of the default prominence size. .

(4) At the start and end points of the accent command, no prominence effect was observed.

(5) At the time when the modification command of the accent modification type (usually increased) prominence is started, the speaker A tends to advance with respect to the reference phoneme boundary, but the speaker B has both advance and delay. Not (Fig. 7). From this result, it is considered that the start time of the prominence deformation command can be set to coincide with the reference phoneme boundary time.

[2] Power As shown in FIG. 8, a high correlation is found between the power P and the fundamental frequency (accent command magnitude Aa) (correlation coefficient ρ≈0.9). The regression line at this time is represented by the following equation.

[0049]

[Equation 12] P = 11Aa (dB) (Equation 12) Therefore, by using the equation 12, the increase amount of power accompanying prominence can be uniquely determined from Aa.

Alternatively, allowing a slight variation

[0051]

[Equation 13]     P = 11Aa ± 4 (dB) (Equation 13) The value may be set within the range.

Since this value is almost equal to the value of the natural increase in power due to the increase in the fundamental frequency, simply the amplitude value of the sound source signal (for example, the prediction residual) is set to the synthesizer as a constant value regardless of the fundamental frequency. It may be a simple process of only sending. As a result, the power of the synthesized speech waveform naturally rises depending on the fundamental frequency.

[3] Temporal Structure The main factors of the temporal structure are the phoneme duration and the pause duration, and the prominence can be expressed by the extension of these durations (the occurrence of a pause is from 0 to the pause duration). The special case of increasing to a positive value). Here, (1)
The effect of pause duration on phoneme duration, and
(2) We investigated from the viewpoint of extension of phoneme duration at the end of the question sentence.

(1) The phoneme duration tended to extend when a pause was generated immediately after. This tendency is common to speaker AB (Fig. 9).

(2) It was found that vowels at the end of a question sentence also tended to expand. This tendency is common to speaker AB (Fig. 10).

FIG. 11 shows the values (stronger or weaker) of each element of the prosody for generating prominence corresponding to each classification of FIG. 4 obtained on the basis of the above-mentioned quantitative analysis result for natural speech. is there. However, the numerical example of FIG. 11 is represented by an increment or an increase rate with respect to each control value when prominence is not added. If a control rule is created according to FIG. 11, natural prominence can be added to the synthesized voice. In FIG. 11, the numerical value on the left side of the symbol "±" is the representative value of the control parameter,
The symbol "±" indicates the range of variation of the numerical value (corresponding to approximately 1σ). That is, it means that natural prominence can be generated as long as the numerical value is set within this variation range. The same effect can be obtained by relatively weakening the part to which prominence is not added with the parameters in Fig. 11. In this case, according to a conventionally known prosody control rule (for example, known example 12), The control parameter value may be obtained, and the prosody control parameter value at the time of adding prominence may be obtained by using the defining expressions of the above equations 7 to 13.

Based on the above analysis of prosodic features of prominence and verification experiment of validity of rules, various realization forms of prominence are summarized as shown in FIG. That is, prominence is realized by any of the following methods.

(1) Increase in fundamental frequency (accent command size) and accompanying increase in power (2) Pause insertion Immediately before a prominence target Immediately after a pause insertion prominence target Immediately before a pause insertion prominence target and Immediately after that, a pause is inserted (3) Speech rate is reduced (4) The combination of (1), (2), and (3) above, the prominence-containing sentence voice synthesized according to FIG. 11 has prominence expression power equivalent to natural voice. However, it is not always the highest level of prominence expression. In order to realize more reliable intention transmission, it is desired that the prominence expression power is higher. Therefore, through probabilistic evaluation experiments, we determine prosodic control parameters for realizing aurally optimal prominence expression. That is, of the various prominence realizations as shown in FIG. 12, for each prominence of each classification, which form is superior in terms of prominence expressing power is evaluated by a listening experiment of rule-synthesized speech. Furthermore, the fundamental frequency (F
₀ ) (+ power) increase prominence,
The relationship between the size of accent command and prominence expression is clarified, and what value is most suitable to be selected in the trade-off with naturalness.

The evaluation is performed based on whether or not various prominence-containing sentence voices synthesized by the prominence generation rule sound like prominences are added as intended.

The test method will be described below.

(1) Among intentional prominences, a speech is used in which 66 kinds of sentences shown in the lists of FIGS. 13 to 15 containing various realization forms of prominences belonging to each classification are synthesized by a rule. Here, regarding prominence due to increase in F ₀ (+ power), in order to examine the relationship between the size of the accent command and the prominence expressing power or naturalness, synthesis is performed by using four increments of the accent command. “Type 1” in the table is a type in which a low accent portion exists between immediately before the target portion of prominence and the target portion when prominence is not placed. That is, the two prosodic words do not undergo accent transformation and the original prosodic words are retained (for example, "rain gear of the elder brother's rain gear"). The second half of the prosodic word (for example, "Chief" in "Chief Nakagawa"), or one beat in the low accent part of one prosodic word (for example, "Baba" in "Kuwabara") is subject to prominence. This is the case. On the contrary, “type 2” is a type in which high accent continues from immediately before the target portion of the prominence to the target portion when the prominence is not placed. That is, the latter half of the two prosodic words that have undergone high-accent transformation to become one prosodic word (for example, "Amegu ha" in "My sister's rain gear" or "Chief" in "Tanaka chief"), or in one prosodic word This corresponds to the case where one beat of the high accent portion of (for example, "Da" of "Nakada-san") is subject to prominence. Since Type 1 and Type 2 have different relative accent heights between the prominence target portion and the portion immediately preceding it, it is expected that the auditory effect of prominence when the fundamental frequency is increased is also different.

