CN101159136A - Low bit rate music signal coding method - Google Patents

Abstract

The invention discloses a method for coding a music signal with low bit rate, which comprises following steps: (1) dividing frame of the music data, adding windows to each frame signal, constructing a harmonic wave model, and selecting each parameter representing the model; (2) acquiring posterior distribution of each parameter according to prior probability distribution and initial value of the parameter in combination with conditional probability distribution of the model and by using the Bayes estimation method; (3) carrying out cyclical iteration according to the posterior distribution and by using reversible jump monte carlo sampling algorithm until the algorithm is convergent, and acquiring the estimation value of each parameter; and (4) transmitting the estimation value of each parameter into a coder to achieve low-bit-rate coding of the audio signal. The method can well fit the music signal, acquire the parameters representing the music signal, and achieve low-bit-rate parameter coding.

Description

A kind of low bit rate music signal coding method

Technical field

The present invention relates to a kind of low bit rate music signal coding method, relate in particular to the low bit rate music signal coding that adopts Bayesian Estimation.

Background technology

The world is just striding forward the information age of digitizing, networking, global integration.Infotech just progressively is penetrated into the every aspect of human society, and people have higher requirement to providing with transmission quality of information.Because the restriction of medium capacity and transmission bandwidth must use various compress techniques to transmit data.Sound is a kind of important transmission information medium, and the audio coding technology of digital audio field mainly contains at present:

1) lossless compress

The development of audio compression techniques is at first from lossless compress, and typical compression method is PCM (pulse code modulation (PCM)).Though PCM can reduce tonequality preferably, the code efficiency of its instantaneous companding technology and piece companding (Block-Compressing) technology is lower, and code check is too high.Along with the development of Digital Signal Processing, the perceptual coding algorithm a series of Frequency Domain Coding algorithms appearred subsequently, i.e..

2) Audio Perceptual Coding:

The target of Audio Perceptual Coding is the perceptual distortion that reduces as much as possible between audio frequency original signal and the encoding and decoding signal, reduces bit rate as much as possible simultaneously and encodes.For music signal, the coefficient that concentrates on mostly the time-frequency representation signal at present carries out transition coding.AAC (the Advanced Audio Coder) coding of typical transition coding such as MPEG-4 can provide gem-pure quality to about 64kbit/s tone signal, but may produce some culture noises under low bit rate.

3) parameter coding:

With respect to perceptual coding, the thinking of parametric audio coding is that sound signal is decomposed according to more adaptive mode.Different compositions represents that with different information source models each model all defines the parameter set of oneself to describe different attributes.Decompose finish after, parameter need be quantized respectively, encodes and be sent to decoding end.For example the SSC of MPEG-4 (Sinusoidal Coding) parameter coding just is based on this mode, and coding quality is better than AAC when 24kbit/s, but it is not suitable for lower code check.

In many instances, under the situation that for example network bandwidth is lower or the capacity of storage medium is very low, the code check of above-mentioned coding method is too high.This just needs one can guarantee better tonequality, and the very low technology of code check.

Summary of the invention

The technical problem to be solved in the present invention provides a kind of music encoding method of low bit rate, and this method is the match music signal preferably, obtains to characterize the parameter of music signal, then parameter is carried out Low Bit-rate Coding.

In order to address the above problem, the invention provides a kind of low bit rate music signal coding method, may further comprise the steps:

1) divides frame with music signal, and, set up harmonic-model, selected each parameter that characterizes this model every frame signal windowing;

2) according to the prior probability distribution and the initial value of described parameter, and in conjunction with the conditional probability distribution of this model, use the Bayesian Estimation method, the posteriority that obtains described each parameter distributes;

3) distribute according to the posteriority that obtains, adopt reversible jump Monte Carlo sampling algorithm, loop iteration reaches convergence until algorithm, obtains the estimated value of each parameter;

Each estimates of parameters that 4) will obtain is sent into scrambler, finishes the Low Bit-rate Coding to sound signal;

Further, method of the present invention, wherein, in the step (1), the described harmonic-model of setting up comprises: with each frame music signal, be decomposed into by different each time multifrequency sinusoid of amplitude and the stack of white Gaussian noise;

Further, method of the present invention, wherein, in the step (1), the parameter of described sign harmonic-model comprises: the amplitude of harmonic number, fundamental frequency, each harmonic and noise variance;

