JPS5981697A - Voice synthesization by rule - 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 speech synthesis method based on rules, and a method for synthesizing two arbitrary word speeches with high-quality accents based on syllable code strings and accent types using rules. Regarding.

[Prior art]

Conventionally, audio according to the rules? Accent when compositing? As a method for giving the pitch frequency, various linear approximation patterns are used, such as a scale-like batan, in which the interval between syllables is a constant value (as shown in Figure 41 when expressed as a batan). However, , Kog) Is the straight line approximation pattern 9 the original real human voice? This pattern is a close approximation of the natural pitch pattern extracted by analysis, and it is inevitable that the resulting accent will feel somewhat unnatural.

[Purpose of the invention]

Is the purpose of the present invention to overcome these conventional drawbacks? In order to improve the performance, a time-varying pattern of pitch frequency that closely matches the pitch pattern extracted by analyzing real human speech is artificially generated, and a pitch pattern corresponding to the accent of any word can be obtained. The purpose of this invention is to quickly provide a speech synthesis method based on the method.

[Summary of the invention]

The speech synthesis method according to the rules of the present invention is able to determine the duration of a phoneme by inputting an accent type number and a syllable code number corresponding to a syllable or phoneme. In a speech synthesizer that synthesizes a speech waveform by determining and generating a time-varying pattern of pitch frequency from the determined accent duration and the accent type number, the time-varying bang of pitch frequency is controlled by critical damping It is characterized in that it is generated as an output of a linear system, and the parameters of the nuclear critical braking quadratic linear system are determined by checking a table registered in correspondence with the accent type and basis weight.

[Embodiments of the invention]

The principle of the present invention and embodiment 21 will be explained below with reference to the drawings.

In the present invention, as a pitch bang generation method.

Is there any quality in using the pitch control mechanism model method?
Improving. Is this model a time-varying pitch frequency bump? This is realized as a step response of a critical braking quadratic linear system. In addition, this model was developed based on the idea that the temporal change pattern of pitch frequency is obtained by smoothing the step-like commands generated in the brain by the mechanical characteristics of the vocal fold vibration control mechanism. The predicted value by this model has a relative error ±
It has been practically confirmed that it matches the actual value measured by analyzing real speech with a high accuracy of 4%. Pitch slam by the above model? It has also been perceptually maintained that a high degree of naturalness is obtained, with the difference in accent feeling being almost indistinguishable between the speech synthesized using the method and the speech synthesized after analysis.

Figure 2 is an example of a time change pattern of pitch frequency generated by the pitch control mechanism model. It is formulated as a ninth-order equation as the sum of (1) local fine changes and (2) the accent component k= (indicated by the solid line in FIG. 2).

J, (f(t)/fv(Gv(−τvl) −G
v(-τ7□) 10Aa(-τa 1) ”a
('−τ3□))・・・・・・・・・(1) Gv(t)
=αte”””μ(t) ------
−−・(2)−()a(t) = (1−(1+βt
)e−”)At) =−−・−(3) Here, AV
is the amplitude of the vocalization component, A, l is the amplitude of the accent component,
α is the natural angular frequency q of the square component.

β is the natural angular frequency of the accent component, τylsτ7□τ
gls τ8□ starts raising its voice and ends raising its voice, respectively.

Parameter *fmin representing the time 2 of accent start and accent end is the lowest frequency, and μ(1) is a unit step function.

Equations (1) to (3) (where, nine types of parameters AV
.

Aa, α, β, τvl, τy2++ τale τa2
By appropriately controlling the fmin and fmin, it is possible to generate time-varying bangs 2 with various pitch frequencies.

Here, for convenience of explanation, the Tokyo accent is used and the accent is fixed in a linear manner. In other words, a type 1 (i = 1, 2..., m) accent with a single m-beat is a type 2 accent when the accent nucleus is on the i-th beat from the back (m+1-i beat from the front). The accent nucleus is the part where the pitch frequency suddenly drops near the end of the accented syllable? Furthermore, a flat accent without an accent nucleus is called type 0, and when a particle is added after it.

