JPH0754436B2 - CSM type speech synthesizer - Google Patents

Description

The present invention relates to a CSM-type speech synthesizer, and more particularly to a CSM-type speech synthesizer with improved synthetic sound quality.

[Conventional technology]

Recently, CSM type voice synthesizers that express voice signals with combinations of sine wave frequencies of about 4 to 6 waves are becoming well known.

As a speech synthesizer, the conventional LPC (Linear Prediction Codin
g, linear predictive analysis) type speech synthesizer) is widely used, but this LPC type speech synthesizer generally has a complicated configuration.
Also, there is a problem that the stability of the characteristics of the LPC synthesis filter used for speech synthesis is often impaired due to an error during the transmission of speech parameters.

On the other hand, the CSM speech synthesizer does not need to use a speech synthesis filter such as an LPC synthesis filter, as will be described later, so that its configuration is extremely simplified and inherently causes stability problems during synthesis. There is no.

[Problems to be solved by the invention]

However, in this kind of conventional CSM type speech synthesizer, it is represented by a sine wave frequency of about 4 to 6 waves.
In order to synthesize speech using CSM, simply linearly combining them is not enough, and it is necessary to perform some special processing in addition to these processings.

These special processes are performed on a set of simple CSM, that is, a line spectrum of about 4 to 6 waves at the maximum, and a spectrum envelope corresponding to the vocal tract transmission characteristic, and an unvoiced voice created by the vocal cord vibration characteristic or the vocal tract narrowing. It is for converting into a spectrum in the form of convolution with the fine structure of the spectrum corresponding to the characteristics of the sound source, and the purpose is to spread the spectrum required for several CSMs at most, as will be described in detail later. It is carried out through a special window processing using. FIG. 11 is an explanatory view of window processing in the conventional CSM. FIG. 11 shows window processing for the case of generating voiced sound. We show a situation in which a time window with a preset window function is used, and this time window is output while resetting the phase of all CSM sine wave signals in response to the pitch period. However, if the sound source is added on the premise of such window processing, the sound source as a fine structure of the spectrum will have an excessive pitch structure.
This is because the real voice has a pitch structure, but its collapse is also large. Therefore, the conventional voiced sound generation method has a problem that the artificial sound is too synthetic to be aurally unnatural or mechanical sound.

The above-mentioned problem is intended for voiced sounds, while unvoiced sounds have the following problems.

That is, the processing of unvoiced sound is generally not always known clearly at present. In particular, the spread spectrum method required for unvoiced speech synthesis using CSM has not been established, and therefore it is difficult to say that CSM-type speech synthesizers are still in practical use.

The object of the present invention is to eliminate the above-mentioned drawbacks and to put it into practical use.
It is to provide an M-type speech synthesizer.

[Means for solving problems]

The CSM type speech synthesizer of the present invention is configured to have a voiced sound generating means including a spread spectrum means that is not restricted by a pitch structure.

Further, the CSM type speech synthesizer of the present invention has unvoiced sound generating means including a random number generating means and a spectrum spreading means utilizing a continuous waveform having a period set based on a random number signal generated by the random number generating means. In the CSM type speech synthesizer having, the spread spectrum means is configured as frequency modulation.

Further, the CSM type speech synthesizer of the present invention has a CSM having unvoiced sound generating means including random number generating means and spread spectrum means using a periodic signal set based on the random number signal generated by the random number generating means. In the type speech synthesizer, the spread spectrum means includes window function generating means having a period different from the period designated by the synchronizing signal.

〔Example〕

 First, the principle of the CSM type speech synthesizer will be explained.

CSM expresses a voice signal as the sum of a specific number of sine waves having amplitude and frequency as freely selectable parameters. A preset number of at most 4 to 6 is used as the specific number. Therefore, in order to perform CSM voice synthesis, it is first necessary to express the voice signal to be analyzed as a sum of a preset number of sine waves by CSM voice analysis.

The CSM speech analysis will be described in detail later, and only the main points will be explained here.

CSM voice analysis also ignores phase information, as in LPC analysis.
The autocorrelation coefficient is used as an intermediate parameter for the purpose of averaging the effects of sound sources and avoiding instability due to miscellaneous components.

That is, in CSM analysis, N low order taps of the sample autocorrelation coefficient directly calculated from the speech waveform to be expressed for each analysis frame are converted into N low order taps of the autocorrelation coefficient of the composite wave. The frequency of each sine wave to be combined and its intensity (power amplitude) are determined so as to match the individual numbers.

Now, assuming that the number of sine waves to be combined is n, the angular frequency of each sine wave is ω _i (i = 1, 2, ... N), and the strength of each sine wave is m _i , the combined wave y _t by CSM Is expressed by the following equation (1).

