CN1760975A - Searching method of fixing up codebook quickly for enhanced AMR encoder - Google Patents

Abstract

The method includes steps: (1) first, defining a threshold value of relativity coefficient between voice frames so as to provide benchmark for searching codebook of voice frame; (2) calculating correlation coefficients between each sub frame and its previous sub frame; then, comparing each correlation coefficient with the threshold value in step (1); (3) for the compared result in step (2), if the result is larger than the threshold value in step (1), then algorithm of pulse substitution is adopted; if the result is smaller than the threshold value in step (1), then algorithm of depth-first tree suggested in AMR standard is adopted.

Description

The AMR scrambler rapid fixed codebook searching method that strengthens

Technical field

The present invention relates to moving communicating field, the AMR scrambler rapid fixed codebook searching method that is specifically related to strengthen among the 3GPP.

Background technology

The adaptive multi-rate speech coding (AMR) that is absorbed as standard by 3GPP is mainly used in the following WCDMA mobile communication system.Scrambler can be supported eight provenance code rates, thereby according to the more intelligent distribution information source coding of the condition of current channel and the proportion of chnnel coding.Therefore, the limited wireless Internet resources obtain more effective and reasonable utilization.

The AMR coding has adopted Algebraic Code Excited Linear Prediction (ACELP) model.Under the sampling rate of 8000Hz, scrambler is handled the speech frame of 20 milliseconds of 160 sampled points, extract the parameter (parameter of LP wave filter, the index of self-adaptation and fixed codebook and gain) of algebraic codebook linear incentive predictive mode, parameter is encoded and transmitted.At receiving end, decoding parametric, and make up the LP composite filter by relevant parameters and come synthetic speech signal.

Yet algebraic codebook search but accounts for whole cataloged procedure 40% calculated amount, directly influences AMR voice coding efficient.Although not only simply but also effective, its calculated amount was still very big for the depth-first codebook searching method that uses in the standard.Now, a large amount of research work focus on that still codebook search algorithm design and DSP realize.Though the pulse alternate algorithm can reduce computation complexity greatly.Yet the pulse alternate algorithm is a non-optimal algorithm, if initial codebook is looked for inaccurately, must not go out more excellent search code book.Therefore, with the AMR standard under the basic the same prerequisite of searching algorithm performance, the quick code book searching algorithm that can greatly reduce computational complexity just has high strategic importance.

Summary of the invention

(1) technical matters that will solve

The purpose of this invention is to provide a kind of effectively and can greatly reduce the AMR scrambler rapid fixed codebook searching method of the enhancing of calculated amount.

(2) technical scheme

For achieving the above object, the present invention has following steps:

1) at first define the threshold value of the relative coefficient of an adjacent speech frame, this threshold value provides benchmark for the search of speech frame code book;

2) all to calculate the related coefficient of it and last subframe to each subframe, compare with the described threshold value of step 1) then;

3) for step 2) middle result relatively, if greater than the described threshold value of step 1), so just adopt improved pulse alternate algorithm; If less than the described threshold value of step 1), just adopt the depth-first tree algorithm of advising in the AMR standard.

Wherein, related coefficient is the fixed codebook degree of correlation of current subframe and last subframe, and its mathematic(al) representation is:

$\mathrm{\δ}=\frac{\underset{n=1}{\overset{40}{\mathrm{\Σ}}}{x}_2^f\left(n\right)x_2\left(n\right)}{\sqrt{\underset{n=1}{\overset{40}{\mathrm{\Σ}}}{\left[x_2^f\left(n\right)\right]}^2\×\underset{n=1}{\overset{40}{\mathrm{\Σ}}}{\left[x_2\left(n\right)\right]}^2}}$

In the following formula, x ₂ ^fBe the fixed codebook search echo signal of last subframe, x ₂Signal for the fixed codebook search target of current subframe.

Wherein, described improved pulse alternate algorithm comprises step: the first step is to choose initialization code book vector, and the best code book result who promptly calls a subframe is as the initialization code book; Second step was the pulse alternative Process, needed to calculate the contribution of each pulse in codebook vectors, replaced the minimum pulse of contribution in the codebook vectors at every turn.