Further, method of the present invention, wherein, in the step (2), the prior probability distribution of described parameter is set to respectively: noise variance is obeyed the distribution of falling the gamma, and harmonic number is obeyed Poisson distribution, and fundamental frequency is obeyed distribution of mean value, whose amplitude obeys Gaussian distribution;

Further, method of the present invention wherein, in the described step (2), comprising: the posterior probability of using resulting each parameter of Bayesian Estimation method distributes, and is proportional to the product of the likelihood function of its prior probability distribution and its conditional probability distribution of description;

Further, method of the present invention, wherein, in the step (3), described reversible jump Monte Carlo sampling algorithm comprises:

I) determine the initial value of each parameter, and select the mode of motion type with certain probability;

Ii) produce recommended value by prior distribution, and begin by 1 to L cycle calculations allowance rate k, greater than 1, each then new parameter prediction value is exactly its recommended value as if allowance rate k; If allowance rate k less than 1, then composes recommended value to new predicted value with the probability of allowance rate k, otherwise new predicted value is a last iteration value constantly;

Iii) loop iteration reaches convergence until algorithm, returns L predicted value;

Further, method of the present invention, wherein, in the step (i), described mode of motion type comprises: harmonic number increase, minimizing, constant three types;

Further, method of the present invention, wherein, step (ii) in, if allowance rate k is less than 1, then, comprising with the recommended value of each parameter of probability calculation of allowance rate k:

For the type that harmonic number increases,, calculate allowance rate g according to reversible jump Monte Carlo sampling algorithm _Birth, judge the value of the predicted value of harmonic number and fundamental frequency, and produce the recommended value of all the other parameters;

For the type that harmonic number reduces,, calculate allowance rate g according to reversible jump Monte Carlo sampling algorithm _Death, judge the value of the predicted value of harmonic number and fundamental frequency, and produce the recommended value of all the other parameters;

For the type that harmonic number reduces,, calculate allowance rate g according to reversible jump Monte Carlo sampling algorithm _Update, judge the value of the predicted value of fundamental frequency, and produce the recommended value of all the other parameters;

Method of the present invention, wherein, in the described step (3), further comprise: distribute according to the posteriority that obtains, adopt reversible jump Monte Carlo sampling algorithm, loop iteration is after algorithm reaches convergence, each parameter all obtains a predicted value ordered series of numbers that contains a L element, select estimation criterion then, respectively the predicted value ordered series of numbers of each parameter is handled, promptly obtained an estimated value of each parameter;

Further, method of the present invention, wherein, described estimation criterion comprises: maximum a posteriori probability is estimated, Posterior Mean is estimated and least mean-square error is estimated;

Further, method of the present invention, wherein, in the step (3), described scrambler adopts following method to encode, and comprising: Differential video coding method, huffman coding method.

Compared with prior art, the method for the invention, match music signal preferably obtains to characterize the parameter of music signal, then parameter is carried out Low Bit-rate Coding.

Description of drawings

Fig. 1 is the low bit rate music signal coding method process flow diagram that adopts the Bayesian Estimation method in the embodiment of the invention;

Fig. 2 is the low bit rate music signal coding system construction drawing that adopts the Bayesian Estimation method in the embodiment of the invention;

Fig. 3 is a reversible jump Monte Carlo sampling algorithm flow chart in the embodiment of the invention;

Fig. 4 is an operation chart level and smooth to the multiframe signal windowing in the embodiment of the invention.

Embodiment

The present invention is in order to solve the drawback that conventional solution exists, further set forth a kind of low bit rate music signal coding method of the present invention by following specific embodiment, below embodiment is described in detail, but not as a limitation of the invention.

Key of the present invention is on the music signal model based, obtains to characterize the parameter of music signal.Music signal has tangible design feature at frequency domain: on frequency domain, music signal is made of two-layer model: at first, each note is that syllable and each harmonic (partials) are formed by fundamental frequency (fundamental frequency), the time luffing degree of harmonic wave has reflected the tone color of music signal, and their frequency is approximately the integer multiple of fundamental frequency; Then, each note is played simultaneously and has been formed chord, i.e. multi-tone signal (polyphony).At these characteristics, the present invention has adopted a kind of harmonic-model, and it regards music signal the stack of a series of multifrequency sinusoids and white Gaussian noise as, and then, the present invention adopts the method for Bayesian Estimation to obtain these parameters.The Bayesian Estimation method provides a class rigorous solution framework for the estimation problem of dynamic system, it utilizes known information to set up the probability density function of system, it is prior distribution, distribute according to condition again, the posteriority that then can obtain known variables or parameter distributes, and distributes the true distribution of approximating parameter with posteriority.The present invention is on the harmonic-model basis, according to the Bayesian Estimation method, distribute the true distribution of each parameter of approximate model with posteriority, adopt reversible jump Monte Carlo sampling (RJMCMC then, Reverse-Jump Markov Chain Monte Carlo) algorithm, loop iteration reaches convergence up to algorithm, has obtained estimates of parameters.At last, estimates of parameters is encoded, promptly realized the low rate encoding of music signal.