The autocorrelation coefficient γ _l of the tap l in the equation (1) can be easily expressed as in the equation (2) by using ω _i , m _i .

On the other hand, if the sample of the speech waveform to be expressed is X _t , the sample autocorrelation coefficient V _l of tap l in a certain frame can be expressed by the following equation (3).

In Expression (3), M is the number of samples per analysis frame.

Now, in the CSM analysis, the above-mentioned γ _l is given V
m _i and ω _i are determined so as to be equal to _l and N low-order ones. That is, m _i and ω _i are determined so that γ _l = V _l (l = 0,1,2, ... N) holds. Here, it is assumed that n sine waves m _i and ω _i are obtained next for each analysis frame corresponding to a given audio signal. FIG. 12 is a voice feature vector pattern diagram showing an example of a voice feature vector pattern based on the CSM parameters m _i and ω _i .

FIG. 13 is a CSM / LPC spectrum contrast diagram showing the CSM line spectrum and the LPC spectrum envelope in contrast. In the case of FIG. 13, analysis frame length is 30 mSEC, analysis order is 9th order (N = 9)
The CSM line spectrum of CSM (n = 5) is compared with the 9th-order LPC spectrum envelope obtained from the same voice sample.

Between the order N and the number n of sine waves, N = 2n−
It is also well known that there is a one relationship.

However, such CSM analysis of resulting n-number of m _i, using the value of omega _i, the m _i, the intensity (amplitude specified by omega _i is By creating n sine waves having angular frequencies and adding them simply, the purpose of reproducing the original voice cannot be achieved by merely hearing as a synthesized sound of the sine waves in the human ear.

This means that even if a sine wave is simply added, the spectrum of the generated signal is only the discretized n line spectra, while the spectrum of the voice signal has a continuous spectral envelope, and Voice sounds are represented by a pitch structure, and unvoiced sounds have a fine spectral structure represented by a stochastic process. Therefore, the simple addition of CSM and the speech signal have completely different spectral structures. So CSM
To synthesize speech using, it is necessary to spread the line spectrum into a continuous spectrum using some method. That is, CSM speech synthesis results in generation of a speech spectrum pattern from a speech feature spectrum pattern represented by a line spectrum as shown in FIGS.

In the present invention, the following method is used to perform the above-mentioned spread spectrum.

That is, the fact that the voiced sound has a clear pitch structure is used to reset the phase of each of the n sine waves at every pitch period. This makes it possible to easily generate a spectrum and a fine pitch structure. Further, as will be described later, a special window process is applied to the above-mentioned phase reset waveform to remove the discontinuity of the composite waveform at the time of phase reset. FIG. 14 is a spectrum envelope and pitch fine structure characteristic diagram of the CSM before spectrum diffusion in the present invention, and FIG. 15 is a spectrum characteristic diagram of the conventional CSM after spectrum diffusion. Experimental results have shown that the contents of FIG. 14 are changed to a spectrum having a spectrum envelope and a pitch fine structure, and the sound quality is audibly sufficient for practical use. On the other hand, in the case of FIG. 14, the purpose of the voice synthesis cannot be achieved because the sine wave can be heard as a synthesized waveform.

The above is the case of voiced sound, but the case of unvoiced sound is performed as follows. That is, in the case of the voiced sound described above,
The phase reset and special window processing performed for each pitch period are randomly generated as a stochastic process instead of the pitch period in the case of unvoiced sound, using a pulse whose distribution width and lower limit value are set. The above-mentioned processing is performed every time the pulse is generated.

However, unlike voiced sound, unvoiced sound originally does not have a clear sound source such as a vocal cord, and the entire vocal tract produces turbulence and becomes a sound source. Therefore, the unvoiced sound does not require a clear phase relationship between frequency components generated by phase initialization. Therefore, in the present invention, a method of FM-modulating each frequency of CSM with noise is used as one unvoiced sound synthesis for performing frequency spreading without having an unnecessary phase relationship. The noise in this case may be white noise or the above-mentioned period for which the distribution width and the lower limit value are set. This eliminates unnecessary phase initialization in unvoiced sounds.

Further, in the present invention, the following method is also used as another synthesis method of unvoiced sound.

That is, targeting a CSM type speech synthesizer having an unvoiced sound generating means including a spread spectrum means that utilizes a synchronizing signal set based on a random number signal generated by the random number generating means, and has a cycle different from that of the synchronizing signal. The unvoiced sound is synthesized by the spread spectrum means including the window function generating means. By using such window function generating means, the continuity in the vicinity of the phase reset time can be remarkably improved.

