JPH0449960B2 - - Google Patents

Description

[Detailed Description of the Invention] [Technical Field] The present invention relates to a residual excitation type vocoder, particularly a means for transmitting spectral envelope information using a pattern matching means, and a means for reproducing high frequency components of a residual waveform on an analysis side. The present invention relates to a residual excitation type vocoder equipped with a residual excitation type vocoder.

[Prior art] The input audio signal is processed by LPC (Linear Prediction).
Coefficient, linear prediction coefficient) obtained by analyzing
Along with the LPC coefficients, information regarding the low frequency components of a preset region of the residual waveform as waveform information is transmitted from the analysis side to the synthesis side, and the synthesis side uses the input low frequency components to restore and enhance the The residual waveform is synthesized with the frequency component, and
RELP (Residual Exciting Linear) is a residual excitation type vocoder that operates a synthesis filter using LPC coefficients and reproduces high-quality input audio signals.
It is also known by the abbreviation ``Prediction Vocoder''.

Compared to a conventional general vocoder, which sends a modeled residual waveform from the analysis side to the synthesis side together with the LPC coefficients, RELP sends a part of the low frequency components of the residual waveform through modeling. Basically, the playback quality is excellent in that part of the waveform is transmitted from the analysis side to the synthesis side without any noise, and it is comparable to conventional vocoders with a speed of about 4.8Kb/s (kilobits per second). CODEC (COder
It has recently become widely used as a DECorder (codec).

However, in conventional RELP of this type, the residual waveform information supplied from the analysis side to the synthesis side is limited to low frequency components in a preset frequency domain due to the available data bit rate. The basic method for high frequency components is to apply this low frequency component to a nonlinear circuit on the synthesis side and simply add the resulting harmonic components, ignoring phase information. It is well known that this method has the drawback that reproduction distortion, so-called RELP sound, occurs in which the waveform of the reproduced audio differs from the input audio signal because the reproducibility of the input audio signal cannot be obtained.

In order to improve such RELP sound, residual excitation type vocoders equipped with high frequency reproduction means on the analysis side have recently been used. This is because the analysis side estimates the tap coefficient of the transversal high-frequency reproduction filter used to reproduce the residual waveform on the synthesis side, and then supplies it to the synthesis side. For a transversal filter with the same configuration as the reproduction filter, set the impulse response characteristics including the phase conditions for the low frequency and high frequency to be synthesized, that is, set the tap coefficient, and then supply this to the synthesis side. It is something.

However, in conventional residual excitation type vocoders of this type, the high frequency components generated on the synthesis side are harmonic components generated by applying the low frequency components supplied from the analysis side to a nonlinear circuit. Since this method utilizes the same frequency range, it is often difficult to obtain the desired high frequency reproduction.

In addition, in conventional residual excitation vocoders of this type, the LPC coefficients are directly quantized and transmitted, making it difficult to allocate a sufficient amount of information to the residual waveform, making it difficult to reproduce high-quality sound. The disadvantage is that it is often difficult to obtain.

[Purpose of the invention]

An object of the present invention is to provide a residual excitation type vocoder which eliminates the above-mentioned drawbacks and has a means for transmitting spectral envelope information by a pattern matching means, and is equipped with a high frequency component reproducing means on the analysis side. Estimating the tap coefficient of a high-frequency reproduction filter using a sample sequence in which a down-sampled value of a waveform low-frequency component is made into a reference sampling conversion sample sequence through interpolation of a zero-level sample and without using a nonlinear circuit. It is an object of the present invention to provide a residual excitation type vocoder that can stably and reliably improve RELP sound to a large extent by analyzing and synthesizing input audio signals.

[Structure of the invention]

According to the present invention, the first step compares the spectral envelope characteristic of an input audio signal with a standard pattern prepared in advance, and transmits the label attached to the standard pattern from the analysis side to the synthesis side as information indicating the spectral envelope characteristic.
means for determining a reference sample sequence of a residual waveform of an input audio signal corresponding to a preset reference sampling frequency; and a downsampling frequency obtained by reducing the reference sampling frequency at a preset rate. Correspondingly, third means for obtaining a down-sampled sequence of low frequency components of the residual waveform, and the residual is formed while setting a zero-level interpolation point in the down-sampled sequence at the timing of the reference sampling frequency. Estimating the coefficients of a transversal high frequency reproduction filter to be provided on the synthesis side based on the reference sampling conversion sample sequence of the low frequency component of the waveform and the reference sample sequence of the residual waveform and without going through a nonlinear circuit. A residual excitation type vocoder is obtained, characterized in that it comprises fourth means for outputting the coefficients to be transmitted to the combining side.

According to another feature of the present invention, the spectral envelope characteristic of the input audio signal is compared with a standard pattern prepared in advance, and a plurality of pattern candidates are preliminarily selected;
A first means for estimating coefficients of a transversal high-frequency reproduction filter to be provided on the synthesis side for each pattern candidate on the analysis side; a second means for determining the /N ratio; and a label attached to a pattern candidate with the best S/N ratio upon receiving the output of the second means is transmitted from the analysis side to the synthesis side as information indicating the spectral envelope characteristic. a third means, a fourth means for determining a reference sample sequence of a residual waveform of an input audio signal corresponding to a preset reference sampling frequency, and downsampling in which the reference sampling frequency is reduced at a preset rate. a fifth means for obtaining a down-sampled sequence of low-frequency components of the residual waveform corresponding to a frequency; and forming a down-sampled sequence while setting a zero-level interpolation point in the down-sampling sequence at the timing of the reference sampling frequency. Coefficients of a transversal high-frequency reproduction filter to be provided on the synthesis side based on the reference sampling conversion sample sequence of the low frequency component of the residual waveform and the reference sample sequence of the residual waveform and without going through a nonlinear circuit. There is obtained a residual excitation type vocoder characterized in that it comprises a sixth means for estimating the coefficients and outputting the coefficients to be transmitted to the synthesis side.