It is distinguished from type 1, in which the particle has a low accent. For four-beat words, the syllables are expressed as ``QJ'', and the height of the accent is represented schematically by the height of the entire straight line.

To give a concrete example, the first order is 0 0 type p 61,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000. "Piage" For example, Asagao type 4 terrorism ○○○ is an example, Kenriyo. A time-varying pattern of pitch frequency can be generated using a small number of control parameters.Moreover, the speech synthesized using this pattern has an extremely natural accent.Specifically, the types of control parameters are as follows ( 1)-(3)
There are 9 types in the ceremony, and the above slam? The realization values of the control parameters that must be provided for generation may be only one for each control parameter for one word. Moreover.

As long as the word and accent type are the same, the same parameter value can be set regardless of the type of word, so overall we only have a small table with only the number of beats and accent type as variables. All you have to do is prepare. in this way.

The reason why the amount of control information can be reduced is because the above f1
) ~ (3) Is the model the basic part of the prosodic information of words? Regardless of the type of phoneme or accent type of the speaker,
The critical damping second-order lag type characteristic patterns y = xe-'' and y = 1 (1 + This is because it only needs to carry a small amount of secondary information such as the relative timing of the descent, the amplitude of the pitch and accent, and the time constant of the accompaniment. If it can be generated in a small-scale and high-speed way, what about the entire pitch bang generation processing section? It should be possible to imitate a small inkstone and also to generate a time-varying pattern of pitch frequency at high speed.

The present invention is capable of separating such prosodic features of speech from phonological/personal features, and furthermore, that these prosodic features can be expressed with a constant bang regardless of other factors (
Focusing on this, the model (this method generates all the time-varying bangs of the pitch frequency.

And how to control model parameters to generate time bangs of arbitrary accent type pitch frequency 2 at arbitrary number of beats using model 2, and make the device smaller and faster? It is something to be achieved. In addition, since ]<tan is constant, 1
In the present invention, it is only necessary to prepare a pattern table,
The realized value can be obtained extremely quickly by modifying the address by making the variable X correspond to the memory address of the table.

Furthermore, if the number of variables is set to about 100, a sufficient numerical value can be obtained, so that the memory capacity can be kept small.

FIG. 3 is a system block diagram of a rule-based speech synthesizer illustrating an embodiment of the present invention.

The speech synthesis device includes speech synthesizer 1. Rules part 2. and academic degree database3.

Rule section 2, further phoneme duration setting processing section 4. Pitch/bang generation processing section 51 Synthesis unit search processing section 6. and a degree connection processing section 7. First, the phonological sequence of the m-beat word given as input is 0□, V, .
02. V2. , song, 40,,, Vm(ko(1:
.

So, OK is the code that shows the consonant of the cuff, ■, is the code that shows the vowel of the cuff? respectively), or [F-node series (OV),, (eV)2・−・−=−(Ov)m((
OV), representing the code indicating the syllable of cuff).

and each phoneme Ok, V from the accent type prefix i.

Phonological duration setting processing unit 4. The phoneme boundary time τ
1τ2..., output as 12m+1. Next, the obtained phoneme boundary time τ1. τ2・・・・・・
... From 12m+1 and the accent type number i, are the pitch frequency time change patterns f(tl), f(t2), shout, group, f(tn) (n minus total number of frames) or the pitch period time change bang? The pitch/bang generation processing unit 5 obtains the pitch frequency time change pattern f(t,),
f(t2)...f(tm) is input to the pitch control terminal of the synthesizer 1 every frame period. On the other hand,
The phoneme sequence o1. which was used as input. vlycm,
v2・dandan・Om.

