JPS59192293A - Pitch frequency generator for uttered language - 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] [Field of Application of the Invention] The present invention relates to a pitch frequency generation device used in a law-based speech synthesis device, and particularly to generate a highly natural pitch frequency with simple processing and small storage capacity. Bi1. The present invention relates to a frequency generation device.

[First Floor] Conventionally, as a method of obtaining the repetition frequency of a driving sound source pulse for speech synthesis, that is, the pitch frequency (hereinafter simply referred to as pitch), a method of extracting it from the actually uttered speech is known.

This method is the most suitable method for speech synthesis of a limited number of words and is widely used. However, in recent years, attempts have been made to utter an infinite number of words using speech synthesis, as typified by the voice response function of terminal devices and the read-aloud function of word processors. To do this, it is necessary to create synthesis parameters according to rules and synthesize the voices using this variation. This method is generally called the law synthesis method.

When producing speech using lawful synthesis, the key to creating a natural, especially accented sound in the synthesized speech is to use simple rules to obtain temporal changes in pitch (hereinafter referred to as pitch patterns) that are similar to natural speech. becomes.

The method of obtaining the pitch pattern for each word is as follows: (1) The generation mechanism of bi-chi is expressed as a step response of a critical damped quadratic linear system (LJ, L, at the accent time at the beginning and middle of pronunciation). A method of inserting a pulse signal and calculating the response to create a pitch pattern (Journal of the Acoustical Society of Japan Vol. 27, No. 9, 1971) (hereinafter referred to as point pitch) is memorized as a pattern (for example, a 3-mora word is expressed as a 6-tone point pitch pattern), and the distance between the pitches is determined by a straight line or a cosine function.
A method of interpolating V between higher-order algebras to create an overall B-7 pattern was considered.

The above method (1) requires a large amount of processing to calculate the step response, and it is difficult to calculate the actual Bi1. A certain amount of delay occurs when the pattern is obtained, and the delay differs depending on the magnitude of the pulse signal value. Therefore, pulsing −@
It was difficult to determine the time to input i.

In the method (2), the method of interpolating between point pitches is greatly simplified if it is a straight line, but it requires a large capacity to store the point pitch pattern.

For example, in Japanese words, φ or n for rumora words.
There are accent types up to 8 bi1. Requires a pitch pattern of the capacity to be asserted at the start.

[Purpose of the invention]

SUMMARY OF THE INVENTION An object of the present invention is to provide an apparatus that eliminates the drawbacks of the prior art described above and generates pitch frequencies necessary for speech synthesis with simple processing and small storage capacity.

[Summary of the invention]

In order to achieve this objective, the present invention takes advantage of the fact that the pitch pattern of Japanese words is almost determined by the number of moras and the accent type, and that the rate of pitch change within a mora is almost constant. We prepare a dictionary that expresses different mora numbers and accent-type pitch patterns in units of mora using encoded data of four or more types of pitch change heights and their rate of change, and use this dictionary and pitch change rate to easily calculate the pitch. It generates frequencies.

[Embodiments of the invention]

Embodiments of the present invention will be described below with reference to the drawings.

Amazumi explains the relationship between pitch patterns and accent types in Japanese words. Standard Japanese word accents are characterized by the fact that there is always a noticeable rise or fall in pitch from the first to the second mora. - The significant decline within a single fetus is limited to one location. This means that a word composed of n moras has (n+1) types of accent types.

FIG. 1 diagrammatically shows, for example, the accent type of a 4- or 5-mole word. The Φ marks in the figure represent the pinch of each mora, and the marks represent the bi and 7 chi of the following particles.

An accent in which B,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,, ki, , and ji and fall significantly between the Lth mora and the i(L+1) mora is called a‐type accent, and an accent that does not fall significantly is called a G‐type accent.

Fujisaki et al. (Journal of the Acoustical Society of Japan, Vol. 27, No. 9 (1971)) welcomed word pronunciations from six men and six women in order to identify bi-chi patterns unique to each accent type, and eliminated all individual differences among the speakers. Then, the same word accent type vs. sand pi 1. We found that the chip patterns were essentially similar except for certain values specific to the speakers.

FIG. 2 shows pitch patterns from type O to type 5, which are composed of 5 moras. This is the average duration of 1 mora (1
20 msec), and the time axis is normalized so that the durations of the actually measured pitch patterns are all multiplied by this value. We also investigated individual differences in pitch frequency V.
In order to do this, the pitch was converted to a gauze value, moved and plotted so that the average pitch of each pitch pattern was 110EZ.