〔Example〕

Next, the present invention will be explained in detail with reference to the drawings.
FIG. 1A is a block diagram showing the structure of the analysis side of a first embodiment of the residual excitation type vocoder according to the present invention, and FIG. 1B is a block diagram showing the structure of the synthesis side.

In the first embodiment, when the spectral envelope information of an input audio signal is transmitted by a pattern matching means, the LPC coefficient obtained by analyzing the input audio signal and a standard pattern having the smallest spectral distance are selected, and the standard pattern is applied to the pattern. This is a format that transmits the corresponding label.

The analysis side shown in FIG. 1A is an A/D converter 1,
LPC analyzer 2, quantizer/decoder 3, LPC inverse filter 4, LPF (Low Pass Filter) 5, down sampler 6, quantizer/decoder 7, filter coefficient estimator (1) 8 , quantizer 9, power calculator 10, quantizer 11 and multiplexer 12
The synthesis side shown in FIG.
4, a multiplier 15, an LPC synthesis filter 16, and a D/A converter 17.

The audio signal input from the input terminal 101 is
After performing low-pass filtering to cut out components of a preset high frequency, in this example, 3.4KHz or higher, which is supplied to the D converter 1, sampling is performed at a preset reference sampling frequency of 8KHz, and a predetermined number of bits is sampled. It is converted into a digital quantity and sent as quantized input audio data to an LPC analyzer 2 and an LPC inverse filter 4. The 8KHz sampling frequency matches the highest frequency of the input audio signal.
This was determined by considering the Nyquist rate under the conditions set at 3.4KHz. The LPC analyzer 2 sequentially performs window processing in which the input quantized input audio data is multiplied by a preset window function for each basic frame period, and then measures the autocorrelation for each basic frame. LPC analysis is performed to extract LPC coefficients of a predetermined order using the well-known Auto Correlation method.

In this way, the output from LPC analyzer 2 is
The LPC coefficients are quantized by the quantizer/decoder 3 and supplied to the multiplexer 12, and this quantized LPC coefficient data is once decoded and supplied to the LPC inverse filter 4. The quantizer/decoder 3 analyzes the LPC coefficients by pattern matching processing, encodes the LPC coefficient data in the form of a pattern label, and further decodes it into LPC coefficients. Figure 10 shows the quantizer/decoder 3
FIG. 2 is a block diagram showing an example of the configuration. The quantizer/decoder 3 will be explained in detail below using FIG.

The LPC coefficients output from the LPC analyzer 2 are supplied to the pattern matching processor 1001. The pattern matching processor 1001 selects a standard pattern having the closest spatial vector distance to the input LPC coefficient from a plurality of standard patterns (4096=2 ¹² in this embodiment) stored in the standard pattern memory 1002. (Hereinafter, this selection process will be referred to as "matching").

Next, the principle of matching will be briefly explained. For this matching, a standard pattern is selected that has generally excellent matching characteristics, that is, one in which the spatial vector distance, so-called spectral distance, is close over the entire frequency range.

Each of the Nth-order LPC coefficients can be considered to represent one spatial vector in the Nth-order parameter space, and therefore, the spectral distance representing the space vector between frequencies by the LPC coefficients can be expressed as the degree of approximation between two LSP coefficients. It is well known that it is used as a matching measure.

The above-mentioned standard pattern is a distribution pattern of standard LPC coefficients obtained by LPC analysis of audio materials in advance, and there are 2 ^{to 12} preset types in this example, and the spectral distance is as follows (1) shown in the formula
Basically indicated by Dij.

Dij=1π/π ₀ {Si(w)−Sj(w)} ² dw...(1) In equation (1), i and j are the frame numbers as processing unit intervals in LPC analysis, and Si
(w) and Sj(w) are logarithmic spectra of frames i and j. Equation (1) is usually converted to the following approximate equation (2).

Dij _N 〓 ^k=1 Wk{Pk(i)−Pk ^(j) } ² ...(2) In equation (2), Pk(i) and Pk ^(j) are the Nth-order LPC coefficients in frames i and j, Wk is the Nth-order LPC spectral sensitivity. The above N is an all-pole type
The order of the LPC digital filter corresponds to 10 in this embodiment, and its LPC coefficients α ₁ , α ₂ , . . . α ₁₀ are shown. Furthermore, the Nth-order spectral sensitivity Wk is Nth-order,
In the case of this example, the degree of spectral change caused by a minute change in the 10th-order LPC coefficient is shown.

This will be explained using FIG. 10 again. When the pattern matching processor 1001 receives the LPC coefficients,
The minimum distance register 1004 is set to a value Do larger than the empirically known maximum value of Dij shown in equation (2) above.
Initialize to . Furthermore, the label register 1005 is initialized to, for example, "4096". Next, the pattern matching processor 1001 selects the start address 0〓10=0(0
〓10 of 10 is equivalent to the order 10 of LPC) is output to the standard pattern memory 1002, the address 0 specifying the spectral sensitivity W ₁ corresponding to the LPC coefficient α ₁ is output to the spectral sensitivity memory 1003, and the spectral distance D is output as “ Set to 0”. Next, the pattern matching processor 1001 calculates α ₁ ^(o) of the standard pattern with the label “0” output from the standard pattern memory 1002 ((2)
P ₁ ⁽ⁱ⁾ in the equation) and the spectral sensitivity W ₁ output from the spectral sensitivity memory 1003
and the first coefficient α ₁ ^(I) of the LPC coefficients supplied from the LPC analyzer 2 (equivalent to P ₁ ⁽ⁱ⁾ in equation (2) above), D=W ₁ (α ₁ ⁽⁰⁾ −α ₁ ^(I) ) ² is calculated. Next, the pattern matching processor 1001
Increments the address data output to the standard pattern memory 1002, that is, 0+1=1, and further increments the address data output to the spectral sensitivity memory 1003, that is, 0+
1=1, and input α ₂ ⁽⁰⁾ and W ₂ . The pattern matching processor 1001 applies W ₂ (α ₂ ⁽⁰⁾ −
Add the quantities calculated in α ₂ ^(I) ) ² and get D=W ₁ (α ₁ ⁽⁰⁾ − _{α 1} ^(I) ) ² + W ₂ (α ₂ ⁽⁰⁾ − α ₂ ^(I) ) ² = Find ₂ 〓 ^K=1 Wk (α _k ⁽⁰⁾ − α _k ^(I) ) ² . Similarly, the pattern matching processor 10
01 gives address data 2, 3, 4, ... 10 one after another to the standard pattern memory 1002, and gives address data 2, 3, 4, ... one after another to the spectral sensitivity memory 1003. D= ₁₀ 〓 ^k=1 Wk ( α _k ⁽⁰⁾ −α _k ^(I) ) ² ...... (3) is calculated.

