JPH0731520B2 - Variable length frame type pattern matching vocoder - Google Patents

Description

The present invention relates to a variable-length frame type pattern matching vocoder, and particularly to frame selection by a piecewise optimal function approximation method and pattern matching by matching with a standard pattern. The present invention relates to a variable length frame type pattern matching vocoder that can remove a frame in which pattern matching distortion is large by performing the processing in association with each other through processing distortion.

[Conventional technology]

The spectral envelope parameter obtained by analyzing the input speech signal is compared with the standard pattern of the spectral envelope to select the optimal standard pattern. Instead of the spectral envelope parameter, the specified code of the standard pattern is analyzed from the analysis side to the synthesis side. It is well known that a pattern matching vocoder that can significantly reduce the required amount of information to be transmitted by transmitting the pattern matching vocoder.

The variable-length frame type pattern matching vocoder is a section made up of continuous K analysis frames, instead of sending all standard patterns corresponding to the analysis frames which are collated and selected for each analysis frame in such a pattern matching vocoder. While selecting L standard patterns corresponding to L representative analysis frames extracted for each, the number of analysis frames represented by these standard patterns together with each of these L standard patterns, that is, a so-called repeat bit (repeat bit). bit) is sent from the analysis side to the synthesis side. In this case, the standard pattern selected for each segment is the representative analysis frame selected for each segment (: the label of the optimum standard pattern, that is, only the designated code is sent to the synthesis side together with the repeat bit. The representative analysis frame selected for is obtained by approximating the distribution of spectral envelope parameters indicated by all analysis frames for each segment by an optimal approximation function. The optimal approximation function is a rectangle, trapezoid, or straight line approximation function, respectively. It is used properly depending on the operation purpose of the vocoder, and usually the function is set via the DP method.

When using a rectangular approximation function as the above-mentioned optimal approximation function,
Specifically, frame selection is performed as follows.

The spectrum envelope parameter sequence analyzed at a frame period of about 10 mSEC from the input speech signal is divided into blocks (including K frames) having a constant time length of about a single syllable, and a fixed number (L Individual) representative frames and the boundaries of the sections represented by the representative frames are selected so as to minimize the sum of the spectral distances from the spectral envelope parameter sequence of the input speech. That is, the contents of the K analysis frames for each section are represented by the contents of the L analysis frames forming the rectangular function and the number of analysis frames represented by each of the L analysis frames.

In this way, the conventional variable frame type pattern matching vocoder is capable of significantly reducing the number of required bits by the variable length frame in addition to the pattern matching.

[Problems to be solved by the invention]

However, in the conventional variable length frame type pattern matching vocoder of this type, the selection of the representative frame required for the variable length frame configuration and the selection of the standard pattern by the pattern matching are performed independently.
Therefore, the spectrum distortion or quantization distortion that accompanies pattern matching and the spectrum distortion that accompanies the selection of variable-length frames using DP, that is, the frame distortion, is generated by substituting the representative frame. The so-called time distortion, which is based on the occurrence of the difference in the spectral distance, is analyzed and synthesized independently of each other. Therefore, there is a drawback that the deterioration of the quality of the voice of the voice cannot be avoided.

[Means for solving problems]

The vocoder of the present invention is a standard pattern selecting means for selecting an optimum standard pattern having a minimum spectral distance by comparing a spectrum envelope parameter obtained by analyzing an input voice signal with a standard pattern relating to the spectrum envelope, and this standard. A frame selection unit that uses the total distortion defined by the scalar addition of the spectral distortion calculated accompanying the standard pattern selection by the pattern selection unit and the spectral distortion calculated accompanying the frame selection using DP as the evaluation value And means for using the spectral distortion calculated from the interpolation parameter of the standard pattern selected as the spectral distortion calculated accompanying the frame selection by the DP and the spectral envelope parameter obtained by analysis. Consists of

〔Example〕

 The present invention will now be described in detail with reference to the drawings.

FIG. 1 is a block diagram showing the configuration of an embodiment of a variable length frame type pattern matching vocoder according to the present invention.

The configuration of the embodiment shown in FIG. 1 comprises an analysis side 1 and a synthesis side 2, and the analysis side 1 has a parameter analyzer 11, a sound source analyzer 12, a pattern collator 13, a standard pattern file 14, and a frame selection. The combination side 2 includes a demultiplexer 21 and a pattern reader 2
2, a sound source generator 23, a standard pattern file 24, and a voice synthesis filter 25.

The audio signal input via the input line 1001 is supplied to the parameter analyzer 11 and the sound source analyzer 12.

The parameter analyzer 11 analyzes the spectral envelope parameter of the input speech signal, and in the case of the present embodiment, LSP (Lin (Lin
We use e Spectrum Pairs, which is another LPC (Lin) that is effective for pattern matching.
Ear Prediction Coffieient) can be used.