Further, in the present invention, the pitch structure of voiced sound is too clarified in the conventional method, and a very unnatural sensation is given, whereas the spread spectrum means not restricted by the pitch structure is used to synthesize the voiced sound. Is also implemented. This is aimed at the effect of enhancing the continuity of the synthesized voice at the boundary portion of the window section, as in the above-mentioned example in which the window function generating means having a period different from that of the synchronizing signal is used in the case of unvoiced sound.

Now, specific embodiments of the present invention will be described in detail with reference to the drawings. FIG. 1 is a block diagram showing a first embodiment of the present invention.

The first embodiment is a CSM type speech synthesizer equipped with unvoiced sound generating means including a spread spectrum means for utilizing a continuous waveform having a period set based on a random number signal generated by random number generation, and further, frequency modulation. It shows a specific example in the case where is the spread spectrum.

The first embodiment comprises an analysis side (transmission side) and a synthesis side (reception side). Although not shown on the analysis side, A / D converter, Hamming window processor, autocorrelation coefficient measuring instrument, CSM analyzer, CSM
It is composed of a quantizer, an electronic quantizer, a pitch extractor, a voiced / unvoiced sound determiner, and a multiplexer.

As shown in Fig. 1, the synthesizing side is a demultiplexer
Decoder 101, interpolator 102, synchronization calculator 103, random number generator 1
04, n variable frequency oscillator with phase reset function 105
(1), 105 (2), ... 105 (n), n variable gain amplifiers 1
06 (1), 106 (n), addition combiner 107, variable length window function generator 108, multipliers 109, 110, n FM modulators 111 (1), 111
(2), ... 111 (n), a sawtooth wave generator 112, voiced / unvoiced switchers 113 (a) to 113 (c), and the like.

On the analysis side, the voice waveform is converted by an A / D converter into digital data whose amplitude and time axis are quantized, and is supplied to a Hamming window processor, a pitch extractor, a voiced sound / unvoiced sound determiner and the like.

The digital data supplied to the Hamming window processor is subjected to weight multiplication by a known Hamming window function for each predetermined analysis frame, and is supplied to the autocorrelation coefficient measuring device for each data of each analysis frame.

The autocorrelation coefficient measuring device obtains N low-order autocorrelation coefficients V _l by the above-described calculation of the equation (3) for each data of each analysis frame thus input. The set of V _l for each analysis frame obtained in this way is supplied to the next CSM analyzer, and V ₀ therein, that is, the autocorrelation coefficient at the delay time of zero is supplied to the power quantizer as the power information in the analysis frame. Supply.

The CSM analyzer supplied with the set of autocorrelation coefficients V _l for each analysis frame specifies the intensity and angular frequency of each of the n sine waves of the CSM of the corresponding analysis frame by performing the operation described later. M _i and ω _i are determined and are supplied to the CSM quantizer.

The CSM quantizer quantizes the set of m _i and ω _i with an appropriate roughness set in consideration of the demand for the reproduced sound quality and the transmission capacity of the transmission line, and then supplies it to the multiplexer. In addition, the power quantizer also quantizes V ₀ with the roughness set from the above viewpoint and then supplies it to the multiplexer. Also,
The pitch extractor, which received the digital data from the A / D converter, extracts the pitch period, quantizes it, and then supplies it to the multiplexer. The unvoiced sound is determined and supplied as a binary signal to the multiplexer.

The multiplexer supplied with the above signals multiplexes these signals in a format suitable for ease of separation on the combining side and for transmission, and transmits the signals to the combining side.

On the synthesizing side, the signal thus transmitted is received by the demultiplexer / decoder 101, decoded and separated to restore the signal state of the multiplexer input on the analyzing side.

In order to simplify the explanation, n FM modulators 111 (1) to 111
(N) will be described assuming that it does not exist.

The angle signal thus restored is an interpolator having a memory function.
After being supplied to 102 and subjected to necessary interpolation, ω _{1 to} ω _i are n variable frequency oscillators with a phase reset function 105 (1)
~ 105 (n) frequency control inputs to set the output angular frequencies of each of these oscillators to ω ₁ to ω _n , respectively.

Further, m _{i to mn} that specify the intensities (power amplitudes) of the n sine waves of the CSM are n variable gain amplifiers 106 (1) to 106, respectively.
It is supplied as a gain control input of (n), so that the oscillation power of each frequency is controlled to a designated value.

The n variable gain amplifiers 106 (1) to 106 thus obtained
The output of (n) is added and synthesized by the addition and synthesis unit 107 and then supplied to the multiplier 109.

The pitch period information output from the demultiplexer / decoder 101 is supplied to the variable length window function generator 108 after the necessary interpolation is performed in the interpolator 102.

