JPH0644199B2 - Variable-length frame speech analysis / synthesis method - Google Patents

Description

Description: TECHNICAL FIELD The present invention relates to a variable length frame voice analysis / synthesis method, and more particularly, to a variable length voice signal for relatively long voice signals of about 10 seconds, which is optimum for variable length frame processing as a whole. The present invention relates to a frame voice analysis / synthesis method.

(Prior Art) For analyzing and synthesizing a voice signal in units of, for example, about 10 seconds, various fields of application such as voice mail, public address, etc. can be considered. When the variable length frame processing is applied to the analysis and synthesis of such a voice signal to achieve the compression of the amount of information or the like, it is generally performed as follows.

That is, for example, a section of 10 seconds is divided into 50 equal parts and 200
A method is used in which sections of about mSEC are created and variable-length frame processing is performed independently for each section.

However, according to this method, optimal variable-length frame processing can be performed for each section of 200 mSEC, but if 10 seconds is viewed as a whole, optimization is not necessarily achieved yet. It has drawbacks.

On the other hand, it is theoretically possible to perform optimization by variable length frame processing using dynamic programming (DP) with the entire 10 seconds as a unit.
In this case, the amount of calculation for performing DP becomes enormous, and the analysis and synthesis apparatus has a large delay, which causes a problem such as lack of resilience against a transmission line error, which is not realistic.

(Object of the Invention) It is an object of the present invention to optimize a voice signal in a relatively long unit of about 10 seconds by variable length frame processing as a whole and realize a realistic variable that enables more efficient compression of information amount. It is to provide a long frame speech analysis and synthesis method.

(Structure of the Invention) According to the method of the present invention, a voice analysis unit that periodically analyzes an input voice signal at a predetermined fixed analysis period to extract a feature parameter vector, and a plurality of predetermined continuous phases. The representative for each section obtained by selecting an arbitrary number of representative parameter vectors from the characteristic parameter vectors in each section for each section consisting of the analysis cycle and performing each piecewise optimal function approximation A piecewise optimal function approximating means for calculating the configuration of the parameter vector and the residual strain due to the optimal approximation in each case in this case, and the piecewise optimal function in a large section consisting of a plurality of predetermined consecutive sections By comparing the residual strains calculated by the approximating means for each of the sections, the configuration of the representative parameter vector of the section having the largest residual strain can be determined by using more representative patterns. Mutual optimum frame selecting means for selecting a predetermined number of all the representative parameter vectors that optimally approximates the large section by repeating the processing step of replacing with the configuration of the representative parameter vector including the parameter vector. Have.

(Example) Next, this invention is demonstrated in detail with reference to drawings. FIG. 1 is a block diagram showing an embodiment of the present invention.

The present embodiment comprises a voice analysis side 1 and a voice synthesis side 2.

The analysis side 1 is further provided with a low pass filter and an A / D converter (L
PF & A / D) 101, window function processor 102, LSP analyzer 103, piecewise optimal function approximator 104, total optimal frame selector 105, quantizer 106, source information analyzer 107, coder 108, and memory 109. In addition, the synthesis side 2 includes a memory 201, a decoder 202, a pulse oscillator 203, a noise generator 204, and a V / UV switch 2
05, power controller 206, LSP synthesis filter 207,
D / A converter and low-pass filter '(D / A & LPF) 2
08 and an interpolator 209 are included.

The operation of this embodiment is as follows.

The audio signal input from the line 1000 is sampled at a sampling frequency of 8 kHz and quantized in the low pass filter and the A / D converter (LPF & A / D) 101 after the frequency band is limited to, for example, 3.4 kHz. Data is converted into digital data, and the data thus obtained is supplied to the window function processor 102.

The window function processor 102 temporarily stores one block (for example, 240 samples) of the supplied data, applies weight multiplication by a predetermined window function to this, and outputs the data of the processing result to the LSP analyzer. 103 and the sound source information analyzer 107. Such processing by the window function processor 102 is repeated, for example, in a cycle of every 10 mSEC. Therefore, the LSP analyzer 103 and the sound source information analyzer 107 are supplied with the data subjected to the window function processing for one block at a cycle of 10 mSEC.

