JPS62999A - Zonal optimum function approximation - Google Patents

Abstract

(57) [Summary] This bulletin contains application data before electronic filing, so abstract data is not recorded.

Description

DETAILED DESCRIPTION OF THE INVENTION [Field of Industrial Application] The present invention relates to a piecewise optimal function approximation method for analyzing human voice signals and extracting feature parameters, and to a device operated using this method.

[Conventional technology]

Analyze the input audio signal and extract its feature parameters,
A speech analysis and synthesis method and an apparatus for transmitting feature parameters consisting of spectral envelope data and sound source data from the synthesis side to the synthesis side via a transmission path, and reproducing an input speech signal based on these feature parameters on the synthesis side, are well known. Are known.

In such a speech analysis and synthesis method and its device, instead of sending the characteristic parameters of the analysis information regarding the input speech from the analysis side to the synthesis side in units of analysis frames, they are divided into sections each consisting of a plurality of consecutive analysis cycles. is optimally approximated to the input audio using a stepwise function such as a rectangular function, and the number of analysis frames and representative feature parameters for each of these sections are supplied together with sound source information from the analysis side to the synthesis side, thereby compressing the amount of transmitted data. The technique has recently become well known due to applications such as variable length frame vocoders. The sections set as variable length frames are obtained by optimally approximating the input audio through rectangular approximation or the like, and DP is used as a method for selecting a plurality of representative analysis frames.

The selection of representative analysis frames for each section by this DP is based on the maximum number M of representative analysis frames set for each section (1<M<K, where K is the number of analysis frames for each section, with the residual distortion The content of the combination that minimizes this as an evaluation scale is determined.

[Problem that the invention seeks to solve]

However, the optimal function approximation using the rectangular function described above has the following drawbacks.

In other words, since a feature parameter vector sequence that originally changes continuously over time is replaced with a step-like function using some representative feature parameter vectors, there is a drawback that the distortion resulting from the replacement becomes large. .

Another object of the present invention is to provide a piecewise optimal function approximation method that eliminates the above-mentioned drawbacks and can better adapt to temporal changes in an actual feature parameter vector sequence.

[Problem 'Means to decide CM']

The method according to the present invention selects a portion of feature parameter vectors representative of the predetermined section from continuous feature parameter vectors of a predetermined section obtained by analyzing an input audio signal at a partial frame period. In the piecewise optimal function approximation method that makes variable the predetermined interval, the predetermined interval is approximated by some representative feature parameter vectors using an optimal trapezoidal approximation method in which the length of the sloped interval is made variable and the length of the non-tilted interval is made short; In addition, the present invention is a method for determining a frame section represented by the representative feature parameter vector.

〔Example〕

Next, the present invention will be described in detail with reference to the drawings.

FIG. 1 is a block diagram showing the configuration of an embodiment of the piecewise optimal function approximation method according to the present invention.

The configuration of the embodiment shown in FIG. 1 is shown as a piecewise optimal function approximator 1, and a sound source information analyzer 100 is also shown. The piecewise optimal function approximator 1 is an LSF analyzer 11. It is comprised of a parameter memory 12, a Drop processor 13, a pre-section selection parameter memory 14, and the like.

When the LSF analyzer 11 receives input audio, it converts the input audio into LPC (LinearPr) for each predetermined analysis frame.
After analyzing the LPC coefficients (linear prediction coefficients) and extracting the LPC coefficients, high-order equations using known techniques such as Newton's iterative method are calculated from the LPG coefficients for each 7L/'-m. LSP (Line 8 pectrnm Pa1rs
, line spectra) and supply these characteristic parameters to the parameter memory 12.

The DP processor 13 performs piecewise optimal function approximation, which will be described later, on the parameters of each analysis frame supplied to the parameter memory 12 under the control of a built-in program using the DP method. In this process, the DP grosser 13 constantly reads out the last two frames selected in the previous section from the parameter memory 12, stores them in the previous section selection parameter memory 14, and stores them in the L of the last selected frame in the previous section.
Piecewise optimal function approximation is performed on a LAP coefficient sequence including SF coefficients.

The selected feature parameters obtained in this way are sent to the synthesis side via the transmission path together with the sound source information data extracted by the sound source information analyzer 100.

The sound source information analyzer 100 extracts and outputs the sound source information of the input speech regarding the strength of the sound source, voiced/unvoiced/silent classification, and pitch period for each analysis frame using known means.

Next, the Dr processor 13, which is the most important part of the present invention.
The operation will be explained using diagrams. In Figure 2, the feature parameter vector analysis period is 10m5ec, and the segment length is 2QQms.
ec (therefore, 20 feature parameter vectors are included in one section) and the number of representative feature parameter vectors is 5. FIG. D
The P processor 13 selects five representative parameter vectors, and determines the section represented by the representative parameter vector, that is, the non-tilted section, and the section that is represented by the interpolated vector of adjacent representative parameter vectors, that is, the tilted section. Yes, and its operation is as follows.

In Figure 2, ■ is the final representative analysis frame of the previous section, ■
~[phase] is the analysis frame number of the current section.

Now, as the first representative analysis frame candidate, any one of the analysis frames ① to [phase] in the temporally preceding order in the classification is targeted. Similarly, analysis frames ① to [phase] are candidates for the fifth frame.

Also, which analysis frame can be used as a second representative analysis frame candidate following the first representative analysis frame candidate.

If analysis frame ■ or [stock] is specified as the representative analysis frame, analysis frame ■～@
In exactly the same way, the fourth frame candidate is any one of the analysis frames ■~0, and the 37th frame candidate is the analysis frames ■~[phase]. It is also self-evident that any one of them is the target.