As shown in Figure 1, be the low bit rate music signal coding method process flow diagram that adopts the Bayesian Estimation method in the embodiment of the invention, may further comprise the steps:

Step 100 is divided into short time frame with music signal;

Step 101 is represented every frame signal with harmonic-model, promptly signal is regarded as by different each time multifrequency sinusoid of amplitude and the stack of white Gaussian noise; Be that the available harmonic number, harmonic amplitude, fundamental frequency, the isoparametric mathematic(al) representation of noise variance of containing characterizes any music signal thus;

Step 102, suppose that the prior distribution of each parameter is as follows:

Noise variance is obeyed the distribution of falling the gamma; Harmonic number is obeyed Poisson distribution; Fundamental frequency is obeyed distribution of mean value; The whose amplitude obeys Gaussian distribution;

Step 103 according to the prior distribution of each parameter, in conjunction with the conditional probability of this model, is used the Bayesian Estimation method, and the posteriority that obtains each parameter distributes;

Step 104 in conjunction with the posteriority distribution and expression formula of each parameter, adopts reversible jump Monte Carlo sampling (RJMCMC) algorithm, loop iteration, up to convergence, and handle the predicted value ordered series of numbers of each parameter that obtains by sampling, obtain the estimated value of each parameter according to estimation criterion;

Step 105 is sent the estimated value of each parameter into scrambler, carries out differential coding and Huffman encoding, the low Bit Rate Audio Coding of final realization.

As shown in Figure 2, be the low bit rate music signal coding system construction drawing that adopts the Bayesian Estimation method in the embodiment of the invention, as can be seen from the figure this is an analysis-synthesis system.In transmitting terminal 1, at first obtain to characterize the parameter of music signal harmonic-model by analysis module 10, encode by 11 pairs of described parameters of scrambler again; Code stream transmits the back and arrives receiving end 2 in channel, decoded by the demoder in the receiving end 2 20, obtains decoded parameter, brings decoded parameter into harmonic-model by synthesis module 21 again and synthesizes, and finishes the reduction to music signal thus.To introduce method of the present invention according to the operating process of various piece correspondence in the system construction drawing below.

(I) at first, in low bit rate music signal coding method, want earlier the music signal that transmits at first to be analyzed.

In analysis process, be divided into following process: divide frame, set up harmonic-model, make up model probability framework and estimation model parameter.

1) divide frame:

Be actually a time varying signal through digitized voice signal, suppose that generally it can regard stably in the short time that at 10ms～30ms common analysis all is the signal analysis at the short time as.In order to obtain voice signal in short-term, voice signal is divided into short time frame, next analyze at the music signal data of each short time frame exactly.

2) set up harmonic-model:

Generally speaking, be periodic by the music signal of instrument playing, can be regarded as by a series of sine-wave superimposed and form.Each note is made of M harmonic wave, and one of them is a fundamental frequency, and remaining M-1 is harmonic wave, is approximately the integer multiple of fundamental frequency.Consider the music signal of one section short time frame, and in this segment signal, do not have the sudden change of note, be expressed as y[t], t=1,2 ..., N, fundamental frequency are ω, sampling rate is ω _s, when being, the amplitude of each harmonic becomes, be respectively a _m[t], b _m[t], noise are υ [t].Model can be expressed as follows:

$y\left[t\right]=\underset{m=1}{\overset{M}{\mathrm{\Σ}}}\{a_m\left[t\right]\cos \left(\frac{2\mathrm{\πm\ω}}{{\mathrm{\ω}}_s}t\right)+b_m\left[t\right]\sin \left(\frac{2\mathrm{\πm\ω}}{{\mathrm{\ω}}_s}t\right)\}+\mathrm{\υ}\left[t\right]---\left(1\right)$

a _m[t], b _m[t] can regard as by a series of basis function φ _iConstitute, i=0 ..., I is expressed as:

$a_m\left[t\right]=\underset{i=0}{\overset{I}{\mathrm{\Σ}}}{a}_{m,i}{\mathrm{\φ}}_i\left[t\right],$ $b_m\left[t\right]=\underset{i=0}{\overset{I}{\mathrm{\Σ}}}{b}_{m,i}{\mathrm{\φ}}_i[t.$ Wherein φ [t] is basis function (basis function), can be any type of non-oscillatory function, as Hanning window.Select Hanning window for use in this present invention, rolloff-factor is 0.5.The initialization function is φ [τ], τ=-N+1 ...., N-1, so, other basis function is respectively:

φ _i[t]＝φ[t-iΔt]i＝0，...，I，t＝0，...，N-1 (2)

Interval of delta t=(N-1)/I wherein.So simplified model can be expressed as follows:

$y\left[t\right]=\underset{m=1}{\overset{M}{\mathrm{\Σ}}}\underset{i=0}{\overset{I}{\mathrm{\Σ}}}\mathrm{\φ}[t-\mathrm{i\Δt}]\{a_{m,i}\cos \left(\frac{2\mathrm{\πm\ω}}{{\mathrm{\ω}}_s}t\right)+b_{m,i}\sin \left(\frac{2\mathrm{\πm\ω}}{{\mathrm{\ω}}_s}t\right)\}+\mathrm{\υ}\left[t\right]---\left(3\right)$

If magnitude matrix is β, length is 2M (I+1), m=1 ..., M, i=0 ..., I, concrete element is:

β(2(Mi+m)-1)＝a _m，i (4)

β(2(Mi+m))＝b _m，i (5)

Nuclear matrix is D, is N * 2M (I+1) dimension matrix, m=1 ..., M, i=0 ..., I, t=1 ..., the concrete element of N is:

D(t，2(Ni+m)-1)＝φ(t-iΔt)cos(mωt) (6)

D(t，2(Mi+m))＝φ(t-iΔt)sin(mωt) (7)

So this model can be reduced to:

y＝Dβ+v (8)

Be noise vector, be white noise, variance is σ _υ ²

3) make up probabilistic framework:

The likelihood function of this model is:

$p\left(y\right|\mathrm{\β},\mathrm{\ω},M,{\mathrm{\σ}}_{\mathrm{\υ}}^2)=\frac{1}{{\left(2\mathrm{\π}{\mathrm{\σ}}_{\mathrm{\υ}}^2\right)}^{N/2}}\exp [-\frac{1}{2{\mathrm{\σ}}_{\mathrm{\υ}}^2}{\left|\right|y-\mathrm{D\β}\left|\right|}^2]---\left(9\right)$

The prior probability of model can be decomposed into:

$p(\mathrm{\β},{\mathrm{\σ}}_{\mathrm{\υ}}^2,\mathrm{\ω},M)=p\left(\mathrm{\β}\right|{\mathrm{\σ}}_{\mathrm{\υ}}^2,\mathrm{\ω},M)p\left(\mathrm{\ω}\right|M)p\left(M\right)p\left({\mathrm{\σ}}_{\mathrm{\υ}}^2\right)---\left(10\right)$

The prior probability of β is the zero-mean Gaussian distribution, and covariance matrix is σ _υ ²/ ξ ²Σ I, I are unit matrix, ξ ²Be super parameter (hyper parameter) the expression signal to noise ratio (S/N ratio).The probability distribution of β represents that formula is as follows:

$p\left(\mathrm{\β}\right|{\mathrm{\σ}}_{\mathrm{\υ}}^2,\mathrm{\ω},M)=N(\mathrm{\β};0_{2M(I+1)},\frac{{\mathrm{\σ}}_{\mathrm{\υ}}^2}{{\mathrm{\ξ}}^2}\mathrm{\ΣI})---\left(11\right)$

ξ ²Obey the distribution of falling the gamma (inverted gamma), be shown below:

$p\left({\mathrm{\ξ}}^2\right)=\mathrm{IG}({\mathrm{\α}}_{\mathrm{\ξ}},{\mathrm{\β}}_{\mathrm{\ξ}})\∝\frac{e^{-{\mathrm{\β}}_{\mathrm{\ξ}}/{\mathrm{\ξ}}^2}}{{\mathrm{\ξ}}^{2({\mathrm{\α}}_{\mathrm{\ξ}}+1)}}---\left(12\right)$

The prior probability of ω is obeyed distribution of mean value, is expressed as follows:

p(ω|M)＝u(ω _m；[0，ω _s/2M]) (13)