D given by equation (3) is input as the standard pattern α ⁽⁰⁾
This is the distance D _0I from the LPC coefficient data α ^(I) .

Next, the pattern matching processor 1001 compares the data D ₀ stored in the minimum distance register 1004 with the D _0I , and if D ₀ > D _0I , rewrites D _0I as D ₀ in the minimum distance register 1004. Set,
Further, the label register 1005 is reset to "0" indicating the label "0". If D ₀ > D _0I , the pattern matching processor 1001 registers the minimum distance register 1.
004 and the label register 1005 are not reset. In the case of D ₀ =D _0I , the resetting may or may not be performed.

Next, the pattern matching processor 1001 stores the start address 1〓10=10, which is equivalent to the label "1" attached to the standard pattern, into the standard pattern memory 1002.
The address “0” specifying the spectral sensitivity W ₁ corresponding to the LPC coefficient α ₁ is stored in the spectral sensitivity memory 1003.
and set the spectral distance D to "0". Furthermore, the pattern matching processor 1001 calculates D= _{W 1 ( α 1 (} ^{1 ) −} α ₁ ^(I) ) Calculate ² . Next, the pattern matching processor 1001 supplies address data 11, 12, . . . to the standard pattern memory 1002.
20 and given one after another, the spectral sensitivity memory 1003
The address data to be supplied to 1, 2, . . . 10 is given one after another, and D= ₁₀ 〓 ^k=1 Wk (α _k ⁽¹⁾ − α _k ^(I) ) ² …… (4) is calculated. D given by equation (4) is the distance D _1I between the standard pattern α ⁽¹⁾ and the input LPC coefficient data α ^(I) .
Next, the pattern matching processor 901 uses the data D ₀ stored in the minimum distance register 904 and the data D _1I
The minimum distance register 1004 and label register 1005 are rewritten as appropriate.

Below, the labels "2", "3", ..., "4095" one after another
Perform comparison with the standard pattern.

Finally, the pattern matching processor 1001 sends the contents of the label register to the multiplexer 1 as encoded data.
The standard pattern data corresponding to the contents of the label register is output to the LPC inverse filter 4 as decoded data.

The explanation will be given by returning to FIG. 1 again.

The LPC inverse filter 4 is a filter whose impulse response characteristic is opposite to that of the LPC synthesis filter 16 on the synthesis side shown in FIG. The quantizer/decoder 3 receives the decoded LPC coefficient data, and then the A/D
Only data related to the residual waveform of the quantized input audio data received from the output of the converter 1 is extracted and supplied to the LPF 5, filter coefficient estimator (1) 8, and power calculator 10.

LPF5 is a low-pass filter with high-frequency cutoff characteristics that is set taking into consideration the data bit rate of the vocoder and other conditions.
Among the residual waveforms received from , 1 KHz is set as a high cut-off frequency, and low frequency components of 1 KHz or less of the residual waveforms are supplied to the down sampler 6 .

The down sampler 6 samples the input low frequency components of 1KHz or less at a preset rate of the reference sampling frequency, in this example, the downsampling frequency is 2KHz, which is down to 1/4, and quantizes/decodes this. It is sent to the converter 7.

The quantizer/decoder 7 quantizes the input residual waveform low frequency component and sends it to the multiplexer 12, and also sends the decoded data to the filter coefficient estimator (1) 8. supply The reason why the downsampled data supplied to the filter coefficient estimator (1) 8 is the data that has been decoded by the quantizer/decoder 7 is to avoid the influence of quantization errors in the high frequency reproduction filter 14 on the synthesis side. This is to make the influence of the quantization error on the filter coefficient estimator (1) 8 almost the same.

Note that, in this embodiment, the quantization and decoding processing of the residual waveform low frequency component in the quantization/decoder 7 is normalized and executed using power data supplied from a power calculator 10, which will be described later. The normalization process efficiently quantizes the low frequency components of the residual waveform.

Filter coefficient estimator (1) 8 is LPC inverse filter 4
A sample sequence of the residual waveform with a reference sampling frequency of 8KHz is input from the quantizer/decoder 7, and a downsample sequence of the low frequency component of the residual waveform with a downsampling frequency of 2KHz is input from the quantizer/decoder 7. In this way, the tap coefficient of the high frequency reproduction filter to be provided on the synthesis side is estimated.

FIG. 2a is a residual waveform reference sample diagram showing an example of the residual waveform obtained by sampling the residual waveform of the input audio signal at the reference sampling frequency in the first embodiment of the present invention shown in FIGS. 1A and B; Figure 2b
Figure 2c is a downsample sample diagram showing an example of a downsample in which the low frequency component of the residual waveform is sampled at the downsampling frequency.
This is a residual waveform low frequency component reference sampling conversion sample sequence obtained by performing reference sampling conversion corresponding to the reference sampling frequency on the down sample sample of the residual waveform low frequency component shown in FIG. The embodiments shown in FIGS. 1A and 1B will be described below with reference to FIGS. 2A, 2B, and 2C.

The residual waveform sample sequence shown in FIG. 2a is the 8KHz sampling output from the LPC inverse filter 4,
The residual waveform low frequency component down-sampled sample shown in Figure 2b is the 2KHz sampling output from the quantizer/decoder 7, which is further converted into a sample with a reference sampling frequency of 8KHz as the second sample.
a ₁ , a ₂ , a ₃ , a ₄ , a ₅ , a ₆ and b ₁ to b ₅ shown in figure c
This is the sample group. Among these sample groups, a ₁ to a ₆ correspond to the samples shown in Figure 2 b, and
b ₁ to _{b 5} are interpolated between samples a ₁ to _{a 6} as a whole
These are zero-level sampling points arranged at equal intervals in order to convert the standard sampling frequency to 8KHz, and in this way, the low-frequency components of the residual waveform can be down-sampled with an image similar to that of a sample using the standard sampling frequency of 8KHz. Standard sampling conversion of samples can be achieved.

The filter coefficient estimator (1)8 is thus 2K
A downsampling sampling sequence with a downsampling frequency of Hz is converted into a sequence of sampling images with a reference sampling frequency of 8KHz, and this is supplied as an input to a built-in filter similar to the transversal filter as a high-frequency reproduction filter that should be provided on the synthesis side. do. The tap coefficients of this transversal filter are determined using the learning identification method or the A-
It is controlled and determined for each analysis frame based on a method such as the b-S (Analysis-by-Synthesis) method.

That is, in the filter coefficient estimator (1) 8,
The sampling sequence of the low frequency component of the residual waveform of 2KHz sampling shown in Figure 2b is input.This is converted into the 8KHz sampling image sampling sequence shown in Figure 2b, which is applied to the built-in transversal filter, and is used as audio material, etc. The error output obtained by comparing the output of the transversal filter to which the initial value of the tap coefficient set in advance is given and the 8KHz sampling sequence of the residual waveform as shown in Figure 2a, which is input separately, is obtained. obtain. In the case of this embodiment, the power of this error output is calculated, and the filter coefficients are estimated so that the error power can be minimized while controlling the transversal filter coefficients for each analysis frame in accordance with the magnitude of the error power. is carried out using methods such as the learning identification method and the A-b-S method.

Figure 3 shows the filter coefficient estimator (1) shown in Figure 1A.
FIG. 8 is a block diagram showing in detail the configuration of FIG.

The filter coefficient estimator (1) 8 includes a transversal filter 81, a subtracter 82, and a power calculator 83, which have basically the same configuration as the high-frequency reproduction filter to be provided on the synthesis side. Out of the output of the quantizer/decoder 7, the down-sampled residual waveform low frequency component as shown in Figure 2b is input, and while receiving the 8KHz sampling frequency, the filter coefficients are estimated as follows. Let's do it.

That is, the transversal filter 81 is
When the residual waveform low frequency component down sample as shown in Figure 2b is input, the second
Convert the residual waveform low frequency component into a reference sampling conversion sampling sequence as shown in Figure c, send the transversal filter output based on the initial setting tap coefficient to the subtracter 82 via the output line 811, and output it to the LPC inverse filter 4. The difference from the residual waveform is taken, and this error output is supplied to a power calculator 83 via an output line 821, and its power is calculated.

The filter coefficient estimator (1) 8 controls the tap coefficient of the transversal filter 81 based on the error power calculated in this way and outputs the tap coefficient as the filter coefficient estimated when the error power can be minimized. After being converted into quantized data by the quantizer 9, it is supplied to the multiplexer 12.

In this way, the 8KHz sampling image sampling sequence of down-sampling of the low frequency component of the residual waveform is created, and the coefficients of the filter can be estimated without using a nonlinear circuit based on the learning identification method or the A-b-S method. It becomes possible.

The power calculator 10 inputs the residual waveform, calculates the short-term average audio power for each analysis frame, sends it to the quantizer 11 and the quantizer/decoder 7, performs predetermined quantization, and then A multiplexer 12 is provided.

This short-time average audio power is used to set the level of the audio to be reproduced when the input audio signal is reproduced by the LPC synthesis filter 16 on the synthesis side. This data transmission is not necessary when the value is approximately constant over time.

The multiplexer 12 multiplexes each of the quantized data thus supplied using a preset method and sends the multiplexed data to the combining side via a transmission path.

On the synthesis side in Figure 1B, each quantized data that has been multiplexed and supplied from the analysis side is demultiplexed and demultiplexed by the demultiplexer/decoder 13, and further decoded, as shown in Figure 2B. The downsampled sampled data of the low frequency component of the residual waveform is supplied to the high frequency reproduction filter 14 via an output line 1301, and the estimated filter coefficient data is also supplied to the high frequency reproduction filter 14 via an output line 1302. In addition, the short-time average audio power data is input to the input line 1303.
The LPC coefficient data, which has been converted into a label format, is sent to the multiplier 16 via the multiplier 16, and is decoded by the same memory as the standard pattern memory 1002 on the analysis side, which is built into the demultiplexer/decoder 13. Reads out the LPC coefficient data and sends it to output line 13.
04 to the LPC synthesis filter 16.

The high frequency reproduction filter 14 has a built-in transversal filter that uses the estimated filter coefficients as tap coefficients, and converts the input downsampling sample into an 8KHz sampling image sample as shown in Figure 2c. As a filter input, residual waveform data of 8KHz sampling as shown in FIG. 2a is reproduced. This residual waveform data is supplied to the multiplier 15 and input line 1303
The level of the reproduced residual waveform data is corrected through multiplication with the short-term average voice power inputted via the LPC synthesis filter 16.

The LPC synthesis filter 16 is connected to the input line 130.
The input audio signal is digitally reproduced using the LPC coefficients input through 4 as filter coefficients and the reproduced residual waveform data as drive sound source information, and is supplied to the D/A converter 17, which converts it into an analog quantity and further incorporates it. Only a predetermined low frequency component is outputted to the output terminal 1701 by the LPF.

FIG. 4 is a block diagram showing the configuration of the analysis side in the second embodiment of the present invention. The configuration of the analysis side shown in FIG. 4 differs from FIGS. 1A and 3 only in the contents of the filter coefficient estimator (2) 18, and the other components with the same symbols are exactly the same, so we will not be concerned about these. Detailed explanation will be omitted.

