JPH0439678B2 - - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION (Technical Field) The present invention relates to CSM parameters, i.e. at most 4 to
CSM (Composite
Simusoidal Modeling: Complex sine wave model) related to parameter quantization.

(Prior Art) Conventionally, LPC-type speech synthesizers have been widely used as speech synthesizers, but LPC-type speech synthesizers generally have a complicated structure. Also used for speech synthesis
The characteristic of the LPC filter is that its stability is compromised due to errors during parameter transmission.

For this, we perform sound synthesis using CSM.
As will be described in detail later, the CSM type speech synthesizer does not have a filter and has a very simple structure, and essentially does not cause stability problems during synthesis. However, regarding the quantization of CSM parameters, conventionally each amplitude of the parameter was quantized separately, and the relationship between the parameters was not considered. Therefore, quantization that takes full advantage of the characteristics of the CSM parameters is not performed, resulting in low quantization efficiency.

(Objective of the Invention) An object of the present invention is to solve the above-mentioned problems when quantizing CSM parameters and to provide an efficient quantization method.

(Structure of the Invention) The quantization method of the present invention uses a set of amplitudes { m ₁ , m ₂ , ..., m _o }, a=max
Normalization coefficient a expressed as {m ₁ , m ₂ , ..., m _o }
It is configured to have means for normalizing by.

(Principle) First, we will explain the principle of the CSM type speech synthesizer.

CSM expresses an audio signal as the sum of a specific number of sine waves whose amplitude and frequency are freely selectable parameters. A predetermined number of 4 to 6 sine waves is used at most.

Therefore, when performing CSM speech synthesis, first,
It is necessary to express the audio signal as a sum of a predetermined number of sine waves using CSM audio analysis.
CSM voice analysis will be explained in detail later.
Only the main points will be explained here.

In CSM analysis, as in LPC analysis,
An autocorrelation coefficient is used as an intermediate parameter for the purpose of ignoring phase information, averaging the influence of sound sources, and avoiding instability due to noise components.

In other words, in CSM analysis, N low-order taps of sample autocorrelation coefficients directly calculated from the speech waveform to be expressed for each analysis frame are combined with N low-order taps of autocorrelation coefficients of a synthesized wave. The purpose is to determine the frequency and intensity (power amplitude) of each sine wave to be synthesized so that they match the N sine waves.

Now, if the number of sine waves to be synthesized is n, the angular frequency of each sine wave is ω _i (i=1, 2,..., n), and the intensity of each sine wave is m _i , then the CSM composite wave yt is yt＝ _o 〓 ⁱ⁼¹ √ _i sin(ω _i t+φ _i ), but the autocorrelation coefficient γ _l of this tap l is ω _i ,
It is easily expressed using m _{i and} γ _l = _o 〓 ⁱ⁼¹ m _i coslω _i .

On the other hand, the sample of the audio waveform to be expressed is
When X _t is assumed, the sample autocorrelation coefficient v _l of tap l in a certain frame is given as v _l =1/M _M-1 〓 ^t=l X _t X _tl . However, M is the number of samples in one analysis frame.

Now, in the CSM analysis, the value of each m _i and ω _i is determined so that the above-mentioned γ _l is equal to the given v _l for N low-order values.

In other words, the values of m _i and ω _i are determined so that γ _l =v _l where l=0, 1, 2, . . . , N holds true.

This specific method will be explained in detail later, but here, m _i and ω _i of the n sine waves mentioned above are obtained one after another for each analysis frame in response to a given audio signal. shall be provided.

FIG. 1 shows an example of a speech feature vector pattern based on the CSM parameters m _i and ω _i obtained in this way.

In addition, the 9th order (N = 9) CSM (number of sine waves n =
5) Figure 2 shows an example of the correspondence between the line spectrum and the 9th-order LPC spectrum envelope (frequency transmission characteristic of the LPC synthesis filter) obtained from the same audio sample.

Note that there is a relationship of N=2n-1 between the above-mentioned order N and the number n of sine waves, as described later.

From these figures, it can be seen that the CSM contains information that extracts the features of the original speech that should be expressed.

