JPS63174100A - Voice rule synthesization system - Google Patents

Abstract

(57) [Summary] This bulletin contains application data before electronic filing, so abstract data is not recorded.

Description

DETAILED DESCRIPTION OF THE INVENTION [Industrial Application Field] The present invention relates to a discriminatory synthesis method for text speech, and more particularly to improving the sound quality of rule-synthesized speech.

[Conventional technology]

A method of synthesizing speech corresponding to an arbitrary sentence or word is called "speech synthesis by rules" or simply "speech synthesis by rules."

Speech synthesized by rules generally differs from natural speech because features such as phonological connections, duration, and changes in pitch (voice height) are externally imparted by rules. Therefore, by rule synthesis, speech becomes
The sound quality is worse than the sound quality of the so-called "analysis-synthesized" sound, which preserves the characteristics of these natural sounds. Possible causes of deterioration in the sound quality of rule-synthesized speech include (1) a decrease in the clarity of phonemes, and (2) an unnatural intonation of sentences.

Regarding the rules that govern the intonation of sentences, that is, the prosodic rules, there are already known examples of rules that generate intonation for declarative sentences, interrogative sentences, imperative sentences, and sentences with various expressions in Japanese (for example, Ichikawa Kuroji , “Experimental Considerations on the Naturalness of Synthesized Speech” Proceedings of the Acoustical Society of Japan 1-3-
8 (Showa 42)).

However, the model used here only provides point pitch information on a syllable basis, so it cannot be used in interrogative or imperative sentences.

It is not sufficient to express the difference between desire sentences. Therefore, the intonation of the synthesized voice with such a pitch pattern sounds unnatural.

In order to fully express the intonation differences of various sentences, it is necessary to clarify the relationship between the fundamental frequency (pitch frequency) within a syllable and time.

Conventionally, as a model that can describe the pitch pattern within a syllable and clearly define the temporal structure, the "pitch control mechanism model J H
, Fujisaki et al., “Analyses
isof voice fundamental fe
quancy contours fordaclat
ive 5 entancas of Japan
ese, “J, Acoust.

Soc, Jpn (E) 5.4 (1984) reference has been used. Here, the pitch control mechanism model is a model as described below.

The pitch control mechanism model assumes that the fundamental frequency that provides information about the pitch of the voice is generated through the following process. The frequency of vocal cord vibration, that is, the fundamental frequency, is controlled by the brain's ■Impact commands issued each time a phrase changes, and ■Step commands issued each time an accent is raised or lowered. At that time, due to the delay characteristics of the physiological mechanism, the impulse command of ■ becomes a gradual descending curve from the beginning to the end of the sentence (phrase component), and the step command of ■ becomes a curve with sharp local ups and downs (accent component). The two ingredients are.

It is modeled as a response of a critical braking quadratic linear system for each command, and the time change pattern of the logarithmic fundamental frequency is expressed as the sum of these circumferential components. FIG. 2 shows a pitch control mechanism model. The model fundamental frequency Fo(t) (t is time) is formulated as follows.

! Here, Fm1n is the lowest frequency, F is the number of phrase commands, A p t is the size of the i-th phrase command, Tot
is the time of the i-th phrase command, J is the number of accent commands, A&J is the thickness of the j-th accent command, Txa
, TI are the start time and end time of the j-th accent command, respectively. Further, apt(t) and aaa(t) are the impulse response function of the phrase control mechanism and the step response function of the accent control mechanism, respectively, and are given by the following equations.

Gpi(t) = a −texp(−att)u(
t) (2) GaJ(t)=Min[1-
(1+β,t)exp(−β−t)u(tL θ,
] Here, α5 is the natural angular frequency of the phrase control mechanism Ia for the i-th phrase command, βj is the natural angular frequency of the accent control mechanism for the j-th accent command, u
(t) is a unit step function.

Also, 04 is the upper limit value of the accent component, for example 0
．． 9 etc.

Note that here, the fundamental frequency (pitch frequency) and pitch control parameter (Aps + As2.Toi eT
sag T'2Jl α1. The unit of the value of β-, F+m1n) is defined as follows. That is, the units of Fo(t) and F@i11 are [Hz, T o t v
T t a and T! The units of J are [sl, α1 and β
The unit of 1 is [s'-1].

Further, the values of A p i and A a i are the values when the units of the fundamental frequency and pitch control parameter values are determined as described above.