By the way, the spectral envelope parameter is usually implemented on the premise of LPC analysis, and the input speech signal is first LPF (Low P
After performing a required low-pass filtering with an ass filter), the signal is quantized with a predetermined number of bits by an A / D (Analog to Digjtal) converter, and then multiplied by a predetermined window function. The multiplication by the window function is performed at preset constant cycles, and the quantized speech signal thus cut out at constant cycles becomes an analysis frame. In this example, LP
The cutoff frequency of F is 3.4KH, and the sampling frequency of the A / D converter is
At 8 KHz, the window function processing stores 30 mSEC portions of the quantized audio signal in the internal memory, reads this at a cycle of 10 mSEC, performs window processing with the Hamming function, and outputs it as a 10 mSEC analysis frame. Further, in this embodiment, 20 consecutive analysis frames, that is, 200 mSEC, is set as one section.

For the quantized speech signal for each analysis frame, after extracting the α parameter of a predetermined order by LPC analysis, a method of solving a higher-order equation using Newton's iterative method, or a zero-point search method is known. The LSP coefficient sequence of a preset order is obtained by using the method of (1) and this LSP is supplied to the pattern collator 13.

The pattern collator 13 collates the LSP as a spectrum envelope parameter input for each segment and each analysis frame with the standard pattern of the spectrum envelope parameter by the LSP stored in the standard pattern file 14 in advance to find the optimum spectrum envelope. Select a standard pattern.

In pattern matching, the spectral envelope pattern by the LSP for each input analysis frame is compared with the spectral envelope standard pattern, and the one with the smallest spectral distance between both patterns is selected as the optimal spectral envelope standard pattern. The minimum spectral distance between patterns can be defined by D _Q (q) in the following equation (1).

W _K is a constant expressing the rate of deformation of the spectrum envelope with respect to a slight change of the Kth element of the Nth order LSP coefficient vector, which is the spectrum envelope parameter, and is called the spectral sensitivity, which has been experimentally predicted. Use what you were asked for.

For spectral sensitivity, see Y. Tohkura, F. Itakura,
“Spectral Smoothing Technique in PARCOR Speech An
alys is Synthesis "IE ³ Trans.on ASSPVol, ASSP−2
VI, “SST and Spect” of 6, NO.6, December 1978, P. 587-596
ral Sensitivity of Parameters "and F. Itakura,“ Op
timal nonlinear Transformation of LPCs to improve
quantization properties JASA Vol.56 (supplemen
t) paper H14, P.516, 1974. N is L
SP analysis order, P _K (Q) is a spectrum envelope pattern of the analysis frame input for each section, Q is the number of each frame included in the block including K frames, and Q = 1,2, ... K. In addition, R = 1 to
M and M are the total number of spectral standard patterns, P _K ^{(S1) to} P _K ^(SM)
Is a spectrum envelope standard pattern from 1 to Mth.

Based on equation (1), the M spectral envelope standard patterns and the spectral envelope patterns of the input analysis frames for each segment are pattern-matched via the LSP distribution, and the minimum D _{Q is obtained.}
The one showing (q) is selected as the standard pattern, and the standard pattern selected in this way, the designation code designating the standard pattern, and D _Q (q) are respectively defined as the standard pattern parameter, label information, and quantization distortion. It is supplied to the frame selector 15.

D _Q (q) represented by the equation (1) indicates a spectral distance between two patterns, and is spectral distortion that should be called pattern matching distortion that occurs in pattern matching, that is, quantization distortion.

Now, the frame selector 15 is also provided with the LSP from the parameter analyzer 11, and the frame selector 15 uses these inputs to select a representative analysis frame for performing variable length framing for each section based on the DP method. .

In this embodiment, the variable length framing is approximated by a rectangle,
This is done by using the method of substituting other parameters by performing zero-order interpolation between the selected parameters. This method uses linear and temporary interpolation or trapezoidal approximation and linear and zero-order interpolation. It does not matter even if other optimal function approximations such as a combination are used.

In this rectangular approximation, a preset number of representative analysis frames are selected from the analysis frames in the division, and all the analysis frames in the division are represented by these representative analysis frames. Using the method, the one that forms a rectangular function that is piecewise optimally approximated to the spectral envelope parameter of the input speech signal is determined as follows.

FIG. 2 is a frame selection explanatory diagram for explaining the basic contents of frame selection by the DP.

The setting of the variable-length frame in the present embodiment is performed in a format in which a piecewise optimum function is set for each section of 200 mSEC consisting of 20 pieces of analysis frame of 10 mSEC, and this section is divided into five representative analysis frames and these representative analysis frames. It is expressed by the repeated information and. That is, each section is expressed by a combination of the selected five representative analysis frames and an analysis frame having each of these representative analysis frames as a representative, and the representative analysis frame is an approximate rectangle set by the representative analysis frame. The one that minimizes the difference in spectral distance from the spectral envelope parameter of
In the format selected by.

The length of the section, the analysis frame length, the number of representative frames, and the like described above can be arbitrarily set in consideration of the operation purpose of the vocoder.

The first representative analysis frame candidate to the fifth representative analysis frame candidate shown in FIG. 2 indicate the first to fifth candidate analysis frames of the five representative analysis frames selected from the 20 analysis frames. It is a thing.

In this embodiment, there are two restrictions on the selection of these first to fifth representative analysis frame candidates.

The first limitation relates to the number of alternative frames. Each of the representative analysis frame candidates precedes in time. Alternatively, the maximum number that can replace other temporally subsequent frames is 6 frames. That is, the analysis frame candidates can represent up to 6 frames each before and after including the candidate itself. Therefore, in the case of the present embodiment, the number of consecutive frames that can be represented by the analysis frame candidates is 13 (= 6 + 1 + 6) to 1 (= 0 + 1).
It is free between +0). In this embodiment, the maximum number of substitutions of the preceding or succeeding frame is set to 6, but this can be arbitrarily set by optimally evaluating the reproducibility of the synthesized voice and the required calculation amount.

The second limitation is the maximum interval between representative analysis frames. In this embodiment, it is set to 7. Regarding the maximum interval, it is possible to optimally evaluate the reproducibility of the synthesized voice, the required calculation amount, etc., and set it arbitrarily.

Now, the first representative frame candidate is limited from the analysis frame (1) (the number of substitutions of the preceding frame is 0) to the analysis frame (7) (the number of substitutions of the preceding frame is 6) due to the restriction that the number of the preceding frames that can be substituted is 0 to 6. Is the target. Similarly, the fifth representative frame candidate has an analysis frame (14) (the number of substitutions of subsequent frames is 6) due to the limitation that the number of subsequent frames that can be substituted is 0 to 6.
From, analysis frame (20) (substituting number of subsequent frames is 0)
Is the target.

The analysis frames that can be the second representative frame candidates following the first representative frame candidates are the analysis frames (2) to (14) for the following reasons.

That is, if the first representative frame is the analysis frame (1), the second representative frame is limited to the second frame because of the limitation on the maximum interval of the representative analysis frames and the possibility that adjacent frames are selected.
Analysis frames that can be representative frame candidates are (2)-
(8). If the first representative frame is the analysis frame (2), the analysis frames that can be the second representative frame candidates are (3) to (9). Similarly, if the first representative frame is the analysis frame (7), the analysis frames that can be the second representative frame candidates are (8) to
(14) Therefore, the analysis frames that can be the second representative frame candidates are (2) to (14).

Similarly, the analysis frames that can be the fourth representative frame candidates are (7) to (19) due to the limitation on the maximum interval of the representative analysis frame with the fifth representative frame candidate.

Now, the analysis frame that can be the third representative frame candidate is
It is constrained by both the second representative frame and the fourth representative frame. That is, it must always exist between the second representative frame and the fourth representative frame.

The analysis frames that can be the third representative frame candidates are (3) to (21) because of the limitation on the maximum interval of the representative analysis frame from the second representative frame candidate and the possibility that adjacent frames are selected. Further, due to the restriction on the maximum interval between the representative frames with the fourth representative frame candidate and the possibility that adjacent frames are selected, the analysis frames that can be the third representative frame candidates are (0) to (18). . Analysis frames (3) to (18) that simultaneously satisfy the above two conditions are analysis frames that can be the third representative frame candidate. In the above description, frames (0) and (21) do not actually exist.

In the frame selection using DP in this embodiment, first, the spectrum distortion due to the analysis frame replacement of the analysis frame replaced by the representative analysis frame, so-called temporal distortion is calculated, and then the quantization distortion included in each analysis frame, that is, pattern matching. Spectral distortion at the time of processing is added, and total distortion obtained by adding these temporal distortion and quantization distortion is used as an evaluation value. In this embodiment, after calculating the time distortion in this way, the quantization distortion is added to measure the total distortion of the evaluation values. However, it can be easily implemented as a processing method for reversing the addition order of these two distortions. It is clear that

Now, consider the case where the analysis frame (1) is selected as the first frame in FIG. On the other hand, the analysis frames that may be the second frame are (2) to (8). Taking the combination of the first and second frame candidates as an example, the time distortion that occurs will be as follows.

The spectral distortion or time distortion due to the analysis frame substitution can be represented by the spectral distance between the representative analysis frame substituted with the optimum standard pattern and the substituted analysis frame, and is represented by the following equation (2).

In equation (2), i, j is the frame number of the representative analysis frame replaced by the optimum standard pattern for measuring the spectral distance d _{i, j} and the frame number of the analysis frame replaced by the representative analysis frame, and P _K ( si) is a parameter of the representative analysis frame i replaced by the standard pattern. The other symbols are the same as in the case of the formula (1). D _{i, j} shown in the approximate expression of (2) is the spectral distance between the frame i substituted by the optimum standard pattern and the frame j substituted by the frame i, and occurs when the frame j is substituted by i. Spectrum distortion, that is, time distortion.

When the analysis frames (1) and (2) become the first and second representative analysis frames, respectively, the time distortion due to the frame substitution does not occur and only the quantization distortions included in each are added. Is the total distortion.

Next, the analysis frame (3) as the second representative analysis frame
Considering the case where is selected, D ₃ shown in the following equation (3)
⁽²⁾ is defined with the minimum total distortion.

In equation (3), D ₃ ⁽²⁾ is the total distortion that occurs when analysis frame (3) is selected as the second representative analysis frame candidate, and D ₁ ⁽¹⁾ and D ₂ ⁽¹⁾ are respectively the ^first distortion. The respective total distortions when the analysis frames (1) and (2) are selected as one representative analysis frame are shown.