(2) The rule-synthesized voices of the above-mentioned 66 kinds of sentences are arranged in random order and presented one by one, and by listening, the answers about the presence or absence of prominence, the position of prominence, etc. are written on the answer sheet. However, the number of presentations of each sentence voice is twice for one type of sentence, and a total of 132 sentences are presented. Between voices, pause for 3 seconds to answer. The subjects were 10 adult men and women with Tokyo accents (type 2). If you hear that there is an emphasized part, answer by enclosing the corresponding part of the sentence written on the answer sheet with a circle. If you hear that nowhere is emphasized, circle the part that says "no emphasis", and if it sounds more unnatural, circle the part that says "unnatural". I was instructed to do so.

(3) The concrete numerical values of the control parameters of the prominence generation rule are as follows unless otherwise specified. DAa or Aa is 1: 0.1 in the order of numbers in FIG.
2: -0.2, 3: 0.0, 4: 0.2, 5: 0.6 (However at the end of the sentence
0.7), 6: 0.3, 7: 0.2, 8: 0.4 (when the last pose is inserted),
0.7 (without pose). The value of TW of the phoneme duration immediately before the pause was 66%, and the other parameters were the values shown in FIG. 11 or the average values (values without ±).

(4) In the prominence due to the insertion of a pause, the phrase is usually rebuilt due to the insertion of a pause. Also, in some cases, the accent type is changed to an accent type different from the original accent type. In this experiment, the following exception rules were adopted in order to minimize the influence on the sound quality due to the change of these fundamental frequencies. These exception rules were created by listening to the synthetic voice while referring to the analysis result of the sentence containing prominence.

A. In the case where the prominence target is a part of a compound word, in the case of "○○ Chief Amendment ○○ Section Manager", if a pause is inserted immediately after "Chief Chief", the fundamental frequency of "Amend" increases due to the effect of phrase re-building. . In order to weaken this effect, the size of the "revision" accent command is set to 0.
Suppressed to value.

B. An example of the subject of prominence is one beat (low accent) of the topic. In "Kuwabara-san" and "Kuwahara-san", if you insert a pose immediately before or after "ba" or "ha", the original prosodic word The specific accent is broken and the usual accent rules cannot be applied. In this case, the following fundamental frequency pattern generation rule was applied.

When a pose is inserted immediately after "ba" or "ha", the magnitude of the accent command of "Rasan Tomo" that follows is suppressed to 0 for the same reason as in the above a.

Similarly, when a pause is inserted immediately before "ba" or "ha", the magnitude of the accent command of "ba (ha) la san tomo" that follows is suppressed to 0 value.

When a pause is inserted immediately before and after "ba" or "ha", the phrase corresponding to the immediately preceding pause is not rebuilt, and the accent level of "ba" or "ha" is set to "low".

C. In the example of the subject of prominence, which is one beat (high accent) of the topic, in the case of "Nakada-san or Nakata-sanka", even when inserting a pose immediately before or after "da" or "ta", the original prosodic word peculiar The accent of is destroyed and the normal accent rules cannot be applied. Also in this case, the following exceptional fundamental frequency pattern generation rule is applied.

When a pause is inserted immediately after "da" or "ta", the size of the accent command for the subsequent "sanka" is suppressed to 0 for the same reason as in the above a.

When a pose is inserted immediately before "da" or "ta", the subsequent "da (san)" accent is a high accent (flat type).

When a pose is inserted immediately before or after "da" or "ta", the phrase corresponding to the immediately preceding pose is not rebuilt, and the accent level of "da" or "ta" is set to "high".

(5) The prominence expressing ability is evaluated by the correct answer rate (prominence correct listening rate) r defined by the equation (14).

[0075]

## EQU00004 ## r = R / N (Expression 14) Here, N is the total number of presentations of all the subjects of the sentence, and R is the number of correct answers of prominence in the N presented speech (provided that the intended prominence is given. The number of audio presentations heard).

Also, the naturalness of the expression format of prominence is evaluated by the unnatural rate u defined by the equation (15).

[0077]

[Mathematical formula-see original document] u = U / N (Expression 15) Here, U is the number of presentations of speech whose expression sounds unnatural in N presentation speeches.

The test results will be described below.

(1) FIG. 16 shows the effect evaluation results of F ₀ increase + power increase. From the figure, in the prominences belonging to any of the categories, the correct answer rate r increases as the size of the accent command increases. However, type 2 (Fig.
(b), (d), (f)) are more type 1 ((a), (c),
(e)) Lower accent command magnitude increment ΔDAa
A sharp increase tendency is seen in the value of r at the value of.

Furthermore, the reason that the correct answer rate r can be brought close to 100% by sufficiently increasing the size of the accent command without significantly impairing the naturalness is that the target of prominence is "entire phrase" (FIG. (A), This is the case of (b)). On the other hand, in other cases, there is a greater tendency to compensate for naturalness in order to obtain a high correct answer rate ((c), (d), (f) in the same figure).