The analysis method uses an optimization method, that is, the pitch control parameters are determined such that the error between the pitch pattern generated by the above pitch control mechanism model and the actual value measured by analyzing and extracting the original speech is minimized. The best approximation of the pitch pattern is made by determining .

As a known example of applying the above pitch control mechanism model to word speech synthesis, Japanese Patent Application No. 57-19086
No. 1 was published in Japanese Patent Application No. 60-278915 as an example of its application to speech synthesis of sentences such as interrogative sentences, imperative sentences, wishful sentences, etc., and it has been recognized that it has a considerable effect of improving sound quality. However, the synthesized voice still has a mechanical feel. Furthermore, phonological clarity is still insufficient. Possible causes of deterioration in phoneme intelligibility include the method of creating (analyzing) synthesis units and unit connection processing, but local changes in pitch patterns within phonemes can also cause problems. It is caused by

[Problem that the invention seeks to solve]

The conventional pitch control mechanism model described above cannot express local fluctuations at the phoneme level that are effective in improving phoneme intelligibility. Just as it is difficult to express subtle changes in the fundamental frequency that are unique to various emotions and facial expressions. contains fluctuations.

The purpose of the present invention is to provide a new pitch control mechanism model that can faithfully express such fluctuations, and furthermore, to provide a method for synthesizing speech with a natural human-like intonation using this model. be.

[Means for solving problems]

FIG. 1 shows a new pitch control mechanism model that is effective for solving the above problems.

The new model is characterized by the addition of three # [ ! This is due to the addition of suri. By introducing these three new control mechanisms (1) to (2), it is possible to add various fluctuation components to the pitch pattern.

For example, equations (i) to (vffl) may be used to easily describe the new pitch control mechanism model. Here, the units of each parameter in equations (i) to (vM) are those of the conventional pitch control mechanism. stipulated in accordance with.

Of course, the formulas to be concretely realized include the above (i) to (vi
), or when it is not necessary to describe all components, such as in affirmative sentences or sentences without emphasis, omit unnecessary terms in (vn) or (4) as appropriate. It's okay.

Estimation (analysis) of model parameters can be performed using an optimization method as in the case of conventional pitch control mechanism models.

[Effect]

The above-mentioned phoneme control mechanism is a mechanism that generates a component of local fundamental frequency fluctuation for each phoneme, for example, voiced consonant/d
/, /m/, /n/, /r/.

/w/ and other local fundamental frequency drops, voiceless plosives /
It is possible to express the descending characteristic from a high fundamental frequency that is often seen at the transition to a subsequent vowel such as l/, /ni/, etc. ■The sentence shape specification control mechanism is a mechanism that generates a component that expresses the rise in the fundamental frequency at the end of an interrogative sentence. The emphasis mechanism is a mechanism that generates components expressing various emotions and facial expressions, such as imperative sentences and desire sentences.

〔Example〕

Embodiments of the present invention will be described below with reference to FIGS. 3 to 5.

FIG. 3 shows the overall configuration of the arbitrary text synthesis method.

In this method, if a text containing kanji and kana is given as input data, the corresponding synthesized speech can be obtained as output. The processing procedure is as follows.

First, the input text is broken down into each word by the morphological analysis means of the Japanese language analysis section 1, the part of speech is determined, and the reading is determined. Next, based on this result, the speech language processing section 2 determines the accent type of each word or phrase. Syllable information, accent information, etc. are obtained as a result of the above syntactic level processing. Note that the breaks between phrases and sentences are determined based on delimiters such as punctuation marks in the input text.The length of pauses within a sentence or between sentences can be specified by the number of spaces after commas or periods. Also, questions,
The type of sentence, such as an imperative sentence or a desire sentence, can sometimes be determined by the conjugation of the ending, or by using street symbols such as r'/), rl tJ, and "!" at the end of the sentence instead of a period, respectively. It can also be specified by using 0. For example, even if the phonetic string "Kawa wo wo wa dō" is the same, ``Kawa wo wo wo wa dō.'' is a declarative sentence, and ``kawa wo wo wo wa dō wa wa wa wa dō wa wa wa wo wa wa dā .'' is a declarative sentence. j is an interrogative sentence.

Above ■Syllable information, ■Accent information, ■Pause information,
0 Phrase/sentence break information and (if necessary, part-of-speech names, etc.) Grammar information is expressed by a series of numbers called a ``syllable code.''

The syllable code is input information to the control parameter generation section 3.