From Vm (or syllable series), the unit code i is synthesized by the unit search unit 6. Find the parameter time series P1P2 of the corresponding synthesis unit from the static position code ic.
...P, (i is the total number of frames in this composition unit)
, and the start time of this composite unit (τck if the composite units are chained Ov) - the end time (τvk in the same case), Cv
By searching the degree database 3, the boundary time τcvk, etc. is output 4'' (k=1, 2. town..., J) o
In addition, the above parameter sunset series is, specifically, p-order P
Parameter p1 representing the characteristics of formant such as ARCOR coefficient. .

p2k, p12. Sound source amplitude Ak,
and voiced/unvoiced information Uk element vector 22 (p, 1゜p2, ......p,, -
Ak − Uk )T (T represents transpose 2). Furthermore, the parameter time series P of the obtained synthetic unit
1. P2...P is the above various boundary times τ. 1. τ. Based on yls τv1......, the phoneme boundary time τ1. τ2...
...Allocate on the time axis according to τ2m+l. The assignment is made for the number of beats as a unit of synthesis.

After that, so that the front and rear boundaries of one composite unit are continuous,
If there is a gap at the boundary, connect interpolation to fill the gap, and if one composite unit data overlaps, select the overlapping part or the appropriate gap after cutting. Replace with interpolated curve (straight line). Through this unit data allocation and connection processing, a time series p of synthesized word parameters whose phoneme connection time is substantially controlled, #p21...Pn(
n is the total number of frames of the word) are created. Finally, the word parameter time series is expressed as the pitch frequency time change pattern f(, generalized f(t2),...
. . f (tn), the frame period and () are inputted to each parameter tffll K terminal of the synthesizer 1. At this time, the synthesized speech S is obtained as the output of the synthesizer 1.

The synthesizer 1 can be realized by one, for example, a PAROO synthesizer, and a unit database 3. Of the rule section 2, the synthesis unit search processing section 6 and the unit suction processing section 7 can all be realized by using corresponding processing sections in the conventional rule synthesis method.

Next, embodiments of the phoneme connection time processing section 4 and the pitch bang generation processing section 5 of the rule section 2 will be described in detail.

First, the phoneme duration processing unit 4 calculates the duration of a consonant, for example. This can be achieved using a consonant length table and a calculation circuit for calculating the vowel duration. Here, the consonant length table contains all naturally uttered fI in advance.
Regarding the consonants of the class, their start time t5 and end time ie1. =Measure.

A value of 1 -1? 1. The measurement of 1a can be performed visually by observing the speech waveform, or by the phoneme boundary time measurement method using FFT (fast Fourier transform), which will be described later. The duration of a vowel can be found, for example, by calculating the following empirical formula.

Smell = λv ('vo-η!. + O, + d ) q ・
...(4) Here, λ7 is a correction coefficient for each type of vowel, for example, λv = 1.00 (in the case of Ial), 0.08 (
(in case of lit), 0.79 (in case of 1 μm), 0.9
6 (in the case of let) and 0.93 (in the case of lql). 1vo is the vowel duration that becomes a waste.For example, 1. . = 136 ms (other than the last syllable) = 16
It is sufficient to set it as 4m5 (last syllable). to is the duration of the consonant of the syllable (ms), and η is a coefficient representing the magnitude of the influence of the duration l of the immediately preceding consonant on the vowel, and may be set to 1, for example, η=O, 43.

O□ is a correction term based on the number of beats of a word, for example 1 beat.

2 beats, 10 beats, 10 beats, 85
, 40°15.0l-10, -15, -20, -23°-25 (unit: In5). In terms of speech rate control, if you want to synthesize at the standard speed, set d=Om8, and if you want to synthesize at a faster or closer speed, set it to the correct value or 1.・You just need to enter the negative value (the amount of moderate change (m S ) per syllable).

