JPH03203798A - Voice synthesis system - Google Patents

Abstract

PURPOSE:To secure the continuity in the coupling of a VCV(vowel-consoant- vowel) phoneme and obtain a smooth synthesized voice by normalizing the power of a voice phoneme based on the mean power value of respective vowels. CONSTITUTION:When voice phonemes are connected, the value of standard power used for power normalization is found and stored in a power normalization data storage part 7 before voice synthesis so as to normalize the power based on the mean power value of the vowels. Then the power of the VCV phoneme parameter which is inputted from a parameter read part 3 is normalized by a power normalization part 6. Consequently, the continuity at the time of the coupling of the VCV phoneme is secured and the smooth synthesized voice is obtained.

Description

DETAILED DESCRIPTION OF THE INVENTION [Field of Industrial Application] The present invention relates to a speech synthesis method using VCV (vowel-consonant-vowel) segment editing.

[Prior Art] Conventionally, there is a speech rule synthesis method for generating speech from character string data. This extracts the feature parameters (LPC, PARCOR, LSP, Melkebstrum, etc., hereinafter simply referred to as parameters) of the speech segments registered in the speech segment file according to the information of the character string data, and uses certain rules. Synthesized speech is obtained by combining parameters and driving sound source signals (impulse train for voiced speech sections, noise for unvoiced speech sections) according to the speaking speed of the synthesized speech, and feeding them to the speech synthesizer. . Here, the types of speech segments are CV (consonant-vowel)
It is common to use a segment, a VCV (consonant-vowel-consonant) segment, or the like.

In order to synthesize speech segments, it is necessary to interpolate parameters, but conventionally, interpolation when parameters change rapidly was simply connecting speech segments with straight lines in the interpolation interval. , there is a risk that the original spectral information will be lost and the synthesized speech will also change. Also, the corresponding VCV section was extracted from human speech uttered as a word or sentence and used as it was.

[Problem to be solved by the invention] In the conventional technology, when interpolating speech segments, the corresponding ■CV section is extracted from human speech uttered as a word or sentence and is used as it is. However, since they were simply connected in a straight line, there was a problem in that the power varied greatly due to differences in the vocal environment, creating gaps, which resulted in sounds that sounded strange.

[Means for solving the problem] In the present invention, by normalizing the power of speech segments based on the average value of the power of each vowel, continuity when combining VCV segments is ensured, The purpose of the present invention is to provide a speech synthesis method that makes it possible to obtain smooth synthesized speech.

In addition, in the present invention, the average value of vowel power is adjusted according to the power characteristics of a word or sentence, and then V
The purpose of this invention is to provide a speech synthesis method that makes it possible to obtain more natural and smooth synthesized speech, such as accented words or sentences, by normalizing the power of CV segments.

[Example 1] FIG. 1 is a diagram for explaining an example of the present invention.
2 is a text input section that inputs words or sentences to be synthesized, and 2 is a control code that analyzes the input text and breaks it down into phoneme sequences, and that is included in the text (a code that controls accent information, speaking rate, etc.) 3 is a parameter reading unit for reading out necessary speech unit parameter number data from the phoneme sequence information from the text analysis unit 2; 4 is a VC
■This is a VCV parameter file that stores speech segments, and this vC■ parameter also includes information on the power of the voice. 5 is a pitch generation unit that generates a pitch from the control information from the text analysis unit 2; 6 is a power normalization unit that normalizes the power of the speech segment read out by the parameter reading unit 5; and 7 is a power normalization unit 6. 8 is a parameter connection section that connects the speech segment data whose power has been normalized; 9 is a connected parameter series and pitch information; 10 is an output means for outputting a speech waveform.

In the present invention, when connecting speech segments, the power is normalized based on the average vowel power value, so before performing speech synthesis, the standard power value used for power normalization is It is necessary to calculate and store it in the power normalized data storage section 7.

The method for doing so will be explained. FIG. 3 is a diagram showing a method for determining the average power value of vowels.

First, the steady part V° is extracted from the power change in the vowel part V, and the feature parameter (b+,) (1≦i≦n
, 1≦j≦k). where k is the analytical order and n is V
is the number of frames of o. Next, feature parameters (b, J)
The average value of the power ring is obtained by adding and averaging terms representing power information (first-order terms in the Melkebstrum coefficients) in the time direction (i-direction). The above operation is performed for each vowel (if necessary, the phlegm is also determined), the average value of the power ring of each vowel is determined, and the average value is stored in the power normalized data storage unit 7 (.

The operation will be explained below along with the flow of data.

Text to be synthesized is input from the text human resource section 1. The text here is assumed to be Roman letters or kana, with control codes inserted to control the accent and speaking speed, but it is assumed that the text is written in kanji or kana, with control codes inserted to control the accent and speaking speed. In such a case, a language analysis section may be provided in front of the text input section 1 to convert sentences containing kanji and kana into readings.

The text input by the text input section 1 is analyzed by the text analysis section 2 and decomposed into information representing the pronunciation (phonological sequence information) and information such as accent position and rate of occurrence (control information). The phoneme sequence information is input to the parameter reading unit 3, and the specified speech unit parameters are input to the VC■
Read from parameter file 4. The power normalization section 6 normalizes the power of the speech segment parameters inputted from the parameter reading section 3 .

Figure 4 is a diagram for explaining the power normalization method of vowels in ■CV fragments, where (a) is the power change of VCV data retrieved from the database, (b) is the power normalization function, and (c ) is the power change of ■Cv data normalized using a normalization function. The VCV data retrieved from the database has large variations in power values even for the same vowel due to differences in the environment in which it occurs.

Therefore, at both ends of the VCV data, as shown in (a), there is a gap with the average power of each vowel stored in the power normalized data storage section 7. Therefore, the gap (ΔX, Δy
), create a straight line that cancels out the gaps at both ends, and use it as a normalization function. Specifically, as shown in (b), the power normalization function is obtained by connecting the gaps (ΔX, Δy) at both ends with a straight line between vC■C-data.

Apply the normalization function created in (b) to the original data in (a), adjust it so that there is no power gap at both ends of the VCv data, and obtain the normalized data shown in (c). At this time, the normalization created in (b) The function can be normalized by a simple process of adding or subtracting the function to the original data in (a). Figure 4 shows the case of Melkebstrum parameters for ease of understanding.

Next, in the parameter connection unit 8, the ■Cv data whose power has been normalized in the power normalization unit 6 are arranged so that the moras are equally spaced, and interpolation processing is performed in the constant part of the vowel to create a parameter series. create.

The pitch generation section 5 generates a pitch sequence according to the control information from the text analysis section 2.

From this pitch series and the parameter series obtained by the parameter connection section 8, a speech waveform is created in the synthesis section 9.

The synthesizing section 9 can be composed of a digital filter or the like. The created audio waveform is output as audio by the output means 10.

This embodiment may be controlled by a program in a CPU (central control unit).

[Embodiment 2] In Embodiment 1, the power normalization unit 6 provides one straight line for one vC■ data interval as the normalization function, but in this method, the C part is also Power changes under the influence of normalization. Therefore, in this embodiment, a method of normalizing only the vowel part will be described.

In the same manner as in Example 1, the average value of the power of each vowel is determined and stored in the power normalized data storage section 7. Also, regarding the VCV data used for connection, I have stored data in which the boundary between ■ (vowel) and C (consonant) is marked in advance (.

FIG. 5 is a diagram for explaining another embodiment in the power normalization unit 6, in which (a) shows the power change of VCV data retrieved from the database, and (b) shows the normalization of the power of the vowel part. (c) is the power change of VCv data normalized using the normalization function.

As in Example 1, ■ the gap (ΔX) between both ends of the Cv data and the average value of the power of each vowel;

△y). In order to cancel the gap in the range of V before CV data, ΔX is connected with ΔxO by a straight line in interval A of (a), and the normalization function of interval A is used. Similarly, for Δy, in order to cancel the gap in the range marked ■ after the VCV data, in the section C of (a), the power normalization function in the section C is set by connecting 0 and Δy with a straight line. No normalization function is set for the consonant part of section B.

In order to actually set the power value, as in Example 1, by applying the power normalization function created in (b) to the original data in (a), the normalization as shown in (c) is obtained. It is possible to obtain converted VCv data. At this time, for parameters given as logarithmic values (for example, Melkebstrum parameters), parameter adjustment can be performed by addition and subtraction, so the normalization function created in (b) is subtracted from the original data in (a). It can be normalized by a simple process of subtracting (.).For ease of understanding, the case of Melkebstrum parameters is shown in FIG.

As explained above, by finding a power normalization function that eliminates the gap between the average vowel power and the power of VCV data and normalizing the cv data, it is possible to obtain more natural synthesized speech. It becomes possible. Furthermore, although two types of power normalization functions have been described in the above embodiments, the following functions may also be considered.

This embodiment may be controlled by a program in a CPU (central control unit).

[Embodiment 3] FIG. 6 is a diagram showing another method of realizing the power normalization function. In Figure 4, the power normalization function was obtained by connecting (ΔX, Δy) with a straight line, but here, VCV
The power normalization function is a quadratic curve whose slope is O at both ends of the data. In the preceding or following interpolation interval of VCV data, the power is not adjusted by the normalization function, so by gradually approaching the slope of the power normalization function to 0 in this way, the power can be adjusted after normalization. An effect is created in which the change in power becomes smooth near the boundary between the ■c■ data and the average vowel power value in the interpolation interval.

The power normalization method is the same as the method described in the first embodiment.

This embodiment may be controlled by a program in a CPU (central control unit).

[Embodiment 4] FIG. 7 is a diagram showing another method of realizing the power normalization function. In Fig. 4, the section A where the power normalization function exists
In section C, a quadratic curve with a slope of O at each boundary is defined as a power normalization function. vC ■
The power of the preceding or following interpolation interval of data is not adjusted by the normalization function, so by gradually approaching the slope of the power normalization function to 0 in this way, the power after normalization can be adjusted. This produces the effect that the change in power becomes smooth near the boundary between the VCV data and the average vowel power value in the interpolation interval. In this way, it is possible to smooth the change in power near the boundary of the VCv data without changing the consonant part.