The patterns in Figure 2 are roughly divided into (1) two rising components that rise rapidly at the beginning of a word (the component that rises steeply toward the end of the first mora seen in type 1 accents, and the component that rises steeply toward the end of the first mora seen in type 1 accents; (2) a component that descends somewhat sharply from the end of the mora indicating each accent type; It can be seen that it can be divided into descending components, and that it is possible to obtain near Of using only these components.

FIG. 3 is a pitch pattern plotted by approximating FIG. 2 using the aforementioned components. In this way, the pitch pattern can be roughly approximated by straight lines with four slopes (a, b, c,
straight line indicated by cl).

It can also be seen that the gradient within the mora can be regarded as constant.

The present invention utilizes the properties found in the above pitch patterns,
Natural visuals with simple processing 1. It generates a pattern.

FIG. 4 is a diagram showing one embodiment of the present invention.

In FIG. 4, 1 is a pitch change memory ROM, 2 is a pitch change pattern memory ROM (accent by mora number),
3 is Bi, Sochi change rate register, 4 is adder/subtractor, 5 is pitch register, 6 is initial value register, 7 is conversion table,
8 is a switch circuit, 9 is a control circuit, 10 is a unit time counter, 11 is a mora counter, 12 is a decoding circuit (A
.

16 is a decoding circuit (Bl), 14 is a mora number input terminal.

15 is an accent type input terminal, and 16 is an output terminal.

A schematic operation of an embodiment of the present invention will be described below with reference to the drawings. In the explanation, we will take as an example a case where a five-mole word, type 1 accent pitch pattern is generated using four pinch change rates.

FIG. 5 shows the contents of the pitch change rate memory RUM1. Pitch change rate memory RUM 1 contains four logarithmic pitch change rates per unit time (+△fα) and (10△fh).

(-thfc) and (-tuft) are address A including the plus and minus signs.

~A, are stored sequentially. FIG. 6 shows the contents of the accent type pitch change car pattern storage ROM 2 for each mora number. As shown in the figure, the accent type B and J chi change rate patterns by number of moras are broadly classified by the number of word moras, and further classified by accent type (in other words, the upper address is determined by the number of moras and accent type. ), within the accent type, the mora position is used as the address unit and address B is
O・°°°・B, B, +1. The pitch change rate is stored as an encoded code in L. In other words, the memory contents are the four pitch change rates of the previous pitch change rate memory RUM 1 (Cane J.), (柁J,), (-△f,), (-△
fd) in 2 bits, α=(00), b=(01),
This is a code encoded as C=(10) and d=(11). (The subscripts α, b, c, d of the straight line in Figure 3 and the fifth
They correspond to the characters a, h, c, and cl in parentheses in the figure. ) Figure 7 shows the number of moras 5. The approximate timing when accent type 1 is input is shown. FIG. 7 shows unit time counter 10.
Using the counter value as a reference unit, unit time counter value, mora counter value, address value of accent type pitch change rate pattern storage ROM for each mora number, bi-chi change rate storage R(-
)M address value, pitch change rate register value, and pitch register value are shown simultaneously.

Mora number input terminal 14. Mora U from the accent type input terminal 15, when the accent type is input, the decoding circuit,
412, and the upper address of the ROM 2 for storing the accent type bi and nochi change car patterns according to the number of moras is determined by the input mora V and accent type. At the same time, the mora number is also input to the control circuit 9, and the control circuit 9 performs initial settings of the present invention. The initial setting is unit time counter 10. The mora counter 11 is reset, the pitch change rate register 3 is cleared, and the initial value register 6 of the switch circuit 8 is selected. The accent type pitch change height pattern storage RUM 2 by mora number is addressed by the upper address value determined by the previously input mora number and accent type and the counter value of the mora number counter 11. Now, mora v=5. If accent type=1 is input and initial setting is performed, the address is set to B7 in FIG.

Mora's law accent type pitch change rate pattern 9th change rate α is in the output state. The encoded pitch change rate α is input to the decoding circuit B1s, and the encoded pitch change rate α is converted into the address value A0 shown in FIG. , pitch change rate storage ROM 1 is at address A. The pitch change rate (1° in the above phrase), which is the content of , is output.