The LSP analyzer 103 performs LSP (line spectrum pair) analysis by a known method using the supplied data for one block, and determines the LSP parameter vector. This LSP parameter vector is an S-dimensional vector having S (even number) components = (P ₁ , P ₂ , ..., P _s ), and each of these components P _{1 to} P _s is a speech for one block. It is data obtained by extracting the information on the form of the vocal tract when uttering a voice in the form of a set of resonance frequencies. As mentioned above, the generation of such an LSP parameter vector is 10 mSEC.
Each of the components of the vector thus obtained is supplied to the next piecewise optimal function approximator 104 every 10 mSEC of each basic analysis period (hereinafter referred to as a basic frame).

Now, the piecewise optimum function approximator 104 collectively handles K parameter vectors which are successively supplied one after another.

That is, each basic frame is represented by a parameter vector belonging to this basic frame, respectively, but this continuous K basic frames are combined into one, and this is divided into one section. The following piecewise optimal function approximation processing is performed. Here, a case where a rectangle approximation is performed as a function used for this piecewise optimal function approximation will be described. It is also assumed that the number of basic frames in one section is 20 (that is, K = 20), and thus the time length of one section is 200 mSEC.

The processing by the piecewise optimum function approximator 104 is as follows.

That is, i is selected from the 20 basic frames in one section.
(I = 1, 2, ..., 20) representative frames are selected, and using the parameter vector belonging to this representative frame,
Parameter vectors belonging to other basic frames in this section are also represented (approximated), and thereby a rectangle is approximated. In this way, the above-mentioned i representative frames (representative vectors) are selected so that the distortion due to this approximation is minimized. At the same time, the minimum value E _{i of the} distortion that can be achieved by approximation with i representative frames at this time is also obtained.

The distortion due to the rectangular approximation in this case is calculated as follows.

For example, supposing that i = 2 and performing the rectangle approximation using the parameter vectors of the two representative frames, the fourth basic frame is selected as the first representative frame, and the parameters belonging to this representative frame are selected. The vector ⁽⁴⁾ is used as the first representative vector to approximate the nine sections from the first basic frame to the ninth basic frame, and then as the second representative frame, the first representative frame is used.
The third basic frame was selected, and the 11th section from the remaining 10th basic frame to the 20th basic frame was approximated using the parameter vector ⁽¹³⁾ belonging to this typical frame as the second typical vector. The distortion due to the rectangular approximation in this case is obtained as follows.

However, W _l (l = 1, 2, ... S) is a predetermined weighting factor for correcting that the influence on the distortion is different due to the difference in the spectral position of each component of the parameter vector.

For example, the optimal rectangle approximation when i = 2 is given is to determine the configuration of the representative parameter vector including the two representative vectors so that the distortion obtained in this way is minimized, that is, This is to determine a representative frame and two basic frame sections that the parameter vector belonging to each representative frame should represent. At the same time, the value of the residual strain reached and obtained by the configuration of the determined representative parameter vector is also obtained as data.

The operations described above are dynamic programming (D
P) can be easily used.

Now, the minimum distortion (residual distortion) that can be reached when the section formed by a basic frames from the beginning of the partition is approximated by b representative frames including the final basic frame (a-th basic frame) Is defined as G (b, a), b =
1, that is, when the number of representative vectors is 1, the configuration of the representative parameter vector ^(a) represents the section of the basic frame from the first to the a-th, and therefore the residual distortion G
(1, a) is Therefore, it is uniquely determined for a = 1 to 20.

However, d _{k, a} is distortion when the parameter vector of the k-th basic frame is represented by the parameter vector (representative vector) of the a-th basic frame (representative frame).

Next, when approximating a section from the x-th basic frame to the y-th basic frame (provided that y> x) as representative frames (that is, this section is represented by a representative vector P
minimum distortion of the D _x obtained reached when) approximated by using ^(x) and a _^(y), D represented by _{y x, y} is Can be easily obtained at. This D _{x, y} is the section represented by the representative vector ^(x) , which is initially the Xth basic frame, the Xth and X + 1th basic frames, and then the Xth basic frame. , The X + 1th and X + 2th basic frame sections, and so on. The distortions when the remaining section is represented by ^(y) are obtained, and the minimum distortion among them is selected. It is a thing.

