JPS5912194B2 - speech synthesizer - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION The present invention relates to a speech synthesizer using linear prediction.

Various methods have been proposed for linear prediction, including advanced methods.

A typical method is LPC (LinearP
redirectiveCodlng) method, PARCOR
(Partial 10 Response and Auto
correlation) method, LSP (LineSp
ectrumPair) method has been proposed. These methods are excellent in terms of speech information compression rate, quality of synthesized speech, etc. These methods are based on modeling 15 the human vocal mechanism as shown in FIG. In FIG. 1, reference numeral 1 represents a voice producing section, which represents the air vibration caused by the vocal cords for voiced sounds, and the turbulent airflow caused by air pressure from the lungs for unvoiced sounds.

Reference numeral 2 is a resonant part, which in biological terms refers to a resonant tube called the vocal tract, such as the throat, nose, tongue, teeth, lips, etc.

The modeled vocal mechanism as shown in FIG. 1 is replaced by an electric circuit system as shown in FIG. In Figure 2, 3
1 is an excitation circuit, which corresponds to the voicing section 1 in FIG. 1. 4 is a resonance filter, which corresponds to the resonance section 2 in FIG. 125.

Next, a speech analysis or synthesis method for modeling a vocal mechanism as shown in FIGS. 1 and 2 will be explained. FIG. 3 shows an example of a conventional speech analysis and synthesis device implementing this type of method. 30 In Figure 3, 11 is a speech analysis device, 12 is a transmission line from the analysis system to the synthesis system (
13 is a voice synthesis device.

14 is an audio analog signal input terminal, 15 is a low-pass filter, 16 is a 35AD converter that converts the analog signal into a digital signal, IT is an autocorrelation circuit that calculates an autocorrelation coefficient, and 18 is an autocorrelation circuit that calculates an autocorrelation coefficient. various parameters (α parameter, P
In the ARCOR system analysis and synthesis equipment, the k parameter, L
In the SP-based analysis/synthesis device, a parameter calculation circuit that calculates LSP parameters), and 19 a sound source extraction filter configured by these parameters (this filter is an inverse filter of the speech synthesis filter used when synthesizing speech). ), 20 is a sound source determination circuit that performs sound source determination;
21 is each parameter output signal, 22 is a sound source information output signal, and 23 is an encoder, which constitute the analysis device 11.

24 is a decoder that decodes the code encoded by the encoder 23 and transmitted through the transmission path 12 (or storage device); 25 is a decoder that decodes the code encoded by the encoder 23 and transmitted through the transmission path 12 (or storage device); 26 is a decoded sound source information signal, 27 is a sound source generation circuit that generates a sound source signal, 28 is a speech synthesis filter, 29 is a DA converter that converts a digital signal into an analog signal,
30 is a low-pass filter, and 31 is an analog signal output terminal for synthesized speech.

The synthesis device 13 is constituted by these. The operation will be explained in detail below with reference to FIG.

A low-pass filter 15 removes high frequency components from the audio signal applied from the audio analog signal input terminal 14. This low-pass filter is used to reduce distortion of the audio signal caused by the AD converter, which will be described later.
The cutoff frequency ωC of the low-pass filter 15 must be set so as to remove frequency components that are greater than % of the frequency sampled by the oil converter. For example, when the sampling frequency of the AD converter 16 is 8 KHz, frequency components of 4 KHz or higher must be removed, so the cutoff frequency of the low-pass filter 15 is selected to be about 3.4 KHz. The output of the low-pass filter 15 is sampled at every sampling period T by the AD converter 16. Here, t=N
The digital signal sampled every T is X. shall be. A
The digital signal Xn, which is the output of the D converter 16, is composed of, for example, 1 bit for the sign and 11 bits for the value. These digital signals are input to an autocorrelation circuit 17 and a sound source extraction filter 19. The autocorrelation circuit 17 calculates the autocorrelation coefficient ρ of the sound source. is calculated. Autocorrelation coefficient ρ. is given by the following formula. (Here, N is the number of sample points for calculating the autocorrelation coefficient.) Autocorrelation coefficient ρ calculated by equation (1).

are input to the parameter calculation circuit 18, and parameters are calculated according to each method. In the parameter calculation circuit 18, LPC method, PARCOR method, LS
α parameters, k parameters, and LSP parameters are calculated according to various P methods.