Noise variance σ _υ ²Obey the distribution of falling the gamma, be expressed as follows:

$p\left({\mathrm{\σ}}_{\mathrm{\υ}}^2\right)=\mathrm{IG}({\mathrm{\σ}}_{\mathrm{\υ}}^2;v_0,{\mathrm{\γ}}_0)---\left(14\right)$

The prior probability of M adopts the Poisson distribution model, is expressed as follows:

$p(M=m|\mathrm{\Λ})=e^{-\mathrm{\Λ}}\frac{{\mathrm{\Λ}}^m}{m!}---\left(15\right)$

Λ is the expectation value of harmonic number, obeys gamma (gamma) and distributes, and expression formula is as follows:

p(Λ)＝G(1/2+α _Λ，β _Λ) (16)

In fact M is limited in certain scope, and minimum value is M _Min, maximal value is M _Max, promptly distributed model is the brachymemma Poisson distribution in fact.So

$p(M=m)=({\mathrm{\Λ}}^m/m!)/({\mathrm{\Σ}}_{m^{\′}=M_{\min }}^{M_{\max }}{\mathrm{\Λ}}^{m^{\′}}/m^{\′}!)---\left(17\right)$

The Bayesian Estimation method shows: posterior probability distributes and to be proportional to the product of prior probability distribution and likelihood function, and in conjunction with prior probability distribution of having given and likelihood function, posterior probability that can this model is expressed as follows:

$p(\mathrm{\β},\mathrm{\ω},M,{\mathrm{\σ}}_{\mathrm{\υ}}^2|y)\∝p\left(y\right|\mathrm{\β},\mathrm{\ω},M,{\mathrm{\σ}}_{\mathrm{\υ}}^2)p(\mathrm{\β},\mathrm{\ω},{\mathrm{\σ}}_{\mathrm{\υ}}^2,M)---\left(18\right)$

Provide the expression of the posterior probability of each parameter below:

$p(\mathrm{\ω},M|y)\∝{({\mathrm{\γ}}_0+y^t\mathrm{Py})}^{\frac{-(N+v_0)}{2}}\det {\left(S\right)}^{\frac{1}{2}}p\left(\mathrm{\ω}\right|M)\frac{{\mathrm{\Λ}}^M}{M!}---\left(19\right)$

P=I-DSD wherein ^t, $S={[D^{t}D+\frac{1}{{\mathrm{\ξ}}^2}I]}^{-1},$ Noise variance σ _υ ²Posterior probability be:

$p\left({\mathrm{\σ}}_{\mathrm{\υ}}^2\right|\mathrm{\ω},M,y)=\mathrm{IG}(\frac{N+v_0}{2},\frac{{\mathrm{\γ}}_0+y^t\mathrm{Py}}{2})\∝\exp (-\frac{{\mathrm{\γ}}_0+y^t\mathrm{Py}}{2{\mathrm{\σ}}_{\mathrm{\υ}}^2}){\mathrm{\σ}}_{\mathrm{\υ}}^{-N-v0-2}---\left(20\right)$

The posterior probability Gaussian distributed of β, expression is:

$p\left(\mathrm{\β}\right|{\mathrm{\σ}}_{\mathrm{\υ}}^2,\mathrm{\ω},M,y)=N(\mathrm{\μ},{\mathrm{\σ}}_{\mathrm{\υ}}^{2}S)---\left(21\right)$

μ=SD wherein ^tY.

ξ ²Posterior probability be:

$p\left({\mathrm{\ξ}}^2\right|\mathrm{\β},\mathrm{\ω},M,{\mathrm{\σ}}_{\mathrm{\υ}}^2)\∝\mathrm{IG}({\mathrm{\α}}_{\mathrm{\ξ}}+M(I+1),{\mathrm{\β}}_{\mathrm{\ξ}}+\frac{{\mathrm{\β}}^{\′}{D}^{\′}\mathrm{D\β}}{2{\mathrm{\σ}}_{\mathrm{\υ}}^2})---\left(22\right)$

The posterior probability of Λ is:

p(Λ|M)∝(M+1/2+α _Λ，1+β _Λ) (23)

4) estimation model parameter:

In Bayes's filtering problem, distribute for general multivariate, can not obtain integral result, therefore must be similar to integration.Monte Carlo method (MCMC) is a kind of statistical test method, and its basic thought is at first to set up a probability model or stochastic process; Statistical nature by the observation or the sampling test of model or process are calculated the parameter of asking then, and with arithmetic mean as the approximate value of being found the solution.Generally, the dimension of variable to be estimated changes, and harmonic number to be estimated among the present invention for example is so the dimension of D and β is all changing.RJMCMC allows to make up the Markov chain of a redirect dimension, and it has defined a series of mode of motion, as renewal, generation, extinction, merging, separation etc., any one mode of motion of picked at random in the process of iteration.