Further, since the composition side in the second embodiment has the same configuration as the composition side in the first embodiment shown in FIG. 1B, detailed explanation thereof will also be omitted.

The filter coefficient estimator (2) 18 in the second embodiment shown in FIG. 4 is different from the filter coefficient estimator (1) 8 shown in FIG.
This method attempts to estimate the filter coefficients of a transversal filter that has an impulse response that minimizes the difference from the residual waveform of the input audio signal by using methods such as the method of changing the filter coefficients sequentially. The difference is that the filter coefficients are estimated by a so-called forward filter coefficient setting method, which attempts to determine a transversal filter with an optimal impulse response.

Figure 5 shows the filter coefficient estimator (2) 1 shown in Figure 4.
FIG. 8 is a block diagram showing in detail the configuration of FIG.

The filter coefficient estimator (2) 18 in FIG. 5 includes a cross-correlation circuit 181, an autocorrelation circuit 182, and a maximum value search circuit 183.

When the residual waveform low frequency component downsampling as shown in FIG. This down-sampling by frequency is converted to an apparent 8KHz sampling sequence while interpolating zero-level samples as shown in Figure 2c, and then the output from the LPC inverse filter 4 as shown in Figure 2a is Calculate the cross-correlation coefficient with the residual waveform of 8KHz sampling and use this as maximum value search circuit 1
83.

Further, the autocorrelation circuit 182 generates a sampling sequence as shown in FIG. Calculate the autocorrelation coefficient sequence. An 8KHz sampling frequency is used to calculate these cross-correlation coefficient sequences and autocorrelation coefficient sequences.

Now, the cross-correlation coefficient sequence supplied to the maximum value search circuit 183 searches for its maximum value within a preset limited delay time. In this case, the tap position corresponding to the delay time is, for example, tap γ, the cross-correlation coefficient at this tap γ is φγ, and the auto-correlation coefficient at zero delay of the auto-correlation coefficient sequence obtained in the auto-correlation circuit 182 is If ψο, then the value of φγ divided by ψο is Cγ = φγ/ψο
is the tap coefficient at tap γ of the transversal filter.

Next, an operation is performed to subtract the autocorrelation coefficient sequence multiplied by Cγ from the cross-correlation coefficient sequence described above so that the tap position γ coincides with the position of the autocorrelation value ψο,
While repeating such calculations, Cγ is successively obtained, all filter coefficients of the transversal filter are determined, and these are output as estimated filter coefficients. In this way, the correction message shown in FIG.
This is the data required to perform subsequent calculations and outputs.

In this way, forward coefficient settings can be performed to obtain the maximum impulse response based on continuous changes in the coefficients of the transversal filter, and the analysis and synthesis processing of the residual excitation type vocoder, which significantly suppresses RELP sound, can be performed stably and reliably. be able to.

FIGS. 6A and 6B are block diagrams showing the configurations of analysis side A and synthesis side B, respectively, in a third embodiment of the present invention.

In FIG. 6A, filter coefficient estimator (3) 19
However, in FIG. 6B, the coefficient regenerator (1) 20 and the high frequency regeneration filter 21 are the same as those in FIG. 1A, respectively.
The only difference from or addition to that in B is that everything else is the same, so detailed explanations regarding these will be omitted.

The filter coefficient estimator (3) 19 shown in FIG. 6A is
In addition to the filter coefficient estimator (1) 8 shown in FIG. Convert it to , remove the zero phase position, and output it.

FIG. 7A is a high-frequency reproduction filter prediction coefficient characteristic diagram formed on the analysis side of the third embodiment shown in FIG. 6A, and FIG. 7B is a zero-phase prediction coefficient characteristic diagram of FIG. 7A. . The embodiment shown in FIGS. 6A and 6B will be described below with reference to FIGS. 7A and 7B.

The filter coefficient estimator (3) 19 first generates estimated filter coefficients as shown in FIG. 7A in substantially the same manner as the filter coefficient estimator (1) 8 shown in FIG. 1A. Next, the estimated filter coefficients are converted from a group of coefficients as a complex spectrum by a transversal filter to a group of coefficients shown in FIG. 7B as a power spectrum by a phase linear filter having substantially the same impulse response characteristics. The coefficient group formed by this conversion loses phase information, but has the same number of taps as the transversal filter coefficient group in Figure 7A, and the energy center position remains unchanged.
Moreover, a symmetrical tap coefficient, that is, an impulse response sequence can be obtained. In this way, if the energy center position remains unchanged, even if phase information is lost, there is practically no problem as a coefficient of a high frequency reproduction filter.

Now, since the phase linear filter coefficient group shown in Figure 7B has left-right symmetry, the whole can be easily reproduced even if only approximately 1/2 of the total, either on the left or right of the energy center position, is sent to the synthesis side as estimated coefficients. It is possible to do so. For this reason, from the filter coefficient estimator (3) 19, only approximately half of the coefficient group of the entire coefficient group is sent to the quantizer 9, where it is subjected to predetermined quantization and supplied to the multiplexer 12. The quantized data is multiplexed with other quantized data and transmitted to the combining side in FIG. 6B.

On the synthesis side shown in FIG.
Supply to 0. The estimated coefficient data thus supplied is approximately 1/2 including the center tap position of the coefficient group shown in FIG. 6B. Coefficient regenerator (1) 20
On the basis of this coefficient group, the symmetrization coefficient group shown in FIG. The reproduced input audio signal is outputted to the output terminal 1701 in substantially the same manner as described with reference to FIG.

FIGS. 8A and 8B are block diagrams showing the configurations of analysis side A and synthesis side B in a fourth embodiment of the present invention.

Figure 8A is the same as Figure 1A and B except for the filter coefficient estimator (4) 22, and Figure 8B is the same as Figure 1A and B except for the coefficient regenerator (2) 23 and the high frequency reproduction filter 24, Detailed explanation will be omitted.

The fourth embodiment shown in FIGS. 8A and 8B is as shown in FIG. 1A.
The tap coefficient group of the high-frequency reproducing filter obtained in the same manner as the filter coefficient estimator (1) 8 shown in Fig. 8 is an estimated coefficient sequence of a triangular envelope array having almost the same power spectrum and the energy center position unchanged. This method aims to reduce the transmission data rate by converting the triangular wave envelope so that be.