Commonly, these parameters provide filter coefficients for configuring the resonant section 2 in the vocal mechanism model diagram of FIG. 1 and the resonant filter 4 in FIG. 2. That is, the transfer functions of the throat, nose, tongue, etc. are determined by these parameters. Let H(jω) be the transfer function of the vocal tract (transfer function of the resonance filter in FIG. 2) determined by these parameters. On the other hand, the spectrum function of the audio output from the AD converter 16 is expressed as Y(j
ω). Using each parameter calculated by the parameter calculation circuit 18, the sound source extraction filter 19 calculates the inverse characteristic of the vocal tract transfer function H(jω), that is, 1/H(
A filter with the characteristic jω) is constructed. Then, the sound source extraction filter 19 receives the signal Y(j
ω) as input, the inverse characteristic of the vocal tract transfer function, i.e.,
This results in a filter having a characteristic of 1/H(jω). The output of the sound source extraction filter 19 is generally called a residual signal,
This corresponds to the vocalization section 1 in FIG. 1, the excitation circuit 3 in FIG. 2, or the biologically speaking vocal cords. The output residual signal of the sound source extraction filter 19 is sent to the sound source determination circuit 20, where the sound source is analyzed. The sound source information analyzed by the sound source determination circuit 20 includes voiced sound fundamental period information T (generally called piptsuchi period), voiced and unvoiced sound determination information V/UV, and audio amplitude information A. There are three types.

Each sound source information output from the sound source determination circuit 20 is the aforementioned three types of information (voiced sound pitch period information T1P, voiced sound/unvoiced sound determination information V/U, and audio amplitude information A). There have also been proposals for transmitting the residual signal, which is the output from the sound source extraction filter 19, as it is, or for transmitting it after passing it through a low-pass filter and encoding it using another encoding method.
This is not common because the amount of information passing through the transmission path 12 (or the amount of information stored in the storage device) is extremely large. The aforementioned three types of sound source information output signals (voiced sound pitch period information T
, voiced sound/unvoiced sound determination information SV/4JV, and p voice amplitude information A) are each encoded into several bits by the encoder 23.

Similarly, the parameters calculated by the parameter calculation circuit 18 are also encoded into several bits by the encoder 23, and passed through the transmission line 12 (or storage device) for 5 milliseconds to
The data is sent to the synthesis device 13 at every cycle called a frame cycle of about 30 milliseconds. Next, in the synthesis device 13,
A decoder 24 decodes each encoded signal sent every frame period Tf of 5 to 30 milliseconds.

Among the signals decoded by the decoder 24, the sound source information signal 26 is input to the sound source generation circuit 27;
Among the signals decoded by 4, parameter 25 (α
parameters (or k parameters or LSP parameters) are input to the speech synthesis filter 28. Sound source generation circuit 27
Now, a signal for exciting the synthesis filter 28 is created using a signal corresponding to the voice generating section in FIG. 1 and the excitation circuit 3 in FIG. 2.

The sound source generation circuit 27 includes a pulse generator and a noise generator, both of which are selected by the voiced sound/unvoiced sound determination information V/UV signal of the sound source information signal 26. Further, the pulse interval in the pulse generator is determined by the pitch period information T in the sound source information signal 26.

These selected signals become signals matched to the amplitude level of the audio according to audio amplitude information A in the audio source information signal 26. The sound source signal generated by the sound source generation circuit 27 is input to a synthesis filter 28.

The synthesis filter 28 constitutes a vocal tract filter for synthesizing speech using each parameter decoded by the decoder 24. The transfer function is given by H(jω) mentioned above. In addition, a synthesis filter 28 based on each parameter
The coefficients are updated every cycle at which parameters are sent through the transmission path 12 (or storage device), that is, every frame cycle Tf. On the other hand, since the input signal from the sound source generation circuit 27 is input every sampling period T (125 μs when the sampling frequency is 8 kHz) corresponding to the sampling frequency of the AD converter 16 in the analyzer 11, The operation of the speech synthesis circuit 28 is performed every sampling period T seconds, and its output is sent to the DA converter 29 every output period T seconds, which is the same period as the sampling period.