The total distortion in the above-mentioned first representative analysis frame candidate t is the analysis frames (1) to (t-
Each time distortion between 1) and the frame (t) replaced by the optimum standard pattern is measured, and the quantization distortion of each analysis frame is added to these measured values.
Expression (4) shows each total distortion when the analysis frames (1) to (7) are selected as the first representative analysis frame.

In equation (4), D ₁ ^{(1) to} D ₇ ⁽¹⁾ are total distortions of analysis frames (1) to (7), respectively, and D ₁ ⁽ q ^{) to} D ₇ ⁽ q ⁾ are analysis frames (1) to ( 7) Each quantization distortion, d _2,1 is the time distortion between the analysis frame (1) and the analysis frame (2) replaced by the optimal standard pattern, and Is the sum of the time distortion between the analysis frame (1) and the analysis frame (3) replaced by the optimum standard pattern, and between the analysis frame (2) and the analysis frame (3) replaced by the optimum standard pattern, Indicates the sum of the time distortions of the analysis frame (7) substituted with the optimum standard pattern and the analysis frames (1) to (6).

Further, D _1,3 in the equation (3) is the first in the analysis frames (1) and (3) as defined by the following equation (5).
And when it becomes the second representative analysis frame, the smaller one of the two types of frame alternative distortion, that is, the time distortion, which differs depending on whether the analysis frame (2) is represented by the analysis frame (1) or (3). Show. Further, in D _2,3, the analysis frames (2) and (3) are the first, respectively.
Also, this is a time distortion that may occur when the frame becomes the second representative analysis frame. However, since there is no substitute frame between the representative frame candidates (2) and (3), D _2,3 = 0.
Becomes It should be noted that D ₃ ⁽ q ⁾ is the quantization distortion of the analysis frame (3).

In the equation (5), d _1,2 is (2) of the analysis frame (1) and the analysis frame (2) substituted by the optimum standard pattern.
Spectral distance by the formula, d _3,2 is the analysis frame (3) and analysis frame (2) replaced by the optimal standard pattern
Shows the spectral distance between and.

The expression (3) means that when the analysis frame (3) is selected as the second representative analysis frame, the analysis frame (1) may be used as the first representative analysis frame. This means that the one with the smaller total distortion in the two cases in which one of the above is selected is selected.

Now, let us consider the minimum total distortion D ₄ ⁽²⁾ when the analysis frame (4) is selected as the second representative analysis frame.

In this case, there is a possibility that the first representative analysis frame may exist, in addition to the analysis frame (1), there are (2) and (3), and the total distortion D ₄ ⁽²⁾ is expressed by the following equation (6). Shown.

In the equation (6), D _1,4 , D _2,4 and D _3,4 each represent time distortion, and, for example, D _1,4 is represented by the following equation (6). D ₄ ⁽ q ⁾ represents the quantization distortion of the analysis frame (4).

In the equation (7), d _1,2 and d _1,3 are analyzes interposed between the analysis frame (1) replaced by the optimum standard pattern and the analysis frame (4) replaced by the optimum standard pattern. The time distortion that occurs when each of frames (2) and (3) is represented by analysis frame (1), and d _4,2 and d _4,3 are analysis frames (2) and (3)
Time distortion when and are both represented by analysis frame (4), d _1,2 is analysis frame (2) is analysis frame (1), and d _4,3 is analysis frame (3) is analysis The respective time distortions represented by the frame (4) are shown. D _2,4 and D _3,4 are also defined by the same policy as that of the equation (7). The expression (6) above means that when (4) is selected as the second representative analysis frame, the first representative analysis frame and the first and second representative analysis frames that give the minimum total distortion. That is, the combination of the representative analysis frames is determined. In this way, for each of the first to fifth representative analysis frame candidates, the total distortion as shown in equations (3) and (6) is obtained up to the fifth representative analysis frame candidate in the same procedure as described below. To go. Such total distortion is a measure for setting an approximate rectangular function that minimizes the residual distortion, which is the difference between the approximate processing and the spectral envelope parameter of the input speech signal.

Thus, for example, when the analysis frame (5) is the second representative analysis frame, the preceding analysis frames (1) to (4) are the first representative analysis frames, and the analysis frame (6) is the second representative analysis frame. In such a case, the preceding analysis frames (1) to (5) reach the fifth representative analysis frame candidate while calculating the total distortion with the setting that can be the first representative analysis frame, and the analysis frame of the fifth representative analysis frame candidate. The following calculation is further performed for (14) to (20).

When any of analysis frames (14) to (20) is selected as the fifth representative analysis frame, Dl represented by equation (8) minimizes the influence of the total distortion due to other analysis frames represented by this. Indicates that one of the analysis frames (14) to (20) is selected as the fifth representative analysis frame, and ▲ D ⁽⁵⁾ ₁₄ ▼ to ▲ D ⁽⁵⁾ ₂₀ ▼ is selected. It is the total distortion that occurs in those analysis frames, and Is the sum of the time distortions of the analysis frame 14 and the analysis frames 15 to 20 substituted by the optimum standard pattern, Is the sum of the time distortions of the analysis frame 15 substituted by the optimum standard pattern and the analysis frames 16 to 20, respectively.
Further, d _19,20 represents the time distortion between the analysis frame (19) _replaced by the optimum standard pattern and the analysis frame (20).