Consider a transition kernel function q (z, z '), be also referred to as the suggestion probability function, it is the function that produces suggestion sample value z ' according to currency z.Destination probability is distributed as π (z), is the posteriority for the treatment of estimated parameter and distributes the allowance rate $k(z,z^{\′})=\min (1,\frac{\mathrm{\π}\left(z^{\′}\right)q(z^{\′},z)}{\mathrm{\π}\left(z\right)q(z,z^{\′})}),$ To determine the value of predicted value according to the size of allowance rate.

The RJMCMC arthmetic statement is as follows:

A) determine initial value (be generally and estimate last sample value in the ordered series of numbers);

B) from possible mode of motion, select the mode of motion type with certain probability;

C) produce recommended value by suggestion probability distribution (probability density function), begin to carry out circulation (by l to L):

Calculate allowance rate k, greater than 1, then new predicted value is recommended value as if the allowance rate, if allowance rate k less than 1, then composes the recommended value that calculates to new predicted value with the probability of allowance rate k, otherwise new predicted value is a last iteration value constantly;

D) L loop iteration finishes, and returns L predicted value.

As shown in Figure 3, be RJMCMC sampling algorithm process flow diagram in reversible jump Monte Carlo in the embodiment of the invention, may further comprise the steps:

Step 300, the initialization model parameter comprises: number of samples N, iterations L, window function number I etc.;

Step 301 begins to carry out reversible jump Monte Carlo RJMCMC sampling iterative algorithm, selects the mode of motion type, i.e. harmonic number M increase, minimizing, constant three types.

Step 302, according to prior distribution, the parameter of each probability distribution function of initialization, nuclear matrix and amplitude vector;

Step 303, the relation of detection loop control variable l and iterations L, if smaller or equal to, then execution in step 304; If greater than, then execution in step 306;

Step 304, in iterative process, difference corresponding harmonic number increase, minimizing and constant type of sports: carry out harmonic number and increase algorithm, harmonic number minimizing algorithm and harmonic parameters update algorithm; RJMCMC has also all been used in these three kinds of calculations, can carry out detailed explanation below;

Wherein, the probability of the type of sports that harmonic number increases is a, and the probability of the type of sports that harmonic number reduces is b, and the probability of the type of sports that harmonic number is constant is 1-a-b, and probability a and b computing formula are as follows:

$a=c\min \{1,\frac{p(M+1|\mathrm{\Λ})}{p\left(M\right|\mathrm{\Λ})}\}$

$b=c\min \{1,\frac{p\left(M\right|\mathrm{\Λ})}{p(M+1|\mathrm{\Λ})}\}$

C=0.15 wherein, p (M| Λ) is the prior probability of M;

Step 305 is through the iteration renewal respectively of above-mentioned three kinds of situations, the predicted value of undated parameter, and record; And return execution in step 303;

Step 306, each parameter tends towards stability, and each parameter all obtains a predicted value ordered series of numbers that contains L element, selects estimation criterion, respectively the predicted value ordered series of numbers of each parameter is handled according to certain estimation criterion, promptly obtained an estimated value of each parameter;

Estimation criterion comprises: maximum a posteriori probability estimation, Posterior Mean estimation and least mean-square error estimation etc.

Need indicate, the algorithm that the harmonic number in the step 304 increases, reduces, upgrades has all used RJMCMC, subscript

With ^*Represent the recommended value of Monte Carlo particle and generation respectively.Elaborate their step below:

Increase harmonic number algorithm (Partials birth)

● recommended value $M^{*}={\tilde{M}}^{(l-1)}+1.$

● make P ^*For containing M ^*Subharmonic P, follow above-mentioned computing formula about allowance rate a, calculate following formula:

$g_{\mathrm{birth}}={\left[\frac{{\mathrm{\γ}}_0+y^T\mathrm{Py}}{{\mathrm{\γ}}_0+y^T{P}^{*}y}\right]}^{\frac{N+v_0}{2}}\frac{M^{*}}{{\tilde{M}}^{(l-1)}}\frac{1}{{(1+{\mathrm{\ξ}}^2)}^{(I+1)}}$

● with probability min (1, g _Birth), receive recommended value: order ${\tilde{M}}^{\left(l\right)}=M^{*},$ ${\widetilde{\mathrm{\ω}}}^{\left(l\right)}={\widetilde{\mathrm{\ω}}}^{(l-1)}.$

● the posteriority according to all the other parameters distributes, and produces the recommended value of all the other parameters.