In the control parameter generation unit 3, accent, intonation, and phoneme duration are determined by rules, and pitch patterns and phoneme parameter time series are generated in accordance with the rules. Here, the accent type can be known from accent information. Accent information is specifically the phoneme with the accent nucleus (the phoneme just before the accent falls)
It is given by inserting a syllable code number indicating the accent immediately after. However, the absence of this syllable code indicates a flat accent. Furthermore, intonation is basically determined based on sentence type information. However, variations due to differences in the arrangement of the final phonemes may also be added, such as 0.

I want to cross the wish list! "I want to cross the river! ”, the intonation pattern is different.

The final pitch pattern is generated based on both accent type and intonation. In the case of consonants, the phoneme duration is determined as a unique length for each type of consonant since it is less affected by surrounding conditions.

In contrast, vowels undergo various transformations depending on surrounding conditions. Therefore, the duration is determined based on the accent type, number of syllables, position within the word, type of consonant immediately before it, type of vowel, etc. Once the phoneme duration is determined in this way, phoneme parameters (spectral envelope parameters and sound source parameters) registered in the file in Cv (consonant-vowel chain) units are extracted and arranged in correspondence with syllable codes. At this time, if it is too long, it is cut to fit within the duration. Thereafter, the Cv units are connected by interpolation (linear interpolation for spectrum envelope parameters, repetition of the same value for sound source parameters) so as to fill in the cut portions or gaps.

Finally, the fundamental frequency and phoneme parameters generated by one or more processes are sequentially sent to the speech synthesis section 4, and a speech waveform is output. Here, as the speech synthesis method, for example, a residual compression method may be used.

In this case, the sound source pulse is basically 1 per frame.
It is generated by extracting residual pulses (representative residuals) corresponding to pitches and arranging the representative residuals at pitch period intervals 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 that length difference, and if it is longer, O is filled in for the section where the representative residual is insufficient. There is.

All of the above processing, except for the intonation generation rules described below, can be configured by known means.

In the following, an example of intonation generation rules in the control parameter generation section 3, which is the most important part of the present invention in the arbitrary sentence synthesis method described above, will be described with reference to FIGS. 4 and 5.

First, the syllable code string obtained from the speech language processing section 2 is input to the sentence type determining means 5. Here, as a first step, the corresponding sentence type is determined by comparing the ending form registered in the ending dictionary in the one-sentence type information dictionary 6 with the sentence ending form of the syllable code string. The final forms in Figure 4 are determined based on the rules of known Japanese grammar, such as verbs in modern sentences ending in the ``tsu'' line, adjectives ending in the ``i'' line, etc. Similarly, the imperative form is determined by the fact that in modern sentences, the conjugated word ends in the ``工'' line. The above sentence type determination becomes more reliable if grammatical information such as part-of-speech information is available. Here, if the conjugation of the ending is determined to be a final form, this sentence is not necessarily a declarative sentence, so the second step is to look at the final time sign (sentence-final sign) of the sentence. The type of sentence is determined by the type of this symbol. An example of the processing of the text type determining means 5 described above is shown in FIG.

Returning to FIG. 4, in addition to outputting the above-mentioned sentence type information, the sentence type determining means 5 separates syllable information from other information, and also outputs a syllable information string. The syllable information is used for (1) determining phoneme boundaries, and (2) generating phoneme components in pitch patterns. That is, for (■), the phoneme duration of each syllable is determined in the phoneme duration rule section 9 based on the syllable information (the above-mentioned known example), and the phoneme boundary time is determined by arranging them in the phoneme boundary determining unit 7. Determined by The phonetic boundary times are used on the one hand to generate phonetic parameters such as LSP parameters. Also, ■ is used in the sentence pitch control parameter generation unit 11 to determine the phoneme control mechanism parameter value.

The previous sentence type information is input into the intonation rule section 8, and according to the sentence type, St semi-intonation (
For example, transformations from declarative sentences are added. There are two types of transformation: time transformation and amplitude (command magnitude) transformation. The former acts on the phoneme boundary determining means 7 and changes the phoneme boundary time. On the other hand, the latter acts on the sentence pitch control parameter generation unit 11, and the size of the command is changed or a new female form designation command or emphasis command is added.