And q is a correction coefficient depending on the height of the accent, 1, for example, q = 0.7 (with i1 beat cuffs.

and if the first beat is a low accent), it may be set to 1.0 (other than the above). The above experimental values are just examples, and there is no need to be restricted to the above values depending on the purpose and required quality. It is also possible to omit it. ] Note 2 (4) 〒(The operation of : is an adder,
Tractor and multiplier configuration 1, - easier to implement. lastly,? r rhyme boundary time τ1. τ2.・・・
......, τ2+n-)] are the consonant duration 'ck - and the vowel duration 'vk determined by the above method.
'Use the following recurrence formula to find.

The suffix in the formula is cuff (k=1-.2°・
......, represents the phoneme duration of m).

τI = IO (initial value) (5) 7□, = τ2 -+ +'ck (
6) +1 to τ2 - +1vk to T2 (7) Here, the initial value l. be determined appropriately (for example, E.

=100mS).

Next, what is the embodiment of the pitch bang generation processing section 5, which is the most important part of the present invention? , which will be explained with reference to FIGS. 4 to 6. FIG. 4 is an internal logic block diagram of the pitch bang generation processing section 5.

First, based on the accent type number i and the number of beats m of the word to be synthesized, select if) from the pitch control parameter table 8 corresponding to the accent type number and number of beats.
~Nine types of parameters AV, Aa, α in equation (3)
, β, ΔTv engineering, ΔTv□.

ΔTa1− ΔTa□, and the value of f? Search and extract. However, ΔTvl−ΔT[l v□, ΔT
a.

Δ′ra2(l−1, respectively prespecified phoneme boundary time τk(k,=1.2,...2m
+1) is the start of raising the voice, the end of raising the voice, and the start of the accent.

Relative time of end of accent. Describe one example of how to specify the reference phoneme boundary time τ.

First, the relative start time Δ'rv□ and end time ΔTv□ are determined by the time τ2 at the boundary between Ov at the beginning and end of the word, respectively. Based on τ2m. That is, let these reference times be τsv and τev, respectively.
If so, a linear equation is established.

τsv=τ2 ・・・・・・・・・ (8
), eV 2m・・・・・・・・・
(9) Next, for the accent start relative time ΔTR1 and accent end time ΔTa2, the beginning of the syllable OV immediately after the accent starts rising from low to high, and the syllable Cv immediately after the accent starts falling from high to low, respectively. Based on the first time of .

In the example of FIG. 2, the former reference time τsa' is τ3, and the latter reference time Tea is τ5. In the 0 type, τ is like this. is not determined, and in this case, τea is determined to be 72m+1. Also.

In the case of type 1, it is assumed that a particle 110 is added after the word, and Tea is set as τ2m+1 as above. The above matters can be expressed as the following formula according to the accent type i.

・・・・・・・・・(10) ・・・・・・・・・(11) Note that the number of beats used to search for pitch control parameters is the number M (=2m+1) of phonetic boundary times τ□. The calculations of the above equations (8) to (12) are executed by the arithmetic circuit 9, and specifically, the calculations of equations (8) to (11) can be obtained using a linear equation. Therefore, equation (12) can be easily executed using a subtracter and a shift register.

Next, the relative time ΔTpv obtained above.

The parameter τv1. τ7□, τale τ
Find 3°.

τ = τ +ΔTv□・・・・・・・・・
・(13) vl sv τ = τ +ΔTv2 ・・・・・・・・
・(14)v2 ev τ =τ +ΔTa□ ・・・・・・・・・
(15) al sa τ = τ +ΔII+, □ ・・・・・・
...(16) a2 ea The above equations (13) to (16) are calculated by the adders 10 and 1, respectively.
1°12.13. Next, τ9.

τ9. ~． α, which was calculated previously, and frame time tk (k=1.2, 3......, n)? Gv(t-τvl) in +11 formula as input to calculation circuit 14
−Gv(t−τ9□) Complete calculation. Similarly, τa
im τ8□, β, t, as input to the calculation circuit 150, 03(-τal) 'a(t-τ2□
) Calculate all. Furthermore, each of the outputs of the one-car calculation circuits 14 and 15 is AV.