(3) beneficial effect

Owing to adopt above technical scheme, the present invention compared with the prior art, under the prerequisite of coding quality that guarantees to provide suitable with prior art and effect, the operand of its codebook search is reduced to the 50%-70% of former algorithm.

1) operand of voice signal processing reduces and can simplify search utility, reduces the total implementation complexity of scrambler, thereby reduces mobile communication equipment cost and power consumption.

2) search procedure of code book is accelerated in the minimizing of operand, thereby has reduced the coding time delay.In voice communication, end-to-end transmission to voice signal has very strong real-time requirement, the coding time delay that audio coder ﹠ decoder (codec) produced is one of significant effects factor, and the time delay that fixed codebook search algorithm is wherein introduced is the major part of coding time delay.Therefore, reduce the coding time delay, the voice quality that guarantees mobile communication is had progressive effect by improving the codebook search algorithm.

Description of drawings

Fig. 1 is the probability distribution graph of related coefficient between the adjacent voice subframe of the present invention;

Fig. 2 is that the depth-first tree algorithm of advising in institute's extracting method of the present invention and the AMR standard compares synoptic diagram.

Among the figure: correlation coefficient, related coefficient; Probability density, probability distribution; Pulse replacement times, pulse substitute number of times; Computationlaod, calculated amount; Algorithm in standardlized algorithm, the AMR standard; Ourproposed method, institute of the present invention extracting method.

Embodiment

Following examples are used to illustrate the present invention, but are not used for limiting the scope of the invention.

In order fully to disclose content of the present invention, before introducing specific implementation method of the present invention, at first introduce the codebook search principle.

In the AMR standard, the target of search optimal fixation code book is to make the square error minimum of input speech signal and synthetic speech signal.Square error is defined as ε _k=‖ x ₂-gHc _k‖ ², x wherein ₂Be fixed codebook search echo signal (be equal to the adaptive codebook echo signal and deduct the adaptive codebook contribution), H is the following triangle Teoplitz convolution matrix that is generated by perceptual weighting composite filter impulse response h (n), and g is the code book gain, c _kFor sequence number is the fixed codebook vector of k.Codebook vectors is one 40 dimension, and each pulse height is+1 or-1 vector.

Based on theoretical analysis, seek the process of best codebook vectors, just seek

$Q_k=\frac{{\left(C_k\right)}^2}{E_k}=\frac{{\left(d^T{c}_k\right)}^2}{c_k^T\mathrm{\φ}{c}_k}$ Peaked process, this equates and ask ε _kMinimum value.Wherein, d is echo signal x ₂(n) and perceptual weighting filter impulse response h (n) carry out the vector that related operation obtains, and φ is the autocorrelation matrix of h (n).Because at codebook vectors c _kIn only some is non-zero pulse, in standard, recommend to use the depth-first tree algorithm as the algorithm of codebook search fast.The definition b (n) for normalization vector d (n) and long-term prediction residual signals res _LTP(n) and, expression formula is as follows:

$b\left(n\right)=\frac{{\mathrm{res}}_{\mathrm{LTP}}\left(n\right)}{\sqrt{{\mathrm{\Σ}}_{i=0}^{i=39}{\mathrm{res}}_{\mathrm{LTP}}\left(i\right){\mathrm{res}}_{\mathrm{LTP}}\left(i\right)}}+\frac{d\left(n\right)}{\sqrt{{\mathrm{\Σ}}_{i=0}^{i=39}d\left(i\right)d\left(i\right)}}$

For simplicity, we are example with the codebook search of 12.2kb/s rate mode, and other rate mode is similar.When code book speed was 12.2kb/s, each codebook vectors included 10 non-zero pulses.The position of these non-zero pulses in vector can be divided into 5 tracks, comprises two non-zero pulses in each track.The structure of code book is as shown in table 1.