FIG. 9A is a high frequency reproduction filter prediction coefficient characteristic diagram formed on the analysis side of the fourth embodiment shown in FIG. 8A, and shows basically the same contents as FIG. FIG. 9B is a triangular envelope distribution transformation characteristic diagram obtained while keeping the estimated coefficient group of FIG. 9A in substantially the same power spectrum.

In FIG. 9B, the coefficient group in FIG. 9A is transformed so that its approximate envelope becomes a desired triangular shape. This transformation is realized by Fourier transforming the coefficient group shown in FIG. 9A, and adjusting the phase of the Fourier transformed value in FIG. 9A so as to match the phase shift of the desired triangular coefficient.

The motivation for making the envelope of the transformation coefficient group triangular is based on the fact that the vocal fold vibration waveform generally takes a triangular waveform, and this kind of transformation causes loss of phase information. However, due to this envelope shape and the characteristic of the coefficient array in which the energy center position remains unchanged, there is almost no effect on the quality of the reproduced audio by synthesis.

In this way, after being converted into a coefficient series that shows a triangular envelope distribution and has almost the same power spectrum, only the difference component from the desired triangular shape is outputted from the filter coefficient estimator (4) 22, and the quantizer 9 Send it to the multiplexer 12 as quantized data via
Now it will be transmitted to the synthesis side.

On the synthesis side shown in FIG.
23, thereby adding the desired triangular shape subtracted on the analysis side to reproduce the coefficient group, and supplying these to the high frequency reproduction filter 24.

The high frequency reproduction filter 24 uses these coefficient groups as its tap coefficients, and inputs the residual waveform low frequency component which has been converted into an 8KHz sampling image via the output line 1302, and generates the total residual waveform including the high frequency component. Play.

The synthesis side then performs the same playback operation as shown in Figure 1B to play back the input audio signal, thus basically removing the RELP sound and also changing the transmission data bits of the tap coefficient of the high-frequency playback filter. Analysis and synthesis processing can be performed at significantly reduced rates.

In the above explanation, the power calculators 10 are provided in each of FIGS. 1A, 4, 6A, and 8A, and the first
Although multipliers 15 were provided in each of FIG. B, FIG. 6B, and FIG. If absolute level quantization and decoding processing are performed without performing quantization processing, the power calculator 10 and multiplier 15 are unnecessary.

FIG. 11 is a block diagram showing the configuration of the analysis side in the fifth embodiment of the present invention. On the analysis side shown in FIG. 11, the same components as in the first embodiment shown in FIG. 1 are denoted by the same symbols, and detailed explanations thereof will be omitted.

Further, since the composition side in the fifth embodiment has the same configuration as the composition side in the first embodiment shown in FIG. 1B, detailed explanation thereof will also be omitted.

The fifth embodiment shown in FIG. 11 transmits the spectral envelope information of the input audio signal using the pattern matching means. 1. LPC coefficients obtained by analyzing the input audio and a standard pattern with the smallest spectral distance; 2 with the standard pattern with the second smallest spectral distance
2. Using each of the two selected standard pattern candidates, each independently calculates the residual waveform and filters the filter coefficients in the first embodiment. It executes a series of analysis processes such as estimation, and further executes audio reproduction processing that would normally be executed on the synthesis side.

3. Furthermore, measure the S/N ratio between each of the two types of reproduced sounds corresponding to the two types of preliminarily selected standard pattern candidates and the input sound.

4. Determine the standard pattern candidate with a good S/N ratio as the standard pattern.

As a result, more appropriate pattern matching can be performed in the residual excitation type vocoder.

A/D converter 1 outputs an output to LPC analyzer 2 and waveform memory 29. LPC analyzer 2
outputs the analyzed LPC coefficients to the quantizer/decoder 25. The quantizer/decoder 25 is, for example, the 12th
The configuration shown in the figure is one in which an auxiliary label register 1201 is added to the quantizer/decoder 3. The quantizer/decoder 25 will be explained below using FIG. 12. Pattern matching processor 10
01 initializes the minimum distance register 1004 to a value D ₀ larger than the empirically known maximum value of Dij shown in equation (2) above when the LPC coefficient is input. Furthermore, the label register 1005 and the auxiliary label register 1201 are initialized to, for example, "4096". Next, the pattern matching processor 1001 transfers the start address 0〓10=0 (10 of 0〓10 is equivalent to the order 10 of the LPC) to the standard pattern memory 1002, which is equivalent to the label "0" attached to the standard pattern, to the standard pattern memory 1002.
The address 0 specifying the spectral sensitivity W ₁ corresponding to the coefficient α ₁ is output to the spectral sensitivity memory 1003, and the spectral distance D is set to "0". Next, the pattern matching processor 1001 calculates α ₁ ⁽⁰⁾ (equivalent to P ₁ ⁽ⁱ⁾ in equation (2) above) of the standard pattern with the label “0” output from the standard pattern memory 1002.
and the spectral sensitivity W ₁ output from the spectral sensitivity memory 1003, and the first coefficient α ₁ ^(I) of the LPC coefficients supplied from the LPC analyzer 2 (P ₁ ( ^{j) in the above equation (2)} ), calculate D=W ₁ (α ₁ ⁽⁰⁾ − α ₁ ^(I) ) ² . Next, the pattern matching processor 1001
Increments the address data output to the standard pattern memory 1002, that is, 0+1=1, and further increments the address data output to the spectral sensitivity memory 1003, that is, 0+
1=1, and input α ₂ ⁽⁰⁾ and W ₂ . The pattern matching processor 1001 applies W ₂ (α ₂ ⁽⁰⁾ −
Add the quantities calculated in α ₂ ^(I) ) ² and get D=W ₁ (α ₁ ⁽⁰⁾ − _{α 1} ^(I) ) ² + W ₂ (α ₂ ⁽⁰⁾ − α ₂ ^(I) ) ² = Find ₂ 〓 ^k=1 W _k (α _k ⁽⁰⁾ − α _k ^(I) ) ² . Similarly, the pattern matching processor 10
01 gives address data 2, 3, 4, ... 10 one after another to the standard pattern memory 1002, and gives address data 2, 3, 4, ... 10 one after another to the spectral sensitivity memory 1003, D= ₁₀ 〓 ^k=1 W Calculate _k (α _k ⁽⁰⁾ − α _k ^(I) ) ² ...(3).

D given by equation (3) is input as the standard pattern α ⁽⁰⁾
This is the distance D _0I from the LPC coefficient data α ^(I) .

Next, the pattern matching processor 1001 compares the data D ₀ stored in the minimum distance register 1004 with the D _0I , and if D ₀ > D _0I , rewrites D _0I as D ₀ in the minimum distance register 1004. Set,
Further, the label register 1005 is reset to "0" indicating the label "0". Next, the pattern matching processor 1001 transfers “4096” stored in the label register 1005 to the auxiliary label register 1201.
Output to. In the case of D ₀ >D _0I , the pattern matching processor 1001 does not reset the minimum distance register 1004 and the label register 1005. In the case of D ₀ =D _0I , the resetting may or may not be performed.

Next, the pattern matching processor 1001 stores the start address 1〓10=10, which is equivalent to the label "1" attached to the standard pattern, into the standard pattern memory 1002.
The address “0” specifying the spectral sensitivity W ₁ corresponding to the LPC coefficient α ₁ is stored in the spectral sensitivity memory 1003.
and set the spectral distance D to "0". Further, the pattern matching processor 1001 uses α ₁ ⁽¹⁾ of the standard pattern labeled “1”, the above W ₁ and the above α ₁ ⁽
D=W ₁ (α ₁ ⁽¹⁾ − α ₁ ^(I) ) ² is calculated from ^I) . Next, the pattern matching processor 1001 supplies address data 11, 12,..., 20 to the standard pattern memory 1002.
are given one after another, and the address data to be supplied to the spectral sensitivity memory 1003 is given one after another as 1, 2, ...10, D= ₁₀ 〓 ^k=1 W _k (α _k ⁽¹⁾ − α _k ^(I) ) ² ... ...(4) is calculated. D given by equation (4) is the distance D _1I between the standard pattern α ⁽¹⁾ and the input LPC coefficient data α ^(I) .
Next, the pattern matching processor 1001 compares the data D ₀ stored in the minimum distance register 1004 with the data D _1I and stores the data in the minimum distance register 100 as appropriate.
5 and the label register 1005, and further saves the contents of the label register 1005 to the auxiliary label register 1201.

Below, the labels "2", "3", ..., "4095" are displayed one after another.
Perform comparison with the standard pattern.

The pattern matching processor 1001 converts the contents of the label register 1005 or auxiliary label register 1201 into coded data to the multiplexer 2 as necessary.
6, and outputs standard pattern data corresponding to the contents of the label register 1005 or the auxiliary label register 1201 to the LPC inverse filter 4 and the LPC synthesis filter 16.

The quantizer/decoder 7 quantizes the residual waveform low frequency component and sends it to the multiplexer 26, and also sends the decoded data to the filter coefficient estimator (1) 8 and the high frequency reproduction filter. 14.
Filter coefficient estimator (1) 8 outputs estimated filter coefficients to quantizer 9. A quantizer 9 quantizes the coefficients and supplies them to a multiplexer 26 and a decoder 28. The decoder 28 decodes the quantized filter coefficients and outputs them to the multiplier 15. Power calculator 10
inputs the residual waveform, calculates the short-term average audio power for each analysis frame, and outputs this to the quantizer 11 and the quantizer/decoder 17. The quantizer 11 quantizes the short-term average voice power and outputs it to the multiplexer 26 and the decoder 28. The decoder 28 decodes the short-time average power and outputs it to the multiplier 15.

The high frequency reproduction filter 14 reproduces the high frequency from the residual waveform low frequency component supplied from the quantizer/decoder 7 and outputs it to the multiplier 15 . The multiplier 15 multiplies the waveform by the short-term average audio power.
Output to LPC synthesis filter 16. The LPC synthesis filter 16 reproduces the audio and outputs it to the S/N ratio calculator 30. The waveform memory 29 is an S/N ratio calculator 30
Outputs the input audio waveform to. S/N ratio calculator 30
calculates the waveform of the difference between the reproduced audio waveform and the input audio waveform (hereinafter referred to as the "error waveform"), and calculates its power P _N
Calculate. Furthermore, the power P _I of the input audio waveform is calculated, the S/N ratio, that is, P _N /P _I is calculated, and the multiplexer 26
Output to.

The above SN calculation process is performed using the label register 1 mentioned above.
005 and the contents of the auxiliary label register 1004 are executed once each. The multiplexer 26 compares the S/N ratio regarding the contents of the label register 1005 and the S/N ratio regarding the contents of the auxiliary label register 1201, and selects a label with a good S/N ratio and a filter coefficient etc. corresponding to the label. Output to the synthesis side.

Of course, the present invention can also be implemented by using the power of the error waveform as the determination parameter instead of calculating the S/N ratio. Further, the quantizer/decoder 25 has at most one auxiliary label register, but the number is not necessarily limited to one. Furthermore, the filter coefficient estimator (1) 8 is the filter coefficient estimator (2) 18 or the filter coefficient estimator (3) 1
9 can be replaced.

In the first to fifth embodiments, the α parameter was used as a pattern matching parameter, but it is clear that this can also be implemented using the K parameter, LSP parameter, etc.

〔Effect of the invention〕

As explained above, according to the present invention, in a residual excitation type vocoder having a means for transmitting spectral envelope information by a pattern matching means and having a high frequency component reproducing means on the analysis side, the residual waveform Means for estimating the tap coefficient of a high-frequency reproduction filter by using a sample sequence in which a down-sampled value of a low-frequency component is made into a reference sampling conversion sample sequence through interpolation of a zero-level sample and without using a nonlinear circuit. By analyzing and synthesizing the input audio signal using the system, it is possible to realize a residual excitation type vocoder that can significantly improve the RELP sound in an extremely stable and reliable manner.

[Brief explanation of the drawing]

FIG. 1A is a block diagram showing the configuration on the analysis side in the first embodiment of the present invention, FIG. 1B is a block diagram showing the configuration on the synthesis side, and FIG. A residual waveform standard sample diagram showing an example of a residual waveform obtained by sampling the residual waveform of an input audio signal at a standard sampling frequency in the first embodiment of the invention, FIG. 2b shows a low frequency component of the residual waveform. A down-sampling sample diagram showing an example of down-sampling sampled at the down-sampling frequency, Fig. 2 c is a reference sampling conversion corresponding to the reference sampling frequency for the down-sampling sample of the residual waveform low frequency component shown in Fig. 2 b. FIG. 3 shows the reference sampling conversion sample sequence of the low frequency component of the residual waveform obtained by performing the following steps.
4 is a block diagram showing the configuration of the analysis side in the second embodiment of the present invention, and FIG. 5 is a block diagram showing the configuration of the filter in the second embodiment of the present invention shown in FIG. A block diagram showing in detail the configuration of the coefficient estimator (2) 18, FIG. 6A is a block diagram showing the configuration of the analysis side in the third embodiment of the present invention,
Figure 6B is a block diagram showing the configuration of the synthesis side, Figure 7
Figures A and B are a high frequency reproduction filter prediction coefficient characteristic diagram A formed on the analysis side of the third embodiment of the present invention shown in Figure 6A, its zero-phase prediction coefficient characteristic diagram B, Figure 8A, B is a block diagram showing the configuration of the analysis side A and the synthesis side B in the fourth embodiment of the present invention, and FIGS. 9A and B are the high regions formed on the analysis side in the fourth embodiment of FIG. The reproduction filter prediction coefficient characteristic diagram A and the triangular envelope distribution transformation characteristic diagram B based on the same power spectrum, FIG. 10 show the configuration of the quantizer/decoder 3 on the analysis side in the first embodiment shown in FIG. Detailed block diagram,
FIG. 11 is a block diagram showing the configuration of the analysis side in the fifth embodiment of the present invention, and FIG. 12 shows the detailed configuration of the quantizer/decoder 25 on the analysis side in the fifth embodiment shown in FIG. FIG. 1...A/D converter, 2...LPC analyzer,
3...Quantizer/decoder, 4...LPC inverse filter, 5...LPF, 6...Down sampler, 7...
...quantizer/decoder, 8...filter coefficient estimator
(1), 9...Quantizer, 10...Power calculator, 11
... Quantizer, 12 ... Multiplexer, 13 ... Demultiplexer/decoder, 14 ... High frequency reproduction filter, 1
5... Multiplier, 16... LPC synthesis filter, 1
7...D/A converter, 18...Filter coefficient estimator (2), 19...Filter coefficient estimator (3), 20
... Coefficient regenerator (1), 21 ... High frequency regeneration filter,
22... Filter coefficient estimator (4), 23... Coefficient regenerator (2), 24... High frequency reproduction filter, 25... Quantizer/decoder, 26... Multiplexer, 27... Decoding encoder, 28... decoder, 29... waveform memory, 30... S/N ratio calculator, 81... transversal filter, 82... subtractor, 83... power calculator, 181... mutual Correlation circuit, 182...
Autocorrelation circuit, 183... Maximum value search circuit, 10
01...Pattern matching processor, 1002...Standard pattern memory, 1003...Spectral sensitivity memory, 1004...Minimum distance register, 1005
... Label register, 1201 ... Auxiliary label register.

Claims (1)

[Claims] 1. A first method that compares the spectral envelope characteristic of an input audio signal with a standard pattern prepared in advance, and transmits a label attached to the standard pattern from the analysis side to the synthesis side as information indicating the spectral envelope characteristic. means for determining a reference sample sequence of a residual waveform of an input audio signal corresponding to a preset reference sampling frequency; and a second means corresponding to a downsampling frequency obtained by reducing the reference sampling frequency at a preset rate. a third means for obtaining a down-sampled sequence of low frequency components of the residual waveform; and the residual waveform is formed while setting a zero-level interpolation point in the down-sampled sequence at the timing of the reference sampling frequency. Estimating the coefficients of a transversal high-frequency reproduction filter to be provided on the synthesis side based on the reference sampling conversion sample sequence of the low-frequency component of and the reference sample sequence of the residual waveform and without using a nonlinear circuit. A residual excitation type vocoder comprising a fourth means for outputting the coefficients to be transmitted to a synthesis side. 2. Compare the spectral envelope characteristics of the input audio signal with a standard pattern prepared in advance, select a plurality of pattern candidates in advance, and determine the coefficients of a transversal high-frequency reproduction filter to be provided on the synthesis side for each pattern candidate. a first means for estimating on the analysis side; a second means for calculating the S/N ratio between the input speech signal and the synthesized speech signal corresponding to each pattern candidate; A third means for transmitting labels attached to pattern candidates with a good S/N ratio from the analysis side to the synthesis side as information indicating spectral envelope characteristics, and a third means for transmitting labels attached to pattern candidates with a good S/N ratio from the analysis side to the synthesis side, and a fourth means for obtaining a reference sample sequence of the difference waveform; and a fourth means for obtaining a downsample sequence of low frequency components of the residual waveform corresponding to a downsampling frequency obtained by reducing the reference sampling frequency at a preset rate. and a reference sampling conversion sample sequence of the low frequency component of the residual waveform formed while setting a zero-level interpolation point in the down-sampling sequence at the timing of the reference sampling frequency and the residual waveform. A sixth means for estimating coefficients of a transversal high-frequency reproduction filter to be provided on the synthesis side based on the reference sample sequence and without going through a nonlinear circuit, and outputting the coefficients to be transmitted to the synthesis side. A residual excitation type vocoder characterized by: