JPH032800A - Intonation control system for voice synthesizer - Google Patents

Abstract

PURPOSE:To obtain a synthesized voice approximated to the pitch pattern of an actual voice by controlling the control component of a fall part of accent type components in accordance with the accent type and the number of moras. CONSTITUTION:Intonation is determined by a sentence processing part 2, and the accent type is determined by an accent processing part 3. Table data is referred to determine a fall part coefficient K in accordance with the accent type and the number of moras, and this table data is preliminarily determined as the value of actual voice because being obtained with the accent type and the number of moras as parameters by a statistical method. With respect to a value F of the fall part, a pitch frequency control component S as the top of accent is multiplied by the fall part coefficient K to perform addition correction of the pitch target value of this fall part at the time of referring to phoneme parameters to determine the pitch target value by an interpolation processing part 5. Thus, the pitch pattern change in the fall part is approximated to an actual voice to obtain a highly natural synthesized voice.

Description

DETAILED DESCRIPTION OF THE INVENTION A. Field of Industrial Application The present invention relates to a speech synthesis device using a rule synthesis method, and more particularly to an intonation control method using an accent type.

B1 Summary of the Invention The present invention provides a speech synthesis device that performs intonation control by adjusting the pitch frequency of a phoneme or syllable depending on the intonation and accent type. By adjusting the pitch pattern, synthesized speech that approximates the pitch pattern of real speech can be obtained.

C1 A speech synthesis device using the conventional technical rule synthesis method divides an input character string into words and phrases by syntactic analysis, determines intonation and accent for each, breaks down the words and phrases into syllables and even phonemes, and converts them into syllables. Alternatively, the sound source wave and the parameters of the articulation filter are determined for each phoneme, and synthesized speech is obtained as a response output of the articulation filter to the sound source wave.

This type of speech synthesis device has, for example, the configuration shown in FIG. The Japanese language processing unit 1 performs pronunciation conversion, etc. on the manually written Japanese text by referring to the segmentation of phrases and a dictionary.

The sentence processing section 2 adds intonation to sentences, and the accent processing section 3 adds accents to syllables constituting sentences and clauses.

For example, as shown in Figure 4, for the sentence input ``The cherry blossoms at school bloomed beautifully,'' the sentence intonation decreases from the rising point according to the number of syllables in a logarithmic manner, and the phrase accent changes depending on the word and phrase. An accent type is determined, and a composite intonation is obtained by synthesizing these intonations and accent types, and adding exhalation intonation, rounding by filter processing, pauses, etc.

The phoneme processing section 4 converts each input syllable data, such as rsAj..., into a phoneme by referring to data in the syllable parameter storage section 4 that defines the correspondence relationship with phonemes, which are units of vowels and consonants. For example, the syllable rsAJ is decomposed into phonemes rsJ and rAJ.

The interpolation processing unit 5 extracts phoneme parameters from the phoneme parameter storage unit 51 for each phoneme from the phoneme string data from the phoneme processing unit 4, and also extracts phoneme parameters from the phoneme parameter storage unit 51. A sound source pattern is extracted, and a sound source waveform and articulatory data are obtained from these data through interpolation processing. For example, as shown in FIG. 5, the phoneme parameters are such that each phoneme for consonants is divided into three utterance time periods 01 to 03, and the duration of each consonant is 1 to 03 for each time period. ~
t1, pitch P1 to P3 which is the repetition frequency of the sound source wave
, the energies E1 to E3 of this sound source wave, the sound source wave patterns G, to G3, and the pitch and energy time constants DP, -D
P*, DEI-DE3 to obtain discrete data of the sound source wave. Further, each vowel is divided into one division O^ and has a pitch time constant D P A , an energy EA% energy time constant DEA, and a sound source wave pattern GA to provide discrete data of the sound source wave. Among these, the sound source wave pattern is, for example, the 6th
The sound source wave patterns G, ~G3, and GA as shown in the figure are associated with each other, and the sound source parameter storage unit 5
! Several dozen sample data strings are prepared and sample data of the sound source wave is extracted. Also, energy E
+~E3, EA defines the magnitude of the level of the sound source wave, that is, the loudness of the sound, and the pitch P I"" P5, PA defines the height of the frequency, that is, the pitch of the sound. The regulation of these sound source wave data is one value for each time period O3 to 03 and OA, and the time constant DP,
~D P 5, DPA1 DEI ~ DE, , DEA are given, and a continuous sound source wave data string is extracted by interpolation processing by the interpolation processing section 5.