The DA converter 29 converts the digital signal from the voice synthesis filter into an analog signal and passes it through the low-pass filter 30.
send to The low-pass filter 30 functions to remove high frequency components generated when converting a digital signal to an analog signal. The output of the low pass filter 30 is
The synthesized voice is listened to by an external audio device (amplifier, speaker) through the analog output terminal 31. In this manner, the speech synthesizer 13 operates. In such a speech analysis and synthesis method, the parameter calculation circuit 18
encodes the parameter signal 21 calculated by the sound source information signal 22 determined by the sound source determination circuit 20,
Information is transferred to the synthesis system at intervals of 5 to 30 milliseconds called a frame period. create a signal,
They were synthesizing synthesized sounds. In such an analysis-synthesis system, the analysis section and the synthesis section can be said to have an inverse relationship to each other. Therefore, if the analysis section transfers all the analyzed speech information to the synthesis section and synthesizes the speech, it is expected that a synthesized speech of fairly high quality will be obtained. However, transferring all the audio information analyzed by the analysis unit to the synthesis unit via the transmission line 12 (or storage device) is inconsistent with the purpose of compressing audio information and reducing the storage area of the storage device. . Therefore, for the purpose of compressing audio information and reducing the storage area in the storage device, the audio information in the transmission path or storage device should be made as small as possible.

For this purpose, the voice information sent from the analyzer 11 to the synthesizer 13 in FIG. The determination information is limited to p information V/UV and audio amplitude information A).

What should be noted here is that instead of encoding and transmitting or storing the residual signal itself, which is the output of the sound source extraction filter 19 in FIG. Each signal is modeled as an impulse or random noise, and the modeling information is transmitted (or stored). FIG. 4 shows a detailed block diagram of the sound source generating circuit 27.

40 is a noise generation circuit, 41 is an impulse generation circuit, 42
is an input terminal for voiced pitch period information T, 43 is a changeover switch, 44 is an input terminal for voiced/unvoiced sound determination information/UV, 45 is an amplitude multiplication circuit, and 46 is audio amplitude information A.
The input terminal 47 is an output terminal for the sound source signal, and the sound source generating circuit 27 is configured as described above.

The function of the sound source generation circuit 27 will be explained with reference to FIG.

In the noise generation circuit 40, random noise is generated every T seconds, and the amplitude value of the noise is +1 or -1.

In the impulse generation circuit 41, the pitch period information T of the voiced sound, the pitch period information T input from the input terminal 42,
Therefore, an impulsive signal is generated every T seconds, and its amplitude is 1. These noise generating circuits 40
The output of the impulse generating circuit 41 and the output of the impulse generating circuit 41 are selected by a changeover switch 43. The selection information for the switch 43 includes voiced sounds input from the switch 44.
Unvoiced sound determination information V/U is used. That is, when the voiced sound/unvoiced sound determination information/UV is '11', the sound is determined to be a voiced sound, and the output from the impulse generation circuit 41 is selected. Further, when the voiced sound/unvoiced sound determination signal /U is "O", the sound is determined to be an unvoiced sound, and the output from the random noise generation circuit 40 is selected and output to the amplitude multiplication circuit 45. The selected sound source is loaded with audio amplitude information A input from an input terminal 46 in an amplitude multiplication circuit 45, and is output as a sound source signal from a sound source signal output terminal 47. The sound source signal obtained in this manner becomes an input to the synthesis filter 28. On the other hand, regarding the amplitude information A, the signal before being multiplied by the audio amplitude information A, that is, the output of the switch 43 as it is, is input to the synthesis filter 28 in order to reduce the number of calculation bits of the audio signal. A method has also been proposed in which the information A is multiplied after the output calculation of the synthesis filter 28, but this is not directly related to the present proposal, so the audio amplitude information A is multiplied before the input to the synthesis filter 28.

Here, a sound source signal in the case of a voiced sound will be considered in particular. As mentioned above, an impulse train is usually used as a sound source for voiced sounds, and its period is given by pitch period information T, in the sound source information signal 25.

The value of this pitch cycle information T, is normally 5 for men.
~10 msec For a woman, it is about 3 to 5 msec. On the other hand, as mentioned earlier, a series of data such as sound source information is updated every 5 to 30 msec called frame period 2, but when the frame period is long, in other words, audio information compression using low bits is If this is done, and the frame period is set to around 30 msec, the result will be a voiced sound source with several impulses in one frame.

The pulse intervals within one frame period are all the same pitch period Tp.

In Fig. 5a, the frame period Tf is 30 milliseconds, the pitch period information T in the nth frame period is 8 milliseconds,
A diagram of the sound p source signal when the pitch period T in p in the n+1th frame period is 10 milliseconds is shown.