However, using n sets of m _i and ω _i values obtained as a result of CSM analysis, the intensity specified by m _i and ω _i (the actual amplitude is m _i as mentioned above)
If you create n sine waves with angular frequencies and simply add and synthesize them, the human ear will hear:
You can simply hear it as a synthesized sound of sine waves,
The purpose of reproducing the original sound cannot be achieved.

This means that even if you simply add sine waves, the spectrum of the generated signal is just a discretized n-line spectrum, whereas the spectrum of the audio signal has a continuous spectral envelope, and also This is thought to be due to the fact that voiced sounds are expressed by a pitch structure, and unvoiced sounds have a fine spectral structure expressed by a stochastic process, and the spectral structures of the simply added CSM and the speech signal are completely different. It will be done.

Therefore, to synthesize speech using CSM,
It is necessary to use some method to spread the line spectrum into a continuous spectrum. In other words
CSM speech synthesis can be thought of as generating a speech spectrum pattern from speech feature spectral patterns expressed by line spectra as shown in FIGS. 1 and 2.

In the present invention, the following method is used to perform the above-described spread spectrum in CSM speech synthesis.

That is, since voiced sounds have a clear pitch structure, the phase of each of the n sine waves specified as described above is reset every pitch period. This makes it possible to easily generate a spectral envelope and a fine pitch spectral structure.

Furthermore, by applying special time window processing to the above-mentioned phase reset waveform as detailed in the explanation of the embodiment, discontinuity in the synthesized waveform at the time of phase reset can be removed and continuity of the audio waveform can be ensured. ing.

Through the above implementation, the CSM line spectrum shown in Figure 2 is diffused as shown in Figure 3A, and changed to a spectrum having a spectral envelope and pitch fine structure, which is sufficiently practical for auditory purposes. Experimental results have shown that durable sound quality can be obtained.

For reference, FIG. 3B shows a CSM spectrum obtained by simple addition without performing the above-mentioned processing. As mentioned above, a waveform with such a spectrum simply sounds like a synthesized sound of sine waves, and the purpose of reproducing speech cannot be achieved.

The above is for voiced sounds, but in the case of unvoiced sounds, it is performed as follows. In other words, in the case of voiced sounds described above, the phase reset and special time window processing performed at each pitch synchronization are performed, and in the case of unvoiced sounds, the synchronization that occurs randomly as a stochastic process is performed instead of pitch synchronization. Using a pulse with a distribution width and a lower limit value set, the above-described process is performed every time this pulse is generated.

By using the above method, it is possible to perform CSM synthesis that is auditorily sufficient for practical use. Note that the above CSM synthesis is a synthesis method that does not use a filter, and therefore does not require consideration of stability on the synthesis side. Therefore, when using a communication method that transmits the information of m _i and ω _i to the synthesis side and reproduces the voice on the synthesis side, the line quality is relatively poor and errors may occur during transmission. In some cases, it may be possible to obtain better sound quality than a vocoder.

(Example) Next, the present invention will be explained in detail using examples.

For convenience of explanation, the present invention will be described using an analytical synthesis system that includes the present invention.

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

This embodiment consists of a transmitting side 1 and a receiving side 2.

The transmitting side 1 further includes an A/D converter 101, a Hamming window processor 102, and an autocorrelation coefficient measuring device 10.
3CSM analyzer 104, CSM quantizer 105, power correction quantizer 106, pitch extractor 107, voiced/unvoiced sound determiner 108, and multiplexer 109.

In addition, the receiving side 2 further includes a demultiplexer and decoder 201, an interpolator 202, a voiced/
Unvoiced sound switcher 203, period calculator 204, random number generator 205, n variable frequency oscillators with phase reset function 206-1, 206-2, ..., 20
6-n, n variable gain amplifiers 207-1, 20
7-2, .

Now, the operation of this embodiment is as follows. The audio waveform to be transmitted is fed via an input line 1000 to an A/D converter 101, where it is converted into digital data whose amplitude and time axis are quantized, and whose outputs are each passed through a Hamming window. The signal is supplied to the input sides of a processor 102, a pitch extractor 107, and a voiced/unvoiced sound determiner 108.

The digital data supplied to the Hamming window processor 102 is subjected to weight multiplication using a known Hamming window function for each predetermined frame, and is supplied to the autocorrelation coefficient measuring device 103 for each frame of data. Ru.

