Test us for second group and selected the different sample rate one group sound signal different to come the coded system that we propose is tested, prove that with this this scheme is in the combination of various sampling rates and quantified precision effect preferably being arranged with quantified precision.

The audio file compression result of different sample rate and quantified precision relatively

Two groups of above experiment proofs all can obtain compression effectiveness preferably for music file the inventive method of different-style, and also can obtain than the suitable result of current main-stream audio frequency lossless compress software for audio file the inventive method of different sampling rates and quantified precision.

Compared with prior art, good effect of the present invention is:

1, native system can carry out the branch frame to signal according to the correlation circumstance self-adaptation of frame before and after the signal, makes that the signal in the frame has strong correlation, handles for the wavelet decomposition of back and predicts with LPC and to bring benefit.

2. native system adopts integer lifting wavelet transform, has avoided the error that deal with data and filter coefficient are blocked generation.

3. native system is majorized function with the optimal compression value, and the integer Lifting Wavelet progression of signal is adjusted according to the unlike signal self-adaptation.

4. the signal after the wavelet decomposition is adopted the LPC prediction processing, increase the compactness of decomposing the back signal.

5. with this sign indicating number of the Lay that is fit to geometric distributions residual signals is encoded, signal is further compressed.

Description of drawings

Fig. 1: coder structure block diagram;

Fig. 2: decoder architecture block diagram;

Fig. 3: divide the frame strategy;

Fig. 4: the decomposition of Lifting Wavelet;

Fig. 5: the reconstruct of Lifting Wavelet;

Fig. 6: integer wavelet transform with and inverse transformation;

Fig. 7: forward prediction device code pattern;

Fig. 8: forward prediction device decoding figure.

Embodiment

Following reference accompanying drawing of the present invention is described most preferred embodiment of the present invention in more detail, describes the technical scheme that how to realize this invention in detail:

Audio frequency according to the inventive method can't harm the coder/decoder system comprising encoder subsystem and Decoder Subsystem two parts.The structured flowchart of system is as depicted in figs. 1 and 2, and wherein Fig. 1 is an audio frequency lossless compression-encoding subsystem structure block diagram, and Fig. 2 is a frequency lossless compress Decoder Subsystem structured flowchart.Below will combine the detailed introducing system structure of accompanying drawing.

1, overall plan:

Encoder section: the coder structure block diagram is as shown in Figure 1: audio file is divided into some frames according to a minute frame strategy earlier, divides frame information to include the side information transmission in; With the every frame individual processing behind minute frame, promptly earlier logical wavelet transformation obtains the approximate signal and the detail signal of each frame, and wavelet decomposition progression is pressed adaptation rule and obtained, and decomposed class is included side information equally in; Approximate signal and detail signal are through the linear prediction module, and the residual signals that in linear prediction module, obtains is through entropy coding, and the LPC parameter is included side information in, at last each road code stream multiplex is formed final compressed bit stream.

Decoder section: the decoder architecture block diagram is as shown in Figure 2: in fact be exactly the inverse process of scrambler, through decoding side information earlier, therefrom isolate the code book of entropy coding; Solve LPC prediction back residual error; Utilize the LPC parametric solution to get the approximate signal and the detail signal of wavelet decomposition, according to the information of wavelet decomposition progression, reconstruct obtains every frame signal again; According to minute frame information each frame is coupled together in order the harmless original audio file that obtains at last.

2, divide the frame strategy:

Owing to have considerable jump signal in the sound signal, if adopt anchor-frame progress row to divide frame, the correlativity between the signal in each frame that obtains can be a greater impact, and then makes compressibility reduce.This programme is according to the related coefficient of consecutive frame, and the signal that correlativity is big merges in the frame, and like this, the compactness of wavelet transformation and linear prediction all can improve, and can obtain higher compression efficiency.So take branch frame strategy in this programme, the consistent signal of characteristic can be handled in a frame like Fig. 3.

Be unit at first, calculate the related coefficient of present frame and former frame, if this coefficient is less than threshold value with minimum frame length; Then this frame of mark and former frame are uncorrelated frame; Become separately a frame, if this coefficient, is then thought present frame and former frame associated frame greater than threshold value; Adjacent associated frame is merged successively, but maximum frame length is no more than the maximum frame length permissible value of setting.