Now, in Figure 2, suppose that analysis frame ■ is the first
Consider the case where it is selected as a frame. On the other hand, the possible analysis claims for the second frame are ■～@・If we take the combination of these first and second frame candidates as an example and consider the distortion that occurs, we get the following. .

Spectral distortion, that is, time distortion due to analysis frame substitution, can be expressed by the spectral distance between the representative analysis frame and the replaced analysis frame, and is expressed by the following equation (1).

DB=1'Wij"-P")"1.
60. ” (1) Ic;=I K K
In Equation K (1), i and j are the frame numbers of two analysis frames in which the spectral distances, . (1)
Dij shown in the formula is the spectral distance between frames, and from a different perspective, the analysis frame jf:i
This is the spectral distortion, that is, the time distortion, that occurs when replacing with .

Now, 1 analysis frame ■ and ■ are the first and second analysis frames, respectively.
In the case of a Yoshizuku analysis frame, time distortion due to frame substitution does not occur.

Next, considering the case where analysis frame ■ is selected as the second representative analysis frame, D(2
) is the minimum distortion when analysis frames ■ to ■ are replaced by two representative analysis frames including analysis frame ■ which is the second representative analysis frame, and the final representative analysis frame ■ of the previous section. Hereinafter, the minimum distortion determined in this manner will be referred to as the "sum".

In the formula, D3 is the approval or disapproval that occurs when analysis frame (3) is selected as the second representative analysis frame candidate, and D and D2 are the approval or disapproval that occurs when analysis frame (1) or (2) is selected as the first representative analysis frame candidate, respectively. ) represents the approval or disapproval of each selection.

As described above, the acceptance or disapproval of one representative analysis frame candidate is determined by the following (Equation 31).

In equation (3), D −'-1) represents the acceptance/disapproval percentage of analysis claim ■～[phase], respectively. DL2～DL 16 indicates the sum of time distortions defined by the following equations (4) and (5). .

D = min(dLs” mu@2 q”*2*L °(iji-L, 1-2), 1 = mlri ・・・・・・・・・・・・ (4) (4) , (5 ), dL □ is the time distortion% between analysis frame ■ and 0 dL i is the time distortion between analysis frame ■.Also, q15, 16.L is the time distortion between analysis frame [phase] 6D is the minimum value of time distortion when the feature parameter vectors of [phase] or [phase] and ■ are replaced by interpolated vectors, and are expressed by the following equation (6).

・・・・・・・・・ (6) In equation (6), d(l-11, 1-16), 15
is the interpolation vector shown by the following equation (7)...(1-L, 1
-16) It is the spectral distance to which frame [phase], that is, the time distortion that occurs due to frame substitution.

Similarly, q14, 16. L is the minimum value of time distortion when analysis frames 0 to [phase] are replaced by linear interpolation vectors of feature parameter vectors of analysis frame [phase] or [stock] and ■. shown.

In equation (8), d(1-L, 1-16), 14
is d(1-1,,1-06) in equation (6), 0 or 1 which can be calculated similarly to 15, d((16-i)-L, (i
-13)-16), i is d(2-L.

The sum of 1w14 and d) 1-16) , 14 (1-L, 2-16) , 1
5 are interpolation vectors (2-L, 1-16), u (i-L, 2-16) shown by the following equations (9) to (10), respectively.
) and yv-*Q.

■This is frame substitution distortion.

Similarly...q3,16. L, 2,16. L
Also, analysis frames [phase], or frames ■ ~ [phase], ■ ~ using linear interpolation vectors of [phase] and ■
It is expressed as the distortion when [phase] is replaced so that the time distortion becomes the minimum.

Returning to the question of equation (2) again. In formula (2), Dl, 3
represents the distortion when frames ■ to ■ are optimally approximated using representative frames ■ and ■, and is expressed by the following equation (11).

Also, sD2@B becomes Dl, s:Q because there is no frame to be substituted between frames ① and ①.

First, let's consider the analysis as the second representative analysis frame.

In this case, 9, which may exist as the first representative analysis frame, is the analysis frame ■, and the approval/disapproval D4 of ■ and ■ is expressed by the following equation (12).

・・・・・・・・・(12) In formula (12), Dltn pDz *41 Rahini D
j, 4 represent time distortions, for example, Dl, 4
is expressed by the following equation (13).

In equation (13), dt, 1, ct, and ss are the analytical frames ■ and ■ that are interposed between the analytical frames ■ and ■.
The time distortion 5qSt4sl that occurs when both are represented by the analysis frame ■ is the minimum time distortion when the analysis frame ■ is replaced by the analysis frame ■ or the interpolation vector of ■ and ■, q 1 # a *t
C Analysis 7 Rem ■～■t-Analysis 7 L'-mu ■ or ■
Regarding 0Js4 and D3.4, which represent the minimum time distortion when substituted using linear interpolation vectors of and ■, (
13) It is defined according to the same principle as Equation 13). As mentioned above (12)
What does the expression mean?

When ■ is selected as the second representative analysis frame.

This means that the combination of the first representative analysis frame that provides the minimum shrinkage distortion and the analysis frames represented by these seven first and second representative analysis frames is determined. In this way, each of the representative analysis frames obtained from the first to the fifth are subjected to the same procedure (2
) and equation (12) are calculated up to the fifth land analysis frame candidate. Such compression distortion is a function of the input audio signal.

This serves as a measure for setting an approximation function that minimizes the approximation process difference with the comprehensive parameter, so-called residual distortion.