The autocorrelation coefficient measuring device 103 calculates the lowest N autocorrelation coefficients v _l (where l=
1, 2,...N).

In other words, data for one frame is X _t (where t
⁼ _{0 , 1 , .}

The set of v _l for each frame obtained in this way is supplied to the next CSM analyzer, and the set of v ₀ in this is supplied to the next CSM analyzer.
(that is, v ₀ =1/M _M-1 〓 ^t=0 X ² _t ) is supplied to the power correction quantizer 106 as power information in this frame.

Now, the autocorrelation coefficient v _l for each frame mentioned above
The CSM analyzer 104, which has been supplied with the set m _i , ω _i ( However, i=1, 2,...n)
and supplies this to the CSM quantizer 105.

The CSM quantizer 105 is a direct part constituting the present invention and will be explained in detail separately, but its outline is as follows.

The CSM quantizer 105 converts the set of values of m _i and ω _i into a=max {m ₁ , m ₂ , ..., m obtained from the set of amplitudes {m ₁ , m ₂ , ..., m _o } _o }, and outputs the a to the power quantizer 106 as correction data, and uses the a to calculate the collection {m ₁ , m ₂ , . . . m _o }. The quantization method includes normalization means, and the number of quantization bits is selected as an appropriate number of bits determined by taking into consideration the requirements for reproduction sound quality and the transmission capacity of the line. The CSM quantizer 105 quantizes the set of values of m _i and ω _i and then supplies it to the multiplexer 109 .

Furthermore, the power quantizer 106 that receives the aforementioned v ₀ and the normalization coefficient a quantizes the v ₀ with an appropriate roughness determined from the above-mentioned viewpoint, and then similarly supplies the same to the multiplexer 109 .

Further, the digital original audio signal is supplied from the A/D converter 101 as appropriately quantized data to the multiplexer 109, and similarly, the voiced/unvoiced sound determiner 108 also determines voiced/unvoiced sound based on the supplied digital data. and supplies it to the multiplexer 109 as a binary signal.

The multiplexer 10 supplied with the above signals
9 combines these signals into a form that can be easily separated on the receiving side and is suitable for transmission over a given transmission path, and transmits the combined signals to the receiving side 2 via a transmission path 1200.

Now, on the receiving side 2, the thus transmitted signals are demultiplexed and separated in the demultiplexer and decoder 201, so that
Each signal at the input side of the multiplexer 109 on the transmitting side 1 is restored.

Each signal thus restored is supplied to an interpolator 202 having a memory function, and after performing necessary interpolation, it is used as follows.

First, ω _i (ω ₁ to _{ω o} ) specifying the angular frequency of each of the n waves of the CSM is added to the frequency control input of the n variable frequency oscillators with phase reset function 206-1 to 206-n. and sets the output angular frequencies of these oscillators to specified angular frequencies ω ₁ to ω _o .

Also, the intensity (power amplitude) of each of the n waves of CSM
m ₁ to _{m o} specified as are supplied to the gain control terminals of the n variable gain amplifiers 207 to 1 to 207-n, thereby controlling the oscillation power of each frequency to a specified value. do.

The n outputs thus obtained are the countable combiner 2
After the countable combination is performed in step 08, the signal is supplied to the next multiplier 210.

Now, the demultiplexer and demultiplexer 201
The pitch period information outputted from the interpolator 202 including a memory performs interpolation as necessary, and is supplied to the voiced/unvoiced sound switch 203 as digital data representing the pitch period.

On the other hand, the random number generated by the random number generator 205 is
It is supplied to the pulse interval calculator 204, where it is converted so that the distribution width of the random number and its lower limit become a specific value, and is sent to the voiced/unvoiced sound switch 2 as a data string that determines the phase reset time interval for unvoiced sounds.
03's other input.

Further, a binary signal (V/U) that distinguishes between voiced and unvoiced sounds output from the demultiplexer and decoder 201 is supplied as a switching control signal to the above-mentioned switch 203, and in the case of voiced sounds, the switch 20
3 selects the digital data representing the aforementioned pitch period output from the interpolator 202 and supplies it to the window function generator 209.