(2) The result of the effect evaluation of the pose insertion is shown in FIG. However, “None” on the horizontal axis in the figure is the reference data showing the result when the maximum correct answer rate is given when prominence is achieved only by increasing F ₀ + power without inserting a pause. Is.
From the figure, it can be seen that the insertion of a pose is the most effective way to improve the prominence expression.
(e), (f)). Regarding the pose position, it is on a case-by-case basis whether the pose should be placed immediately before or after the subject of prominence. However, when placed both immediately before and after (“front and back” in the figure), stable and good results are obtained compared to when placed on only one side.

(3) FIG. 18 shows the evaluation result of the combined effect of F ₀ increase + power increase and pause insertion. In the figure, “F ₀ ” or “F ₀ + pause” on the horizontal axis is an abbreviation for “F ₀ increase + power increase” and “F ₀ increase + power increase + immediately before / after pause insertion”, respectively. is there. In these two cases, the size increment ΔDAa of the accent command was set to the same value in the same sentence in order to see the effect of the pause insertion under the same condition. Further, in the figure, the word "pose" on the horizontal axis indicates F ₀
In addition, it is a case where the prominence is to be realized only by inserting the pose immediately before and after the prominence target portion without increasing the power. In the example of the figure, as the prominence due to the increase of F ₀ + the increase of power, the case where the degree of emphasis is relatively weak is referred to. Therefore, in prominences belonging to any of the categories, the expression of prominences is greater when the pose is inserted. The pose insertion effect increases as the size of the sentence unit targeted for prominence decreases as the whole phrase → part of the compound word → one beat of the topic. Furthermore, if "F ₀ increase + power increase" and "immediately before and after pause insertion" are combined, the prominence expression power is further enhanced.

It can be said that the above results show that even if the increase in F ₀ + increase in power alone has little effect, more effective prominence can be realized by combining with pause insertion. In addition, an example in which the naturalness is greatly reduced as a cost of the improved prominence expression ability due to this combination effect can be seen only in an example in which the target of prominence is a part of a compound word (FIG. 18 (d)). , Not seen elsewhere (other than (d) in the figure).

Next, an embodiment of the speech rule synthesizing apparatus and method according to the present invention will be described with reference to FIGS. 1 and 19 to 28.

FIG. 19 shows the overall construction of an embodiment of a speech rule synthesizing apparatus applicable to speech synthesis of an arbitrary sentence. In this embodiment, if the text of a kanji / kana mixed sentence is given as input data, the corresponding synthetic speech can be obtained as an output. The processing procedure is as follows.

First, the input text is decomposed into each word by the morphological analysis means of the Japanese analysis section 1 (see Japanese Patent Laid-Open No. 59-98236), the part of speech is determined, and the reading is further determined. Next, based on this result, the speech language processing unit 2 (Japanese Patent Publication No. 59-13040, Japanese Patent Publication No. 59-081697, Japanese Patent Publication No. 61-
In Japanese Patent No. 6693), the accent type of each word or phrase is determined.

As the processing result of the syntax level as described above, syllable information, accent information, prominence information, etc. are obtained. The delimiter between phrases and sentences is determined based on delimiters such as punctuation marks in input text. The pause length in a sentence or between sentences can be specified by the number of spaces after the reading and punctuation. In some cases, the type of sentence such as question sentence, imperative sentence, and desire sentence can be determined by utilizing the ending of the sentence, or "?", "!", And "!" Instead of punctuation at the end of the sentence. It can also be specified by using a terminator such as ". For example, even if the same phoneme sequence "cross the river", "cross the river." Is a plain text,
"Across the river?" Is a question.

The above syllable information, accent information,
Pause information, phrase / sentence break information, grammatical information (for example, part of speech if necessary), and prominence information are expressed by a series of numbers called "syllable chords". The syllable code is input information of the control parameter generation unit 3.

In the control parameter generator 3, the accent, intonation, phoneme duration, and sound source power (amplitude) correction value are determined by the rule, and the pitch pattern and the phoneme parameter time series are generated according to the rules.
Here, the sound source power correction value is a coefficient for increasing / decreasing the standard sound source power value depending on the presence / absence of strength. This sound source power correction value may be given as a magnification as compared with the case where there is no strengthening, or as an absolute value. The accent type can be known from the accent information.
Specifically, the accent information is given by inserting a syllable code number indicating an accent immediately after a phoneme with an accent nucleus (a phoneme immediately before the accent descends). However, the absence of this syllable chord indicates that it is a flat accent. Further, the intonation is basically defined by the sentence type information and the prominence information. However, a modification is also added due to the difference in the phoneme sequence of the ending. For example, the intonation pattern is different between the wish sentences “I want to cross the river!” And “I want to cross the river!”. The final pitch pattern is generated based on both accent type and intonation. However, a sentence containing prominence, which will be described later, may be accompanied by accent transformation. The phoneme duration is determined as a unique length for each type of consonant, since the influence of ambient conditions is small for consonants. On the other hand, vowels undergo various transformations depending on the ambient conditions. Therefore, accent type, number of syllables, position within a word,
The duration is determined from the type of consonant immediately before, the type of vowel, etc. (see Japanese Patent Laid-Open No. 59-081697). When the phoneme duration is determined in this way, the phoneme parameters (spectrum envelope parameters and sound source parameters in the case of the generation source method, and the voice in the case of the waveform synthesis method) registered in the file in CV (consonant-vowel concatenation) units. Element) corresponding to the syllable code and extracted and arranged. At this time, if it is too long, it is cut to fit within the duration. After that,
Interpolate between CV units so as to fill the cut or gap (generation method: linear interpolation for spectral envelope parameters,
The sound source parameters are connected by repeating the same value, and the waveform synthesis method: interpolation of the maximum value of the segment extraction window) (see FIG. 26 for details). Finally, the fundamental frequency and the phoneme parameters generated by the above processing are sequentially sent to the speech synthesizer 4, and the speech waveform is output. Here, as a voice synthesis method, for example, a residual compression method (Japanese Patent Laid-Open No. 60-150100,
JP-A-61-296390) may be used. In this case, the sound source pulse is basically generated by extracting a residual pulse (representative residual) for one pitch for each frame and arranging the representative residual at intervals of a pitch cycle given from the outside. . At this time, if the pitch period given from the outside is shorter than the length of the representative residual, the end of the representative residual is truncated by the length difference, and conversely, if it is longer, 0 is filled only in the section where the representative residual is insufficient. There is. FIG. 19 shows an example in which the residual compression method is used in the speech synthesis unit 4, but of course the speech synthesis method is not limited to the residual compression method. For example, a waveform synthesizing method, especially a segment editing method may be used.