By multiplying Aa by multipliers 16 and 17 and adding the results by adder 18, the calculation on the right side of equation (1) is executed. Finally, using the output of the adder 18 as input, the calculation circuit 191 calculates the exponential function eXy, and then multiplies frrl and n by the 0 multiplier 2o.
(t,) is obtained.

If the above calculation explained based on FIG.
) , -... f(in) is obtained. If you want to find the pitch period T(t,)k instead of the pitch frequency, use the calculation circuit 19 to calculate the reciprocal of the pitch frequency. What is necessary is to modify it so that it can be executed and change the parentheses multiplier 202 to a divider. Also, pi.

The chi frequency can also be determined using a first-order circuit configuration. In other words, 1 multiplier 2o is omitted, and 'n fmi is written instead of f in the pitch control parameter table 8.
Even if you register n\ and change the 1st shift to be added together with the outputs of multipliers 16 and 170 by adder 18,
You can get the same functionality as before the change. Moreover, an inexpensive adder instead of an expensive multiplier? Therefore, the production cost is reduced.

According to such a change circuit, in order to obtain the pitch period 2, it is only necessary to modify the calculation circuit 19 to perform the reciprocal calculation when obtaining the pitch frequency, and changes to other parts are unnecessary. .

In the embodiment shown in FIG. 4, the parameter AV is registered in the pitch control parameter table 8.

Aa, α, β, ΔTv1. ΔTv2s ΔTa1−
The values of ΔTa2 and 'min are determined by analyzing and extracting the waveform of one human's natural voice, and the time mean square error of the pitch frequency time change pattern and the pitch frequency time change pattern according to the model of equation (1) is minimized. It can be obtained by determining the above parameter values so that

Here, since there are several known methods for extracting the pitch frequency from the audio waveform, it can be easily realized, and the above error t'! Since there are various known optimization methods for reducing the size, this can also be easily realized. In addition, for the above ΔTv1゜ΔTv□, lTa□, ΔTa□, τ7□, τ7□, R1゜τ8, which can be directly obtained by the above optimization τ method.
□ and further determine by the following method.

In other words, on the same time axis as the pitch frequency time axis,
FFT (Fast Fourier Transform) is applied to the audio waveform for each frame to obtain a time change pattern of the frequency spectrum of the audio signal. By reading the time at which this bang suddenly changes, the phoneme boundary time can be found, and from this, according to the accent type, according to the rules of equations (8) to (11), τ@9. τevs τsa, τ63 are obtained, and ΔI[v□, ΔTv□, ΔTa□, ΔT8□ can be determined based on the relationships of equations (13) to (16).

Next, examples of calculation circuits 14 and 15 are shown in FIGS. 5(a) and (
b) will be explained with reference to FIG.

'Wz 511 (a) is a circuit that fully calculates Gv (-τv1) - Gv (-τ9□) in equation (1), and figure (b) is a circuit that calculates Ga (t-τal) - in equation (1). G
This is a circuit that fully calculates a(t-τ8°). Figure 5(a)
First, let us consider the starting and ending times τ9□, v2゜τ and frame time tk (k=1.2...n
), t−τyl from the subtractor 21.22i, respectively.
st-τv2 is calculated, and the output thereof is input to unit step function calculation circuits 23 and 24, respectively.

Multiply those outputs by the outputs of the subtracters 21 and 220 obtained above using 1 multipliers 25 and 26, respectively.

The results are further multiplied by multipliers 27 and 28, respectively α
Multiply by Next, each of these dekashi.

Function g (x) = xe-x\ Function table 2 to calculate
9゜30, and its output is Gv(t-τ7□)
and Gv(t-rv2), obtained. Further, the subtractor 31 subtracts the latter from the former to obtain a desired calculation result. Fifth
In Figure (b), each circuit 32 to 42 corresponds in turn to each circuit 21 to 31 in Figure 5(a), and their functions are the same as in Figure 5(a) except for the following three points. . ■Subtractor 32.33
are τ31s and τ8 instead of τ7□ and τ7□, respectively.
゜ is input. 0 multipliers 38 and 39 have β instead of α.
is input. (2) Function tables 40 and 41 are both tables that give the function h(x)=1-(1+x)e-''.