Furthermore, if the binary signal (V/U) specifies an unvoiced sound, the switch 203 selects the data string side representing a random time interval generated in the stochastic process of the output of the period calculator 204. This is then supplied to the window function generator 209 in place of the above-mentioned example of digital data representing the pitch period.

Now, the window function generator 209 is for generating a window function that ensures the continuity of the audio waveform except for discontinuities that occur in the output waveform due to phase reset, and furthermore, the window function A related phase reset pulse is also generated.

As mentioned above, the window function generator 209 includes a switch 2.
03, a data string specifying the interval between the next phase reset pulses is input, and the window function generator 209 successively generates impulses having the time intervals specified by this data. , is supplied via line 2090 to the phase reset terminals of variable frequency oscillators with phase reset function 206-1 to 206-n, thereby resetting the phases of these oscillators. Also add this to line 2090
and is used as a timing signal for interpolating the angular frequency data ω _i and the intensity data m _i .

Now, the window function generator 209 generates a variable length window function w _(x) as shown below in synchronization with the generation of the above-mentioned phase reset pulse.

That is, if T is the value of the interval between phase reset pulses at that point specified by the input data, and x is the elapsed time since the previous phase reset pulse was generated, w _(x) = 0.5 + 0. .5cos(πx/T) However, a window function expressed as 0<xT is generated. This window function w _(x) is shown in FIG. 5A. The above-mentioned value of T represents the pitch period in the case of a voiced sound, and represents a variable that occurs in a stochastic process in the case of an unvoiced sound, so it changes over time. Therefore, this window function w _(x) has a variable length and is synchronized with the generation of the phase reset pulse described above in the relative time relationship shown in FIG. 5B (the start and end points of the window function (This almost coincides with the time point at which the phase reset pulse is generated.)

The window function thus generated is provided to multiplier 210 via line 2091. As a result, the multiplier 210 generates n sine waveforms whose phase is reset for each phase reset pulse synthesized by the summing synthesizer 208, and the above-mentioned window function generated in synchronization with each phase reset pulse. The product with w _(x) is obtained. The waveform obtained in this way is continuously converged to 0 by multiplying the window function w _(x) just before each sine wave is phase reset, and since each sine wave rises from 0 at the time of phase reset, the waveform Continuity is ensured, and thus discontinuity occurring in the phase reset waveform by multiplication by the window function w _(x) can be removed.

The output of the multiplier 210 from which discontinuities have been removed is supplied to the next multiplier 211, where it is weighted by the power information of each frame sent from the transmitting side 1, and is output from line 2000 as synthesized speech. be done.

As explained above, on the receiving side 2 of this embodiment, the CSM synthesis necessary for the above-mentioned speech synthesis is executed, and as a result, the reproduction of the original voice inputted to the transmitting side 1 is transmitted using the information on the transmission path 1200. This is done with relatively good sound quality despite volume compression and transmission errors.

The interpolation for each transmission data by the interpolator 202 described above is performed using various combinations (for example, only ω _i , or only ω _i , m _i , etc.) ), and it is also possible to use linear interpolation or interpolation using a higher-level function. Regarding interpolation for ω _i and m _i ,
It is advantageous to select interpolation points so that interpolated data can be obtained at each time point when the above-mentioned phase reset pulse occurs, and in order to update the values of ω _i and m _i at this timing, the phase A reset pulse is provided to interpolator 202 via line 2090.

In order to perform such interpolation, since the interpolated data is obtained after the required subsequent data arrives or occurs, it is necessary to reset the phase and set the frequency ω _i for the oscillator 206, and also adjust the intensity for the amplifier 207. Actual processing such as the setting of m _i will be executed with a necessary fixed time delay from the actual time. For this reason, interpolator 202 includes a memory for storing necessary information until a necessary point in time.