Algebraic codebook structure under the table 1 12.2kb/s pattern

Track	Pulse	The position
			1	i0，i1	0，5，10，15，20，25，30，35，
2	i2，i3	1，6，11，16，21，26，31，36
			3	i2，i3	2，7，12，17，22，27，32，37

4	i2，i3	3，8，13，18，23，28，33，38
			5	i2，i3	4，9，14，19，24，29，34，39

The depth-first tree algorithm of advising in the AMR standard comprises two stages.Phase one is all positions on each track of search, finds out b (n) local maximum and the corresponding position thereof of each track, selects global maximum from 5 local maximums, and its pulse position is as i0.Subordinate phase is that the position of pulse i1 is set to all the other four local maximums, repeats this and operates 4 times to guarantee that four local maximums are all got.After setting up i1 at every turn, with other pulse positions paired, sequencedly in nested loop, find promptly corresponding respectively { i2, i3}, { i4, i5}, { i6, i7} and { i8, i9}.Draw four code books of i1 at last,, therefrom draw optimum fixed codebook by comparison to them at four diverse locations.

Described pulse alternate algorithm is exactly to calculate the contribution of each pulse and replace a minimum pulse of contribution.In the method, at first select an initial codebook, then every non-zero pulses is changed to the local optimum position of each track, then carry out the pulse alternate algorithm.In each step, what at first select is the minimum pulse of contribution, and substitutes with the pulse seat that other contribution is big in the same track.Identical before new search code book and iteration, iterative process finishes.

Below specific implementation method of the present invention is described.

Under normal conditions, because the variation of voice signal has certain continuity, therefore the correlativity of two adjacent speech frames is very strong, and related coefficient may be very big.A speech frame is made up of four voice subframes, so also can obtain similar conclusion for adjacent voice subframe.Like this, we can utilize the optimum code book result of a subframe when carrying out the codebook search of current speech frame, thereby reduce the complexity of searching algorithm.Yet rapid variation can take place in voice signal sometimes, and the correlativity between the adjacent sub-frame is just smaller at this moment.In this case, still use the depth-first tree algorithm of advising in the AMR standard.

In view of above analysis, codebook searching method of the present invention is described below: the thresholding that at first defines a relative coefficient; All to calculate once the relative coefficient of it and last subframe to each subframe, compare with this threshold value then; If the threshold value greater than definition so just adopts improved pulse alternate algorithm; Otherwise, just adopt the depth-first tree algorithm of advising in the AMR standard.

If x ₂ ^fBe the fixed codebook search echo signal of last subframe, x ₂Be the fixed codebook search echo signal of current subframe, the fixed codebook degree of correlation that then defines current subframe and last subframe is

$\mathrm{\δ}=\frac{\underset{n=1}{\overset{40}{\mathrm{\Σ}}}{x}_2^f\left(n\right)x_2\left(n\right)}{\sqrt{\underset{n=1}{\overset{40}{\mathrm{\Σ}}}{\left[x_2^f\left(n\right)\right]}^2\×\underset{n=1}{\overset{40}{\mathrm{\Σ}}}{\left[x_2\left(n\right)\right]}^2}}$

The value of δ is big more, illustrates that the correlativity of adjacent sub-frame is big more, and the optimal fixation code book of two subframe correspondences is similar more.In this case, the codebook search of present frame can utilize the codebook search result of former frame.Therefore, we define a δ ₀, and δ ₀Value by obtaining by great number of statistic data.If δ＜δ ₀, still adopt the depth-first tree algorithm of advising in the AMR standard; Otherwise current subframe adopts pulse replacement algorithm to carry out best codebook search, and the code book initial value is the optimal fixation code book of previous frame.

Threshold delta ₀Determine to be based upon on the test basis of a large amount of speech datas, 22 voice documents that we carry the AMR standard are as test data.

At first, adopt the searching algorithm that provides in the AMR standard to obtain the codebook search result of each voice subframe in the voice document.The codebook search algorithm that uses us to propose then: set δ earlier ₀Value be a smaller value, search out the codebook vectors of each voice subframe of these voice documents, and this vector compared with the result who adopts canonical algorithm to obtain, add up the identical sub-frame number of codebook vectors that two kinds of algorithms obtain, with the accuracy rate of likening to of the quantity of all subframes in its quantity total value and 22 voice documents into our algorithm.If accuracy rate, thinks then that accuracy rate is too low less than 90%, increase δ with a certain step-length ₀, the algorithm that adopts us to propose carries out codebook search again, the result of Search Results and canonical algorithm is compared obtain new accuracy rate, more than rate of accuracy reached to 90%, stops circulation, determines current δ ₀Threshold value for needs.

Search for δ at us ₀Process in, our initial value of choosing is 0.4, step-length is 0.05, the threshold value that final search obtains is 0.6, this moment our algorithm rate of accuracy reached to more than 90%, obtained the communication efficiency suitable with canonical algorithm.In the reality,, this threshold value is adjusted dynamically according to the characteristics of speech data under the concrete environment with to the requirement of communication quality.

Be that example illustrates method of the present invention with AMR 12.2kb/s coding mode below.Concrete code book search step is as follows:

(1) the relative coefficient δ of current subframe of calculating and last subframe, and and threshold delta ₀Compare, if δ＞δ ₀, execution in step (2); If δ＜δ ₀, execution in step (3);

(2) codebook search to current subframe adopts the depth-first tree algorithm of advising in the AMR standard.

(3) adopt improved pulse alternate algorithm that this voice subframe is carried out codebook search.

The first step of improved pulse alternate algorithm is to choose initialization code book vector, and we call the best code book result of a subframe as the initialization code book.Second step was the pulse alternative Process, needed to calculate the contribution of each pulse in codebook vectors, replaced the minimum pulse of contribution in the codebook vectors at every turn.The pulse alternative steps is as follows:

(a) suppose m pulse cast out after, total contribution of its after pulse is Q _{K, m}, Q _{K, m}Can pass through Q _kExpression formula obtain.The contribution of this pulse is more little so, Q _{K, m}Value big more on the contrary.Therefore for the code book that 10 non-zero pulses are arranged (i0, i1, i2, i3, i4, i5, i6, i7, i8 i9), calculates 10 Q _{K, m}Value makes wherein Q _{K, m}Maximal value be Q _{K, m0}, then the contribution minimum of m0 pulse is replaced it with other positions on this track.

(b) calculate the Q that obtains after replace all available positions on this track _k, find out maximum Q _kValue and the position that it is corresponding are made as the reposition of m0.New code book vector and Q have been obtained like this _kIncrease that can be steady.Turn to step (a) then, replace iteration once more.If there is no than original big Q _k, then stopping pulse is replaced program, and this codebook vectors is the best code book of being asked.

Below in conjunction with accompanying drawing effect of the present invention is analyzed.

In this section, we will contrast the depth-first tree algorithm of advising in the AMR standard and the computation complexity of method proposed by the invention.Suppose the computation complexity of the algebraic codebook search of only considering a subframe.Analysis code speed is the coding mode of 12.2kb/s, and other speed has identical conclusion.

During the depth-first tree algorithm of in adopting the AMR standard, advising, for each voice subframe, if i ₀Given, i ₁Four kinds of different selections can be arranged.If i ₀And i ₁Simultaneously given, four pulses are to (i ₂, i ₃), (i ₄, i ₅), (i ₆, i ₇), (i ₈, i ₉) will be carried out to nested form search, so operand is (8 * 8) * 4=256.Add i ₁Four kinds of different values, the total operand is 256 * 4=1024.

According to codebook searching method proposed by the invention, each subframe can be selected two kinds of algorithms arbitrarily.When adopting improved pulse alternate algorithm, because our initialization code book vector is given, so operand is embodied in the alternative Process of pulse.Whenever substitute a pulse, the calculated amount that needs is (10+10), supposes to carry out N time and substitutes, and then the amount of calculation is (10+10) * N=20N.In general, the value of N is not too large, so operand is little more a lot of than the operand of canonical algorithm.Suppose δ＞δ ₀Probability be p, then the operand of a subframe internal fixation codebook search is 1024 (1-p)+p * 20N.Obviously, the size of operand can change and change along with p and N value.Relatively being described canonical algorithm and algorithm complex that we carry in the table 2.