For example, the pitches P1 to P3 of consonants are given as target values for each interval O8-03 as shown in Fig. 7, and the pitch P in each interval is expressed as a solid line or Interpolation processing is performed n times resulting in changes as shown by broken lines. This interpolation calculation is performed using the following recurrence formula Pnk=DP (P
,,P. -+) +P.

However, Pnk is the pitch control value DP for the second time, pitch time constant Po is the current pitch target value Pn-+: calculation is performed n times using the previous pitch target value, and P Hh, P Hk+1.
The pitch Pnk is determined as follows.

Next, as shown in FIG.
A3 and DA^ are also stored. This parameter gives the parameter of the vocal tract articulation equivalent filter, and the human vocal tract (approximately 17 cm in the case of a male) is
As an individually connected articulation model, the cross-sectional area of each acoustic tube for each time period A I-+ ~ A+? -IN A+-t~A+,-*
, A I-3 to Al7-3. These parameters are given to the articulation calculation unit 6 together with the acoustic tube time constant, and articulation calculations are performed on the sound source wave.

The articulation calculation unit 6 obtains a radiated sound waveform data string when a sound source wave is applied to an acoustic tube having a cross-sectional area parameter,
This waveform data is converted into an analog signal by the D/A converter 7, and synthesized speech is obtained from the speech output device 8.

Here, for the intonation of the synthesized speech, the sentence intonation and accent type are added or multiplied by the phoneme pitches (Fig. 5) PI to P3 and PA, respectively, and the pitch target value is determined by the result of this calculation, and the interpolation processing unit 5 The above-mentioned interpolation process is performed in , and the pitch frequency is calculated.

D0 Problems to be Solved by the Invention Conventional intonation control involves uniform adjustment of pitch frequency based on intonation and accent type, and depending on the word or phrase that is a combination of phonemes or syllables, it may be difficult to control the intonation due to natural human speech. Sometimes things got out of control.

For example, as shown in Fig. 8, for intonation control in which accent type component A is superimposed on intonation component ■ for the phrase ``Konosochi'', the pitch target value of each mora will be indicated by a circle, The pitch frequency interpolated between the target values becomes a pitch pattern P shown by a solid line. In such a pitch pattern P, the adjustment of the pitch target value by accent type A is made into two levels, high and low. Return to level O. In this pitch pattern P, the adjustment amount due to the accent component of the mora "u" below the accent waterfall is level O, but in the pitch pattern of the actual voice, the accent component at the accent waterfall part is not zero but has a certain value, and the The intonation in some parts becomes unnatural synthesized speech.

An object of the present invention is to provide an intonation control method that can obtain highly natural synthesized speech by bringing the pitch pattern of an accent-type waterfall closer to the pitch pattern of real speech.

91 Means and Effects for Solving the Problems In order to achieve the above object, the present invention adjusts the pitch target value of each phoneme or syllable of an input sentence using intonation and accent type components to inflect the synthesized speech. a first means for obtaining a coefficient according to the accent type and mora number of a word or phrase; and a first means for obtaining a coefficient according to the accent type and mora number of the word or phrase; and a second means for determining a pitch frequency adjustment amount F for the waterfall part of the accent type, the pitch target value of the waterfall part of the accent type is added and corrected by the pitch frequency adjustment amount, and each phrase is adjusted according to the accent type and the number of moras. A value is given to the pitch target value of the accent waterfall part, and the pitch pattern change at the accent waterfall part is brought closer to that of the actual voice.