Next, variable frequency oscillator 2 with phase reset function
A circuit example of 06 is shown in FIG. Frequency control terminal 2
The voltage applied to 061 causes constant current power supply 206
2 and 2063. The charging/discharging current value for the capacitor 2064 is controlled, thereby making the oscillation frequency variable. The oscillation voltage waveform at point v is a triangular waveform that linearly rises and falls between +Vr and -Vr of the reference voltage. When an impulse is applied to the phase reset terminal 2065, the point v is momentarily grounded and forcibly pulled back to 0 potential, and oscillation is restarted from there to perform a phase reset. This triangular wave oscillation output at point v is input to a sine wave converter 2066, converted to a sine wave, outputted from a terminal 2067, and used as the output of the oscillator 206. The sine wave converter 2066 can be easily realized, for example, by reading out a sine function value stored in a ROM using an input waveform.

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

Next, a circuit example of the variable gain amplifier 207 is shown in FIG. By applying a signal to be amplified to a terminal 2071 and a control signal to a terminal 2072, the amount of negative feedback is controlled, and an output having a controlled amplitude is obtained at an output terminal 2073.

In addition to this, it can also be realized by using an analog multiplier, or by using an analog waveform input as the reference voltage of the D/A converter and using control information expressed in digital quantities as the digital input. It can also be easily realized by a method.

Next, an example of a circuit of the random number generator 205 is shown in FIG. The 15-stage left register 2051 and one middle adder 2052 have a period of 2 ¹⁵ -1.
Generate the next M series of pseudo-random numbers. By applying a shift pulse to clock terminal 2053 at the required time, the next random value is obtained.

Next, the block diagram of the period calculator 204 is shown in FIG. 9A.
Shown below. This is suitable for using the random numbers uniformly distributed in the range of 0 to ²¹⁵ -1 output from the random number generator 205 mentioned above as random numbers for specifying the time interval of the phase reset pulse during unvoiced speech. It converts into a distribution and consists of a constant multiplier 2041 and a constant adder 2042. By this, Figure 9B
As shown in the figure, the random number distribution width D and the lower limit L can be set to appropriate values.

Next, an embodiment of the window function generator 209 is shown in FIG. This includes a register 2091, a presettable down counter 2092, and a counter 209.
3. Contains read-only memory (ROM) 2094.

Data T specifying the phase reset pulse interval supplied from the switch 203 is stored in the register 209.
It is ranked as 1. The down counter 2092 is a counter that counts the high speed clock CLK of a fixed period. First, the content T of the register 2091 is preset, and this is down counted using the clock CLK. When the contents of the counter 2092 reach 0, a pulse is generated from the output terminal, thereby presetting the contents of the register 2091 again and starting counting down this value. Thus, a pulse train having a period proportional to T (for example, T/k) is generated at the output 2092-1 of the down counter 2092. This pulse train is added as a clock to counter 2093. The count output 2093-1 of the counter 2093, which is incremented by this clock, is
The data of the window function w _(x) , which is added to the ROM 2094 as an address designation signal and written therein, is sequentially read out and output to the line 2091.
When the contents of the counter 2093 reach k, ROM2
The last data of the window function w _(x) of 094 is read out,
At the same time, counter 2093 is reset and outputs a reset pulse on line 2090. This reset pulse is applied to the oscillators 206-1 to 206
-n phase reset terminal and the aforementioned phase reset pulse supplied to the interpolator 202, and is also used to set the next input data in the register 2091. Also ROM
Data of the window function w _(l) previously stored as k samples in line 2094 is output to line 2091 and supplied to multiplier 210 . In this way, phase reset pulses whose inter-pulse intervals have successively specified values and a variable-length window function w _(x) synchronized with these pulses as shown in FIG. 5B are generated.

Next, we will explain CSM analysis.

As mentioned above, CSM analysis uses, for each analysis frame, N low-order tap values of sample autocorrelation coefficients calculated directly from the speech waveform to be represented;
The frequency ω _i of each sine wave to be synthesized and its intensity (power amplitude) m _i are determined so that the N low-order tap values of the synthesized wave (sum of n sine waves) match. That's true.

Now, if the autocorrelation coefficient of tap l of the composite wave is γ _l , then as mentioned above, γ _l = _o 〓 ⁱ⁼¹ m _i coslω _i .

On the other hand, the sample of the audio waveform to be represented X _t
Therefore, the sample autocorrelation coefficient v _l of tap l in a certain frame is v _l =1/M _M-1 〓 ^t=l X _t X _tl (1).