Reduce harmonic number algorithm (Partials death)

● recommended value $M^{*}={\tilde{M}}^{(l-1)}-1.$

● remove the

Subharmonic.

● make P ^*For removing

The P of subharmonic follows above-mentioned computing formula about allowance rate a, calculates following formula:

$g_{\mathrm{death}}={\left[\frac{{\mathrm{\γ}}_0+y^T\mathrm{Py}}{{\mathrm{\γ}}_0+y^T{P}^{*}y}\right]}^{\frac{N+v_0}{2}}\frac{M^{*}}{{\tilde{M}}^{(l-1)}}{(1+{\mathrm{\ξ}}^2)}^{(I+1)}$

● with probability min (1, g _Death), receive recommended value: order ${\tilde{M}}^{\left(l\right)}=M^{*},$ ${\widetilde{\mathrm{\ω}}}^{\left(l\right)}={\widetilde{\mathrm{\ω}}}^{(l-1)}.$

● the posteriority according to all the other parameters distributes, and produces the recommended value of all the other parameters.

Harmonic parameters is upgraded (algorithm for partials parameter updates)

● sampling u, obey distribution of mean value, i.e. u=U _[0,1]

● as u＜λ, adopt suggestion distribution q _Init(ω | y), and with probability min (1, g _Update) the reception recommended value.

● otherwise adopt suggestion to distribute

And with probability min (1, g _Update) the reception recommended value.

● the posteriority according to all the other parameters distributes, and produces the recommended value of all the other parameters.

In above-mentioned harmonic parameters update algorithm, two kinds of suggestion distribution functions are arranged.

When u＜λ, the suggestion distribution function is q _Init(ω | y), it is the distribution that the Fourier transform according to signal y produces, and is expressed as:

$q_{\mathrm{init}}\left(\mathrm{\ω}\right|y)\∝\underset{l=0}{\overset{N-1}{\mathrm{\Σ}}}{p}_l\mathrm{II}(\mathrm{l\π}/N,[l+1)\mathrm{\π}/N]$

p _lBe the Fourier transform amplitude of signal at l π/N place.

When u＞λ, the suggestion distribution function is Be expressed as:

$q_{\mathrm{RW}}\left({\mathrm{\ω}}^{*}\right|{\widetilde{\mathrm{\ω}}}^{(l-1)})=N({\mathrm{\ω}}^{(l-1)};{\mathrm{\σ}}_{\mathrm{RW}}^2)$

For with last sample value

Be average, σ _RW ²Be the Gaussian distribution of variance, i.e. recommended value ω ^*Be

Contiguous place samples.

Allowance rate corresponding to top two formulas is $g_{\mathrm{update}}={\left[\frac{{\mathrm{\γ}}_0+y^T\mathrm{Py}}{{\mathrm{\γ}}_0+y^T{P}^{*}y}\right]}^{\frac{N+v_0}{2}},$ P wherein ^*For using ω ^*Matrix P after the renewal.

(II) according to the estimated value of analyzing each parameter of the harmonic-model that obtains in the previous step, encode.

Analysis part has obtained to describe the parameter of music signal, i.e. harmonic wave number M, harmonic amplitude β, fundamental frequency ω, noise variance σ _υ ²The design of scrambler is such: to the harmonic wave number M, and fundamental frequency ω, noise variance σ _υ ²Quantize huffman coding.Wherein harmonic amplitude β first ties up the amplitude energy maximum of corresponding fundamental tone, and the amplitude of all the other harmonic waves is used difference value, also carries out Huffman encoding then.

So far, low bit rate music signal coding flow process of the present invention is finished.Reduction is through the music signal of the method for the invention coding below.

(III) obtain each parameter of harmonic-model by decoding, and bring decoded parameter into harmonic-model and synthesize.

The work of composite part is simple relatively, as shown in Figure 4, is operation chart level and smooth to the multiframe signal windowing in the embodiment of the invention.