The power normalization method is the same as that described in the first embodiment.

This embodiment may be controlled by a program in a CPU' (central control unit).

[Example 5] In the above example, the average value of the power of each vowel is a constant value for each vowel regardless of the connection point of CV data, but when synthesizing words or sentences, ,vC
■Varying the power of the vowel depending on the position of the elemental piece will result in more natural synthesized speech. Considering that power is linked to pitch, it is also possible to manipulate the average power value of each vowel (hereinafter referred to as the reference value for each vowel) in synchronization with pitch. In this case, the rate at which the reference value is raised or lowered (referred to as a power characteristic) is determined according to the pitch pattern added to the synthesized speech, and the power is adjusted by changing the reference value at that rate. An example in that case is shown in FIG.

In FIG. 8, blocks 1 to 10 have the same functions as the same blocks in FIG.

Reference numeral 11 denotes a power reference value creation unit that changes the power reference value of the power normalized data storage unit 7 in accordance with the pitch pattern created by the pitch generation unit 5.

Here, a power reference value creation section 11 is added to the first embodiment, so this part will be explained using FIG. 9.

In FIG. 9, (a) shows the change in power when VCv data is arranged in the time axis direction according to the input phoneme sequence and the reference value of power for each vowel in the interpolation interval, and (b)
is the power characteristic required according to the pitch pattern, (c
) Reference value changed according to power characteristics, (d) is (c)
This is the power value obtained by normalizing the VCV data in (a) based on .

When pronouncing a sentence or word, the power is high at the beginning of the word (and the power gradually decreases as it approaches the end. This is due to the number of moras, which is the number of syllables in the sentence or word, and It can be determined by the number of mora in the mora.Also, the power temporarily increases at the accent position in the word.Therefore, it is possible to assume the power characteristics from the number of mora in the word and the accent position. Now, hypothetically,
Assuming a power characteristic as shown in (b), the reference value of the vowel in the interpolation interval of (a) is corrected according to this characteristic. In the case of Melkebstral coefficients, the parameters are logarithmic values, so the reference value can be changed by adding or subtracting the correction amount, as shown in (c). Using this modified criterion, each VC in (a), as shown in (d),
Normalize the power of V data. The normalization method is described in Examples 1 to 4.

This embodiment may be controlled by a program in a CPU (central control unit).

[Effect] In the present invention, by normalizing the power of speech segments based on the average value of the power of each vowel, continuity is ensured when VCv segments are combined, and smooth synthesized speech is obtained. It has now become possible to provide a speech synthesis method that enables this.

In addition, in the present invention, by adjusting the average value of vowel power according to the power characteristics of a word or sentence, etc., and then normalizing the power of a VCV segment, it is possible to make the accent of a word or sentence more natural. , obtaining smooth synthesized speech. Figure 1 is a block diagram showing the configuration of the first embodiment of the present invention. Figure 2 is the power gap in vC ■ segment connection. Figure 3 is the method for determining the average power value of vowels. Figure 4 shows a vowel power normalization method in a VCV unit. Figure 5 shows a vowel power normalization method in a VC ■ unit. Part 2. Figure 6 shows another example of the power normalization section 6, part 1. Other Embodiment 2 of the Power Normalization Unit 6 FIG. 8 is a block diagram showing the configuration of the fifth embodiment of the present invention. FIG. 9 is a method for changing the power reference value of a vowel.
Text input section, 2... Text analysis section, 3... Parameter reading section, 4... vC ■ parameter file, 5... Pitch generation section, 6... Power normalization section, 7... Power normalization storage unit, 8... Parameter connection unit, 9... Synthesis unit, lO... Speaker, 11... Power reference value creation unit.

A3 ■ #Sound of H0Wa 乎j 转い还语 ouurochishi tama 7re 87 flute diagram ((L') Nogwa change before regularization Cb) ノ\°Wakoji 3J Basification V (C) Genuine nohi 5 torture power/variation song diagram VCV piece IZ;F;17. The power change before normalization of the power change before normalization (the positive RLI of the power change whisper 6 figure) and the famous funeral 61S is the same as the power change before normalization. Hi3darimo 1st year wa figure

Claims (4)

[Claims] (1) Read out the feature parameters and driving sound sources registered in the VCV (vowel-consonant-vowel) speech segment file according to the phonetic sequence of the speech to be synthesized, and use the read parameters and sound source information to In a speech synthesis method that sequentially combines the vowels based on rules and feeds them to a speech synthesizer to output speech, the average value of the power of each vowel is stored in advance, and the power at both ends of each VCV segment is calculated as the power of each vowel. A speech synthesis method characterized by normalizing VCV segments so that they match the average value of . (2) The speech synthesis method according to claim (1), characterized in that the power of the VCV element is normalized for the entire VCV element. (3) The speech synthesis method according to claim (1), characterized in that the power of the VCV segment is normalized only for vowels. (4) The speech synthesis method according to claim 1, characterized in that the power of the VCV element is normalized by adjusting the average value of the power of each vowel according to the power characteristics of the word or sentence.