F. Embodiment FIG. 1 is a flowchart showing an embodiment of the present invention. The intonation determination in step S1 is performed by the sentence processing section 2 as in the prior art, and the accent type determination in step S2 is performed by the accent processing section 3.

The determination of the Takibe coefficient in step S3 is determined by referring to table data according to the accent type and the number of moras, and this table data is calculated in advance by a statistical method using the accent type and the number of moras as parameters, so it is determined in advance as the value of the actual voice. It can be decided. In addition, the coefficient is a positive real number O<
The range is K<1.

Next, when the interpolation processing unit 5 determines the pitch target value with reference to the phoneme parameters, the value F of the waterfall part in step S4 is calculated by multiplying the pitch frequency adjustment amount S at the top of the accent by the waterfall coefficient. The pitch target value of is added and corrected. To establish the sound source wave in step S5, the interpolation processing unit 5 synthesizes the sound source wave in the same manner as in the past, but at this time, the pitch frequency target value is interpolated with respect to the corrected pitch target value as described above to establish the sound source wave. performs articulation calculation etc. by the articulation calculation section 6.

As described above, by correcting the pitch target value in steps S3 and S4, it is possible to bring the pitch pattern at the waterfall portion of the accent closer to that of the actual voice. For example, the second
As shown in the figure, for each mora of the phrase "Konosochi", the waterfall value FF is calculated by multiplying the adjustment amount S at the top of the accent by a coefficient determined based on the accent type and the number of moras.
'=SXK is set as the adjustment amount for the pitch target value of the syllable "u", and the accent component is left in the pitch target value to obtain a pitch pattern P' that is close to the pitch pattern of the actual speech.

G1 Effects of the Invention As described above, according to the present invention, the pitch target value based on the accent type component is added and corrected at the waterfall part of the accent type according to the accent type and the number of moras. This has the effect of making it possible to obtain synthesized speech that is close to real speech and has a high degree of naturalness.

[Brief explanation of drawings]

FIG. 1 is a flowchart showing one embodiment of the present invention, and FIG.
The figure is an intonation waveform diagram of the embodiment, Figure 3 is a configuration diagram of the speech synthesis device, Figure 4 is an intonation waveform diagram, Figure 5 is a data diagram of phoneme parameters, Figure 6 is a waveform diagram of the sound source wave pattern, and Figure 6 is a diagram of the intonation waveform. FIG. 7 is a pitch characteristic diagram obtained by interpolation processing, and FIG. 8 is a conventional intonation waveform diagram. 1... Japanese language processing section, 2... Text processing section, 3...
Accent processing unit, 4... Phoneme processing unit, 4. ... syllable parameter storage unit, 5... interpolation processing unit, 5. ...
Phoneme parameter storage unit, 5. ... Sound source parameter storage unit, 6... Articulation calculation unit, 7... D/A converter, 8.
...Audio output device. Figure 2 Intonation waveform diagram of the embodiment □ Mora Figure 1 Flowchart of the embodiment Figure 3 Configuration diagram of the speech synthesizer 8-Speech output device Figure 4 Intonation waveform diagram Figure 6 Waveform diagram of the sound source wave pattern Fig. 7: Pitch characteristics due to interpolation processing Fig. 5: Data diagram of phoneme parameters Fig. 8: Conventional intonation waveform diagram Control method 3, relationship with the person making the amendment

Claims (1)

[Claims] (1) In a speech synthesis device that adjusts the pitch target value of each phoneme or syllable of an input sentence using intonation and accent type components to obtain the intonation of synthesized speech, depending on the accent type and mora number of the word or phrase. and second means for multiplying the pitch frequency adjustment amount S by the top of the accent type by the coefficient K to obtain the pitch frequency adjustment amount F of the waterfall portion of the accent type. . An intonation control method for a speech synthesizer, characterized in that a pitch target value of an accent type waterfall portion is additionally corrected by the pitch frequency adjustment amount.