The variable-length window function generator 108 generates a window function for ensuring the continuity of the speech waveform, excluding the discontinuity that occurs in the output waveform due to the phase reset, and at the same time, the window function close to this window function. A related phase reset pulse is also generated. As described above, the variable length window function generator 108
Next, the pitch period data string that specifies the interval of the phase reset pulse is input to the variable length window function generator 10.
8. This data next generates an impulse having a time interval specified by this data, which is then transmitted via a voiced / unvoiced switch 113 (a) to a variable frequency oscillator 105 (1) with a phase reset function. .About.105 (n), which are supplied as phase reset inputs to reset the phase of each oscillator. It is also fed to the interpolator 102, where ω _i and m _i
Is used as a timing signal for interpolating.

Now, the variable length window function generator 108 generates the following variable length window function W (x) in synchronization with the above-described phase reset pulse generation. That is, if the phase reset pulse interval at that point in time specified by the supplied data is T and the elapsed time from the occurrence of the previous phase reset pulse is t, then W (t) is given by the following equation (4). Indicated by.

FIG. 16 is a variable length window function characteristic diagram of the embodiment shown in FIG.
In the case of voiced sound, the value of T in this case represents a pitch period and changes with time. Therefore, W (t) has a variable length. FIG. 17 is a characteristic diagram of the relative time relationship between the variable length window function and the phase reset pulse of FIG. As is clear from FIG. 17, the starting point and the ending point of the window function almost coincide with the phase reset pulse.

The output of the variable length window function generator 108 is the voiced / unvoiced switch 113.
It is supplied to the multiplier 109 via (b). As a result, the multiplier 109 obtains the product of the n sine waveforms synthesized by the addition synthesizer 107 and the window function W (t) generated in synchronization with each phase reset pulse. The waveform thus obtained has a window function W just before each sine wave is reset in phase.
It is continuously converged to zero after being multiplied by (t), and each sine wave rises from zero at the time of phase resetting, so that the continuity of the waveform is secured, and thus the multiplication of the window function W (t) is performed. Discontinuities that occur in the phase reset waveform can be eliminated.

The output of the multiplier 109 is supplied to the multiplier 110, weighted by the electric power V ₀ output through the interpolator 102, and output as synthesized speech.

The above is the operation at the time of speech synthesis in the first embodiment, but it is intended for voiced sound, and therefore the voiced / unvoiced switch 113 (a), (b), (c) is the first. It is connected to the (V) side shown in FIG. 1, that is, the voiced sound side.

The above is the case of voiced sound, but in the case of unvoiced sound, the voiced / unvoiced switch 113 (a), (b), (c) is (UV).
It is switched to the side and processed as follows.

The random number generator 104 generates a random number signal and outputs it to the period calculator 103.
Output to. The period calculator 103 outputs the period data calculated based on the provided random number signal to the sawtooth wave generator 112.

The sawtooth wave generator 112 generates a sawtooth wave whose period is controlled by the supplied period data. FIG. 18 is a sawtooth wave output waveform diagram in the embodiment of FIG.

The sawtooth wave generator 112 has the periods T ₁ , T ₂ , T ₃ , T ₄ shown in FIG.
A sawtooth wave such as ... is output, and a voiced / unvoiced switch 113 (c)
Supply to.

The FM modulators 111 (1) to 111 (n) FM-modulate the sine waves provided from the variable frequency oscillators 105 (1) to 105 (n) with a phase reset function, and the modulated signals are in the unvoiced state. In case of, the output of the sawtooth wave generator 112 is the voiced / unvoiced switch 1
13 (c), and in the case of voiced voice, the "0" voltage is supplied via the voiced / unvoiced switch 113 (c). In other words, variable frequency oscillator with phase reset function 105 (1)
The output waveforms of up to 105 (n) are FM-modulated by a sawtooth wave for unvoiced sounds and not for voiced sounds. It is well known that a sine wave is spread in frequency by FM modulation, and a description thereof will be omitted. The modulation index is set empirically from the viewpoint of hearing.
The voiced / unvoiced switching units 113 (a) to 113 (c) are switched simultaneously by the V (voiced) / UV (unvoiced) signal output from the demultiplexer / decoder 101. In this way, the voice synthesis is synthesized by using the phase resetting pulse for voiced sound, and the voice is synthesized by FM modulation for unvoiced sound.

In the case of unvoiced sound synthesis, the variable frequency oscillators 105 (1) to 105 (n) with a phase reset function do not reset the phase, and a constant DC signal is applied to the multiplier 110 to shape the waveform. Absent. In addition, the interpolator 102 is a variable frequency oscillator with a phase reset function 105 (1) to 105 only when voiced.
Interpolation processing is performed in synchronization with the phase reset pulse supplied to (n), and during unvoiced sound, interpolation processing is performed at a constant cycle such as 5 mSEC.