In this way, for example, if analysis frame ■ is the second representative analysis frame, then the preceding analysis frames ■ to ■ are the first representative analysis frame, and if analysis frame ■ is the second representative analysis frame, then the preceding analysis Frames ■ to ■ are each the first representative analysis frame, and the contraction distortion is calculated with the 9-channel setting to the fifth representative analysis frame candidate, and the analysis frame ■ to [phase] of this fifth representative analysis frame candidate is further Perform the following calculations.

II expressed by equation (14) minimizes the influence of shrinkage distortion caused by other analysis frames represented by this when any of the analysis frames ■ to [phase] is selected as the fifth representative analysis frame. D5 to D2° are shrinkage distortions that occur in the analytical claims when any of the seven analytical frames ■ to [phase] is selected as the fifth representative analytical frame. Σ
dB, i is the sum of time distortions from analysis frame ■ and analysis frame ■ to QD master, respectively, from analysis frame ■ and analysis frame ■ to @, and alues is the sum of time distortions between analysis frame ■ and analysis frame ■ to @, respectively. shows the time distortion of

When Dj, which is determined by equation (14), is determined for each section, immediately select the five representative analysis frames that determine the r path with the minimum shrinkage distortion among the combinations of the first to fifth representative analysis frame candidates, and these representatives. An analysis frame or an analysis frame represented by a linear interpolation vector of one representative frame is determined, and in this way, variable-length frames can be easily created by piecewise optimal function approximation in which both the tilted section and the non-tilted section are made variable.

The above is regarding the operation of the DP processor 13 shown in FIG. This is a detailed explanation using FIG. 2.

Next, the Dr processor 13 will be explained in detail using the drawings. FIG. 8 is a block diagram showing the configuration of the DP processor 13), in which the parameter memory 12 and the previous section selection parameter memory 14 are also shown. The Dr processor 13 shown in FIG. 8 includes a time distortion calculator 131. It is composed of an interpolator 132 and a control calculator 133. The control processor is, for example, a calculation system based on a microprocessor, and is at the base of IchisukeM. Of these, @RAM is used for convenience of explanation.

We will express the area. These areas are FO-RT
This is expressed in the RN program as follows.

DIMENSION ALSP(N), BLsP(2
0, N), IDP (5, 20, 3).

l QDP(5,20), IIdAB
l (2 sardines), Ql (20).

Q2 (20).

2 1BIAB2(2+2*N), DMA
B3(N), QB(20) Now, the Nth L8F of frame ■ stored in the previous segment selection parameter memory 14
Parameters are supplied to the control calculator 133 via the input line 141 in response to address signals supplied from the address line 142. The control calculator 133 stores this data in area AL8P. Next, the N-order LSF of frames ■ to ■ stored in the parameter memory 12
Parameters are supplied to the control calculator 133 via the output line 121 in response to the at L/black signals supplied via the address line 122.

The control calculator 133 calculates this data to the area BLSP and determines the corresponding path. Since DI is ``O'', the area QDP (ltl) for storing compression distortion is set to 90''. Also, since there is no frame to be replaced by frame ■ or frame to be replaced by frame ■, D
Area IDP (1, 1, 1) for storing P path =
0. IDP (1°1.2) = O, IDP (1, 1, 3)
=1.

Note that IDP (1, 1, 2) indicates the range of frames substituted for 7 frames ■ when the first representative analysis frame candidate is frame ■ by frame number.

IDP (1, 1, 3) is indicated by the frame number fc7, which is a range of frames to be substituted for frame ■. IDP('1.1,
1) is a Dr pass, and the S card is always ■.

This is expressed here as @O”.

Next, Ds (""DL, 2)] is calculated using the formula. First, d is calculated as follows. Area DMAB I
Addresses (1) to (1) of L, 1 to (between addresses (1) to (1) of ALS F
N) data.

BL at address (N÷1) ~ (2*N) of DjilAB 1
Data at IP addresses (111) to (1,N) is transferred. The control calculator 133 stores the data of DMAB1 at address (2).
*N), then (1) Continuously output line 2 □ 3 at t
4 tRL, ctIMi:31:ff15131 wt
bka fb. : The control calculator 133 also sends 2*N pulses synchronized with this data to the clock line 135.
9 is a block diagram for explaining the time distortion calculator 131 in detail. In FIG. 9, the time distortion calculator 131 includes registers 1311-1 to 1312-N, 1312-1 to 1312-N, subtracters 1313-1 to 1313-N, 1314-1-N, and a multiplier 13.
15-1 to 15-N, and an achievator 1316. The registers 1311 and 1312 are, for example, 16-bit registers, and the stomach line 135t
- storing data synchronously with pulses supplied through the The data supplied via the output line/134 is stored one after another by the aforementioned 2+N pulses, and is finally stored in register 1311-1 as DMABI(1), register 1311-NKDMABI(M), and register 1311-NKDMABI(M). 1312
-1 stores the contents of DMABI (N+1), and register 1312-N stores the contents of DMABI (2*N). That is, the register 131-1 of frame 1 is stored in registers 1311-1 to 1311-N
N-th LAF data of frame (2) is stored in 2-1 to 2-N. Subtractor 1313-1 Calculates the difference between the parameter P1 of frame ■ stored in the a register 1311-1 and the parameter Pl of frame ■ stored in the register 1312-1, and outputs it to the multiplier 1314-1. do. Multiplier 1314 calculates the square of this difference, and multiplier 1
315-1 to one input terminal. Multiplier 131
The other input terminal of 5-1 has spectral sensitivity W as a constant.
1 is applied. Therefore, the output of multiplier 1315-1 is W, (Pl-Pl). Similarly 1315-N
The output of is WN(P' - P'').As a result, the output of the aN key emulator 1316 is the time distortion dL9 calculated between frames ■ and ■ by the time distortion shown in equation (1) above. The calculator 131 outputs the calculated dL, □ to the input line 136.