The above-mentioned processing can be configured by any known means except for the prominence generation rule described below.

Below, the most important part of the present invention, that is, the description of the prominence generation rule in the control parameter generation unit 3, will be mainly described with reference to FIGS. 1 and 20 to 28.

The prominence information is the following (1) to (5)
It can be extracted from the information.

(1) From a sentence type such as a plain text / an interrogative sentence (excellence peculiar to the sentence form) (2) Syntax information (see Reference 10).

(3) Old information / new information (see Reference 11),
Idiomatic tone.

(4) From text information (square brackets, bold letters, underlines, etc.)

(5) Semantic information (eg, strengthening the answer part to the preceding question sentence).

In the above (1), the parameter that realizes prominence can be generated from the sentence type information, whereas in (2) to (5), the prominence information (syllabic code) is used by the speech language processing unit 2 or the like. Expression) must be generated. For example, when the prominence information is obtained from the brackets in (4) above, when the opening of the brackets is detected, the start time of the accent command and the size information (or the classification information of the prominence (for example, information as shown in FIG. 4)) are input. The contained syllabic code is issued, and when closing of the bracket is detected, the syllabic code containing the information at the time when the accent command ends is issued. In the case of (5), a semantic analysis means is required. If the semantic analysis means is not used, (4) will be substituted. That is, what the person wants to strengthen may be designated in the text by the above-mentioned brackets and the like.

First, an example of a rule for realizing prominence peculiar to a sentence pattern will be shown. First, in FIG. 20, the syllable code string obtained from the speech language processing unit 2 is input to the sentence type determining means 5. Here, as the first step, the corresponding sentence type is determined by matching the ending forms registered in the ending dictionary in the sentence type information dictionary 6 with the ending forms of the syllable code strings. The final type in Fig. 20 is
In the case of modern sentences, it is determined based on the rules of national grammar, such as the ending of the line "u" for verbs and the ending "i" for adjectives. In the case of imperatives as well, in the case of modern sentences, the inflection ending is defined as "e" line. The above sentence type determination becomes more reliable if there is grammatical information such as part-of-speech information. If the inflection of the ending is determined to be the final form, this sentence is not necessarily a plain sentence. So, as the second step, in this case, go to the end symbol of the sentence (end-of-sentence symbol) and determine the sentence type by the type of this symbol (for example, "." Or "." Is a plain sentence, "?"
If it is an interrogative sentence, if it is "!!" it is an imperative sentence, if it is "!" It is a wish sentence, etc.).
FIG. 21 shows an example of the processing of the sentence type determining means 5 described above.

Returning to FIG. 20, the sentence type determining means 5 selectively outputs only the sentence type information described above.

The syllable information extracted from the syllable code string by the syllable information extracting means 16 (for example, the kind of syllable such as "A", "I", "U" represented by a number) determines a phonological boundary. And for generating phoneme components in the pitch pattern. That is, with respect to, the phonological element duration determining unit 7 determines the phonological element duration time of each syllable based on the syllable information, and the phonological element boundary time is determined by the phonological element boundary determination unit 7 in the form of arranging them. The phoneme boundary time is used, on the one hand, to generate phoneme parameters such as LSP parameters. See also
It is used in the sentence pitch control parameter generation unit 11 to determine a phoneme control mechanism parameter value.

The above sentence type information includes the intonation regulation section 8 and the sound source power (amplitude) correction value calculation means 15.
, And a transformation from the standard intonation (eg, Hirajo sentence) is added according to the type of sentence. There are three types of deformation: time deformation, pitch amplitude (command magnitude) deformation, and sound source power or amplitude deformation. The time deformation acts on the phoneme boundary determining means 7 to change the phoneme boundary time. On the other hand, the modification of the size of the command acts on the sentence pitch control parameter generation unit 11 to change the size of the command or add a new sentence pattern designating command or an emphasis command. At this time, the standard intonation control parameters are supplied from the accent rule unit 10.
Note that the sentence pitch control parameter generation unit 11 obtains a reference phoneme boundary time (timing reference information) from the phoneme boundary determining means 7 in order to achieve temporal matching with the phoneme information. Further, the deformation of the sound source power acts on the sound source power (amplitude) correction value calculation means 15, and the correction value of the sound source power value is calculated and sent to the sound source generation unit. The correction value of the sound source power value can be calculated by using Expressions 12 and 13, but if the natural increase in power due to the increase in the fundamental frequency is used, the correction process may be omitted.