From this, γ _l =v _l ...(2) l=0, 1, 2,...N However, if N=2n-1, the following matrix expression is obtained.

However, the above equation cannot be solved by simple matrix operations because ω _i and m _i are unknown. Therefore, let ω _i = cos ^-1 x _i ...(4) and perform the substitution coslω _i = cosl cos ^-1 x _i ≡T _l (x _i ) ...(5). This Tl(x) is a Tchebycheff polynomial. When this substitution is performed, equation (3) is converted as follows.

However, in general, x ^l is T ₀ (x), T ₁ (x)...T _l
It can be expressed as a linear combination of (x).

That is, x ^l = _l 〓 ^j=0 S ^(l) _j T _j (x) ... (7) where S ^(l) _j is the inverse Tchebycheff coefficient.

Using this S ^(l) _j , the sample autocorrelation coefficient v _j
Define the linear combination A _l as shown below.

A _l = _l 〓 ^j=0 S ^(l) _j v _j ...(8) However, l=0, 1, 2, ..., 2n-1 In this way, on the left and right sides of equation (6), respectively
By using the relationships of equations (7) and (8), the following relational expression is established.

Now, here, _the nth degree polynomial Pn( _x )≡ _o 〓 _k ⁼⁰ p ⁽ⁿ⁾ _k x ^k ＝ _o 〓 ⁱ⁼¹ (x− x _i ), and using this Pn(x), create _o 〓 ⁱ⁼¹ m _i Pn(x _i )x ^l _i , which is similar to the left side of equation (9), and consider this. View. It is clear that the above equation is 0, but it can be further rewritten as follows.

0 = _o 〓 ⁱ⁼¹ m _i Pn(x _i ) x ^l _i = _o 〓 ⁱ⁼¹ m _io 〓 ^k=0 p ⁽ⁿ⁾ _k x ^k+l _i = _o 〓 ^k=0 p ⁽ⁿ⁾ _ko 〓 ⁱ⁼¹ m _i x ^k+l _i = _o 〓 ^k=0 p ⁽ⁿ⁾ _k A _k+l From the above, the following formula can be obtained with l=0, 1, 2,...n.

However, since p ⁽ⁿ⁾ _o = 1 holds true. The matrix defined by A _i on the left side is generally called the Hankcl matrix. As described above, each A _i is given by equation (8) from the sample autocorrelation coefficient v _j of the speech waveform to be expressed and is known.

Therefore, by solving equation (10), p ⁽ⁿ⁾ ₀ , p ⁽ⁿ⁾ ₁ , ...
The value of p ⁽ⁿ⁾ _o-1 can be found.

Once each p ⁽ⁿ⁾ _i is found, as a solution to the n-dimensional equation p _o (x) = x ⁿ p ⁽ⁿ⁾ _o-1 x ^n-1 +... p ⁽ⁿ⁾ ₀ = 0, {x ₁ , x ₂ ,..., x _o } are found.

From this, each CSM frequency ω _i is obtained from ω _i =cos ⁻¹ x _i in equation (4), and the CSM intensity m _i is obtained using the following equation derived from equation (9).

Note that the matrix on the left side of the above equation is generally Vander
This is called the Monde (Juare del Monde) matrix.

To summarize the above, the analysis algorithm for CSM analysis is as follows.

(1) Calculate the sample autocorrelation coefficient v _l = 1/M _M-1 〓 〓 _t=l X _t X _tl (2) Define A _l using the inverse Tievisiev coefficient.

A _l = _l 〓 ^j=0 S ^(l) _j v _j (3) Solve the Hankel matrix equation by A _l to find p ⁽ⁿ⁾ _i (4) p ⁽ⁿ⁾ Find n x _t by solving an n-dimensional algebraic equation with _i as a coefficient.

p _o (x)≡x ⁿ +p ⁽ⁿ⁾ _o-1 x ^n-1 +p ⁽ⁿ⁾ _o-2 x ^n-2 +… +p ⁽ⁿ⁾ ₁ x+p ₀ =0 (5) Perform inverse cos transformation Find the CSM angular frequency {ω _i }.

ω _i =cos ^-1 x _i (6) Solve the Van del Monde matrix equation to find the CSM intensity {m _i }.