At this time, standard intonation control parameters are supplied from the accent rule section 10. Note that the sentence pitch control parameter generation unit 11 uses a reference phoneme boundary time (timing reference information) in order to achieve temporal consistency with phoneme information.
The intonation rules of 0 or more obtained from the phoneme boundary determining means 7 are stored in the rule table by the intonation rule section 8.
This can be achieved by setting it up and refer to it. In addition, phoneme control parameters are obtained by analyzing the command magnitude, natural angular frequency, relative time from the boundary, upper limit value, etc. for each phoneme in advance, and are provided in the phoneme rule section 13 as a table corresponding to syllable information. Just leave it there. From here, the phoneme control parameter string is sent to the sentence pitch control parameter section 11 in accordance with the order of the syllable information string. Here, the phoneme start or end time (relative time) is converted to absolute time based on timing reference information. The pitch control parameters thus created by the sentence pitch control parameter generation section 11 are sent to the pitch pattern generation section 12, where a sentence pitch pattern is generated using the new pitch control mechanism model (Equations (i) to (4)). .

According to this embodiment, there is an effect of giving a natural, human-like expression to the speech synthesized from the text containing kanji and kana.

〔Effect of the invention〕

FIG. 6(a) is an example of the result of estimating a pitch pattern actually measured using a conventional pitch control mechanism model for a part of a sentence. In contrast, figure (b) is the result of estimating the pitch pattern of the same sentence using the new pitch control mechanism model.A comparison of the two figures shows that the phoneme control mechanism is able to faithfully express subtle changes in the fundamental frequency. It is clear how important a role it plays.

FIG. 6 (Q) is an example of estimating the pitch pattern for an interrogative sentence, and it can be seen that a component is required to express the rise at the end of the sentence. The female role designation control mechanism plays this role.

Furthermore, FIG. 6(d) is an example of estimating a pitch pattern for a command sentence, and unless components other than the conventional phrase component and accent component are added, a good approximation to the actual value cannot be obtained. In this figure, the approximation to the actual measurement value is improved by adding a new emphasized component.

These pitch pattern estimation accuracy improvement effects have been confirmed not only for the above example but also for various sentences.

When we calculated the average error between the actual pitch pattern values generated by the conventional pitch control mechanism model and the new model, we found that the error was 4% for the conventional model, whereas the error was 4% for the conventional model.
In the case of the new model, it is less than ±1%, and a large improvement effect has been quantitatively confirmed.

As described above, according to the present invention, it is possible to generate an intonation pattern that closely approximates the original speech. In other words, it is not only possible to faithfully express the delicate fluctuations peculiar to humans, which was impossible with conventional methods, but also to clearly express the differences between interrogative sentences, imperative sentences, and wishful sentences. In this way, not only is it possible to obtain an extremely natural intonation feeling that is typical of humans, but it is also effective in improving phoneme clarity, and the effect of improving sound quality is remarkable.

[Brief explanation of the drawing]

Fig. 1 shows the basic part of the present invention, Fig. 2 shows the basic part of the conventional system, Figs. 3 to 5 show the embodiment of the present invention, and Fig. 6 shows the basic part of the present invention. Detailed illustrative diagram. 3... Control parameter generation section, 8... Intonation rule section, 10... Accent rule section, 11...
Sentence pitch control parameter generation unit, 12...Pitch pattern\mi Figure 1 Figure Z Figure A. Figure 3 Figure 4

Claims (1)

[Claims] 1. A means for generating a time-varying fundamental frequency pattern (hereinafter simply referred to as a "pitch pattern"), the pitch pattern generating means comprising: (1) a phrase control mechanism; (2) a phrase control mechanism; A speech rule synthesis method characterized by having five control mechanisms: an accent control mechanism, (3) a phoneme control mechanism, (4) a sentence designation control mechanism, and (5) an emphasis control mechanism. 2. The pitch pattern generation means has (1) a phrase control mechanism, and (2) an accent control mechanism, and further has (
Claim 1, characterized in that it has at least one of the following three control mechanisms: 3) a phoneme control mechanism, (4) a sentence designation control mechanism, and (5) an emphasis control mechanism.
Speech rule synthesis method described in section. 3. The speech rule synthesis method according to claims 1 and 2, further comprising means for setting values of input commands corresponding to each of the control mechanisms. 4. The speech rule synthesis method according to claims 1 to 3, wherein each of the control mechanisms is described by a critical damping quadratic linear system. 5. The pitch pattern generation means has a phoneme control mechanism,
Claims 1 to 4 are characterized in that the phoneme control mechanism outputs either a response function of a critical braking quadratic linear system or an exponential function in response to an input command, depending on the type of phoneme. The speech rule synthesis method described.