Now, using G (1, a) and D _{x, y} obtained above,
G (2, a) when the number of representative frames is 2 can be easily obtained as follows.

That is, the distortion when the (a-1) th basic frame is selected as the first representative frame (of course, the second representative frame is the ath basic frame) is obviously the first distortion on the right side of the above equation. G (1, a-1) shown in FIG.

Next, the first representative frame is moved forward by one, and the a-2
Minimum value of the resulting strain reached if chosen th, G shown in the second upper right side of equation (1, a-2) + D a-2, a
Becomes That is, G (1, a-2) represents the distortion when the section from the 1st to the a-2nd section is represented by the representative frame (representative vector ^(a-2)) of the ^a-2nd basic frame. The symbol D _a−2, a represents the section from the a−2nd section to the ath section, which is a representative frame (representative vector P based on the a−2nd and ath basic frames at both ends thereof).
It represents the minimum distortion that can be reached in the case represented by ^(a-2) and ^(a) , and it is clear that the minimum distortion in this case is the sum of both (a-2nd basic frame). Since the distortion of 0 becomes 0, the section representative of G (1, a-2) and D _a-2 may be doubled with this basic frame.) Similarly, the first representative frame is forwarded one by one. The next step is to find the minimum distortion that can be reached at this time.

When the first representative frame is advanced to the front and selected as the first basic frame, the minimum distortion that can be reached is G (1,1) + D _{1, a} shown at the bottom of the right side of the above equation. It is clear that Of course, G (1,1) = 0.

From the above, the minimum distortion that can be reached when the first to a-th sections are represented by two representative frames (one of which is the a-th basic frame as a representative frame) The (residual strain) G (2, a) is obtained by selecting the minimum strain among all the strains obtained above, and this is shown by the above equation. Thus residual strain G
(2, a) is obtained, and together with this, the configuration of the representative parameter vector when giving this residual distortion, that is, the two representative vectors ^(X) , ^(a) and the section width B _x , which each representative vector represents, a−B _x is determined. In this way, the configurations of G (2, a) and the corresponding representative parameter vector are also obtained for a = 1 to 20.

G (3, a) when the number of representative frames is increased by one
Is calculated by using G (2, a) obtained above and D _{x, y.}
It is obtained by the following equation in exactly the same manner as (2, a).

Thus, when G (3, a) is obtained for a = 3 to 20, G (4,4), which is the number of representative frames, is increased by one.
a) is obtained in the same manner for a = 4 to 20.

In this way, by using DP, the number of representative frames is increased next and the minimum value of the distortion that can be reached is obtained, whereby any i, j (where j = 1, 2, ... 20:
It is possible to determine G (i, j) for ij) and the configuration of the representative parameter vector at that time, that is, all the sets of i representative vectors and the interval width of the basic frame represented by each representative vector.

G (i, 20) thus obtained is the basic frame 2
It represents the minimum value of the distortion that can be reached when approximating one section consisting of 0 pieces by i pieces of representative frames. As described above, one of the i pieces of representative frames is the 20th frame. There is a constraint that the second basic frame is used.

Excluding this constraint, the minimum value of the distortion (residual distortion) that can be reached when optimal i approximation is performed by selecting any i representative frames in one section consisting of 20 basic frames is E _i.
Then, E _i is obtained as follows using (i, j) obtained above.

Now, the distortion when approximating the section from the kth basic frame to the 20th basic frame using the kth basic frame as a representative frame (that is, using ^(k)) is represented by D _k When, As easily requested.

With this D _k . For example, E ₁ is And any i (where i = 1, 2, ... 2)
E _i for 0) is Can be asked as Thus, when E _i is obtained,
As described above, the configuration of the representative parameter vector that gives this residual distortion, that is, the set of i representative vectors and the set of i numbers representing the section width of the basic frame represented by each of these i representative vectors, , Are determined.