By performing each of the above steps, CSM
The angular frequencies {ω ₁ , ω ₂ ...ω _o } and the intensities of each wave {m ₁ , m ₂ , ...m _o } can be determined.

Note that, as an efficient method for solving the above-mentioned Hankel matrix equation, a method is known in which initial conditions are given and solutions are sequentially obtained.

Furthermore, since it has been proven that the above nth-order algebraic equation has only real roots, the roots can be found using the Newton-Raphson method or the like.

Further, as an efficient method for solving the above-mentioned Vander Monde matrix equation, it is possible to use a method of sequentially obtaining solutions by performing triangular matrix formation. The above analysis method is based on the paper “Speech spectrum analysis using a composite sine wave model” by Mr. Sagayama et al.
Journal of the Institute of Electronics and Communication Engineers '81/2 Vol.J64-ANo.
Details are given on 2P.105-112.

Finally, it is a direct part constituting the present invention.
CSM quantizer 105, power correction quantizer 106
will be explained in detail using the drawings. Figure 11 is
CSM quantizer 105, power correction quantizer 106
FIG. 2 is a block diagram for explaining in detail.

m _i , ω _i (where i
= 1, 2, ...n) is temporary memory (1), 105
1. Temporary memory (1), 1051 is the above
m _i is output to the normalization coefficient searcher 1052 and the CSM intensity normalizer 1053. Normalization coefficient searcher 1
052 searches for the normalization coefficient a and the maximum frequency number according to the following procedure.

(1) Set the initial state a=m _i and I=1.

(2) Investigate the magnitude relationship between a and m ₂ . If a
If m ₂ , execute (4) next. If a<m ₂ , perform (3) next.

(3) Set a=m ₂ and I=2.

(4) Investigate the magnitude relationship between a and m ₃ and perform the same processing as in (2) above.

(5) From here on, perform the same processing as in (4) up to m ₄ . . . m _N.

The normalization coefficient searcher 1052 applies the searched a to the power corrector 1061 and the CSM intensity normalizer 105.
3, and also convert the above I to the CSM intensity quantizer 105
Output to 4. The CSM intensity normalizer 1053 uses the normalization coefficient a to convert the m _i supplied from the temporary memory (1) 1051 into m′ _i =m _i /a (where i
=1, 2,...n). Furthermore, the CSM intensity normalizer 1053 calculates the square root √′ _i of the calculated m′ _i
(i=1, 2, . . . n) is obtained and output to the CSM intensity quantizer 1054. CSM intensity quantizer 105
4 is the maximum frequency number I supplied from the normalization coefficient searcher 1052 and the aforementioned √′ _i (i=1, 2, . . .
. Next, the format of quantization will be explained with reference to FIG. FIG. 12a shows the bit allocation for quantizing the CSM _intensities m ₁ , m ₂ . The highest CSM strength is designated in correspondence with the number I. Here, when the strongest CSM has a number I of 1, "0" is given to the bit shown at the left end in the figure, as shown in FIG. 12 b-a, and I is 2, 3, 4,
In the case of 5, "1" is given to the bit shown at the left end in the figure, as shown in FIG. 12 b-b-e.

By the way, when we investigate the distribution of CSM strength, we find that the strongest
Very often m ₁ has CSM intensity. Figure 13 is a distribution diagram when the CSM intensities m ₁ , m ₂ , ... m ₅ obtained as a result of the 9th order CSM analysis (n = 5) are normalized by calculating the normalization coefficient a. ``Number of frames'' is the number of frames at which the intensity is the strongest. Note that the total number of analysis frames is 6963. That is, the ratio of m ₁ having the strongest CSM strength is 5895/6963 = 0.847, and as shown in Figure 12b, when m ₁ is the strongest (I = 1), I
The configuration is such that the specification can be made using the least number of bits.

Note that the strongest CSM strength itself is normalized, so it is always 1.0 and does not require information transmission.