Function calculation using function table 29.30 or 40.41, specifically, the above g (x), h (x)
Register all the results of all calculations, and enter the address a of the registration location.
1. a2. −． -an(N is g (x) or h (
The registration of the value of x) is made to correspond to X, and the calculation is performed by searching the table by modifying the address.

According to this method, functional calculations can be performed easily and quickly. ak (k=1.2...

N) and X, for example, the range of X1≦X
Suppose we create a table where ≦X.

This can be done using a conversion formula. Here, [ ] is a Gauss symbol. Note that M umbrella (aN-a,)/(X, -X,) is Xl,
If XM and N are determined, they are constant values, so the parameter α
Alternatively, if Mα or IMβ (direction) is registered in the pitch control parameter table 8 instead of β, (17)
The conversion of the expression.

a, = (x - x,) + a, (1,
7) It becomes easy. In real terms, X, = 0.

In the end, formula (17)' means that as long as X is converted into an integer, it is sufficient to directly modify the address of the memory area at the first address a1 with (X).

O FIG. 5(a) 9. Substantially no conversion is required. In (b), the same circuit is used in the upper and lower halves of each figure. For example, 21 and 22 degrees, 23 and 24 degrees, 25 degrees and 26 degrees, 27 degrees and 28 degrees, 29 degrees and 30 degrees,
Each is an identical pair. Therefore, the same circuits are combined into one and the device scale is reduced to about 172 in FIG. 6, and one function is the same as in FIG. 5(a) or FIG. 5(b). The functions of the circuit q) shown in FIG. 6 will be explained below.

Basically, parallel 1 in Figure 5 (a) or Figure 5 (b)
The processing of the upper half and lower half performed here is 21 times in one frame.
The method shown in FIG. 6 is to perform the DI separately. That is,
Is the clock signal OL pulsed twice in one frame? Occurrence and start of processing? Multiplying by 0, using the first pulse of the clock signal OL as a trigger, the shift register 43°4
τvt (or τa1) and τv2 (or τ8□) are written to buffer register 51 at the same time.
Clear to zero. Next, go to the frame time (k=1, 2.
..., n) to the output τ7□ (or τ8□) of the shift register 44 t! ! The result is subtracted by the subtractor 45, inputted into the unit step function calculation circuit 46, and multiplied by the output of the subtracter 45 obtained above and the multiplier 47, and then! The multiplier 48 further adds α (or β) to the fruits and vegetables. Multiply. Next, this output is converted to the function g (
x) (or h (X) ) into the function table 49 that calculates Gv(t-τ9□)(
Alternatively, oR(t-τ8□)) can be obtained. Furthermore, from this value, the output of the buffer register 51 (that is, 0'' at the first startup)?
Write to buffer register 51. After the above processing is completed,
Clock signal OL generates a second pulse. As this pulse 2 trigger, the contents of the shift register 43 are shifted to the shift register 44.

However, at this time, no operation is performed on the buffer register 511 (no zero clear f is performed).
) is stored, so the lower τ9□ (or τa
Regarding l), is it the same process as the 11th cil? I'm going. As a result, the buffer register 51 (is the second
The second processing result Gv(t-τ7゜)(or(is Ga(
t-τ81)) to the first processing result Gv(t-τ7
□) (or Ga (-τ32)) is stored, and eventually the desired GV (-τvl) -Gv(t-τv2) (or Ga(t-τa, ) is output as the output of the buffer register 51. -(), (1-τ3□)) are obtained.