Now, the piecewise optimal function approximator 104 uses the basic frame 2
For each division consisting of 0 (time length 200 mSEC),
The above calculation is performed to determine the configuration of the representative parameter vector including any i representative vectors and the residual distortion in the case of adopting this configuration. That is, the piecewise optimum function approximator 104 performs the above-mentioned calculation for each section to obtain E _i
(However, i = 1 to 20) All {E} and each E
_{The set of i} representative vectors {} corresponding to i and the set of widths B of the i sections representative of these representative vectors {B} are determined, and these data are sent to the next total optimum frame selector 105. Supply.

The total optimum frame selector 105 is a processor that performs optimum frame selection processing for a large section (for example, a time length of 10 SEC) that is obtained by collecting, for example, one section consisting of 20 basic frames described above, for example, 50 sections. .

The selector 105 has a memory capable of storing the data {E}, {} and {B} for each section supplied as described above for at least one large section (for 50 sections). When the supply of the above-mentioned data for one large section is completed, the comprehensive optimum frame selection processing as described below is started using these data.

Now, {E} for each section is stored in the work area of the memory of the selector 105 as a matrix table as shown in FIG. However, the suffix (j) above ▲ E ^(j) _i ▼ indicates that this is the distortion of the j-th section, and the suffix i below indicates that this section has i representative frames (i This represents the minimum value of the distortion (residual distortion) that can be reached when the optimum rectangle approximation is performed using the representative vector). Therefore, for the same value of (j), the larger i becomes (in the same column ^), ▲ ^(j)
It is clear that _i ▼ becomes small.

In the present embodiment, as described above, the piecewise optimal function approximator 104 performs the piecewise optimal function approximation for each segment.
Although all the residual strains E _i of the optimum approximation in the case of selecting arbitrary i pieces as the representative frame are obtained and the data related thereto are supplied, from among these as the representative parameter vector configuration for each section, It has not decided on how to choose that one.

On the other hand, the total optimum frame selector 105 calculates the total number of representative frames (representative vectors) representing each large section as
It has a function of selecting and deciding the representative frame configuration in this large section so that the total distortion in this large section is balanced and small in each section as much as possible when fixed to N in advance.

As the total number N of representative frames (representative vectors) in this large section, the average value of the number of representative frames in each section is
For example, if there are five, N = 5 × 50 = 250.

The overall optimum frame selection process is performed according to the following algorithm using the table shown in FIG.

(A0): First, N = 50 is set assuming that one representative frame is selected for each section. In this case, the first row ▲ E ⁽ⁱ⁾ ₁ ▼ of the above matrix
The content of (where j = 1 to 50) represents the distortion of each section. Of course, optimal approximation is performed in each section when one representative frame (representative vector) is used.

(A1): First row ▲ E ⁽ⁱ⁾ ₁ ▼ (j = 1 to ₁ ) described above
Compare the contents of 50) and take the maximum value among them ▲ E
^(I) Select ₁ ▼. That is, the section that produces the maximum distortion when the representative frame configuration of the large section as described above is taken is selected.

(A2): Let the maximum value selected above be ▲ E ^(m) ₁ ▼. In other words, if a distortion belonging to the m-th column (m-th section) is selected, that column has all ▲ E.
^(M) Shift _i ▼ upward by one. That is,
▲ E ^(m) _i ▼ is replaced with ▲ E ^(m) _{i + 1} ▼ (however, i = 1, 2, ..., 19).

Obviously, ▲ E ^(j) ₂₀ ▼ is always 0 regardless of the value of (j), so performing such an upward shift will add only 0 to the tail of the shifted column. It will be.

The processing by the above (A1) and (A2) finds a section that generates the maximum distortion in the representative frame configuration of this large section, increases only the number of representative frames in this section, and The process is to update to an approximation.

(A3): The value of N is increased by 1 in response to the increase in the number of representative frames in the large section by 1 due to the processing in (A2). As a result, when the value of N reaches the total number of representative frames 250 in a predetermined large section, the following (A
If the process of 4) is not performed, and if this is not reached, the process returns to the process of (A1) again, finds the section in which the maximum distortion occurs in the representative frame configuration of this large section, and sets only the number of representative frames of this section to 1 The process of increasing the number and updating the optimum approximation is repeated.