Table 2 algorithm complex when N=30 compares

p	0	0.2	0.4	0.5	0.6	0.8	1
								The AMR canonical algorithm	1024	1024	1024	1024	1024	1024	1024
Institute of the present invention extracting method	1024	940	854	812	770	685	600

Notice that when the value of δ was smaller, if the code book of next subframe is selected as initial codebook, after using the pulse alternate algorithm, the possibility of result of codebook search was not optimum.Through after a large amount of voice signals tested, we find when δ greater than 0.6 the time, the performance of improved pulse alternate algorithm and AMR canonical algorithm is much at one.Therefore, when the value of δ was made as 0.6, codebook searching method proposed by the invention just can provide the performance identical with the AMR canonical algorithm.

Simultaneously, we have obtained the statistics of related coefficient between the adjacent sub-frame from a large amount of voice signals.Based on this statistics, we can obtain the probability distribution of related coefficient δ, as shown in Figure 1.

Because p is defined as δ＞δ ₀Probability, we can obtain its expression formula $p=\underset{\mathrm{\δ}>{\mathrm{\δ}}_o}{\∫}f\left(\mathrm{\δ}\right)\mathrm{d\δ}.$ Therefore, according to the probability distribution of δ, work as δ ₀=0.6 o'clock, p=0.66.Calculated amount-1024 (1-p)+p * 20N that so just can guarantee our algorithm can obtain good performance.Can know that based on a large amount of experiments the value of N usually can be greater than 40.Fig. 2 has provided the comparison of the calculated amount of method of the present invention in this case and AMR canonical algorithm.Obviously, the present invention proposes the calculated amount of method will be much smaller than the AMR canonical algorithm, and its calculated amount is reduced to AMR canonical algorithm 50%-70%.When adopting the coding mode of other speed, also can obtain similar conclusion.

Claims (3)

1, a kind of AMR scrambler rapid fixed codebook searching method of enhancing is characterized in that following steps are arranged:

1), at first define the threshold value of the related coefficient of an adjacent speech frame, this threshold value provides benchmark for the search of speech frame code book;

2), all will calculate the relative coefficient of it and last subframe, compare with the described threshold value of step 1) then each subframe;

3), for step 2) in relatively result, if greater than the described threshold value of step 1), so just adopt improved pulse alternate algorithm; If less than the described threshold value of step 1), just adopt the depth-first tree algorithm of advising in the AMR standard.

2, the AMR scrambler rapid fixed codebook searching method of a kind of enhancing as claimed in claim 1, it is characterized in that: related coefficient is the fixed codebook degree of correlation of current subframe and last subframe, and its mathematic(al) representation is:

$\mathrm{\δ}=\frac{\underset{n=1}{\overset{40}{\mathrm{\Σ}}}{x}_2^f\left(n\right)x_2\left(n\right)}{\sqrt{\underset{n=1}{\overset{40}{\mathrm{\Σ}}}{\left[x_2^f\left(n\right)\right]}^2\×\underset{n=1}{\overset{40}{\mathrm{\Σ}}}{\left[x_2\left(n\right)\right]}^2}}$

Wherein, x ₂ ^fBe the fixed codebook search echo signal of last subframe, x ₂Signal for the fixed codebook search target of current subframe.

3, the AMR scrambler rapid fixed codebook searching method of a kind of enhancing as claimed in claim 1, it is characterized in that, described improved pulse alternate algorithm comprises step: the first step is to choose initialization code book vector, and the best code book result who promptly calls a subframe is as the initialization code book; Second step was the pulse alternative Process, needed to calculate the contribution of each pulse in codebook vectors, replaced the minimum pulse of contribution in the codebook vectors at every turn.

Images