When Dl determined by Eq. (8) is determined for each section, immediately, five representative analysis frames that determine the DP path with the minimum total distortion among the combinations of the first to fifth representative analysis frame candidates and these representative frames are selected. An analysis frame represented by the analysis frame is determined, and thus variable length framing by piecewise optimal rectangle approximation is easily implemented.

In this way, five representative analysis frames selected by performing frame selection with the total distortion obtained by scalar-adding the quantization distortion in the pattern matching and the time distortion caused by the frame selection using DP as the evaluation value, and these representative analysis frames The number of analysis frames represented by (i.e., repeat bits) is determined, and the representative analysis frame is supplied to the multiplexer 16 together with the repeat bit information after replacing the corresponding spectrum envelope standard pattern with the label information that specifies it.

The quantization distortion as the pattern matching distortion is significantly larger than the frame substitute distortion due to the frame selection normally performed through the setting of the DP path. In this way, the variable length frame is set while excluding the frame with the large pattern matching distortion. The pattern matching information can be output in a format.

Next, a specific configuration and operation of the frame selector 15 will be described with reference to the drawings. FIG. 3 is a block diagram for explaining the frame selector 15 in detail.

The frame selector 15 shown in FIG. 3 is the LSP parameter memory 5
1, standard parameter memory 52, quantization distortion memory 53, label memory 54, DP controller 55, temporal distortion calculator 56, temporal distortion temporary memory 5
7, frame boundary determiner 58, node distortion memory 59, path memory
60, node distortion calculator 61, node distortion temporary memory 62, path determiner 6
3, a frame determiner 64, a total distortion calculator 65, and a timer 66.

The timer 66 generates a frame period signal and a division signal, and outputs the frame period signal every 10 msec.
Output line signal via 661 and every 200 msec
2 to the DP controller 55.

The DP controller 55 is a microprocessor and operates by a built-in program to control the entire frame selector. The DP controller 55 is initialized for each division signal. By this initialization, the LSP parameter memory 51, the standard pattern memory 52, the quantization distortion memory 53, the label memory 54, the node distortion memory 59, and the path memory 60 are cleared.

Now, the 10th-order LSP parameter analyzed by the parameter analyzer 11 is supplied to the LSP parameter memory 51 every frame period. The LSP parameter memory 51 is a 256w RAM.
The LSP parameter memory 51 has a capacity for storing 10th-order LSP for one segment length, that is, 20 frames (minimum 10 × 20 = 20).
0w) is taken into consideration. LSP parameter memory 5
An address signal 1 is supplied from the DP controller 55 via the address line 551 with an address signal corresponding to the frame number based on the division signal, and stores the LSP parameter at a desired address.

The standard pattern parameter memory 52, the quantization distortion memory 53, and the label memory 54 are RAMs of 256, 32w, and 32w, respectively, and standard pattern parameters ▲ P ^(SR) _K ▼ supplied from the pattern collator 13.
(K = 1, ..., 10), quantization distortion ▲ D ^(q) _Q ▼, and standard pattern label R information are supplied from the DP controller 55 via address lines 552, 553, and 554, respectively. The address signal is stored at a desired address.

Now, it is assumed that the seventh frame cycle signal is supplied from the timer 66 to the DP controller 55 based on the division signal. DP
The controller 55 calculates the distortion corresponding to the first representative analysis frame candidate by the following procedure and stores it in the node distortion memory 59. Size node distortion memory 59 for ease of explanation (5,20)
2D area. Quantization distortion ▲ D ^(q) ₁ ▼ of analysis frame (1), that is, D ₁ ⁽¹⁾ is read from the quantization distortion memory 53 as is clear from the equation (4), and output line 531 and DP controller 55 Then, it is stored in the address (1, 1) of the node distortion memory 59 via the input / output line 555. The address is specified by an address signal supplied via the address line 556. Next, the quantization distortion D ₂ ⁽ q ^{) of the} analysis frame (2) is read from the quantization distortion memory 53 and supplied to the node distortion calculator 61 via the DP controller 55 and the quantization distortion input line 557. To be done. The standard pattern parameter of the analysis frame (2) and the LSP parameter of the analysis frame (1) are the standard pattern memory 52-the output line 521-the DP controller 55-the standard pattern parameter input line 562 and the LSP parameter memory 51-, respectively.
Output line 511-DP controller 55-LSP parameter input line
It is supplied to the time distortion calculator 56 via 561. The time distortion calculator 56 calculates the time distortion d _2,1 generated when the LSP parameter of the analysis frame (1) is replaced with the standard pattern parameter of the analysis frame (2) by using the equation (2).
Next, the time distortion calculator 56 outputs the time distortion d _2,1 to the output line 563, DP
Node distortion calculator via controller 55 and time distortion input line 558
Supply to 61. The node distortion calculator 61 uses the quantization distortion D ₂ ⁽ q ⁾
And the time distortion d _2,1 sum D ₂ ⁽¹⁾ is calculated by the second line of the equation (4), and this D _{2 is} output via the output line 611 and the DP controller 55.
⁽¹⁾ is supplied to the node distortion memory 59. Node distortion memory 59
Stores this D ₂ ⁽¹⁾ at address (1,2).