Next, the control circuit 9 sends an acquisition signal to the pitch register 5, and the pitch register 5 receives the cleared content zero of the pitch change rate register 3 and the content f of the initial value register 6. is selected by the switch circuit 8, and the value f is added by the adder/subtractor 4. is taken in.

Next, the control circuit 9 sends a capture signal to the pitch change rate register 3, and the pitch change rate register 3 captures the output (+tsf, ) of the pitch change rate memory RUM1. At the same time, the control circuit 9 sends a signal to the switch circuit 8 to select the bias register 5 and connect it to the power subtracter 4.

Next, the current content f of the pitch register 5. The value (fo+tsf-), which is obtained by adding the current contents (+thf, ) of the tohi and sochi change high registers in the adder/subtractor 4, is taken into the pitch register 5 by the take-in signal from the control circuit 9.

At this point, the contents of the pitch register 5 are updated. At the same time, the control circuit 9 sends an increment signal to the unit time counter 10 to count up. Similarly, next, add or subtract the current contents of pitch register 5 (fo+1d-) and the current contents of pitch change register 3 (匈,). Added by calculator 4, Cfn+2△f,),! : Naribi 1. At the same time, the unit time counter 10 is also counted. After repeating the same operation, the contents of pitch register 5 become (fo+6△fa), (fo+
4△f6).

It increases linearly as (fo+5fa).

The control circuit 9 monitors the counter value of the unit time counter 10, and when the counter value reaches 5, the pitch register 5 (jo+
When the value of 5△f, ) is taken in, the unit time counter 1
At the same time, an increment signal is sent to the mora counter 11 to count up the mora counter 11. The address of the turn storage RUM2 for accent type pitch change rate by mora number has the mora number and accent type in the upper part, and the lower part is determined by the mora counter 11, so at this point the address changes from Bn to Bn+1 in FIG. 6. Therefore, the accent type pitch change height pattern storage RUM2 by mora number outputs the coded pitch change rate C, which is the content of the address B7+1. The encoded pitch change rate C is input to the decoding circuit B13 and addresses the pitch change rate memory RUM 1, which is the pitch change rate (-X(,)) which is the content of the address A. becomes the output state.

Next, the control circuit 9 sends a capture signal to the pitch change rate register 6, and the pitch change rate register 3 receives the pitch change rate (-
1,f,).

If the same operation as when the mora counter value is zero is repeated, the contents of the pitch register 5 will be lost.

+5△f6 吃fha, Cfn+5△f4-2 phantom 1°),・
It continues to decrease.

Finally, the control circuit 9 has a unit time counter 10. When the counter values of the mora counter 11 reach 5 and 4 (the value obtained by subtracting 1 from the mora V input 5), the operation is stopped, and the next mora number and accent type input state is entered.

The contents of the pitch register 5 are inversely logarithmized by a conversion table 7 to become a normal pitch and output to an output terminal 16.

For example, if the values of Bi and Vchi are expressed by 8 pits, the pitch change rate (that is, equivalent to the difference value of Bi and ·chi) can be expressed by a smaller number of pits since the pitch does not change suddenly. It is clear that For example, in our study, about three pits were sufficient.

Here, we need pitch change deer memory RUM 1 and accent type peach change rate pattern memory ROM 2 for each mora number (
(assuming 1 to 10 mora words), the pitch change rate memory ROM = 3 pitches x 4 = 12 bits Accent-type bibuti change history pattern by number of moras,
A total of 876 pits would be sufficient.

Compared to the previous method of memorizing point pitch patterns, the capacity is approximately 1
It turns out that /4 is sufficient.

In the point pitch method, in order to linearly interpolate between point pitches, subtract the difference between the point pitches, and then divide that difference by the time difference between Point Bi and Sochi to calculate the pitch change rate of the present invention. However, in the present invention, since the pitch change rate is stored from the beginning, such subtraction and division are not necessary.

Furthermore, in the present invention, it is easy to raise or lower the entire output pitch pattern by changing the contents of the initial value register 6, but in the 5-point pitch method, in order to raise or lower the entire output pitch pattern, it is necessary to raise or lower the output pitch pattern as a whole. It is necessary to prepare multiple pitch patterns.

Furthermore, the conversion table 7 does not need to be a highly accurate table; it may be a simple polygonal line approximation table. Furthermore, even if the conversion table 7 is deleted, there is no significant effect on the synthesized sound quality, and the conversion table 7 can be deleted in low-cost devices.