(A4): As a result of the above processing, the residual distortion for each section when this large section is optimally approximated using 250 representative frames is shown in the first row of the above-mentioned ⁽ E ^{) (j)} _i table. It is shown. Also, each j in this table
Assuming that the number of 0s included in the tail of the column for (j = 1 to 50) is M _j , this number M is set for the reason described above.
_j represents the number of representative frames for each j-th segment when approximation is performed for this large interval.

For the above reasons, the configuration of the number of representative vectors in each section, the value of the component of each representative vector, and the section width represented by each representative vector when the large section is optimally approximated using 250 representative frames Since it is determined, the selector 105 reads the value of the component of the representative vector of each of these sections and the value of each section width (the number of basic frames) represented by this representative vector from the stored memory area, The values of the components of each representative vector are supplied to the quantizer 106 one after another, and the value of the interval width (the number of basic frames) represented by each representative vector is designated as the repetition of each basic frame of this vector. It is supplied to the coder 108 as a number.

The quantizer 106 requantizes the supplied components of each representative vector with the roughness determined by the required transmission path and transmission quality, and then supplies them to the coder 108.

On the other hand, the sound source information analyzer 107 extracts pitch information, voiced sound / unvoiced sound information (V / UV), sound volume information, etc. from the sound data supplied from the window function processor 102 using a known means, and extracts these. Information is provided to the coder 108.

The coder 108 composite-encodes each information supplied as described above into a form suitable for transmission, and supplies it to the memory 109.

The memory 109 temporarily stores the supplied data in order to store and transmit the voice, and sends it to the synthesizing side 2 in accordance with the empty state of the transmission path 1200.

Now, on the synthesis side 2, the data transmitted via the transmission path 1200 is once stored in the memory 201, and is flown out of the memory 201 as needed for voice generation.
The sound is reproduced by the following processing.

That is, the data read from the memory 201 is decoded by the decoder 202, whereby the analysis side 1
The data supplied to the input side of the coder 108 is restored.

The pitch information from the sound source information analyzer 107 in the restored data is supplied to the pulse oscillator 203, and the oscillation frequency is controlled so as to become the fundamental frequency of the pitch. In addition, voiced / unvoiced information (V / UV) is V / UV switcher 2.
It is supplied as a switching control signal of 05, and this is a voiced sound (V)
When designating, the switching unit 205 sets the pulse oscillator 203
When the unvoiced sound (UV) is designated by selecting the output side of, the switching unit 205 controls so as to select the output side of the noise generator 204.

Furthermore, the volume information is supplied as the control information of the power controller 206, whereby the power controller 206 variably amplifies the selected output of the switch 205 and controls it so that the output has a specified power amount. To do.

The output of the power controller 206 thus obtained is supplied to the LSP synthesis filter 207 as a sound source signal for driving the LSP synthesis filter.

On the other hand, the information on each component of each representative vector and each section width represented by each representative vector decoded by the decoder 202 is supplied to the LSP synthesis filter 207 via the interpolator 209.

The interpolator 209 calculates each component of each supplied representative vector as
Interpolation for the rectangular approximation is performed by repeatedly reproducing for each basic frame by the section width represented by each of these representative vectors, each component of the LSP parameter vector for each basic frame is generated, and this is used for the LSP synthesis filter 207. Supply to.

The LSP synthesis filter 207 determines the LS thus supplied.
A sound signal is synthesized by a known means using each component of the P parameter vector and the sound source signal, and the synthesized sound signal is output to the D / A converter and the low pass filter 208.

Thus, the synthesized digital voice signal is converted into an analog voice signal, unnecessary frequency components are removed, and the resultant digital voice signal is output from the output line 2000.

As described above, according to the present embodiment, the transmitted voice information is not only the optimum approximation for the number of representative frames assigned to each section of about 200 mSEC, but also 50 of these sections. 10S consisting of about 10 pieces
Even in the large section up to EC, the distortion is well balanced in the optimum approximation.

That is, in the section where the voice information changes drastically,
By using a larger number of representative frames, a higher degree of optimal approximation is performed, while a rough approximation is performed by using a small number of representative frames for sections with little change in audio information, and the total amount of information to be transmitted is calculated. Optimal approximation is performed so as to balance and reduce the distortion for each section as much as possible when the value is limited to a fixed value. As a result, it is possible to more faithfully follow the fluctuations in the audio information amount for each segment in a large section, as compared to the case where the number of representative frames in each segment is fixed, and thus more efficient compression of the information amount. Alternatively, a higher quality audio reproduction is achieved.

Moreover, for example, if a large section of 10 seconds including 1000 basic analysis frames is directly subjected to the optimum approximation by the method described in the piecewise optimal approximation, a large amount of calculation is required. It becomes almost impossible to realize it. In this embodiment, this large section is
It is divided into commonly used sections of about 200 mSEC, and by piecewise optimal function approximation for each section,
First, when an arbitrary number of representative frames are assigned to each segment, optimal approximation is performed for each segment, distortions in those cases are calculated, and by using this distortion optimally for large sections can be realized. It is supposed to be.

It should be noted that the above shows one embodiment of the present invention, and it is obvious that the present invention is not limited to the above embodiment.

For example, in the above embodiment, the basic frame length is 10 mSEC, the number of basic frames in one segment is 20 (hence, the one segment length is 200 mSEC), and the number of segments in the large section is 5.
0 (therefore, the time length of the large section is 10 SEC, and the number of basic frames included in the large section is 1000) and the number of representative frames in the large section are 250, etc., and the explanation is given using specific values. Is shown and nothing need be limited to these values.

The dynamic programming method for performing the piecewise optimal function approximation is also an example, and need not be limited to this.

Furthermore, LS is used as a voice feature parameter vector.
The method using P (line-spect pair) has been described,
It is also clear that this need not be limited to the LSP parameter vector, and can be implemented using, for example, the LPC parameter vector or other characteristic parameter vector.

Further, in this embodiment, the rectangular approximation is used as the function used for the piecewise optimal function approximation, but linear approximation or trapezoidal approximation may be used instead.

Linear approximation is to connect the tip of each selected next representative vector with a straight line, thereby determining the vector of each representative basic frame by linear interpolation, and to determine this by the actual of the representative basic frame. Is used instead of the parameter vector of, and the distortion in the case of performing such an approximation can be easily performed in the same manner as described above from the difference of each component between the actual parameter vector of each basic frame and the vector used instead. Therefore, the method used in the present embodiment can be applied almost as it is, and piecewise optimal function approximation and total optimal frame selection can be performed.

That is, the piecewise optimal function approximator is used to find all distortions when optimal linear approximation is performed by arbitrarily changing the number of representative vectors (the number of representative frames) of each section within a required range, and the total optimal frame is calculated. This result is used in the selector to perform the total optimum frame selection exactly as described above.

In other words, virtual setting is performed so that the same number of minimum representative vectors is first given to each section in the large section. Next, the section that produces the maximum distortion in this setting is found using the above results, and the number of representative vectors in this section is increased by one to reduce the distortion. Next, the section which generates the maximum distortion again with respect to the updated setting is found, the number of representative vectors of this section is increased by one, and the distortion is further reduced. In this way, by finding the section that produces the maximum distortion in each setting step, and increasing the number of representative vectors of this section, the number of representative vectors of the entire large section is increased by one, until the number reaches a predetermined number. Is repeated to select a total optimum frame.

When linear approximation is performed, the interpolator 2 on the synthesis side 2
Reference numeral 09 denotes a next-order representative parameter vector supplied from the decoder 202 and the number of basic frames between these parameter vectors to perform linear interpolation to generate a parameter vector for each basic frame, and supplies this to the synthesis filter 207. To do.

In addition, the trapezoidal approximation is a feature of voice information, in which a transient portion of the voice information that changes drastically is approximately 20 mSEC.
A characteristic parameter vector of a voice is obtained by performing optimal approximation using a trapezoidal function in which the time length of a change portion is set to a predetermined constant time length (for example, two basic frames) by having a time length of about Is particularly effective for the optimal approximation of. By using such a trapezoidal approximation, it is possible to reduce adverse effects such as reverberant sound that accompany a sudden change in the parameter vector.

Also in the case of using such a trapezoidal approximation, the distortion due to the approximation can be obtained almost in the same manner as described in the present embodiment, and therefore, the piecewise optimum function approximation described above and this result are used. It is clear that the method of the present invention based on the total optimum frame selection can be applied as it is.

Further, in this embodiment, the memory 10 is provided on the analysis side 1.
9 is provided, whereby voice information arranged in a form suitable for sending to the transmission path is stored, and this is transmitted to the synthesizing side 2 by utilizing a convenient time of the transmission path. Two
In the memory 20, the transmitted voice information is directly stored in the memory 20.
Although the example of applying the method of the present invention to a so-called voice storing and transmitting apparatus such as a voice mail that stores the data in No. 1 and reproduces it when it is convenient for the user has been shown, other normal vocoders and the like can be used. Obviously, it can be applied to a voice analysis / synthesis device. In this case, the memory 109 on the analysis side and the memory 201 on the combining side shown in this embodiment can be omitted.

Furthermore, the present method can also be applied to, for example, a public address device or the like that generates various short speech units stored in advance in combination according to a designation. That is, the present method can be applied to the generation of each speech unit used in such a speech synthesizer to reduce the amount of information and / or improve the sound quality.

According to the method of the present invention, although the optimum approximation is performed for a large section of up to about 10 SEC, for example, the time of the section time width (about 20 mSEC) or more is required for reproducing the voice on the synthesis side. It has the feature that no delay is required. This is particularly effective in that when applied to a device such as a voice mail or a public address, the voice can be reproduced without delay according to the user's request.

As described above, according to the present invention, the basic frame is
It is possible to realize a variable-length frame speech analysis / synthesis method capable of optimally approximating a speech block including a huge number of basic frames by a variable-length frame to a speech block including a plurality of basic frames.

This enables more effective compression and / or
Alternatively, the sound quality can be improved, and the performance of the voice analysis / synthesis device, the voice storage / transmission device, and the voice synthesis device can be improved.

[Brief description of drawings]

FIG. 1 is a block diagram showing an embodiment of the present invention and FIG.
The figure is a diagram for explaining the contents of the work area of the memory in the comprehensive optimum frame selector of the above embodiment. In the figure, 1 ... voice analysis side, 2 ... voice synthesis side, 1
01 …… Low-pass filter and A / D converter (LPF & A /
D), 102 ... Window function processor, 103 ... LSP analyzer, 104 ... Piecewise optimal function approximator, 105 ... Total optimal frame selector, 106 ... Quantizer, 109 ...
Memory, 201 ... Memory, 202 ... Decoder, 20
3 ... Pulse oscillator, 204 ... Noise generator, 205 ...
… V / UV switch, 206 …… Power controller, 208 ……
D / A converter and low-pass filter (D / A & LPF), 2
09 ... Interpolator.

Claims (1)

[Claims] 1. A voice analysis means for periodically analyzing an input voice signal at a predetermined constant analysis cycle to extract a characteristic parameter vector, and a plurality of predetermined continuous analysis cycles. The configuration of the representative parameter vector for each section obtained when selecting an arbitrary number of representative parameter vectors from the feature parameter vector in each section for each section and performing the piecewise optimal function approximation In this case, the piecewise optimum function approximating means for calculating the residual strain by the optimum approximation for each section, and the piecewise optimum function approximating means for the large section consisting of a plurality of predetermined continuous sections are calculated. Comparing the residual strain for each segment, the configuration of the representative parameter vector of the segment with the largest residual strain is determined by adding more representative parameter vectors. A total optimum frame selecting means for selecting a predetermined number of all the representative parameter vectors that optimally approximate the large section by repeating the processing step of replacing with the configuration of the representative parameter vector including A variable length frame speech analysis and synthesis method characterized by: