JPH0632037B2 - Speech synthesizer - Google Patents

Description

Description: TECHNICAL FIELD The present invention relates to a PARCOR-type speech synthesizing apparatus using rule synthesizing means.

[Background Art] Generally, a rule-based synthesis method that can obtain only a phoneme, a syllable symbol, or a character sequence data as an input and obtain an arbitrary synthesized sound output corresponding to the data is not limited in the number of vocabularies of the synthesized sound. It is an excellent speech synthesis method. However, it is not easy to create synthetic sounds with high naturalness and clarity by the rule synthesis method. Therefore, what is thought to be practically usable has been developed by limiting the applications. PARCOR, where relatively inexpensive LSI is provided here
Method [Voice synthesis data consisting of an amplitude parameter, a pitch parameter, and a partial correlation coefficient (κ parameter), which is a spectral parameter, is extracted from a stationary frame in which a voice signal is sampled, and a voice is generated using this voice synthesis data. As a rule synthesizing method using a voice synthesizing means of a method of synthesizing, a rule synthesizing method using (vowel + consonant + vowel) syllables (VCV syllables) and (consonant + vowel) syllables (CV syllables) is a unit. Proposed. Here, since there are about 770 types of VCV syllables and about 100 types of CV syllables, the rule synthesis method using CV syllables as a unit is superior in terms of storage capacity. However, in the method using the VCV syllable as a unit, it is sufficient to connect the vowels to each other (V
₁₁ C ₁ V ₁₂ + V ₂₁ C ₂ V ₂₂ ), but CV
In the syllable-based method, vowels and consonants or vowels (C ₁
Since V ₁ + C ₂ V ₂ or C ₁ V ₂ + V ₂ ) has to be connected, there is a problem in that the interpolation characteristic in the connected portion is deteriorated and the sound quality of the synthesized sound is deteriorated. That is, parameters are interpolated when connecting vowels and vowels and vowels and voiced consonants, but the κ parameter of the PARCOR method has poor interpolation characteristics when linear interpolation is performed, resulting in poor quality of synthesized speech. However, there is a problem in that the sound lacks naturalness and clarity.

Therefore, as a solution to such a problem, the inventors of the present invention have proposed a speech synthesizer which is designed to interpolate the parameter of the connection part of the syllable based on the vocal tract cross-sectional area.
No. 77 has been filed (hereinafter referred to as a basic example). That is, in a voice synthesizing apparatus equipped with a PARCOR-type voice synthesizing means for synthesizing a voice by the voice synthesizing data processed by the rule synthesizing means, after linearly interpolating a parameter in a syllable connecting portion on the vocal tract cross-sectional area Ai. , The above-mentioned rule synthesizing means is formed so as to be converted into the .kappa. Parameter, so that a synthesized voice with good naturalness and clarity can be obtained. Here, in the above basic example, each vocal tract cross-sectional area Ai is a straight line when the vocal tract is subjected to a straight-cylindrical cascade connection approximation (for details, refer to Chapter 10 of Reference "Basics of Speech Information Processing" (Ohmsha)). Based on the idea that it will change over time, the quality of synthesized speech will be higher than when linear interpolation is performed on the κ parameter.
The relationship between i and the vocal tract cross-sectional areas Ai, Ai _{+ 1} is κi = (Ai-Ai _{+ 1} ) / Ai + Ai _{+ 1} ) (1).

However, in such a basic example, although the quality of the synthesized voice is high, in addition to the calculation of the linear interpolation on the vocal tract cross-sectional area, the division for converting the vocal tract cross-sectional area into the κ parameter is performed. However, there was a problem that this division would be a relatively heavy burden. Therefore, there is a method in which linear interpolation is performed on the vocal tract cross-sectional area, Ai and Ai ₊₁ are obtained, and then the κ parameter is obtained using a conversion table. However, in this case, although division is unnecessary and the arithmetic processing is simplified, 2
A dimensional conversion table is required, and the storage capacity required for this conversion table is very large (generally, the storage capacity required to form a two-dimensional table is 2 times the storage capacity required to form a one-dimensional table). There was a problem that it became difficult to realize because of the multiplication.