In FIG. 5a, there are four impulses every 8 milliseconds within the n-th frame period, three impulses every 10 milliseconds within the n+1-th frame period, and the impulse generating circuit 41 in the sound source generating circuit 27. generated by.

The interval between these impulses, 8 or 10 milliseconds, ie, the pitch period T, is considered to be the average value of the pulse interval of the residual signal in the analyzer 11 within the same frame period. On the other hand, the vocal part of the actual voice, that is,
It seems that the pulse interval of the excitation waveform emitted from the vocal cords etc. slowly changes as shown in Fig. 5b. Therefore, the impulse generating circuit 41 in the synthesizer 13 generates frames in such a manner that all pulses are generated at 8 ms intervals within the n-th frame period, and pulses are generated at 10 ms intervals within the n+1 th frame period. Rather than using a pulse train that discontinuously changes every cycle,
As shown in FIG. b, if the synthesis filter is driven using an impulse train whose pulse intervals gradually change, the similarity with the actual residual signal will increase, and the sound quality will be improved.

The present invention aims to improve the quality of synthesized speech, and serves as an input to the synthesis filter 28.

The interval of the pulse train of the voiced sound source signal is the frame period Tf
By gradually changing the pulse train interval within the frame, discontinuity in the pulse train interval between frames is prevented, and the sound source of the voiced sound is given in a form that is closer to the residual signal in the analysis unit 11. This will be explained in detail below. As mentioned earlier, the various information sent every frame period is an average value within that frame period, so there is likely to be considerable discontinuity within the frame period. .

On the other hand, various types of actual information change slowly and continuously, and in order to reduce discontinuities within the frame period, the frame period must be set short.

However, setting the frame period short causes an increase in the amount of information. Therefore, when setting a long frame period,
A process called "interpolation" is usually used to eliminate discontinuities in various types of information. FIG. 6 describes the interpolation process.

According to FIG. 6, 50 is various types of information sent within each frame period, and 51 is various types of information as a result of interpolation.

This process of interpolation usually involves various parameters 25 (LPC
This process is performed for the α parameter of the PARCOR method, the k parameter of the PARCOR method, the LSP parameter of the LSP method), or the amplitude information A of the sound in the sound source information signal 26, but interpolation is performed for the pitch period T. No ivy synthesizer has been proposed. FIG. 7 shows a p-pitch period interpolation processing circuit that performs interpolation processing of pitch period information T used in the voiced sound source generating section of the speech synthesizer of the present invention.

In FIG. 7, 60 is the SP input terminal of pitch cycle information T, 61 is a subtracter, 62 is a shifter, 63 is an interpolation processing unit register, 64 is an adder, 65 is a pitch cycle register, and 66 is the timing of pulse generation. A counter that obtains 6
7 is a pulse generation timing signal, 68 is a pulse or pulse train generation circuit, and 69 is an output terminal for a voiced sound source signal (signal before being multiplied by audio amplitude information A).

In addition, Fig. 8a shows a time chart of these interpolation processes.
FIG. 8b shows a partially enlarged time chart of FIG. 8a.

The first clock signal 70 is a signal generated every frame period Tf, and provides the timing for subtraction processing in the difference register 63 and the subtracter 61, and the timing for reading in the difference register 63.

The second clock signal 71 represents a signal generated every interpolation processing period T (same as the output clock of the synthesis filter and the sampling period of the AD converter in the analyzer), and is generated in the interpolation processing unit register 63 and the pitch period register 65. Acts as a driving pulse. The period Tf of the first clock signal 70 and the period T of the second clock signal 71 are usually selected to be a power of 2 times Tf in order to simplify the circuit configuration. Therefore, in the description of this embodiment, as an example, the interpolation processing period T=125 microseconds, the frame period Tf-32 milliseconds, and the ratio of T and Tf is T:Tf=1:256=1:
28.

Therefore, pulses 70,000 and 7 of first clock signal 70
There are 256 pulses of the second clock signal 71 between 0001 and 0001. Change the number of those pulses from 71000 to
71255. It is also assumed that the pulse 70000 in the first clock signal 70 and the pulse 71000 in the second clock signal 71, and the pulse 70001 in the first clock signal 70 and the pulse 71256 in the second clock signal 71 have the same timing. A stepped line 72 is a graph representing the contents of the pitch period register 65 in an analog manner at each timing of the second clock signal, and a stepped line 73 is a graph representing the contents of the counter 66 in an analog manner. Pulse generation timing signal 74
indicates the timing of the impulses issued by counter 66. The operation of the pitch period interpolation processing circuit p which performs interpolation processing of different pitch period information T in adjacent frame periods Tf will now be described with reference to FIGS. 7 and 8A and 8B.

The pitch period information T of the voiced sound input from the input terminal 60 at the beginning of each frame period is composed of a 7-bit p digital code.

It should be noted that this digital code is the output of the decoder 24 of the synthesizer 13 in FIG. 3, and is not the digital code itself obtained through the transmission line 12 (or storage device). The reason why the pitch period information T, inputted from the input terminal 60 is set to a 7-bit digital code is as a result of taking into account the range of human pitch periods, and the reason will be briefly explained. The pitch period T is set with the output period T of the synthesis filter as p reference, and is 6
When the digital code input from 0 is 1, the pitch period is T = T p 2, T = 2T p 3, T = 3T p .

Therefore, when T = 125 microseconds, the maximum value of the 7-bit digital code input from 60 is 127, so 127 x 125 microseconds = 15.875 milliseconds, which is a human pitch period of 3 to 10 Can reproduce milliseconds.

A digital code with a pitch period T from the input terminal 60 is input at the timing of the pulse 70000 in FIG. 8a. Next, in the subtracter 61, the digital code of the pitch period T, input this time and the pitch period register 6
Subtraction with the contents of 5 is performed.

The pitch period 65 stores the final value of the pitch period information of the interpolation process in the previous frame period. The pitch period register 65 has a total of 15 bits, 7 bits above the decimal point and 8 bits below the decimal point. The 7 bits above the decimal point correspond to the 7 bits of the digital bit configuration of the pitch period information T input from the input terminal 60. The subtracter 61 calculates the difference between the final pitch cycle information in the previous frame cycle and the current frame cycle information. The output of the subtracter 61 is reduced by 1/256 by a shifter 62. This shifter 62 may be implemented using a shift register or the like, but it can also be realized by devising wiring. That is, a connection is made between the subtracter 61 and the interpolation processing unit register 63 to shift the output of the subtracter 61 to the right by 8 bits. Therefore, there is no need to provide any elemental component corresponding to the shifter 62. shifter 6
2 is 1/256 because the pitch period T in the adjacent frame period Tf obtained by the subtractor 61 is
, into an interpolation processing unit per interpolation processing period T, and the processing in which the ratio of one frame period Tf and the interpolation processing period T is selected as a power of 2 (256 in this processing) is this conversion processing. This is to simplify the configuration. shifter 62
The output of the subtracter 61 shifted to the right by 8 bits is stored in the interpolation processing unit register 63 as an interpolation processing unit. The contents of the interpolation processing unit register 63 mean the difference in frame period per interpolation processing period T, as described above. The digital code structure of the interpolation processing unit register 63 consists of a 1-bit code and 8 bits below the decimal point. Next, in the adder 64, the interpolation processing unit register 63 and the pitch period register 65 are added, and the result is read into the pitch period register 65 at the timing of the pulse 71001.

From pulse 71001 to pulse 71002, the contents of the interpolation processing unit register 63 and the contents of the pitch period register 65 are added by the adder 64, and the result is added to the pulse 71000.
The process of reading into the pitch period register 65 is performed at timing 2.

At this time, the contents of the interpolation processing unit register 63 do not change. Similar processing is performed for pulses 71002 to 71003. The same process is repeated 254 times, and the final interpolated data within the frame period is read into the pitch period register 65 by pulse 71256.

The change in the interpolated pitch cycle information stored in the pitch cycle register 65 is shown by the stepped line 72 in FIGS. 8A and 8B.
Indicated by Of the digital output of the pitch period register 65, seven bits above the decimal point serve as a preset input to the counter 66.

The counter 66 performs a down-count operation by the clock 71, and when the counter output value reaches O, the value of 7 bits above the decimal point of the register 65 is loaded. Also,
At this time, an impulse is generated from the output terminal 67. The operation of the counter 66 will be explained according to the time chart shown in FIG. 8b.

The contents of the counter 66 when the pulse 71050 comes in are 00000101, and the contents of the register 65 are 00010.
00.00000000, set the contents of difference register 63 to 0
．． 01000000.

Then, the contents of the pitch period register 65, the contents of the counter 66, and the operation of the pulse generation timing signal 67 at the time of each pulse are as shown in Table 1 below. FIG. 8b shows how the contents of the counter 66 change during the operation according to Table 1.
The pulse generation timing signal 67 obtained from the counter 66 is shown in FIG. 8b. The pulse generation circuit generates a pulse or a pulse train according to the pulse generation timing signal 67, and outputs a voiced sound source signal from the voice signal output terminal 69. Various signals have been proposed for the pulses generated by the pulse generation circuit 68, such as an impulse, a Barker series signal, or a sine square wave, but in the present invention, the interval of the pulse generation timing signal that gives the generation of these pulses is It is characterized by interpolation, and the type of pulse or pulse train does not matter. As described above, the voiced sound source generation circuit operates. By performing pitch period interpolation processing according to the present invention, discontinuity in the pitch period between frames can be reduced, resulting in smoother intonation of the synthesized sound and improved sound quality of the synthesized sound.

Since the present invention has a circuit that eliminates the discontinuity of the pitch period Tp in the frame period, it can be used in a speech synthesis device that uses pulses or pulse trains for each pitch period as a sound source for voiced sounds. can.

[Brief explanation of drawings]

Figure 1 is a diagram showing the modeling of the generation mechanism, Figure 2 is a model diagram in which the voice generation mechanism is replaced with an electric circuit, Figure 3 is a conventional speech analysis and synthesis device, and Figure 4 is an explanatory diagram of the sound source generation circuit. Figure 5a is a diagram explaining the discontinuity of the pitch period T of the voiced sound in the nth and n+1th frames, Figure 5b is a diagram showing the excitation waveform during an actual voiced sound, and Figure 6 7 is a diagram showing the interpolation process, FIG. 7 is a diagram showing the pitch period interpolation processing circuit, FIG. 8a is a time chart for explaining the pitch period interpolation process, and FIG. 8b is the same as FIG. This is a partially enlarged time chart. 11...Speech analysis device, 12...Transmission path or storage device, 13...Speech synthesis device,
14...Audio analog signal input terminal, 15.
...Low pass filter, 16...AD converter, 17...Yumi correlation circuit, 18...
Parameter calculation circuit, 19... Sound source extraction filter, 20... Sound source determination circuit, 21...
Parameter (α parameter, or k parameter, or L
SP parameter), 22... sound source information signal (T
,,V/UV. A), 23... Encoder, 24...
... Decoder, 25 ... Decoded parameters, 26 ... Decoded sound source information signal, 27
......Sound source generation circuit, 28...Synthesis filter, 29...DA converter, 30...
Low pass filter, 31...Synthesized voice output terminal, 40...Noise generation circuit, 41...Impulse generation circuit, 42...Pitch period information T
Input terminal, 43...Switch, 44...
・Voiced・P・)
Silence determination signal V/UV input terminal, 45... Amplitude multiplier circuit, 46... Input terminal for amplitude A, 47.
... Sound source signal output terminal, 60 ... P input terminal for pitch period information T, 61 ... Subtractor, 6
2...Shifter, 63...Difference register, 64...Adder, 65...Register, 66...Counter, 67... . . . Pulse generation timing signal, 68 . . . Pulse or pulse train generation circuit, 69 . . . Voiced sound source signal output terminal, 70 . . . 1
Clock signal, 70000, 70001...Pulse of first clock signal 70, 71...Second clock signal representing interpolation processing period T, 71000-71
256...Pulse of second clock signal 71, 7
2...A stepped line representing the contents of the register 65,
73...A stepped line indicating the contents of the counter 66,
74... Timing of the impulse emitted from the counter 60.

Claims (1)

[Claims] 1. In the voiced sound source generation section of the linear prediction type speech synthesizer, the first register 65 stores the interpolation processing result.
and a subtracter that takes the difference between the pitch period information of the voiced sound sent every frame period and the interpolation processing result stored in the first register 65, and converts the difference into an interpolation processing unit corresponding to the interpolation processing period. A second register 63 stores a unit of interpolation processing and a second register 63 for storing a unit of interpolation processing.
an adder that sequentially adds the interpolation processing unit stored in the register 63 and the interpolation processing result stored in the first register 65, and stores the addition result in the first register 65 again; It is also connected to the pulse generation circuit, and counts down every interpolation processing cycle, and when the stored value reaches a specific value, updates the stored value with the contents of the first register 65, and starts the pulse generation circuit. 1. A speech synthesizer comprising a pitch period interpolation processing circuit having a counter 66 that generates a signal for activating the circuit.