Similarly, the quantization distortion D ₃ ⁽ q ⁾ is supplied from the quantization distortion memory 53 to the node distortion calculator 61. Next, the standard pattern parameter of the analysis frame (3) is supplied from the standard pattern parameter memory 52 to the time distortion calculator, and the LSP parameter of the analysis frame (1) is supplied from the LSP parameter memory 51 to the time distortion calculator. The time distortion calculator 56 calculates d _3,1 from these parameters by the equation (2) and outputs it to the node distortion calculator 61. Node distortion calculator 61 obtains the sum of the quantization distortion D ₃ ^(q) and the time distortion ^{_{d 3,1 D 3 (q) +}} d 3,1. Next, the standard pattern parameters of analysis frame (3) and analysis frame (2)
The time distortion _d3,2 with the LSP parameter is calculated by the time distortion calculator 56 and accumulated in the addition result. The result of this accumulation is D ₃ ⁽¹⁾ shown in the equation (4). D ₃ ⁽¹⁾ is stored in the address (1,3) of the node distortion memory 59 similarly to D ₂ ⁽¹⁾ .

Similarly, D ₄ ^{(1) to} D ₇ ⁽¹⁾ are accumulated in the node distortion calculator 61 and stored in the address (1,4)-(1,7) of the node distortion memory 59.

Now, it is assumed that the 14th frame synchronization signal is supplied from the timer 66 to the DP controller 55 with reference to the division signal. DP
The controller 55 calculates the distortion corresponding to the second representative analysis frame candidate, the DP path, and the frame boundary by the following procedure, and stores the distortion in the node distortion memory 59 and the frame boundary data in the path memory 60.

The quantization distortion D ₂ ⁽ q ^{) of the} analysis frame (2) is the quantization distortion memory 5
It is supplied from 3 to the node distortion calculator 61. When the second representative analysis frame is (2), the first representative analysis frame is (1)
Except for Therefore, the DP path is 1-2, the frame boundary is between 1-2 because there is no substitute frame between the representative analysis frames, and the total distortion D ₂ ⁽²⁾ is D ₂ ⁽²⁾ = D ₁
⁽¹⁾ Calculated easily with + D ₂ ⁽ q ⁾ . In this embodiment, the DP path 1-2 is represented by the preceding frame 1, and the frame boundary 1-
2 is represented by the section 1 represented by the preceding frame. Size the path memory 60 for ease of explanation (5,20,2)
It will be a three-dimensional area.

Now, the total distortion when the first representative analysis frame is (1)
D ₁ ⁽¹⁾ is the node distortion calculator 59 from the node distortion memory 59 via the input / output line 555, the DP controller 55, and the quantization distortion input line 557.
Is supplied to. The node distortion calculator 61 converts the quantization distortion D ₂ ⁽ q ⁾ into D ₁
Add ⁽¹⁾ . The addition result D ₂ ⁽²⁾ is stored in the address (2, ²⁾ of the node distortion memory 59. DP controller 55 is input / output line 601
The data "1" is written in the address (2,2,1) of the path memory 60 through the and the data "1" is written in the address (2,2,2). The address is designated through the address line 602.

Next, the total distortion D ₃ ⁽²⁾ when the second representative analysis frame is frame (3) is calculated as follows.

The time distortions d _3,2 , d _1,2 are calculated by the time distortion calculator 56 in the above-described manner, and the time distortion temporary memory 57 is successively output via the output line 564.
Is output to. The temporal distortion temporary memory 57 is a two-dimensional area of size (20,2). d ₃ , ₂ , d ₁ and ₂ are stored in the addresses (2, 1) and (2, 2) of the time distortion temporary memory 57, respectively. The address designation is performed via the address line 571. The frame boundary determiner 58 compares the magnitude of the above-mentioned d _3,2 and d _1,2 supplied via the output line 572, and selects a small time distortion. This distortion is D _1,3 shown in the equation (3), and d
_{If 3,2} <d _1,2 , then D _1,3 = d _3,2 . The calculated distortion D _1,3 is sent to the node distortion calculator 61 via the distortion output line 581,
If the frame boundary candidate, d _3,2 <d _1,2 , the frame (2) is replaced by (3), so “1”, data is supplied to the bus memory 60 via the boundary data output line 582. It The bus memory 60 writes this data at the address (2,3,2). Next, D ₁ ⁽¹⁾ from the node distortion memory 59 and D ₃ ⁽ q ⁾ from the quantization distortion memory 52 are respectively supplied to the node distortion calculator 61,
It is added to the time distortion D _1,3 . The result of this addition, D ₁ ⁽¹⁾ + D _1,3
+ D ₃ ⁽ q ⁾ is sent to node distortion temporary memory 62 via output line 612.
Is supplied to. The node distortion temporary memory is a 20w RAM, and an address is designated through an address line 621. The addition result D ₁ ⁽¹⁾ + D _1,3 + D ₃ ⁽ q ⁾ is stored in the address (1).