3, integer lifting wavelet transform:

Promote strategy: people such as Swledens are verified, and any existing wavelet transformation can be realized through the lifting step cascade of limited number of time, and in addition, promoting the conversion implementation framework can realize with integer wavelet transform.Fig. 4 has provided a lifting scheme the most basic realizing wavelet transformation, comprises three basic steps: separate, predict and upgrade.

Separate (split): refer to through down-sampling x [n] is separated into odd sequence xo [n]=x [2n+1] even sequence xe [n]=x [2n].This separation also is known as the Lazy wavelet transformation.

Prediction (predict): in general, have very strong correlativity between sequence of parity, can be with even number collection prediction odd number collection (P is a predictive operator) thus remove the redundancy between data, obtain the detail of the high frequency of signal.

d[n]＝x_o[n]-P(x_e[n]) (1)

If signal is local smooth, then the value of prediction residual can be very little.

Upgrade (update): this step can obtain the low-frequency information of x [n], i.e. the general picture of signal, or be called approximate signal.Corresponding computing is that prediction residual d [n] is added among the even sequence xe [n] through upgrading operator U,

s[n]＝x_e[n]+U(d_o[n]) (2)

More than be the decomposition algorithm of lifting scheme, the restructing algorithm that can be derived lifting scheme by the character that promotes matrix is following, comprises anti-renewal, anti-prediction, three steps of amalgamation, and is as shown in Figure 5

Anti-upgrade (undo update):

x_e[n]＝s[n]-U(d[n]) (3)

Anti-prediction (undo predict):

x_o[n]＝d[n]+P(x_e[n]) (4)

Synthetic (merge):

x[n]＝Merge(x[2n+1]，x[2n])?(5)

The conversion of structure different wavelet is depended on selection different predicting operator and upgrades operator that these operators can be constant coefficient (simply multiplying each other), also can be the shock responses (convolution algorithm) of wave filter.Predictive operator is corresponding with the scaling function in the traditional wavelet, can obtain different scaling functions by the different predicting interpolation operator.Same, it is also corresponding with the wavelet function in traditional wavelet transformation to upgrade operator.

No matter traditional transform method is Fast Fourier Transform (FFT) or wavelet transformation, and input signal is an integer, and the coefficient after the conversion that obtains is a floating number, and there is round-off error in computing machine when handling, can not realize the lossless compress of data.

Review and adopt lifting scheme to calculate wavelet transformation, in lifting process, just can add quantization operations, owing to promote the requirement that matrix satisfies completely reversibility, thus just can realize the wavelet transformation from the integer to the integer.Concrete situation is shown in following formula: (Q is quantificational operators)

y(i)＝x(i)+Q[αx(j)]，y(k)＝x(k)，k≠i (6)

x(i)＝y(i)-Q[αy(j)]，x(k)＝y(k)，k≠i (7)

According to above-mentioned quantification lifting step, just can obtain the way of realization of integer wavelet transform and inverse transformation thereof, as shown in Figure 6, each step lifting step all adds quantization operations, has guaranteed that each step is output as integer.At first, any existing first generation small echo, the lifting structure cascade that can pass through limited number of time is achieved; Secondly; Can promote conversion through extra four times and make that gain factor is 1, like this, the LP and the BP of output also are integer at last; Because the integer transform that each step is all reversible obviously can accurately recover original signal by LP and BP when inverse transformation.

Viewpoint with multiresolution analysis or BPF.; Wavelet decomposition is not limited to above-mentioned one-level decomposes, and the approximate signal after can also decomposing one-level continues to do wavelet decomposition, further removes its correlativity; But difference because unlike signal distributes on frequency; It is influential to adopt different decomposed classes that the result who compresses is understood, and the present invention adopts the different wavelet decomposed class that signal is decomposed, according to the result after decomposing; Select the highest progression of ratio of compression, and the decomposed class information of the best is recorded in the side information.

4, LPC prediction:

What this programme adopted is the method for forward prediction, thinks that the value of discrete-time signal current point in time can predict that expression formula is as follows by the linear combination of preceding K the value of putting:

\hat{x} (n) = Σ_{k = 1}^{K} h_{k} \cdot x (n - k) - - - (8)

Wherein k is the exponent number of fallout predictor.If predictor predicts result is near current demand signal, the variance ratio original signal of its residual error approaches zero more so, thereby has reduced code length effectively.The structure of forward prediction device is as shown in Figure 7, and its corresponding decode procedure is as shown in Figure 8.