The above intonation rule can be achieved by providing a rule table (see Document 5) in the intonation rule section 8 and referring to it. Thus, among the prominences, the prominence peculiar to the sentence pattern is realized by the above means.

On the other hand, intentional excellence ((4) and (5) above)
The prominence information for other default prominences ((2), (3), etc.) is obtained from the prominence information code extracted by the prominence information extracting means 14 from the syllable chord string. The prominence information acts on the intonation regulation section 8 and the sound source power (amplitude) correction value calculation means 15.

Here, a specific example of a method for extracting sentence type information, syllable information, and prominence information from a syllable code string will be described. For example, if the information content is defined according to the syllable code number as shown in FIGS. 22 and 23, the sentence type determining means 5 and the syllable information extracting means 1 are defined.
6. By providing each of the prominence information extracting means 14 with a numerical magnitude judgment function, it is possible to judge whether or not the information is the corresponding information. That is, if the syllable code is 1 to 400, it is determined to be syllabic information, and if it is 9004 to 9020, it is information that gives a sentence type, and thus the sentence type information can be determined by the above-described method. If the syllable code is 9030 to 9039, it is determined to be prominence information, and for example, the accent command value information may be assigned to the last one digit. As an example, when the last digit of the syllable code is represented by I, the accent command increment value DAa by prominence with respect to the magnitude of the accent command when prominence is not added can be given by the following equation.

[0105]

[Equation 16] DAa = 0.1I (Equation 16) Using Equation 16, the syllable code can increase the magnitude of the accent command in the range of 0.0 to 0.9 in 0.1 steps. Of course, when it is desired to change the size of the accent command in smaller steps, the syllable code is assigned to another value range (for example, 9100 to 91).
99), the accent command value information may be assigned to the last two digits. Further, the timing for increasing / decreasing the accent command by prominence can be determined as follows, for example. First, a syllable boundary that determines the accent command start time is specified between the syllable code having the prominence information (for example, 9030 to 9039) between the syllable code immediately before the boundary and the syllable code immediately after the boundary. Can be achieved by inserting into Next, the designation of the syllable boundary that determines the end point of the accent command is, for example, 9 as a code indicating the end of prominence.
Similarly, 030 can be achieved by inserting it between the syllable code corresponding to the syllable immediately before the boundary and the syllable code corresponding to the syllable immediately after the boundary. Also, when the start or end of prominence occurs in a high-accent area, that is, in the case of accent deformation type, the specification of the syllable boundary that causes accent modification is likewise done between the syllable chords corresponding to the syllable immediately before and after the boundary. This can be accomplished by inserting the start or end code of. Thus, when the timing reference time for setting the accent command start / end time point by the prominence is determined, the actual start / end time points can be obtained by searching the timing table as the deviation amount from this reference time. FIG. 24 shows an actual example.

Next, a specific example of a method for controlling power and generating a pose will be shown. Although FIG. 20 shows an example in which a generation source method (for example, residual compression method + LSP combiner) is used in the speech synthesis unit, the present invention is not limited to the specific generation source method described below. Of course, even in the waveform synthesizing method, the power of the waveform amplitude can be controlled by the same idea.

FIG. 25 shows an example in which the residual compression method is used in the speech synthesizer 4. As the spectral envelope parameter, any parameter such as LSP parameter and PARCOR coefficient can be used. Incidentally, the connection interpolation processing in the figure can be realized by the processing as shown in FIG. 26, for example. The square root of the power value obtained by the sound source power (amplitude) correction value calculation means 15 (FIG. 20) (the value as it is if given by the amplitude value) is given to the voiced sound source generation unit or the unvoiced sound source generation unit, and left. The difference (source) amplitude is corrected. When the correction value is given as an actual value, for example, in order to prevent time discontinuity, a power measurement value (for example, Japanese Patent Application No. 1-214799) is calculated for each frame.
No. 2, Japanese Patent Application No. 2-183947, Japanese Patent Application No. 2-250172), the amplitude envelope curve (for example, FIG. 27) approximates to the square root. If the correction value is given as a scale factor, the natural voice's amplitude envelope of the synthesis unit can be used, so the sound source amplitude value of the synthesis unit is multiplied by the specified scale factor only between the frames corresponding to the emphasis section. You just have to shift. Further, when a pause having a predetermined duration is generated, a silence generation command may be issued only during that time to output silence (zero value).

FIG. 28 shows an example in which the voice synthesis section 4 uses a waveform synthesis method. In this case, the sound source power (amplitude) correction value calculation means 15 of FIG. 20 is replaced with the waveform power (amplitude) correction value calculation means, but the processing content is exactly the same as that of the sound source. The only difference is the realization values. The square root of the power value obtained by the waveform power (amplitude) correction value calculation means (the value as it is if it is given as an amplitude value) is given to the element window generator, and the element amplitude is corrected at the time of editing the element. . The time-varying pattern of the correction value is given by the same idea as in the residual compression method. In addition, the pause generation method can be realized by outputting a zero-amplitude waveform of a predetermined time length as in the residual compression method.

Also in the case of the other synthesizing methods, the power (amplitude) control can be realized by a completely similar method according to each waveform amplitude control means.

[0110] The prosody (pitch, power,
FIG. 11 shows an example of a control method of (time length).
It should be noted that the reference value in the case where prominence is not included in FIG. 11 may be determined by a known accent component generation rule (see References 3 and 14) with respect to, for example, the size of the accent command and the start and end times. Alternatively, as a simpler method, the reference value Aa = 0.3 for the accent command size and the relative value ΔT ₁ = ΔT ₂ = ΔT ₁₂ = 0 from the reference syllable boundary at the beginning and end of the accent command can be used for practical sound quality. Almost no hindrance. Since FIG. 11 is obtained based on the quantitative analysis result of natural speech (FIGS. 5 to 10), if the speech is synthesized in accordance with the prominence generation rule of FIG. Is obtained. Of course, FIG. 11 is an example of the parameter realization values, and is not limited to these numerical values. In reality, there are various kinds of strong deformations, so that there is an infinite number of possibilities of corresponding numerical deformations. Among such numerical transformations, the parameter values can be selected so as to have excellent performance in terms of intention transmission (that is, excellent prominence expressing power).
An example of such parameter selection will be described below.

Based on the results of the listening evaluation experiment described in detail above, FIG.
FIG. 1 shows which of the various realization forms of prominence shown in FIG. 2 is superior in terms of prominence expressing power. Further, regarding the fundamental frequency, it is also described what kind of accent instruction magnitude ΔDAa should be given.

If it is assumed that accurate intention transmission is most important in conversations such as telephone calls, it is practically meaningful to adopt a realization with high prominence expression, even if a certain amount of naturalness is sacrificed. There is. With respect to the special example in FIG. 1, the optimal prominence realization mode is selected from this point of view.

The value of r in the table is the value of the prominence correct listening rate (correct answer rate) when the optimal prominence realization form is adopted in the above meaning, that is, the maximum value of r, and is generally used. An average of 98% was obtained in the example, and an average of 96% in the special case. These results indicate that the prominence expressiveness was improved by about 20% compared to before the optimization of the realization form and the prosody control parameters (about 77%). It should be noted that the prominence expression format using the reduction of the speech rate is not very effective as a method for enhancing the prominence expression power.

As described above, FIG. 1 is the center of the present invention, and provides means for remarkably improving the prominence expressing power.

FIG. 24 shows a specific example for actually implementing the prosody control shown in FIGS.

In this embodiment, an example is shown in which the prominence pitch is strengthened or weakened by increasing or decreasing the accent command, but of course, as described above, the emphasizing component may be used. In this case, for example, the parameter values may be converted by using Equations 4 to 6, or a new parameter table may be recreated.

On the other hand, the phoneme control parameter is obtained by analyzing in advance the magnitude of the command, the natural angular frequency, the relative time from the boundary, the value of the bottom, etc. for each phoneme, and the phoneme rule is used as a table corresponding to the syllable information. It may be provided in the section 13. From here, the phoneme control parameter sequence is sent to the sentence pitch control parameter unit 11 in the order of the syllable information sequence. Here, the phoneme start or end time (relative time) is converted into an absolute time based on the timing reference information. Thus, the pitch control parameter created by the sentence pitch control parameter generation unit 11 is sent to the pitch pattern generation unit 12, where the new pitch control mechanism model (Equations 17 to 24 below).
Produces a sentence pitch pattern.

Phrase control mechanism:

[0119]

[Equation 17]     Gp (i, t) = α (i) t exp (-α (i) t) u (t) (Equation 17)         t: time α (i): i-th natural angular frequency u (t): Unit step function Accent control mechanism:

[0120]

[Equation 18]     Ga (j, t) = Min [1- (1 + β (j) t) exp (-β (j) t) u (t), θ (j)] (Equation 18)         β (j): jth natural angular frequency θ (j): j-th upper limit value Phoneme control mechanism:

[0121]

[Formula 19]     Gf (k, t) =-Min [1- (1 + γ (k) t) exp (-γ (k) t) u (t), φ (k)] (Equation 19)     Or

[0122]

[Equation 20]     Gf (k, t) = exp (-γ (k) t) u (t) (Equation 20)         γ (k): kth natural angular frequency φ (k): k-th base value Sentence control mechanism:

[0123]

[Equation 21]     Gt (l, t) = Min [1- (1 + ζ (l) t) exp (-ζ (l) t) u (t), θt (l)] (Equation 21)         ζ (l): l-th natural angular frequency θt (l): l-th upper limit value Emphasis control mechanism:

[0124]

[Equation 22]     Gs (m, t) = Min [1- (1 + η (m) t) exp (-η (m) t) u (t), θs (m)] (Equation 22)         η (m): m-th natural angular frequency θs (m): m-th upper limit value Pitch pattern:

[0125]

[Equation 23]

Or

[0127]

[Equation 24]

Here, Fmin is the minimum frequency, I is the number of phrase commands, Ap (i) is the size of the i-th phrase command, T
₀ (i) is the time of the i-th phrase command, J is the number of accent commands, Aa (j) is the size of the j-th accent command, T
₁ (j) and T ₂ (j) are the start and end points of the jth accent command, K is the number of phoneme commands, Af (k) is the size of the kth phoneme command, and T ₃ (k) , T ₄ (k) are the start and end times of the k-th phoneme command, L is the number of pattern designating commands, At (l)
Is the size of the l-th sentence pattern directive, T ₅ (l) and T ₆ (l) are the start and end points of the l-th sentence pattern directive, M is the number of emphasis commands, As (m) Is the size of the m-th emphasis command, T
₇ (m) and T ₈ (m) are the start point and the end point of the m-th emphasis command, respectively.