In this way, CSM synthesis necessary for speech synthesis is executed,
A relatively good sound quality can be obtained regardless of the amount of information compression or transmission error in the transmission path.

In the above description, the interpolation for each data in the interpolator 102, each data analysis side can be performed in various combinations depending on the roughness when quantizing, for example omega _i only, or omega _i, Only m _i can be considered, and the interpolation method can be linear interpolation or high-level interpolation. Regarding the interpolation for omega _i, m _i, it is advantageous to select the interpolation points so interpolation data is obtained for each time phase reset pulse generator described above, the update of omega _i, m _i in this time In order to perform this, it is also supplied to the phase reset pulse interpolator 102 as described above. In order to perform such interpolation, the interpolated data is obtained after the necessary post-data arrives or is generated, so the interpolator 102 stores the necessary information up to the time required for the interpolating operation. It has a memory to store.

FIG. 2 is a circuit diagram showing in detail a portion of the variable frequency oscillator with the phase reset function shown in FIG.

The current value flowing through the constant current sources 1051 and 1052 is controlled by the voltage applied to the frequency control terminal. These constant currents are controlled by the capacitor 10 which is switched by the switch 1050.
The charge / discharge current of 53 will be controlled. The output of the V point becomes linearly up and down triangular waveform between the reference voltage + V _gamma and -V _gamma. The constant currents i ₁ and i ₂ respectively determine the rising and falling characteristics of this triangular waveform and thus also its period. For example, if this current value is small, the cycle becomes large.
The switch 1050 is controlled by the Q output of the RS flip-flop circuit 1059, but the RS flip-flop circuit 1059
Receives the output of the comparator 1056 at the reset terminal R,
In addition, the output of the OR circuit 1058 which takes the output of the comparator 1057 and the input of the phase reset terminal as two inputs is set terminal S
In response to this, the switches are switched to form the triangular waveform. That is, when an impulse is applied to the phase reset terminal, the point V is forcibly returned to the zero potential by the switch 1054, the oscillation is restarted from this zero potential, and the switch 1050 is connected to the charging side and the phase is reset.

The triangular wave oscillation output at point V thus obtained is supplied to the sine wave converter 1055 and converted into a sine wave. This sine wave converter can be easily implemented, for example, by reading the sine function value stored in ROM as an input waveform.

Further, such a variable frequency oscillator with a phase reset function can be easily implemented by using a computer program.

FIG. 3 is a circuit diagram showing in detail a portion of the variable gain amplifier shown in FIG. The variable gain amplifier shown in FIG.
1, the gain of the operational amplifier by the resistor 2073 is made variable according to the voltage applied to the field effect transistor 2072, the signal to be amplified is added to the input terminal, the control signal is added to the control terminal, and the output terminal Obtain an output with a controlled amplitude.

In addition to this, it is also possible to realize by using an analog multiplier. Furthermore, the analog waveform input is used for the reference voltage of the D / A converter, and the control information expressed in digital quantity is used as the digital input. It can be easily realized by a method such as

FIG. 4 is a circuit diagram showing in detail the portion of the random number generator shown in FIG. The exclusive OR of the signals applied to the two input terminals A and B is obtained from the output S terminal by using the combination of the shift register 1041 of 15 stages and the half adder 1042 shown by D _{1 to} D _15. A 15th-order M-series random number signal having a period of 2 ¹⁵ -1 is generated in the form of being fed back to the input of. A shift pulse is applied to the CLK (clock) terminal at a necessary timing to advance to obtain the next random number value.

FIG. 5 is a block diagram showing in detail the portion of the period calculator 103 of FIG. Random numbers output from the random number generator 104 and uniformly distributed in the range of 0 to 2 ¹⁵ -1 are converted into a distribution suitable for use as a random number that specifies the time interval of the phase reset pulse when unvoiced. Thing with constant multiplier 1031
And a constant adder 1032, and the upper and lower limits of the random number and the distribution width can be set to appropriate values through these constant multiplications and additions.

FIG. 6 is a block diagram showing in detail the portion of the variable-length window function generator shown in FIG. 1, and comprises a register 1081, a presettable down counter 1082, a counter 1803, a ROM 1084 and the like.

Pitch cycle data specifying the phase reset pulse interval supplied from the interpolator 102 is stored in the register 1081.