[Object of the Invention] The present invention has been made in view of the above points, and an object of the present invention is to obtain a synthetic sound with good naturalness and clarity, and to calculate and store a large amount of calculation and memory. An object of the present invention is to provide a speech synthesizer that can reduce the capacity, simplify the configuration, and reduce the cost.

DISCLOSURE OF THE INVENTION (Embodiment) FIG. 1 shows an embodiment of the present invention, in which a syllable synthesis parameter storage unit 1 for storing syllable synthesis parameters in units of CV syllables composed of consonants and vowels, and a rule. A rule data storage unit 2 for storing data, a synthesized input data D ₀ such as character sequence data, a syllable synthesis parameter D ₁ and a rule data D ₂ are used to calculate a voice synthesis data D ₃ and a syllable connection unit. The parameter synthesizing means 4 is formed by the arithmetic unit 3 for interpolating the parameters of 1. and the PARCOR type speech synthesizing means 5 for synthesizing the speech by the speech synthesizing data D ₃ processed by the rule synthesizing means 4 is provided. In the speech synthesizing apparatus, the vocal tract cross-sectional areas of the connecting portions of the preceding syllable and the following syllable in connecting the syllables are ▲ A ^s _i ▼, ▲ A
^{If s} _{i + 1} ▼, ▲ A ^e _i ▼, ▲ A ^e _{i + 1} ▼, The arithmetic unit 3 is formed so as to perform the interpolation by obtaining the value of, and selecting the interpolation pattern of the κ parameter stored in advance from this value. In the embodiment, the analyzing means 10 for obtaining the syllable synthesis parameter of each syllable is equipped with a low-pass filter and A / D for A / D converting the original voice V _0.
The conversion circuit 11, the difference circuit 12, the amplitude of the original voice V ₀ ,
It is composed of an analysis circuit 13 which analyzes a period and a spectrum to form a synthetic parameter (amplitude parameter, pitch parameter and κ parameter). By performing inverse filtering such as difference before analysis, analysis accuracy is improved. It is designed to improve the quality of synthesized speech. Further, a D / A conversion circuit 15 provided with an inverse difference circuit 14 and a low-pass filter for canceling the preprocessing by the inverse filtering is provided at the subsequent stage of the voice synthesizing means 5.

The operating principle of the embodiment will be described below. Now, in the case of considering the linear interpolation on the vocal tract cross-sectional area, the vocal tract cross-sectional area of the starting point of the interpolation, that is, the vocal tract cross-sectional area of the last section of the preceding CV syllable is ▲ A ^s _i ▼, ▲ A ^s _{i + 1.} ▼, If the vocal tract cross-sectional area of the end point of interpolation, that is, the first section of the subsequent CV syllable is ▲ A ^e _i ▼, ▲ A ^e _{i + 1} ▼, the κ parameter of the desired point is the start point of interpolation. Let l be the length from the point to be obtained and L be the length from the start to the end of interpolation, Becomes Here, when dκi / dx and dκi ² / dx ² are calculated, Both numerator in both equations are constant. Also, 0 <▲
A ^s _i ▼ ≤ ▲ A ^s _i ▼ + Δix ≤ ▲ A ^e _i ▼ or 0 <▲ A ^e _i ▼ ≤ ▲ A ^e _i ▼ + Δix ≤ ▲ A ^s _i ▼ ...
(8) 0 <▲ A ^s _{i + 1} ▼ ≦ ▲ A ^s _{i + 1} ▼ + Δi ₊₁ × ≦ ▲ A ^e _{i + 1} ▼ or 0 <▲ A ^e _{i + 1} x ≦ ▲ A ^e _{i + 1} ▼ + Δi ₊₁ x ≦ ▲ A
^s _{i + 1} ▼ (9) (However, since Ai, Ai, Ai ₊₁ and Ai ₊₁ are vocal tract cross-sectional areas, they take positive values.) Therefore, x (0 ≦ x ≦ 1) ) Regardless of whether it always takes a positive or negative value.

The fact that dκi / dx is always a positive or negative value here means that κi is a monotone function with respect to x, and d ²
The fact that κi = dx ² is always a positive or negative value means that κi does not have an inflection point with respect to x and is always a function that is convex upward and convex downward.

Therefore, the form of the function of κi with respect to x is It can be predicted to some extent by looking at the value of. FIG. 2 shows an example of changes in the κ parameter when linear interpolation is performed on the vocal tract cross-sectional area.

Therefore, some patterns are prepared as the function form of κi, and which pattern belongs to is determined by the value of dκi / dx | _{x = 0} , thereby obtaining an effect similar to that of linear interpolation on the vocal tract cross-sectional area. It will be possible.

The operation of the embodiment will be specifically described below with reference to FIGS. 3 to 6. FIG. 3 is a flowchart showing a voice synthesizing operation. In the arithmetic unit 3, a syllable to be read out is extracted from the input character string data of the generated sound, and the voice synthesizing data corresponding to the extracted syllable is stored in the voice synthesizing parameter storage. Read from Part 1. Next, interpolation calculation (described later) is performed based on the read voice synthesis parameter, and the syllable length is determined by the rule data, and the κ parameter sequence,
Determine a pitch parameter sequence and an amplitude parameter sequence. The voice synthesis parameter created in this way is appropriately data-compressed and stored as voice synthesis data. Then, the next syllable is extracted from the character sequence data and the above-described operation is repeated until the end of the sentence, and at the end of the sentence, the voice synthesizing means 5 performs voice synthesis. Next, the interpolation calculation of the κ parameter in the syllable connection will be described. Now, the parameters of each CV syllable or V syllable up to the point of reaching the stationary part are stored in the voice synthesis parameter storage unit 1, and the syllable length of the synthesized voice is adjusted by the number of repeats of the parameter in the last section. It is like this. At the syllable connection, the κ parameter is interpolated from the vocal tract cross-sectional area of the last section of the preceding syllable and the vocal tract cross-sectional area of the first section of the next syllable to connect. FIG. 4 shows κ parameter data ▲ P ⁱ ₁ ▼ to ▲ P ⁱ ₅ ▼ of the non-stationary part of each syllable, and κ of the stationary part.
Parameter data P ₅ and vocal tract cross-sectional area data ▲ A ⁱ ₁ ▼ corresponding to the κ parameter in the first and last sections of each syllable,
The memory state of ▲ A ⁱ ₅ ▼, Fig. 5 shows examples of connection of CV syllables, respectively, and κ parameter ▲ A ⁱ _{5 at the} connection part.
Interpolation is performed based on ▼ and ▲ A ^{i + 1} ₁ .
Needless to say, voiced / unvoiced determination parameters, amplitude parameters, pitch parameters and the like are required as the voice synthesis data. Next, in the interpolation of the κ parameter, dκi / dx | _{x = 0} is obtained from the equation (10), and the interpolation pattern is selected by referring to the interpolation table 7 from the value to perform the interpolation. FIG. 6 shows, for example,
The pattern data of the interpolation table when interpolation is performed in four steps and there are three patterns for increase and three patterns for decrease is shown. Here, in order to make the calculation for selecting the interpolation pattern a little easier, dκi
Instead of / dx, -1/2 (dκi / dx) is calculated, and the threshold value corresponding to that is facilitated. At that time, If, the pattern 3 is used, and the value of the first step κ = ▲ κ ^s _i ▼ + (1/16) Δκ _i the value of the second step κ = ▲ κ ^s _i ▼ + (3/16 ) Δκi 3 th step value κ = ▲ κ ^s _i ▼ + a (8/16) Δκi. _{^{However, Δκi = ▲ κ s i ▼}} - ▲ κ e it is a ▼.

Regarding the interpolation pattern, the increment at the j-th step is Δ ▲ κ ^j _i ▼ = (n ^j / a) Δκ _i n is an integer, and the value is repeated n times when actually calculating Δ ▲ κ ^j _i ▼. The calculation is easy because all you have to do is add. Furthermore, if a is a power of 2, the calculation of Δκi / a becomes very easy.

Comparing the calculation complexity in the case of interpolating in this way and in the case of actually linearly interpolating on the vocal tract cross-sectional area and converting into the κ parameter, in the former case, once −1/2 (d
κi / dx | _{x = 0} ) needs to be calculated, and the latter is κ = (Ai−Ai _{+ 1} ) / (Ai + A) for each step.
i _{+ 1} ) needs to be calculated. Also, in the former case, it is necessary to perform three multiplications and one division for the calculation of −1/2 (dκi / dx | _{x = 0} ) from the equation (10), which seems to be a complicated calculation at first glance. However, since this calculated value is only compared with the threshold value for selecting the interpolation pattern, it does not require so much accuracy. On the other hand, in the latter case, κ
Since i is an actual κ parameter and it is considered that accuracy of about 10 bits is required, the calculation accuracy of the κ parameter is required to be quite high, and the interpolation calculation in the former is considerably higher than that in the latter. It will be easier. Needless to say, the effect of linear interpolation on the vocal tract cross-sectional area is sufficiently maintained even when the interpolation method is used. Further, in the present embodiment, the synthesis parameter is compressed, and when such data compression is performed, the amount of calculation is large, but the storage capacity can be reduced, and the speech synthesizer is compact and inexpensive overall. Will be obtained. Of course, it goes without saying that data compression need not be performed. In addition, the interpolation process may be performed only for low-order κ parameters that have a large effect on the quality of synthesized speech, and linear interpolation may be performed for high-order κ parameters.

[Advantages of the Invention] As described above, the present invention has a syllable synthesis parameter storage unit that stores syllable synthesis parameters in units of CV syllables composed of consonants and vowels, a rule data storage unit that stores rule data, and a character. The rule synthesizing means is formed by calculating speech synthesizing data based on synthetic input data such as sequence data, syllable synthesizing parameters, and rule data, and forming a rule synthesizing unit with an arithmetic unit that interpolates parameters of syllable connection parts. In a voice synthesizing apparatus comprising a PARCOR type voice synthesizing means for synthesizing a voice by means of voice synthesizing data processed by the means, a vocal tract disconnection between a preceding syllable and a succeeding syllable when connecting syllables. Area is ▲
If A ^s _i ▼, ▲ A ^s _{i + 1} ▼, ▲ A ^e _i ▼, ▲ A ^e _{i + 1} ▼, The calculation unit is formed so as to perform the interpolation by selecting the value of γ parameter and selecting the interpolation pattern of the κ parameter stored in advance from this value, and the same natural interpolation as in the case of linear interpolation on the vocal tract cross-sectional area is performed. It is possible to obtain a synthetic sound with good sex and clarity.
Moreover, since the amount of interpolation calculation can be reduced, the calculation unit can be formed by using an inexpensive CPU, and the cost can be reduced.

[Brief description of drawings]

FIG. 1 is a block circuit diagram of an embodiment of the present invention, and FIGS. 2 to 6 are operation explanatory diagrams of the same. Reference numeral 1 is a voice synthesis parameter storage unit, 2 is a rule data storage unit, 3 is a calculation unit, 4 is a rule synthesis unit, 5 is a voice synthesis unit, and 7 is an interpolation table.

Claims (1)

[Claims] 1. A syllable synthesis parameter storage unit for storing syllable synthesis parameters in units of CV syllables consisting of consonants and vowels, a rule data storage unit for storing rule data, and synthetic input data such as character sequence data. , The synthesizing parameter and the rule data are used to calculate the voice synthesizing data, and the rule synthesizing unit is formed by the calculating unit that interpolates the parameter of the syllable connection part, and the voice synthesizing process is performed by the rule synthesizing unit. PA that synthesizes voice with data
A speech synthesizing apparatus comprising RCOR type speech synthesizing means, wherein a preceding syllable at the time of connecting between syllables,
The vocal tract cross-sectional areas of the connecting parts of the following syllables are respectively ▲ A ^s _i ▼, ▲
If A ^s _{i + 1} ▼, ▲ A ^e _i ▼, ▲ A ^e _{i + 1} ▼, The speech synthesizer is characterized in that the arithmetic unit is formed so as to perform the interpolation by obtaining the value of, and from this value, select the interpolation pattern of the κ parameter stored in advance in the interpolation table.