Finally, referring back to FIG. 4, the calculation circuit 19 will be explained. The calculation circuit 19 is a circuit that performs all calculations of the exponential function eX, and can be easily and quickly calculated and executed using the same method as when calculating g (x) or h (x). However, the eXO value must be registered in the function table.

According to the embodiments described above, it is small, has excellent real-time performance, and has a more natural accent feel than before. have Word speech synthesis based on rules becomes possible.

Finally, we will give specific examples regarding equipment scale and calculation time.

First, regarding the equipment scale, there are 9 types of control parameters, so are there actual values for all 9 types for each word? Assuming that the maximum number of beats for a word is 210 beats, the total capacity of the required pager table is $-1-.

This means that it is sufficient to prepare 585 actual values.

If it is 2 byte data, the required memory capacity is Bt approximately IKBy.
It is te. Also, some parameters can be treated as constant values, so by omitting such parameters, the table can be made even smaller.Furthermore, the function table can also be used to store all function values corresponding to, for example, 100 variables. Preparing these functions is sufficient for accuracy, so if three types of functions are prepared, a memory capacity of 600 bytes is sufficient.

Next, regarding the calculation time, it takes 8 days to calculate one sample of pitch frequency data.

This mainly consists of the time required for table search and the time required for about 10 arithmetic operations, and if two fixed-point operations are performed, the total of these times will be on the order of several hundred microseconds. Since it is one to two orders of magnitude smaller than the frame interval (5 to 20 m5), real-time processing is sufficiently possible.

〔Effect of the invention〕

As explained above, according to the present invention, an extremely natural Japanese word accent can be created by simple control in which only one realized value is given for each control parameter per word. Moreover, the actual value of the above control parameters is more than the number of beats of the word and the accent type glide? It can be obtained by a simple method of searching. Furthermore, what is the time change pattern of pitch frequency?

This can be easily generated by realizing a function value by referring to a table and combining a small number of four arithmetic operations.

Therefore, the present invention can improve the quality, size, and speed of a speech synthesis device for arbitrary words. It is possible to provide useful means for realizing this.

[Brief explanation of drawings]

Figure 1 is a characteristic curve diagram of a conventionally used linear approximation pitch pattern, Figure 2 is a characteristic curve diagram of a time-varying pitch frequency bump generated by a pitch control mechanism model, and Figure 3 is a characteristic curve diagram of a pitch frequency change over time generated by a pitch control mechanism model.
The figure is an overall block diagram of a speech synthesis device according to the rules shown in the embodiment of the present invention, FIG. 4 is a circuit configuration diagram of the pitch bang generation processing section in FIG. 3, and FIG. 5 is a circuit configuration diagram for performing function calculations. , FIG. 6 is a block diagram of a modified circuit of FIG. 5. DESCRIPTION OF SYMBOLS 1... Speech synthesizer, 2... Rule section, 3... Unit database, 4... Phoneme duration setting processing section, 5...
- Pitch bang generation processing section, 6... synthesis unit search processing section. 7... Unit connection processing section, 8 pitch control parameter table, 9... Arithmetic circuit, 14.15.19... Calculation circuit, 23, 24, 34, 35, 46... Unit step function calculation circuit , 29, 30. 40, 41, 49...
・Function table, 43, 44...shift register, 5
1...Buffer register. 84 Figure 2θ fttn no 5th country ((2) ty, t2. tyl (b) tt, tp, tn 甐6 amount

Claims (2)

[Claims] (1) By inputting the accent type number and the syllable code number corresponding to the syllable or phoneme. By determining the phoneme duration and generating a time change pattern of pitch frequency from the determined phoneme duration and the above accent type number, a speech waveform is generated. In the synthesizer,
By generating the above-mentioned pitch frequency time change pattern as the output of the critical damping quadratic linear system, and searching the parameters of the critical damping quadratic J stem in a table registered corresponding to the accent type number. A method of speech synthesis using rules that has all the desired characteristics. (2) The speech synthesis method according to claim 1, wherein the critical damping quadratic linear system has a function table search means as a main part.