Since the prosody control by the prominence generation rule in this embodiment (FIG. 1) is obtained as the highest score of the listening evaluation experiment result, if the prosody is controlled by this prominence generation rule, the kanji and kana mixed sentence text will be obtained. A very effective enhancement effect can be brought to the speech synthesized from.

In the above embodiments, the method of realizing the strengthening or weakening of the prominence by the pitch by the pitch control mechanism model or the modified pitch control mechanism model has been shown, but the prominence realizing method is not limited to these models. . Any model may be used.
For example, a point pitch (pitch pattern approximated to a broken line) can be used, or a stepwise pitch pattern can be used without any problem.

[0131]

Industrial Applicability As described above, the present invention provides means and method for realizing the strengthening and weakening included in a natural sentence voice uttered by a human in rule synthesis.
According to the present invention, it is possible to realize strengthening and weakening in almost all cases that can occur in a real sentence voice. Therefore, since the user can easily understand the utterance content without paying special attention, the burden on the user can be significantly reduced. In particular, the effect of reducing fatigue when working for a long time such as newspaper editing is remarkable, and the benefit obtained by improving work efficiency is great.

[Brief description of drawings]

FIG. 1 is a diagram showing a basic part of the present invention.

FIG. 2 is a chart showing a configuration example for realizing the present invention.

FIG. 3 is a chart showing a configuration example for realizing the present invention.

FIG. 4 is a chart supplementing the basic part of the present invention.

FIG. 5 is a chart illustrating the concept of the present invention.

FIG. 6 is a chart illustrating the concept of the present invention.

FIG. 7 is a diagram illustrating the concept of the present invention.

FIG. 8 is a chart illustrating the concept of the present invention.

FIG. 9 is a diagram illustrating the concept of the present invention.

FIG. 10 is a diagram illustrating the concept of the present invention.

FIG. 11 is a chart illustrating the concept of the present invention.

FIG. 12 is a diagram illustrating the concept of the present invention.

FIG. 13 is a chart illustrating the concept of the present invention.

FIG. 14 is a chart illustrating the concept of the present invention.

FIG. 15 is a diagram illustrating the concept of the present invention.

FIG. 16 is a chart illustrating the concept of the present invention.

FIG. 17 is a diagram illustrating the concept of the present invention.

FIG. 18 is a diagram illustrating the concept of the present invention.

FIG. 19 is a chart showing an example of the present invention.

FIG. 20 is a chart showing an example of the present invention.

FIG. 21 is a chart showing an example of the present invention.

FIG. 22 is a chart showing an example of the present invention.

FIG. 23 is a chart showing an example of the present invention.

FIG. 24 is a chart showing an example of the present invention.

FIG. 25 is a chart showing an example of the present invention.

FIG. 26 is a chart showing an example of the present invention.

FIG. 27 is a chart showing an example of the present invention.

FIG. 28 is a diagram showing an example of the present invention.

[Explanation of symbols]

3 ... Control parameter generation part, 8 ... Intonation rule part, 10 ... Accent rule part, 11 ... Sentence pitch control parameter generation part, 12 ... Pitch pattern generation part, 14 ...
Prominence information extraction means, 15 ... Sound source power (amplitude)
Correction value calculation means.

Claims (16)