Returning to Figure 11 again, the quantized
CSM intensity parameter is temporary memory (2), 1056
Output to. The CSM frequency quantizer 1055
A set of ω _i (where i = 1, 2,...n) specifying the angular frequencies of n sine waves of CSM is temporarily stored in
Each ω _i supplied from 051 and investigated in advance
Linear quantization is performed in consideration of the distribution range of , and the quantized data is output to temporary memory (2), 1056.
A temporary memory (2) 1056 sends the quantized CSM intensity and CSM angular frequency data to multiplexer 1.
Output to 09. The power corrector 1061 multiplies the power data supplied from the autocorrelation coefficient measuring device 103 by the coefficient a supplied from the normalization coefficient search device and outputs the result to the power quantizer 1062. The power quantizer 1062 raises the result to the 1/2 power and converts it into amplitude information, then performs nonlinear quantization, such as that used in μ255 PCM, and outputs it to the multiplexer 109.

Note that the denormalization on the synthesis side is automatically performed by the multiplier 211.

(Effects of the Invention) As described above, the present invention has the effect of increasing the efficiency of quantization by quantizing the CSM intensity parameters in consideration of the relationship between the CSM intensity parameters.

[Brief explanation of drawings]

Figure 1 shows an example of a voice feature vector pattern based on CSM parameters, and Figure 2 shows a CSM line spectrum and a pattern obtained from the same voice sample.
Diagram showing an example of correspondence with LPC spectrum envelope, 3rd
Figure A is a diagram showing the spectral envelope of the diffused CSM and the fine structure of the pitch, Figure 3B is a diagram showing the CSM spectrum obtained by simple addition, and Figure 4 is an implementation of the analysis and synthesis system including the present invention. A block diagram showing an example, FIG. 5A is a diagram showing the functional form of a variable length window function,
FIG. 5B is a diagram showing the relative time relationship between the variable length window function and the phase reset pulse, FIG. 6 is a diagram showing an example of a circuit of a variable frequency oscillator with a phase reset function, and FIG. 7 is a diagram showing a variable gain amplifier. FIG. 8 is a diagram showing an example of a circuit of a random number generator, FIG. 9A is a block diagram of a period calculator, and FIG. 9B is a distribution of random numbers output from the period calculator. and FIG. 10 are block diagrams showing an example of a variable length window generator, FIG. 11 is a block diagram for explaining in detail the direct parts constituting the present invention, and FIG. 12 is a block diagram showing an example of a variable length window generator. FIG. 13, which is a diagram showing an example of the format, is a diagram showing an example of the distribution of CSM intensity. In the figure, 1...sending side, 2... receiving side, 1
01... A/D converter, 102... Hamming window processor, 103... Autocorrelation coefficient measuring device, 104...
...CSM analyzer, 105 ...CSM quantizer, 10
6...Power quantizer, 107...Pitch extractor,
108... Voiced/unvoiced sound determiner, 109... Multiplexer, 201... Demultiplexer and decoder, 202... Interpolator, 203... Voiced/unvoiced sound switcher, 204... Period calculator, 20
5... Random number generator, 206-1 to 206-n...
Variable frequency oscillator with phase reset function, 207-
1 to 207-n...variable gain amplifier, 208...
Addition synthesizer, 209...Variable length window function generator, 2
10, 211... Multiplier, 1051... Temporary memory (1), 1052... Normalization coefficient searcher, 1053
...CSM intensity normalizer, 1054...CSM intensity quantizer, 1055...CSM frequency quantizer,
1056...temporary memory (2), 1061...power corrector, 1062...power quantizer.

Claims (1)

[Claims] 1. A CSM that expresses the spectral envelope of speech as a set of a predetermined number n of sine waves with free frequencies and amplitudes.
In parameter quantization, the set of amplitudes {m ₁ ,
A CSM parameter quantization method characterized by using means for normalizing m ₂ , ..., m _o } by a normalization coefficient a expressed as a=max{m ₁ , m ₂ , ..., m _o }. 2 CSM that expresses the spectral envelope of speech as a set of sine waves with a predetermined number n of free frequencies and amplitudes.
For expressing parameters with a small number of bits
In the CSM parameter quantizer, from the set of amplitudes of the sine wave {m ₁ , m ₂ , ..., m _o }, the maximum value a=max of the set {m ₁ , m ₂ , ..., m _o } is calculated. a means for searching and a maximum value a searched by said means;
and means for normalizing the set of amplitudes by.