Parameter substitution harmonic-model with decoding obtains restores music signal, then again each frame is carried out the windowing smooth operation and gets final product.For the processing of multiframe signal, be not directly a plurality of single frames music signals to be stitched together among the present invention, because every like this frame signal adjacent can produce sudden change.The present invention selects Hamming window for use, has effectively realized level and smooth.

Certainly; the present invention also can have other various embodiments; under the situation that does not deviate from spirit of the present invention and essence thereof; those of ordinary skill in the art can make various corresponding changes and distortion according to the present invention, but these corresponding changes and distortion all should belong to the protection domain of the appended claim of the present invention.

Claims (11)

1. a low bit rate music signal coding method is characterized in that, may further comprise the steps:

1) divides frame with music signal, and, set up harmonic-model, selected each parameter that characterizes this model every frame signal windowing;

2) according to the prior probability distribution and the initial value of described parameter, and in conjunction with the conditional probability distribution of this model, use the Bayesian Estimation method, the posteriority that obtains described each parameter distributes;

3) distribute according to the posteriority that obtains, adopt reversible jump Monte Carlo sampling algorithm, loop iteration reaches convergence until algorithm, obtains the estimated value of each parameter;

Each estimates of parameters that 4) will obtain is sent into scrambler, finishes the Low Bit-rate Coding to sound signal.

2. the method for claim 1 is characterized in that, in the step (1), the described harmonic-model of setting up comprises:

With each frame music signal, be decomposed into by different each time multifrequency sinusoid of amplitude and the stack of white Gaussian noise.

3. the method for claim 1 is characterized in that, in the step (1), the parameter of described sign harmonic-model comprises:

The amplitude of harmonic number, fundamental frequency, each harmonic and noise variance.

4. method as claimed in claim 3 is characterized in that, in the step (2), the prior probability distribution of described parameter is set to respectively:

Noise variance is obeyed the distribution of falling the gamma, and harmonic number is obeyed Poisson distribution, and fundamental frequency is obeyed distribution of mean value, whose amplitude obeys Gaussian distribution.

5. the method for claim 1 is characterized in that, in the described step (2), comprising:

The posterior probability of using resulting each parameter of Bayesian Estimation method distributes, and is proportional to the product of the likelihood function of its prior probability distribution and its conditional probability distribution of description.

6. the method for claim 1 is characterized in that, in the step (3), described reversible jump Monte Carlo sampling algorithm comprises:

I) determine the initial value of each parameter, and select the mode of motion type with certain probability;

Ii) produce recommended value by prior distribution, and begin by 1 to L cycle calculations allowance rate k, greater than 1, each then new parameter prediction value is its recommended value as if allowance rate k; If allowance rate k is less than 1,, and composes and give new predicted value then with the recommended value of each parameter of probability calculation of allowance rate a;

Iii) loop iteration reaches convergence until algorithm, returns L predicted value.

7. method as claimed in claim 6 is characterized in that, in the step (i), described mode of motion type comprises: harmonic number increase, minimizing, constant three types.

8. method as claimed in claim 7 is characterized in that, step (ii) in, if allowance rate k is less than 1, then recommended value is composed to new predicted value with the probability of allowance rate a, comprising:

For the type that harmonic number increases,, calculate allowance rate g according to reversible jump Monte Carlo sampling algorithm _Birth, judge the value of the predicted value of harmonic number and fundamental frequency, and produce the recommended value of all the other parameters;

For the type that harmonic number reduces,, calculate allowance rate g according to reversible jump Monte Carlo sampling algorithm _Death, judge the value of the predicted value of harmonic number and fundamental frequency, and produce the recommended value of all the other parameters;

For the type that harmonic number reduces,, calculate allowance rate g according to reversible jump Monte Carlo sampling algorithm _Update, judge the value of the predicted value of fundamental frequency, and produce the recommended value of all the other parameters.

9. method as claimed in claim 6 is characterized in that, in the described step (3), further comprises:

Distribute according to the posteriority that obtains, adopt reversible jump Monte Carlo sampling algorithm, loop iteration is after algorithm reaches convergence, each parameter all obtains a predicted value ordered series of numbers that contains a L element, select estimation criterion then, respectively the predicted value ordered series of numbers of each parameter is handled, promptly obtained an estimated value of each parameter.

10. method as claimed in claim 9 is characterized in that, described estimation criterion comprises: maximum a posteriori probability is estimated, Posterior Mean is estimated and least mean-square error is estimated.

11. the method for claim 1 is characterized in that, in the step (3), described scrambler adopts following method to encode, and comprising: Differential video coding method, huffman coding method.

Images