It is also clear that the hardware configuration of the embodiment shown in FIG. 4 can be sufficiently implemented by software processing of a microcomputer.

In the description of the embodiment of the present invention, the pitch pattern shown in FIG. 2 is approximated by four types of straight lines, but in order to improve the accuracy of the approximation, it may be approximated by four or more types, for example, eight types of straight lines. Even in this case, if the inclination, that is, the pitch change part within the mora does not change, there is almost no need to change the configuration shown in FIG. The capacity increases and only the decoding circuit B13 needs to be changed. For example, in the case of 8 types, the birich change rate memory RUM 1 increases from 12 bits to 3 bits x 8 = 24 bits, and the accent type pitch change pattern memory RUM 2 by number of moras increases from 864 pits to y (3 is: 1 pit x (?L+1)): Increases to 1296 bits.

The decoding circuit B13 can be changed from 2-man power and 4 outputs to 3-man power and 8 outputs.

Approximate with 8 types of straight lines, and if the slope changes within the mora and the time of each slope is equal, the accent type P1 by number of moras. The change rate pattern storage ROM 2 is changed from a storage format corresponding to one address and one mora to a storage format corresponding to one mora of an address (an integer of A(t':)-1) (one mora is subdivided into two). and mora counter 11, mora counter A,
The mora counter A is divided into two parts, and the mora counter A controls one address, and the mora counter B controls t addresses in units.The counter, which is a reference for control, is arranged in a hierarchy of unit time counter → mora counter A → mora counter B. , set the address of accent type Hirichi change rate storage ROM 2 by mora number to Mora Tsuyoshi, accent type, mora counter A,
The determination may be made using the mora counter B.

The above explanation has been made assuming that the time lengths of the approximated straight lines (one-molar time length or divided time length within L mora) are all the same, but there is a drawback that control becomes complicated even when the time lengths are different. It is obvious that it is only necessary to provide information on the time length of the mora bedding accent type pitch change rate pattern memory RUM K and to have the control circuit 9 control the operation of the present invention based on this information.

〔Effect of the invention〕

According to the present invention, a natural pitch pattern fI:i can be created with a small storage capacity and simple processing.

[Brief explanation of drawings]

Figure 1 is an explanatory diagram showing the accent type diagram of 4- and 5-molar words, Figure 2 is a waveform diagram showing the pitch pattern of 5-molar words, and Figure 3 is the pitch shown in Figure 2. A waveform diagram showing the linear approximation φ of the pattern, FIG. 4 is a block diagram showing an embodiment of the present invention, FIG. 5 is an explanatory diagram showing the contents and addresses of the pitch change rate memory ROM, and FIG. 6 is a mora bedding accent. FIG. 7 is an explanatory diagram showing the contents and addresses of the pitch change rate pattern storage ROM, and FIG. 7 is a station diagram showing the approximate timing of pitch patterns in one embodiment of the present invention. 1...Pitch change rate memory ROM 2...Mora bedding accent type bi1. CH change rate pattern memory ROM B... Pitch change rate register 4... Addition/subtraction unit 5... Pitch register 6...
...Initial (tfE register 7...Conversion table 4
1 figure, 2 figures, Masao Katsui, and Lj: β Telt 4 Utsumizu figure, Masao Rokumi, Otsugareta θ strange question (wa) 44 figure, 45 figure, 16 flashes

Claims (1)

[Claims] A plurality of temporal pitch change t-patterns classified according to the number of moras constituting the language to be uttered and their accent types, and within the range of each unit time length, a change in pitch frequency, A memory for storing a take in the form of a line graph with one constant value selected from a plurality of predetermined values, and a memory for selecting one of the plurality of patterns according to the specified number of moras and accent type. , Bi in the pattern. The present invention comprises means for reading out the pitch change data in chronological order for each unit time length, and means for outputting the read pitch change data as an initial value and individually for each pitch frequency. A bivchi frequency generation device for spoken language characterized by. 2. In the bias frequency generation device according to claim 1, a plurality of bias frequencies are prepared as the initial value, and one of them is selectively used. Characteristic pitch frequency generation device. 3) Statement in claim 1 or 2.'
2. A pitch frequency generating device characterized in that the unit time length is equal to a mora time length or equal to an integer of the mora time length. 4) The pitch frequency generation device according to claim 3, wherein the mora time, length, and language are the same for all moras constituting the language.