The explanation will be given again using FIG. 8. The control calculator 133 stores 'L, 1 supplied through the input line 136 at address (1) of area Q1.

Next, qlmlsL is calculated as follows. As mentioned above. d3,1 is calculated as follows in the same way as d, and 1 is obtained. B to addresses (1) to (N) of area DMAB1.
The data of LSF (2,1) to (2,N) is
BLSF address (N+1) ~ (2*N) of (1°1)
~(1,N) data is transferred. Control calculator 133
outputs DMABI data to the time distortion calculator 131. The time distortion calculator 133 calculates dltl and the control calculator 1
Output to 33. The control calculator 133 writes DUEL to address (1) of the area Qs.

Further, d (1-I,, 1-2), 1 is calculated as follows. d(l-L, 1-2), 1 is the time distortion calculated from frames ■ and ■ (0m interval parameters and frame ■ parameters. First, BLSF (2, 1) ~
The data at (2,N) is at address (2+N) in area DMAB2.
+1) to (2+2 N), data of AL8F (1) to (to) is transferred to addresses (241) to (2+N) of DMAB2. Next, an interpolation weight of ``1'' is written to addresses (2) and (1) of %DMAB2. The control calculator 133 sequentially sends the data of DMAB 2 to the interpolator 1 via the output line 137 in order from address (2+2*N) to address (1).
Output to 32. The control calculator 133 also uses a clock line 1 divided into (2+2*N) pulses synchronized with this data.
38 to the interpolator 132.

FIG. 10 is a block diagram for explaining the interpolator 132 in detail. In FIG. 10, the interpolator 132 includes registers 1321-1 to N, 1322-1 to N, 1323-1 to
2. Multipliers 1324-1 to N, 1325-1 to N, 13
27-1-N, adders 1326-1 to N, 1328. R
t) Constructed with M1329° registers 330-1 to 330-N. Registers 1321-1~N, 1322-1~
N, 1323-1 and 1323-2 are, for example, 15-bit registers that store data in synchronization with pulses supplied via the clock life 138. The data supplied via the output line 137 is stored one after another by the aforementioned 2+2+N pulses, and is finally stored in the register 321.
-1 to NK DMAB2 (2+1) to (24N), register 1322-1 to NK DMAB2 (2+N+1) →
+2*N), DMAB2(
1+2) is stored. That is, register 321
7 frames for -1 to N ■Register 322-1 to N
The N-th LSP data of frame ■ is stored in . Interpolation weight ``1'' is written in registers 323-1 and 323-2. As a result, the interpolation weight "1" is supplied to all the multipliers 1324-1 to N1323-1 to N1323-N.
6-1~...N),! : Added value (P' +
P'') (k=1...-N)kk.The adder 1328 adds the weights 1'' and ``1'' and outputs @2'' to the ROM 1329.R(7M132
9 is a 16W OM, and each address has the reciprocal of the address 1.
329 receives 12'' supplied from adder 1328 as address data, outputs α5, and multiplier 1327-1
-%-N supply. Multipliers 1327-1 to N generate N) and output it to L/ and registers 130-1 to N.

This calculation result is input to the controller 1 in synchronization with N pulses supplied from the control calculator via the clock line 139.
40. That is, registers 1330-1 to 1330-N take in data from multipliers 1327-1 to 1327-N at the first pulse among N pulses. By taking in this data, the output of multiplier 1327-1 becomes register 1330-1.
is output to input line 140 via. Also, N
The second to N pulses cause the register to transfer data one after another, resulting in input line 140
Output the calculation results to.

The explanation will be given again using FIG. 8. The control calculator 133 converts the data supplied via the input line 140 into an area DMA.
Write in order from address (1) to address (N) of B5.

This written data is an interpolation parameter for 7 frames (2) and (2). Next, address (1) ~ ( of area DMABI)
The data of DMAB 3(1) to (N) is 111 in N).
BLSF (1,1) to address 0 (N+1) to (2*2)
Data (1, N) is transferred. The DMABI data is output to the time distortion calculator 131, and d(x-L.

1-2) 1 is calculated. This d (1-L, 1-2)
is written to address (2) of area Q2 of the control calculator 133 ◎ As mentioned above, Q 2 (1) contains d Yuki, Oniga, and Q2 (
2) stores d (1-L, 1-a) and x. Now, the control calculator 133 uses the data of Q 2 (1) and Q
Compare the size with the data in 2(2) and select the smaller data by q
l, 2. Let it be L. This q is specially for area Q1
Number 1.2. Write in L address (2). This address (20) corresponds to address (0) on the image. That is, dL1□ is written in Ql (1), ql, 2.L is written in Ql (20). The control calculator 133 further writes the data of Ql(1) and the data of Ql(20).
Compare the data with DL, 2 and at the same time the smallest one is dl, 1, d(1-1, □-2), 1. Area IDP (1, 2, 1) for storing the DP path based on this finding result
The next data is written at ~(x, 2°3). ID
P(1,2゜1) is written with O" corresponding to ■ as the DP path. IDP(1,2,2) and IDP(1,2゜3) are replaced by representative frames ■, ■. Indicates the range of the frame and is written as follows, corresponding to the minimum distortion.

Next, the control calculator 133 calculates D3 (DL 3). First, when frames ■ and ■ are substituted for frame ■, the distortion i 孟□”+, i is calculated using the following procedure. dL 1
is calculated using the above-described procedure and stored in area Q1(2). Next, dL,2 is calculated in the same way as d,□. The control calculator 133 calculates the sum of d, □ and the contents of Qt(2), that is, ΣdL,i, and stores this again in ERI1=1AQ1(2).

Next, dL 1"q2 3 L is calculated like an arrow. dL □ is calculated, and area Q 1 (1)
written to. Furthermore, q2,3. L is the aforementioned ql, 2. L
can be obtained using the same procedure as . In addition, %q23 d3. calculated in the process of L calculation. ! is temporarily Q2
(2) has Q2 (2) has d(1-L, 1-3), 2 is Q
2(3).

The control calculator 133 uses this q23L and Ql(1)
The sum of the contents, that is, 'L, 1''q2, 3.L, is determined and stored again in area Q 1 (1).

Next 1 cql, 3. L is required. ql, 3. L is mi
I get scolded. First, d3.1 is calculated, and area Q2(1
) is written. Furthermore, “8mm is calculated, Q2 (1)
is added to the contents of and written to Q2(1) again.

Next, dmsz is calculated and written to Q2 (2).

Furthermore, d(1-L, 1-3), 1 is the above d(1-L, 1-
It is calculated in the same way as a) and z, and this and the contents of Q 2 (2) are added and written to Q 2 (2) again. Next, the control calculator 133 calculates d(2-L, 1-3), 1.

■L stored in area AL8P(1) to (to)
8P parameter goes to area DMAB2 (2+1) to (20N).

The L8P parameter of ■ stored in areas BLSF (3, 1) to (3, N) is
1) to (2+2*N), respectively. Next D
The interpolation weight @2” is set at address (1) of M-AB2.
2), an interpolation weight of 11" is written in the control calculator 13.
3 outputs the data of DMAB2 to the interpolator 132.

Interpolator 132 receives this data. As for the received data, in FIG. 10, L8F parameters (2) are stored in registers 1321-1 to 1321-N, and L8F parameters (2) are stored in registers 1322-1 to N. Also, the weight "1" is stored in the register 1323-1, and the weight "1" is stored in the register 323-2.

The multiplier 1324-1-N has a weight @2”, and the multiplier 132
As a result of the weight @1'' being supplied to 5-1 to N, the adder 13
The outputs of 26-1 to 26-N become frame (2). Adder 1328
adds the weight “2” and @1, and stores the sum @3” in ROM1
Output to 329. The ROM 1329 outputs 1) to α333, and supplies it to multipliers 1327-1 to N. The multipliers 1327-1 to 1327-N have a multiplication result of 1! -zo2・P
(IJ+1・P@) (k=t,...-...tN)3
kk is generated and output to register 330-1-w. This result is output to input line 140 via registers 330-1 through N.

In FIG. 8, the control calculator 133 converts this data into DM.
Received at AB5. The control calculator 133 calculates d(, -b, 1-
g), d(2-L, 1-s)1 is calculated in the same manner as l and written in Q2(3). The control calculator 133 further includes d.

-L, 2-s) and z, calculate this result and the full content of Q2(3), and write it to Q2(3) again.

The control calculator 133 searches for the minimum value among the contents of Q 2 (13 to (3)) and sets the result as q.
Write to l(20). Further, a control calculator 133. is Ql(
1), Ql(2), Ql (The smallest of the contents of 2 is P(1,3,1)~(1°3.3
) is written in ◎Heat theory ID D, P (1, 3, 1), 1o'' corresponding to ■ is written as the DP path.

IDP (1, 3, 2) and (1, 3, 3) indicate the range of frames substituted for the representative frames ■ and ■.

Corresponding to the minimum distortion, it is written as follows.

Calculate D1s (=DL, □6) and calculate the shrinkage strain by QDP (
1゜4) to (1,16), transfer the DP path data to IDP (
1e 4 sμ) ~ (1,,1,,,6,,,μ),(
μ=1°...3). The details of the processing of the L)P processor 13 regarding the first representative analysis frame candidate have been described above.

Subsequently, the OP processor 13 performs processing regarding the second representative analysis claim candidate. As described above, the second representative analysis frame candidates are O to @. First, processing is performed on the 7th frame ■.

When ■ is the second representative analysis frame candidate, only ■ is the first representative analysis frame candidate that is the target of the bus. Also, since there are 7 relays between frames ■ and ■, Ds
= o. The control calculator 133 QDPs the bruise distortion 10''
Write to (2,2). Furthermore, the control calculator 133 inputs @1'' to IDP (2.2.1) as Dr path data (262
Write @1” to 12) and @2” to (2, 2, 3).

Next, processing regarding frame (2) is performed. If ■ is the second representative analysis 7 frame candidate, the first
Representative analysis 7 frame candidates are ■ and ■. First, time distortion D1 when claims ■ and ■ are taken as representative frames and frame ■ is replaced with one of the representative frames or their interpolated data. s is calculated in the same manner as DL □ and written to area QB (1). Next, when the target of the pass is frame ■, the time distortion n,os is QB
Write in (2). The thermal theory, D,, s is 10''. Furthermore, the control calculator 133 adds QDP(1) to the contents of Q B (1).
, 1) and write this again to QB(1). Similarly, the contents of QDP (1, 2) are added to the contents of QB (2), and this is looked into Q B (2) again. Furthermore, the control calculator 133 compares the magnitude of the contents of QB(1) and QB(2), selects the smaller one, writes it to QDP(2,3) as Ds, and writes the corresponding DP path information to I.
Write to DP (2, 3, 1) to (2, 3, 3). In addition,
The above process for calculating Ds is performed by executing equation (2) above.

Next, processing regarding frame (1) is performed with frames (2) to (2) being passed. D l using the same procedure as above.
*4 s n, t4m DI *4 is Q B (1),
QB (2).

QB(3) is then written to Qntt)(i=1.2.
3) contents and QDP (1#i) (i=1.2.
The addition result of the contents of 3) is again Q B (*) (x=1
t 2 3 ).

The minimum value of QB(i) (i=L...3) is searched, this is written as D4 to QDP(2,4), and the corresponding Dr
Pass information as IDP (2, 4, 1) ~ (2', 4゜3)
write to. Note that the processing related to D4 above is performed in steps 5) to (2) above.
, 17). Thermal Dr path information is also written to the corresponding address of the IDP.

Subsequently, the Dr processor 13 performs processing related to the third representative frame candidate for frames ■~[phase], processes related to the fourth representative frame candidate for frames ■~0, and processes related to the fifth representative frame candidate for frames ~[phase].
This will be carried out targeting the following.

Finally, the DP processor 13 executes the process shown in equation (14) above in the following procedure. The control calculator 133 has ds, -
- is output, this result is added to the contents of QDP (5, 5), and the addition result is written to QDP (5, 5) again. then d
I ;? is calculated and similarly added to the contents of QDP (5, 5), and the % result is written to QDP (5, 5). QIDP is calculated from dm g8 @ d8sl e""...d@@*@ one after another. Furthermore, the control calculator 133 is a QDP
(5°5) to Ql) The minimum value of the contents of P(5, 20), ie, as shown in equation (14), is determined to determine the fifth representative frame.

When the fifth representative frame is determined, the fourth to first representative frames are determined at the same time from the DP bus master stored in the IDP, and the first to fifth representative frames directly substitute for other frames; A section in which the interpolated data of the representative frame replaces other frames is determined. These section information are output to encoder 201 as repeat bits.

Further, the parameters of the first to fifth representative frames are output to the BLSP encoder 201. Further, the parameters of the fifth representative frame are outputted to the previous section selection parameter memory 14 as the next section (■).

In the above explanation, the tamed general in the first representative analysis 7 frame candidates ■ ~ [phase] was explained as a distortion of the frame that is replaced by the previous division representative analysis frame ■ and the representative frame candidate ■ ~ [phase]. It is not always necessary to use frame ■. In this case, instead of the above equation (3), the following (15
) formula is used.

When using the thermal theory equation (15), the pre-section selection parameter memory 14 is not required.

Figure 3 shows an example using the piecewise optimal function approximation method shown in Figure 1.
FIG. 2 is a block diagram showing the configuration of an embodiment of a 7'C variable length frame type vocoder.

The variable length frame vocoder shown in FIG. 3 is comprised of a variable length frame vocoder analysis side 2 and a variable length frame vocoder synthesis side 3t. Further, the variable length frame vocoder analysis side 2 includes a piecewise optimal function approximator 1. Sound source information analyzer 100. The variable length frame vocoder synthesis side 3 includes demultiplexers 301 . Pitch pulse generator 302° noise generator 303. Switcher 304. Variable amplifier 305° interpolator 306. LSF synthesis filter 307
, D/A converter 308 and LPF (Low P
a5sFilter) 309.

Piecewise optimal function approximator 1 of analyzer and sound source information analyzer 100
output selected feature parameter data and sound source information data, which are encoded by encoders 201 and 202, and then supplied to a multiplexer 203, where they are multiplexed in a predetermined format and sent to the synthesis side via a transmission line 2001. will be sent to.

In this embodiment, the piecewise optimal function approximator 1 performs piecewise optimal function approximation and outputs LSP coefficients for frame compression as selected feature parameters. That is, outputting information between a number of representative frames equal to or less than a preset maximum number and a variable length slope interval of the number of frames represented by these representative frames for each division with a preset number of analysis frames as a unit; The sound source information analyzer 100 also outputs data regarding the strength of the sound source, voiced/unvoiced/silent, and pitch period.

On the synthesis side, the demultiplexer 301 performs demultiplexing, and among the decoded data, the selected feature parameter data is sent to the interpolator 306, and the pitch period data of the sound source information data is sent to the pitch pulse generator 302.
The silent/silent discrimination data is supplied to a switch 304, and the sound source strength data is supplied to a variable gain amplifier 305.

The interpolator 306 interpolates and reproduces the LAF coefficients for all analysis frames for each section based on the LSF coefficient sequence based on the representative frame selected for each section and the information regarding the analysis frame specified by these 7 representative frames. It is supplied to the synthesis filter 307 and used as its filter coefficient.

On the other hand, when the input voiced/unvoiced/silent discrimination data specifies voiced, the switch 304 switches the pitch pulse generator 30
The output of the noise generator 303 is switched so that the output of the noise generator 303 is supplied to the variable gain amplifier 305 when silent or no sound is specified. Therefore, when the signal is voiced, a bibchi pulse with a repetition frequency corresponding to the pitch period is supplied to the variable gain amplifier 305, and when there is no voice or no sound, white noise generated by the noise generator 3030 is supplied to the variable gain amplifier 305.

The variable gain amplifier 305 uses a gain setting corresponding to the sound source data to amplify the bitch noise or white noise and supplies them as a driving sound source to the LSP synthesis filter 307. Thus, the L8F synthesis filter 307 inputs the digital quantity. Play the audio and then D/A/Bata 3
08. The signal is output as analog audio via the LPF 309.

In this way, a variable length frame type vocoder using the piecewise optimal function approximation method shown in FIG. 1 can be realized.

FIG. 4 is a block diagram showing the configuration of an embodiment of a speech synthesizer that uses the piecewise optimal function approximation method shown in FIG. 1 and synthesizes input speech while accumulating analysis data.

All other components of the speech synthesizer 4 shown in FIG. 4 except for the memory 310 are the same as those of the variable length frame vocoder synthesis side 3 shown in FIG. 3 with the same symbols, so detailed explanations regarding these will be omitted. .

The memory 310 of the speech synthesizer 4 stores encoded data regarding feature parameters and sound source information obtained by subjecting preset various audio materials to piecewise optimal function approximation according to the present invention. Each time a read command signal received via the read command signal 4001 is input, the feature parameter information and sound source information regarding the audio material specified by the read command signal are sent to the demultiplexer 3.
Output to 01.

The demultiplexer 301 decodes the thus supplied input, and sends the feature parameter data to the interpolator 306, the pitch period data of the sound source information to the pitch pulse generator 302, and the voiced/unvoiced/silence discrimination data to the switch 304.
Then, the sound source strength data is supplied to variable gain amplifiers 305, respectively.

The switch 304 outputs the output of the pitch pulse generator 302 when input voiced/unvoiced/silent discrimination data specifies voiced, and outputs the output of the pitch pulse generator 303 when input voiced/unvoiced/silent discrimination data specifies voiced/no voice.
is switched to supply the output of the variable gain amplifier 305 to the variable gain amplifier 305.

When the LSF synthesis filter 307 receives the output of the variable gain amplifier 305, it operates using this as a driving sound source and uses the characteristic parameters received from the interpolator 306 as a filter coefficient to reproduce the input audio signal.

This digital reproduction signal is converted to a D/A converter 308°L.
It is converted into a desired analog quantity via PF3Q9 and output.

In the embodiment shown in FIG. 4, the analysis data stored in the memory 310 uses feature parameters extracted on the analysis side based on the piecewise optimal function approximation means shown in FIG. In the example, an LSF coefficient sequence is used as a feature parameter.

FIG. 5 is a block diagram showing the configuration of an embodiment of a waveform encoding device that uses feature parameters of input speech extracted using the piecewise optimal function approximation method shown in FIG. 1. Fifth
The waveform encoding device 5 shown in the figure is a piecewise optimal function approximator 1.
Noise weighting unit 501, encoding/decoding unit 502. Interpolator 503° Correlation coefficient calculator 504. Autocorrelation coefficient calculator 505° multipulse searcher 506. It is configured with an encoder 507 and a multiplexer 508, and among these components, the parts other than the piecewise optimal function approximator 1 and the multiplexer 508 input using the feature parameters extracted by the piecewise optimal function approximator 1. This is a portion for encoding the waveform of speech, and in this embodiment, multipulses as a sound source waveform are obtained from these components using a known correlation region evaluation method.

The input speech is supplied to a piecewise optimal function approximator 1 and a noise weighter 501.

The noise weighting device 501 has a noise filter with a transfer function determined based on the order of the feature parameters extracted by the piecewise optimal function approximator 1, audio materials, etc. Perform convolution multiplication.

Piecewise optimal function approximator 1 extracts LPC coefficients of a predetermined order by a piecewise optimal function approximation method, and supplies these feature parameters to noise weighting unit 501. This feature parameter is also supplied to an encoder/decoder 502, encoded, and supplied to a multiplexer 508 as feature parameter data. The encoded feature parameters are decoded again and then supplied to the interpolator 503.

The interpolator 503 uses the characteristic parameters supplied from the piecewise optimal function approximator 1 via the encoder/decoder 502 to represent seven representative analysis frames selected for each section and information on the analysis frame specified by this representative analysis frame. Interpolate the feature parameters between representative analysis frames using
After each frame is reproduced, the impulse response of the vocal tract filter is determined and supplied to a cross-correlation coefficient calculator 505 and an autocorrelation coefficient calculator 505.

The cross-correlation coefficient calculator 504 performs convolution integration of the impulse response of the vocal tract filter and the noise weighting supplied in this way with the subsequent input speech data, calculates the cross-correlation between the two, and multiplies the obtained cross-correlation coefficient. Pulse searcher 5
Supply on 04.

The autocorrelation coefficient calculator 505 receives data regarding the impulse response of the vocal tract filter from the interpolator 503, calculates this autocorrelation coefficient, and uses the multipulse searcher 5 to calculate the autocorrelation coefficient.
Supply to 06.

The multi-pulse searcher 506 uses the thus input cross-correlation coefficients and auto-correlation coefficients to search for a multi-pulse train using a method based on publicly known phase-open region evaluation, encodes it in the encoder 507, and then encodes it. It is supplied to multiplexer 508 as sound source data. This sound source data can be said to be sound source waveform information itself, and in this way a waveform encoding device that utilizes characteristic parameters can be realized.

In this case, the waveform information to be encoded uses the multi-pulse obtained from the feature parameters, but other waveform information extraction means, such as an LPC inverse filter whose frequency response characteristic is opposite to that of the synthesis filter, is used. It is clear that waveform encoding can be similarly performed by means such as receiving the LPG parameters from the piecewise optimal function approximator 1, generating a residual signal from this and the input audio signal, and encoding this waveform information. It is.

FIG. 6 is a block diagram showing the configuration of an embodiment of the compressed Dr type word speech recognition device using the piecewise optimal function approximation method shown in FIG.

The compressed DP type word speech recognition device 6 shown in FIG. 6 includes a piecewise maximum function approximator 1 as well as a switch 601. Standard button memory 602. It is configured to include a slam matcher 603 and a minimum distance searcher 604.

The compressed Dr type word speech recognition device 6 shown in FIG.
The basic operation is as follows.

That is, as shown in FIG. 1, the piecewise maximum function approximator 1 converts the LPG coefficients into LSP parameters using a known method in order to extract the LPG coefficients at a predetermined order for each analysis frame of the input audio signal, and converts them into LSP parameters using a known method. 7-frame compression is performed using a thousand methods in preset segment units. This frame compression is obtained through the approximation process explained in FIG. In this way, the optimal combination of representative frames that do not exceed the maximum number set in advance for each category and analysis claims specified by these 7 representative frames is selected, and the variable length 7 frames obtained in this way are first sent to the switch 6010.
It is stored in the standard meta/memory 602 depending on the connection made at the time of Kamai.

Next, the switching device 601 is switched to the recognition side, and when the specific speaker utters the word sound stored in the standard bang memory 602, it undergoes exactly the same processing as at the time of registration, and the bang matching device 60
3 is supplied with LSP parameters.

Battan matching device 603 is equipped with a spectral distance measuring device, an interpolator, etc., and is a spectral distance measuring device that sets interpolated values between representative analysis frames for each category of human power compressed by the piecewise maximum function approximator 1. After performing this for all standard patterns, the spectral distance of the characteristic parameter pattern for both hands is determined between corresponding points in a preset range, and then supplied to the minimum distance search unit 604 together with the standard pattern designation number.

The minimum distance searcher 604 selects the standard button with the minimum spectral distance and outputs its designated number as a recognition result, thus compressed DP type word speech recognition processing can be realized using the piecewise optimal function approximation method.

3 to 6 described above are all devices that utilize the piecewise optimal function approximation method explained with reference to FIG. It can be realized.

〔Effect of the invention〕

As explained above, according to the present invention, it is possible to realize a piecewise optimal function approximation method and apparatus that can significantly improve the degree of approximation versus processing amount.

FIG. 7 is a piecewise optimal function approximation characteristic diagram showing an example of the above effect due to the piecewise optimal function approximation in the embodiment of FIG.

[Brief explanation of drawings]

FIG. 1 is a block diagram showing the configuration of an embodiment of the piecewise optimal function approximation method according to the present invention, FIG. 2 is an explanatory diagram for explaining in detail the piecewise optimal function approximation method according to the present invention, and FIG. is a block diagram showing the configuration of an embodiment of a variable length frame type vocoder using the piecewise optimal function approximation method shown in FIG. 1, and FIG. FIG. 5 is a block diagram showing the configuration of an embodiment of a speech synthesizer that synthesizes input speech while accumulating analysis data. A block diagram showing the configuration of an embodiment of a waveform encoding device using feature parameters, FIG.
A block diagram showing an example of a compressed DP type word speech recognition device using the piecewise optimal function approximation method shown in the figure.
FIG. 7 is a piecewise optimal function approximation characteristic diagram showing an example of the effect of piecewise optimal function approximation in the embodiment of FIG.
The figure is a detailed block diagram of an example of the interpolator 132 shown in FIG. 8, which is an example of the Dr processor 13 shown in FIG. 1... Piecewise optimal function approximator, 2...
Variable length frame vocoder analysis side, 3...Variable length frame vocoder synthesis side, 4...Speech synthesizer,
5... Waveform encoding device, 6... Compression D
P-type word speech recognition device, 11... LAF analyzer, 12... Parameter memory, 13...
・DP7 four processors, 14... Previous section selection parameter memory, 201... Encoder, 202...
... Encoder, 203 ... Multiplexer, 3o1 ... Demultiplexer, 3o2 ...
...Pitch pulse generator %303...Noise generator, 3o4・mountain...switcher, 305...Variable gain amplifier, 3o6・mountain...interpolator, 307...・・・
・LSP synthesis filter, 3o8...D/A converter, 3o9...LPF, 310...
・Memo! 7.501・River・Noise weighted vessel, 5o2
... Encoder decoder, 5o3 ... Interpolator, 5o4 ... Cross correlation coefficient calculator, 5o5.
... Autocorrelation coefficient calculator, 506 ... Multi-pulse search 4ft, 507 ... Encoder,
508・Mountain...Multiplexer, 6o1...Switcher, 602・River...Standard meta/memory, 6o3...
...Bang matching device, 6o4...Minimum distance search device, 100...Sound source information analyzer, 131
. . . Time distortion calculator, 132 . . . Interpolator, 133 . . . Control calculator, 1311-1 to N.
...Register, 1312-1~N・mountain...Register, 1313-1~N...Subtractor%1314
-1~N... Multiplier, 1315-1~N...
... Multiplier, 1316 ... Accumulator. 1321-1~N River...Register, 1322-1~N
...Lucista, 1323-1~2...Lucista, 1324-1~N...Multiplier, 1325
-1~N... Multiplier, 1326-1~N...
...Adder, 1327-1~N... Multiplier,
1328...Adder, 1329...R
OM, 1330-1~N...Register. Agent Susumu Uchihara Patent attorney

Claims (1)

[Claims] Piecewise optimal function approximation that varies the frame period by selecting some feature parameter vectors representative of the predetermined section from continuous feature parameter vectors in a predetermined section obtained by analyzing an input audio signal at a constant frame period. In the method, the predetermined section is approximated by some representative feature parameter vectors using an optimal trapezoidal approximation method in which a slope section length is variable and a non-slope section length is variable, and the representative feature parameter vector is representative. A piecewise optimal function approximation method characterized by determining a frame interval.