[Claims] 1. A language processing unit for morphologically analyzing an input sentence; a control parameter generation unit for determining a type of the input sentence based on an output of the language processing unit and generating a control parameter according to the type; Is a first prosody control unit having a pitch pattern generation unit for generating a temporal change pattern of a fundamental frequency (hereinafter, abbreviated as a pitch pattern) according to the above. The prominence classification is based on the output of the language processing unit. According to the change amount of the control parameter previously obtained in correspondence with the classification of the prominence based on the analysis result of the natural voice,
First prosody control means for controlling the control parameter;
A voice synthesizing means for generating a phoneme parameter string corresponding to the input sentence based on the output of the language processing means, and for sequentially synthesizing a voice by the phoneme parameter string and the pitch pattern generated by the first prosody control means; In the speech rule synthesizing device having the :, the pitch pattern generation unit has at least an accent control mechanism for controlling the size of the accent component and the start and end times thereof, and the change of the accent component as the change amount of the control parameter. If the sentence unit to which the prominence is added is the entire phrase, the amount of change in the above control parameter is defined as an increment to the size of the accent component when the prominence is not added The difference between "increment" and "accent command increment" for the adjacent accent component If a certain amount of change ΔDAa is set to a positive value and the sentence unit to which prominence is added is a part of a word and the accent type is not the accent deformation type, set ΔDAa to a positive value and the sentence unit If a sentence unit to which prominence is added is a part of a word and the accent type is an accent-deformed type by inserting a pause immediately before and immediately after, the ΔDAa is set to a positive value and prominence is added. The sentence unit is a beat of the topic (syllable)
If it is, the speech rule synthesizing device is characterized in that the ΔDAa is set to a positive value and a pause is inserted immediately before and immediately after the sentence unit. 2. When the sentence unit to which the prominence is added is the entire phrase and the accent type is not the accent-deformed type, the change amount of the control parameter is compared with the size of the accent component when the prominence is not added. The difference ΔDAa, which is the difference between the "accent command increment" defined as an increment and the "accent command increment" for the adjacent accent component, is set to a value of 0.8, and the sentence unit with prominence is the entire phrase. If the accent type is the accent deformed type, the ΔDAa is set to a value within the range of 0.7 ± 0.1, and the sentence unit with prominence is a part of the word and the accent type is not the accent deformed type. Sets the ΔDAa to a value of 0.1 and inserts a pause immediately before and immediately after the sentence unit, and prominence is added. Blocks if it and the accent type is part of the word is accented variant type, the ΔDAa was set to a value of 0.6, prominence is the added sentence topic one heartbeat (syllable)
In the case of, the ΔDAa is set to a value of 0.3 and a pause is inserted immediately before and after the sentence unit.
The described voice rule synthesizer. 3. The voice rule synthesizing apparatus according to claim 1, further comprising second prosody control means for controlling the power of the voice synthesized by said voice synthesizing means. 4. The second prosody control means is a decibel (d
4. The voice rule synthesis according to claim 3, wherein the value of the power P defined in B) is set to a value obtained by the equation P = 11Aa ± 4 (dB) from the magnitude Aa of the accent component command. apparatus. 5. The speech rule synthesizing apparatus according to claim 3, wherein the second prosody control means uses a change in power caused by a change in pitch pattern by the first prosody control means. 6. The voice rule synthesizing apparatus according to claim 1, further comprising a third prosody control unit for controlling a time length of a voice synthesized by said voice synthesizing unit. . 7. The speech rule synthesizing apparatus according to claim 6, wherein said third prosody control means comprises means for controlling the duration of a phoneme corresponding to said phoneme parameter sequence. 8. The third prosody control means, if there is a pause immediately after the sentence unit to which the prominence is added, the duration of the vowel at the end of the sentence unit is not emphasized and the vowel duration is maintained. If the sentence pattern is an interrogative sentence, the duration of the vowel at the end of the sentence is 78 ± 22% of the duration of the vowel in the case of a plain sentence. The speech rule synthesizing apparatus according to claim 7, wherein the speech rule synthesizing apparatus expands the value. 9. A step of morphologically analyzing an input sentence and expressing it as a syllable code string; determining a type of the input sentence based on the syllable code string, generating a control parameter according to the type, and setting the control parameter as the control parameter. A step of generating a pitch pattern according to the method, determining the classification of the prominence based on the syllable code string, according to the change amount of the control parameter obtained in advance corresponding to the classification of the prominence based on the analysis result of natural speech. A step of controlling the control parameter; a phonological parameter sequence corresponding to the input sentence is generated based on the syllable code sequence, and voices are sequentially synthesized by the phonological parameter sequence and the pitch pattern generated by the prosody control means. In the speech rule synthesizing method including the step of The size of the xent component,
With parameters that control its start and end times,
The change amount of the accent component is set as the change amount of the control parameter, and the change amount of the control parameter is set as the accent component when the prominence is not added when the sentence unit to which the prominence is added is the entire phrase. The difference ΔDAa, which is the difference between the "accent command increment" defined as an increment for the magnitude of the and the "accent command increment" for the adjacent accent component, is set to a positive value, and the sentence unit with prominence added If the word is a part of a word and the accent type is not the accent deformed type, the ΔDAa is set to a positive value and a pause is inserted immediately before and after the sentence unit so that the sentence unit to which the prominence is added is the word unit. If it is a part and the accent type is the accent modified type, set ΔDAa to a positive value and add prominence. Sentence that was the topic of one heartbeat (syllable)
If ΔDAa is set to a positive value, a pause is inserted immediately before and immediately after the sentence unit. 10. When the sentence unit to which the prominence is added is the entire phrase and the accent type is not the accent-deformed type, the change amount of the control parameter is compared with the size of the accent component when the prominence is not added. The difference ΔDAa, which is the difference between the "accent command increment" defined as an increment and the "accent command increment" for the adjacent accent component, is set to a value of 0.8, and the sentence unit with prominence is the entire phrase. If the accent type is the accent deformed type, the ΔDAa is set to a value within the range of 0.7 ± 0.1, and the sentence unit with prominence is a part of the word and the accent type is not the accent deformed type. Sets the ΔDAa to a value of 0.1 and inserts a pause immediately before and immediately after the sentence unit to add prominence. If sentence is part of a word and the accent type is accented variant type, the ΔDAa was set to a value of 0.6, prominence is the added sentence topic one heartbeat (syllable)
10. If it is, the ΔDAa is set to a value of 0.3 and a pause is inserted immediately before and after the sentence unit.
The voice rule synthesis method described. 11. The voice rule synthesizing method according to claim 9, wherein the power of the synthesized voice is controlled. 12. The power control is performed by setting the value of the power P defined in decibel (dB) unit to a value obtained by the equation P = 11Aa ± 4 (dB) from the magnitude Aa of the command of the accent component. The method of synthesizing a voice rule according to claim 11, characterized in that. 13. The voice rule synthesizing method according to claim 11, wherein the power control utilizes a change in power associated with a change in the pitch pattern. 14. The voice rule synthesizing method according to claim 9, wherein the time length of the synthesized voice is controlled. 15. The speech rule synthesizing method according to claim 14, wherein the control of the time length of the speech is performed by controlling the duration of a phoneme corresponding to the phoneme parameter sequence. 16. The control of the time length of the voice is such that when there is a pause immediately after the sentence unit to which the prominence is added, the duration of the vowel at the end of the sentence unit is not emphasized and the vowel duration is maintained. Extend by a value within the range of 66 ± 33% of the time,
The speech according to claim 15, wherein when the sentence pattern is a question sentence, the duration of the vowel at the end of the sentence is extended by a value within the range of 78 ± 22% of the duration of the vowel in the case of a normal sentence. Rule composition method.