Next, D ₂ ⁽¹⁾ from the node distortion memory 59
D ₃ ⁽ q ⁾ is supplied to the node distortion calculator 61, respectively, and D ₂ ⁽¹⁾
+ D ₃ ⁽ q ⁾ is calculated. The calculated D ₂ ⁽¹⁾ + D ₃ ⁽ q ⁾ is supplied to the node distortion temporary memory 62 and written in the address (2) of this memory.

The two distortion amounts temporarily stored in the node distortion temporary memory 62 are supplied to the bus determiner 62 via the output line 622.
The bus determiner 63 compares the two distortion amounts and selects the smaller one. The selected distortion is D ₃ ⁽²⁾ shown in the equation (3). The bus determiner 63 outputs this D ₃ ⁽²⁾ to the distortion output line 631, DP
It is supplied to the node distortion memory 59 via the controller 55. The node distortion memory 59 stores D ₃ ⁽²⁾ at address (2,3). The bus determiner 63 further determines the path data “1” or “2” that gives the minimum distortion to the frame (3) that is the second representative frame candidate.
Is output to the DP controller 55 via the path output line 632. D
The P controller 55 transfers the path data to the address (2,3,
Write in 1). If the pass data is "2", the data "2" is changed to the pass memory to correct the boundary data written in the address (2,3,2) of the pass memory 60.
Send to 60.

Similarly, the total distortion D ₄ ⁽²⁾ when the second representative analysis frame is frame (4) is calculated as follows. First, the total distortion is calculated assuming that the corresponding first representative analysis frame is the frame (1), and the address (1) of the node distortion temporary memory 62 is calculated.
Written in. The path data “1” and the frame boundary data “1”, “2”, or “3” are the addresses (2, 4,
1), written in (2,4,2). Next, the total distortion when the first representative analysis frame corresponding to the second representative analysis frame candidate (4) is assumed to be the frame (2) is calculated and written in the address (2) of the node distortion temporary memory 62. Then the path determiner
63 is 2 temporarily stored in the node distortion temporary memory 62
Compare the three distortions and select the smaller distortion. If the distortion corresponding to the frame (2) is small, the addresses (2,4,1) and (2,4,2) of the path memory 60 are rewritten. Next, the same processing is performed on the frame (3). as a result,
The path determiner 63 determines D ₄ ⁽²⁾ shown in the equation (6). This D ₄ ⁽²⁾ is stored in the address (2, 4) of the node distortion memory 59.

D ₅ ^{(2) to} D ₁₄ ⁽²⁾ are obtained one after another by the same procedure and are stored in the addresses (2,5) to (2,14) of the node distortion memory 59.
Further, the path and frame boundary data determined in the process of calculating the node distortion are respectively the addresses {(2,5,1),
It is stored in (2,5,2)} to {(2,14,1), (2,14,2)}.

In this way, the distortion corresponding to the third representative frame candidate, the DP path, and the frame boundary are calculated after the 18th frame synchronization signal is supplied from the timer 66 to the DP controller 55 on the basis of the division signal, and the node distortion memory 59 And stored in the path memory 60. Similarly, the corresponding distortion, DP path, and frame boundary of the fourth representative frame candidate of the 19th frame synchronization signal and the fifth representative frame candidate of the 20th frame synchronization signal are calculated, and the node distortion memory is calculated. 59 and the path memory 60. As a result, the time generated as a result of frame substitution when the analysis frames (14) to (20) are selected as the fifth representative frame candidates at the addresses (5,14) to (5,20) of the node distortion memory 59, respectively. The total sum of the distortion and the quantization distortion of the representative frame is stored.

However, for example, when the analysis frame (14) is selected as the fifth representative frame candidate, the analysis frames (15) to
The time distortion resulting from the substitution of (20) by the standard pattern parameters of frame (14) is not included in D ₁₄ ⁽⁵⁾ . Therefore, the processing shown in the equation (8) is required. In this embodiment, this is carried out as follows.

First, Is calculated as follows. D ₁₄ ⁽⁵⁾ stored in the node distortion memory 59 is input to the total distortion calculator 65 via the node distortion input line 651. Next, the standard pattern parameter memory
The standard pattern parameter of the frame (14) is supplied from 52 to the time distortion calculator 56. Next, the LSP parameter of the frame (15) is supplied from the LSP parameter memory 51 to the time distortion calculator 56. The time distortion calculator 56 calculates the time distortion d _14,15 . The calculated d _14,15 are input to the total distortion calculator 65 via the time distortion input line 652. Follow the same procedure one after another d
_14,16 , _d14,17 ... _d14,20 are input to the total distortion calculator 65.
The total distortion calculator 65 calculates the sum of these distortions, that is, Is calculated and the result is output to the frame determiner 64 via the output line 653. The frame determiner 65 has a 20w RAM built-in and writes this data at address (14).

As well Is written in the RAM (15) in the frame determiner 64. Thereafter, up to D ₁₉ ⁽⁵⁾ + d _19,20 are written up to the address (19) of the RAM one after another. Finally, D ₂₀ ⁽⁵⁾ is the node distortion memory 59
Is directly supplied to the frame determiner 64 via the total distortion calculator 65, and written in the address (20) of the RAM. Frame determiner 6
4 determines Dl based on the equation (8) and outputs the corresponding frame number to the DP controller 55 via the output line 641. Based on the frame number and the path data and the frame boundary data stored in the path memory 60, the DP controller displays five representative frames, which represent all 20 frames,
The section represented by the five representative frames is determined, and the number of frames equal to the section length is output to the output line 151 in the form of repeat bits. Further, the standard pattern number corresponding to this representative frame is output to the label memory 54. The label memory 54 outputs the label to the DP controller 55 via the output line 541.
Output to. DP controller 55 outputs this on output line 152.

The sound source analyzer 12 inputs the quantized speech signal for each analysis frame from the parameter analyzer 11, extracts sound source strength, voiced / unvoiced / unvoiced data, and pitch period data by a known method, and outputs these sound source information. Supply to the multiplexer 16.

The multiplexer 16 encodes and multiplexes various inputs in a predetermined format and sends them to the synthesizing side 2 via the transmission line 101.

The synthesizing side 2 demultiplexes and multiplexes the multiplexed signals thus sent out, and supplies the label information and the repeat bit information to the pattern reader 22 and the sound source information to the sound source generator 23, respectively. .

The pattern reader 22 reads the spectrum envelope standard pattern corresponding to the label information from the standard pattern file 24, and transmits it to the speech synthesis filter while repeating the number of times specified by the repeat bits.

In the present embodiment, the standard pattern file 24 has the same contents as the pattern collator 13, and thus the spectral envelope parameters are supplied to the speech synthesis filter 25 in units of analysis frames.

The sound source generator 23 inputs sound source information and generates a sound source modeled by a pulse train corresponding to the pitch period for voiced / unvoiced / voiced voices, and white noise when there is no voice, and outputs this as a sound source strength. It is supplied to the voice synthesis filter 25 while being amplified correspondingly.

The speech synthesis filter 25, which is configured as an all-pole digital filter, uses the spectral envelope parameter received from the pattern reader 22 as a filter coefficient and synthesizes a digital amount of speech driven by the sound source received from the sound source generator 23,
/ A (Digital to Anlog) converter to analog, remove unnecessary high frequency components by LPF, and send to output line 2001 as output audio signal.

Although the LSP is used as the analysis parameter in the above-mentioned embodiment, other LPC coefficients or the like that can be processed with substantially the same efficiency may be used. Further, it is clear that the sound source analysis and generation can be easily performed by using a sound source waveform transmission such as a multi-pulse train instead.

〔The invention's effect〕

As described above, according to the present invention, in the variable-length frame type pattern matching vocoder, it is possible to basically remove a frame having a large pattern matching distortion by processing the spectrum distortions associated with the frame selection and the pattern matching in association with each other. There is an effect that a long frame type pattern matching vocoder can be realized.

[Brief description of drawings]

FIG. 1 is a block diagram showing the configuration of an embodiment of a variable length frame type pattern matching vocoder of the present invention, and FIG. 2 is a frame selection explanation for explaining the basic contents of frame selection in the embodiment of FIG. 3 and FIG. 3 are detailed block diagrams of the frame selector 15 shown in FIG. 1 …… Analysis side, 2 …… Synthesis side, 11 …… Parameter analyzer, 12
...... Sound source analyzer, 13 …… Pattern collator, 14 …… Standard pattern file, 15 …… Frame selector, 16 …… Multiplexer, 21 …… Demultiplexer, 22 …… Pattern reader, 23
...... Sound source generator, 24 …… Standard pattern file, 25 …… Speech synthesis filter, 51 …… LSP parameter memory, 52 …… Standard parameter memory, 53 …… Quantization distortion memory, 54 …… Label memory, 55… … DP controller, 56 …… Time distortion calculator, 57 …… Time distortion temporary memory, 58 …… Frame boundary determiner, 59 …… Node distortion memory, 60 …… Path memory, 61 …… Node distortion calculator, 6
2 ... node distortion temporary memory, 63 ... path determiner, 64 ... frame determiner, 65 ... total distortion calculator, 66 ... timer.

Claims (1)

[Claims] 1. A standard pattern selecting means for selecting an optimum standard pattern having a minimum spectral distance by comparing a spectrum envelope parameter obtained by analyzing an input speech signal with a standard pattern relating to the spectrum envelope, and this standard pattern selecting means. The total distortion defined by the scalar addition of the spectral distortion calculated by the standard pattern selection by the method and the spectral distortion calculated by the frame selection using DP (Dynamic Programming) A pattern matching vocoder comprising a frame selecting means as an evaluation value, the time distortion calculated from the interpolation parameter of the parameter of the standard pattern selected for each analysis frame and the parameter analyzed for each analysis frame, , Select a frame that has the smallest sum of quantization distortions that accompany the selection of the standard pattern. A variable length frame type pattern matching vocoder characterized by having DP means.