The key problem of LPC prediction is how to obtain predictor coefficient according to input signal, and the exponent number of establishing fallout predictor is p, is then got by the Yule-Walker equation:

(\begin{matrix} r_{x} (0) & r_{x} (1) & r_{x} (2) & . . . & r_{x} (p) \\ r_{x} (1) & r_{x} (0) & r_{x} (1) & . . . & r_{x} (p - 1) \\ r_{x} (2) & r_{x} (1) & r_{x} (0) & . . . & r_{x} (p - 2) \\ . & . & . & . \\ . & . & . & . \\ . & . & . & . \\ r_{x} (p) & r_{x} (p - 1) & r_{x} (p - 2) & . . . & r_{x} (0) \end{matrix}) (\begin{matrix} 1 \\ - a_{1} \\ - a_{2} \\ . \\ . \\ . \\ - a_{p} \end{matrix}) = (\begin{matrix} σ^{2} \\ 0 \\ 0 \\ . \\ . \\ . \\ 0 \end{matrix}) - - - (9)

Wherein, rx (i), i=0,1,2 ..., p is the sampling point autocorrelation value, can be estimated by the autocorrelation function of p point before the current sampling point in the fallout predictor.A1, a2, a3 ..., ap is a predictor coefficient, σ 2 is the least error power of forward prediction.When known autocorrelation matrix, make the minimum predictor coefficient of error power be institute and ask.This programme adopts classical Levinson-Durbin algorithm to find the solution this coefficient.

Make reflection coefficient am (m)=km, the recurrence equation that can be obtained coefficient by the Levinson-Durbin algorithm is following:

k_{m} = (- Σ_{k = 1}^{m - 1} a_{m - 1} (k) r_{x} (m - k) + r_{x} (m)) / ρ_{m - 1} - - - (10)

a_m(k)＝a_m-1(k)-k_ma_m-1(m-k) (11)

ρ_{m} = ρ_{m - 1} (1 - k_{m}^{2}) - - - (12)

Each rank coefficient recursion obtains in this algorithm, so can relatively select suitable fallout predictor exponent number and each rank coefficient.Because km2 is the number greater than 0, so ρ m＜ρ m-1 is arranged forever, promptly along with the carrying out of iterative process, predicated error will reduce step by step.

Because the predictive coefficient that calculates according to above-mentioned optiaml ciriterion all is a floating number; So the fallout predictor reflection coefficient that in side information, keeps is the fixed-point number through quantizing; Though given up certain optimization criterion, guaranteed the complete restructural characteristic of signal.

5, entropy coding

Because residual error is obeyed certain distribution, so generally use entropy coding for residual error.This programme adopts the Rice sign indicating number that residual signals is encoded.The Rice coding is applicable to that its distribution is similar to the digital coding of geometric distributions.Its distribution function is:

P_r{r_θ＝p}＝(1-θ)θ^p，θ∈(0，1) (13)

But these coding requirement data of Lay be necessary on the occasion of.In order to reach this requirement, at first need do a mapping to residual signals, with negative value be mapped on the occasion of.This process is as follows:

r_{i}^{*} = \{\begin{matrix} {2 r}_{i}, & r_{i} > 0 \\ - 2 r_{i} - 1, & r_{i} < 0 \end{matrix} - - - (14)

For a data p, at first use M (M=2n generally speaking) to remove this number, obtain a merchant q and corresponding remainder r.Promptly

r＝p-q·M (16)

Then the coded system of these data is following: 1 of q bit is used for representing the merchant; 0 of 1 bit is used for distinguishing the zone bit of quotient and the remainder; [log2r] individual bit is used for representing remainder r.

M value suitable in the formula is confirmed by following method:

u_{N} = \frac{1}{N} Σ_{i = 1}^{N} | r_{i} | - - - (17)

Constant c1=0.97 wherein.

Although disclose specific embodiment of the present invention and accompanying drawing for the purpose of illustration; Its purpose is to help to understand content of the present invention and implement according to this; But it will be appreciated by those skilled in the art that: in the spirit and scope that do not break away from the present invention and appended claim, various replacements, variation and modification all are possible.Therefore, the present invention should not be limited to most preferred embodiment and the disclosed content of accompanying drawing.