The down counter 1082 is a counter that counts the high-speed clock CLK having a constant cycle, presets the content T of the register 1081, and down counts it by using CLK. When the content of the down counter 1082 becomes zero, a pulse is generated from the output terminal, whereby the content of the register 1081 is preset again and the down counting is started. In this way, the output of the down counter 1082 generates a pulse train of a period proportional to T, for example, T / K. This pulse train is provided as the clock of the counter 1038, and the count output of the counter 1083 which is incremented by this clock is applied to the ROM 1084 as an addressing signal to sequentially output the data of the window function W (t) written at this address. To read and output. When the content of the counter 1083 becomes K, the last data of the window function W (t) of the ROM 1084 is read out, and at the same time, the count 1083 is reset and the ROM 1084 outputs the phase reset pulse. In addition, K in ROM1084
The window function data stored in advance as individual samples is sent to the multiplier 109 via the voiced / unvoiced switch 113 (b).
Is supplied to. In this way, the phase reset pulse having the next specified pulse spacing and the variable length window function W (t) synchronized with the pulse are generated.

For the CSM analysis and its concrete configuration, refer to “Speech spectrum analysis using complex sine wave model” Sagayama et al., IEICE Transactions 81/2, Vol.J64-A No.2 P.105-112, and others. This is detailed in Japanese Patent Application No. 59-143045.

Note that the CSM analysis on the analysis side in the above-described first embodiment uses a method of solving an equation in which the sample autocorrelation coefficient and the CSM autocorrelation coefficient are equal, but instead of this, LP
It is also possible to use the method of calculating the line spectrum frequency by making the C coefficient lossless and calculating the residue. In this embodiment, the interpolator performs parameter interpolation at the time of resetting the phase, but this interpolation can be omitted depending on the operating conditions.

Further, the variable length window function used in the present embodiment is an example, and other function forms may be used.

Further, the random number generator period calculator and the like also show an example thereof, and need not be limited to this.

Next, a second embodiment of the present invention will be described. FIG. 7 is a block diagram showing a second embodiment of the present invention. In the second embodiment, the spectrum spreading means in the CSM type speech synthesizer having the voiced sound producing means including the spread spectrum means not restricted by the pitch structure is frequency-modulated. That is, in the first embodiment, in the synthesis of unvoiced sound
Although FM modulation is used, FM modulation can also be used for voiced sound, and this is used.

In FIG. 7, components related to the same symbols are exactly the same as those in FIG. 1, and therefore detailed description thereof will be omitted.

Voiced / unvoiced switch 114 is demultiplexer / decoder 101
When the voiced (V) is specified while receiving the V / UV information supplied from, the pitch period information from the interpolator 102 is supplied to the variable length window function generator 108, in the same manner as in the case of FIG. The generated phase reset pulse is a variable frequency oscillator with phase reset function 105 (1) to 105 (n) and an interpolator.
Supply to 102. The variable length window function is also provided to the multiplier 109, and thus frequency modulation is performed as a spread spectrum means even in the case of voiced sound. On the other hand, in the unvoiced case, the provision of the pitch period data to the variable length window function generator is stopped, and the unvoiced sound is generated as in the case of FIG. In this way, even in the synthesis of voiced sound, FM modulation is performed, and it becomes possible to synthesize a voice closer to a real voice. This means the following effects. That is, the spectrum of speech synthesized by simple CSM phase initialization gives a pitch harmonic structure that is too clear, whereas actual speech has a pitch structure that is somewhat unclear due to fluctuations in the vocal cords. Therefore, it is possible to get closer to the real voice by using FM modulation together with voiced sound synthesis.

Next, a third embodiment of the present invention will be described. FIG. 8 is a block diagram showing a third embodiment of the present invention, which includes a demultiplexer / decoder 101, an interpolator 102, and a period calculator 10.
3. Random number generator 104, etc. In addition to the above-mentioned contents, voiced / unvoiced switch 114, counter 115, fixed window length CSM synthesizer 116 (1), 1
16 (2), ... 116 (n), an adder / combiner 117, and the like.

The third embodiment is a CSM type speech synthesizer having a voiced sound generating means including a spread spectrum means not restricted by a pitch structure, wherein the spread spectrum means has a window function generating means having a period different from the pitch period length. And a CSM type speech synthesizer having unvoiced sound generating means including a spectrum spreading means using a periodic signal set based on a random number signal generated by the random number generating means, the spectrum spreading means being Shows an example of a case in which a voiced sound and an unvoiced sound are generated by combining two cases in which the window function generating means having a cycle different from the cycle specified by the above-described periodic signal is included. is there.

The voiced / unvoiced switch 114 is a demultiplexer / decoder 1
Counter 115 switched by V / UV information received from 01
When voiced, the pitch period, and when unvoiced, the period calculator 103
Data about the random number cycle output from the counter 115
Supply to.

The counter 115 outputs a phase reset pulse to the fixed window length CSM synthesizers 116 (1) to 116 (n) while stepping each time the pitch period data or the random number period data input in this manner is input. Fixed window length CSM combiner 1
Supply to 16 (1) to 116 (n).

The n fixed window length CSM synthesizers 116 (1) to 116 (n) respectively interpolate data m _{1 to mn} , ω ₁ to from the interpolator 102.
ω _n and power information V ₀ are provided, and CSMs having different fixed window lengths are combined and output to the adder / combiner 117.

FIG. 9 is a block diagram showing in detail the part of the fixed window length CSM synthesizer 116 of FIG. 8, and n variable frequency oscillators 1161 (1) to 1161 (n) with a phase reset function and fixed length window function. Generator, n variable gain amplifiers 1163 (1) -1163 (n),
It is configured to include an adder / combiner 1164, a multiplier 1165, and a multiplier 1166.

n frequency oscillators with phase reset function 1161 (1) to 11
Reference numeral 61 (n) receives ω ₁ to ω _n from the interpolator and oscillates a sine wave frequency while resetting the phase with the phase reset pulse and outputs the sine wave frequency to the variable gain amplifiers 1163 (1) to 1163 (n). These variable gain amplifiers are subjected to variable gain control corresponding to m _{1 to mn,} and then added by an adder / combiner.
In the case of voiced sound, the phase reset of the waveform thus output is performed in correspondence with the pitch period, and in the case of unvoiced sound, it is performed in accordance with the random number period.

Multiplier 1165 multiplies the output of addition combiner 1164 and the fixed length window function received from fixed length window function generator 1162, and supplies the output to multiplier 1166. The fixed length window functions are different for each of the n fixed window length CSM combiners, and thus a treatment may be chosen such that these fixed length windows overlap or are separated from each other. The purpose of using such a fixed-length window function is nothing but the purpose of further improving the continuity of synthesized sound quality near the phase reset time.

FIG. 10 is a block diagram showing in detail the part of the fixed length window function generator 1162 of FIG. 9, and the RS type flip-flop circuit 11
62-1, AND circuit 1162-2, counter 1162-3, ROM1162-4,
It is configured by including a gate circuit 1162-5 and the like.

In the RS type flip-flop circuit 1162-1, the Q output terminal becomes “0” level each time the S terminal receives the phase reset pulse from the counter 115, the counter 1162-3 is also reset, and the AND circuit 1162-2 outputs The high speed clock is output to the counter 1162-3. The counter 1162-3 stores the data of the value corresponding to the time length of the fixed length window function to be formed for each fixed window length CSM synthesizer, and when it is read by the input high speed clock, this is the address of the ROM 1162-4. The data regarding the fixed-length window function stored in the corresponding address of the ROM 1162-4 is output to the multiplier 1165. Further, the counter 1162-3 outputs the address data thus output to the gate circuit 1162-5, and the reset pulse of a predetermined format is output using the logic gate.

The reset pulse output from the gate circuit 1162-5 is RS
Applied to the R terminal of the flip-flop circuit 1162-1 to invert the polarity of the output terminal and stop the counting operation of the counter 1162-3.

In this way, a predetermined window function is output from the ROM 1162-4 for each phase reset pulse input and provided to the multiplier 1165.

The multiplier 1165 multiplies the fixed-length window function thus input and the output of the adder / combiner 1164 and supplies the result to the multiplier 1166.

The multiplier 1166 multiplies the output of the multiplier 1165 by the power information V ₀ and outputs it to the adder / combiner 117.

The additive synthesizer 117 thus additively synthesizes the CSM synthesized sound output from each fixed window length CSM synthesizer and outputs the synthesized sound.

In this way, voiced sound that is not constrained by the pitch structure and that uses a window function with a period different from the pitch period length, and voiceless sound that uses a window function with a period different from the period specified by the random number signal Synthesis is easy.

〔The invention's effect〕

As described above, according to the present invention, in the CSM type speech synthesizer, a voiced sound generating means including a spectrum spreading means that is not restricted by a pitch structure, or a spectrum using a continuous waveform of a period set based on a random number signal An unvoiced sound generating means including spread spectrum means having a window function generating means having a period different from a period specified by a periodic signal set based on a random number signal Due to
This has the effect of realizing a CSM-type speech synthesizer capable of implementing a speech synthesis method that is remarkably close to the real voice.

[Brief description of drawings]

FIG. 1 is a block diagram of a first embodiment of the present invention, FIG. 2 is a block diagram showing in detail a portion of a variable frequency oscillator 105 with a phase reset function in the embodiment of FIG. 1, and FIG. FIG. 4 is a block diagram showing in detail the portion of the random number generator 104 in the embodiment of FIG. 1, and FIG.
FIG. 6 is a block diagram showing in detail the portion of the period calculator 103 in the embodiment of the figure, FIG. 6 is a block diagram showing in detail the portion of the variable length window function generator 108 in the embodiment of FIG. 1, and FIG. FIG. 8 is a block diagram of a second embodiment of the invention, FIG. 8 is a block diagram of a third embodiment of the present invention, and FIG. 9 is a detailed view of a portion of the fixed window length CSM synthesizer 116 in the embodiment of FIG. FIG. 10 is a block diagram showing in detail the fixed length window function generator 1162 of FIG. 9, FIG. 11 is an explanatory view of window processing in a conventional CSM, and FIG. 12 is a CSM parameter m _i , Voice feature vector pattern diagram by ω _i, FIG. 13 is a CSM / LPC spectrum contrast diagram showing the CSM line spectrum and the LPC spectrum envelope in contrast, and FIG. 14 is a spectrum envelope and pitch after spectrum spread of the CSM in the present invention. Fig. 15 is a microstructure characteristic diagram of the conventional CSM after spread spectrum. FIG. 16 is a characteristic diagram of the spectrum, FIG. 16 is a characteristic diagram of the variable length window function in the embodiment of FIG. 1, FIG. 17 is a characteristic diagram of the relative temporal relation between the window function and the pulse for phase reset in the embodiment of FIG. FIG. 18 is a sawtooth wave output waveform diagram in the embodiment of FIG. 101 ... Demultiplexer / decoder, 102 ... Interpolator,
103 …… Period calculator, 14 …… Random number generator, 105 (1) ～ 10
5 (n): Variable frequency oscillator with phase reset function, 106
(1) to 106 (n) ... Variable gain amplifier, 107 ... Addition synthesizer, 108 ... Variable length window function generator, 109, 110 ... Multiplier, 111 (1) -111 (n) ... FM modulation 112, sawtooth wave generator, 113 (a), (b), (c), 114 ... voiced / unvoiced switch, 115 ... counter, 116 (1) -116
(N): Fixed window length CSM combiner, 117: Additive combiner.

Claims (9)

[Claims] 1. A CSM type speech synthesizer having a random number generating means and an unvoiced sound generating means including a spread spectrum means using a continuous waveform having a period set based on a random number signal generated by the random number generating means. CS characterized in that the spread spectrum means is frequency modulation
M-type speech synthesizer. 2. A CSM type speech synthesizer having a random number generating means and an unvoiced sound generating means including a spread spectrum means utilizing a periodic signal set based on a random number signal generated by the random number generating means, A CSM type speech synthesizer characterized in that the spread spectrum means includes window function generating means having a plurality of fixed time lengths which are set independently of the cycle specified by the periodic signal and have mutually different cycles. . 3. A CSM having voiced sound generating means including a spread spectrum means not restricted by a pitch structure, and including a frequency modulation means as the spread spectrum means.
Type speech synthesizer. 4. A voiced sound generating means including a spectrum spreading means which is not restricted by a pitch structure, wherein a plurality of fixed time lengths having mutually different periods set independently of the pitch period length are provided as said spectrum spreading means. A CSM type speech synthesizer characterized by including window function generating means having the same. 5. The CSM type speech synthesizer according to claim 1, wherein the continuous waveform having a period set based on the random number signal generated by the random number generating means is a sawtooth wave. 6. A CSM according to claim 1, further comprising a random number generating means for generating a random number signal having a distribution range set by a finite lower limit value and an upper limit value set in advance. Type speech synthesizer. 7. A sine having means for inputting or accumulating the strength and frequency parameters of a plurality of separately extracted sine wave signals representative of a voice signal and outputting a plurality of sine wave signals using the parameters. Wave generation means, superposition means for superposing a plurality of sine wave signals output by the sine wave generation means, and random number generation for generating a random number signal having a distribution range set by preset finite lower and upper limit values. And a means for resetting the phase of the sine wave signal corresponding to the pitch period of the voice signal when the voice signal is voiced, and modulating the sine wave signal with a continuous waveform having a period set based on the random number signal when the voice signal is unvoiced. A CSM type speech synthesizer according to claim (1), characterized in that it comprises a modulation means. 8. A means for extracting intensity and frequency parameters of a plurality of sine wave signals representing a voice signal, and a sine wave generating means for outputting a plurality of sine wave signals having the extracted intensity and frequency parameters. Superimposing means for superimposing a plurality of sine wave signals output by the sine wave generating means, random number generating means for generating a random number signal having a distribution range set by a preset finite lower limit value and upper limit value, When the voice signal is voiced, the phase of the sine wave signal is reset corresponding to the pitch period of the voice signal, and when the voice signal is unvoiced, the sine wave signal is modulated with a continuous waveform having a period set based on the random number signal. The CSM type speech synthesizer according to claim (1), characterized by comprising: 9. A random number generating means having a distribution range set by a finite lower limit value and an upper limit value set in advance, and the random number generating means according to claim (2).
CSM type speech synthesizer.