CN1165036C - Blocks length selection method based on adaptive threshold and typical sample predication - Google Patents

Abstract

The present invention relates to a block length selection component in the transition coding of voice frequency signals, particularly to a block length selection method on the basis of the prediction of self-adaptive thresholds and typical samples in the compression of voice frequency signals. The provided novel method obtains high coding efficiency at the same time of complete elimination of pre-echo noise with small required operation quantity. Two criteria relevant to the measurement of jump signals are suggested. The proposed 'local maximum masking analysis method' can prevent the overuse of short block lengths for optimizing the rate-distortion property of encoders. The present invention also provides a rapid operation scheme on the basis of a sub-sampling mechanism.

Description

A kind of length selection method based on adaptive threshold and typical sample prediction

Technical field

The present invention relates to the data compression in the Audio Signal Processing, in more detail, it is in order to eliminate the pre-echo noise (pre-echo) in the high product encode/decode audio signal of the low bit rate process; The present invention be more particularly directed to predict the length selection method that carries out based on adaptive threshold and typical sample in a kind of audio signal compression.

Prior art

As everyone knows, some linear transformations can cause approaching zero high frequency coefficient, in other words, the most information that time-domain signal comprises can be converted or focus on a son of frequency domain or time-frequency domain coefficient and concentrate, so the audio signal compression technology adopts conversion as the means that improve code efficiency widely; These audio signal compression technology are included in the following document:

(1)R.N.J.Veldhuis，“Bit?Rates?in?Audio?Source?Coding”，IEEE?J.SelectedAreas?in?Communications，vol.10，pp.86-96，Jan.1992.

(2)J.D.Johnston?in?“Perceptual?Transform?Coding?of?Wide-band?StereoSignals”，ICASSP’89，Glasgow，Scotland，pp.1993-1996，May?1989.

(3)“Information?Technology¨CCoding?of?Moving?Pictures?and?AssociatedAudio?for?Digital?Storage?Media?at?up?to?About?1.5?Mbit/s?Part?3：Audio(ISO/IEC?11?172-3：1993)”.

(4)“The?MD?system?description?document(Sony?Inc.：Sept.1992).

(5)“MPEG-2?AAC，ISO/IEC?13818-7：1997(E)”

And the encode/decode audio signal scheme that present major part is developed recently all adopts Modified DiscreteCosine Transform (MDCT) as its bank of filters, MDCT proposed (" Subband/Transform Coding Using Filter BankDesigns Based on Time Domain Aliasing Cancellation " by Princen and Bradley in 1987, Proceedings of theICASSP 1987, pp 2161 ¨ C2164), it becomes isometric subband to a signal decomposition with overlapping block with dynamic window, because the speciality of cosine series function and the quantizing noise in the encoding-decoding process, the use meeting of coefficient reconstruct and reverse MDCT forms ripple around the hop signal that synthesizes, human auditory system's backward masking timeliness is longer than the forward masking timeliness, the rear end of PCM frame is in the backward masking timeliness usually, therefore, ripple after the prominent exceeding signal can not be heard, if the forward masking timeliness can not cover the front end of PCM frame, the ripple of prominent exceeding signal front can form appreciable noise (as shown in Figure 1), and this noise is called as pre-echo.

For suppressing or eliminating this pre-echo, less MDCT block length should be used so that the ripple of prominent exceeding signal front is limited in the forward masking timeliness, because block length is directly proportional with code efficiency, a kind of algorithm of selecting the MDCT block length automatically and accurately can be eliminated or suppress pre-echo and optimize the distortion performance of codec.

In the past few years, some are used to suppress or the length selection method of eliminating pre-echo is suggested and is used in different systems, and Fig. 2 has described the length selection method in the MD system (ATRAC).This method is proposed in September, 1992 by Sony Corporation.As shown in Figure 2, step 601MD block length is selected beginning, in the voice data decomposition framing of step 602 with input, step 603 detects each peak value by 32 subframes that constitute, the peak value of continuous subframe compares in step 604, and step 605 is for the decision-making part, if the difference of the peak value of continuous subframes is greater than 18dB, short block selected (represented by step 606), MD block length select to finish (step 607); Otherwise, current number of sub frames is added 1 (by step 611 expression), if subframe does not surpass the upper limit (step 610), then being circulated to step 604 iteration selects, otherwise long piece selected (represented by step 609), to be applied to different frequency bands, the MD block length selects to finish (step 608).In this block length selection course, the peak value of contiguous PCM subframe is extracted out to be used as the wave mode feature, then, is performed based on the classification of waveform character, obviously, has lost a large amount of shape informations in the process of feature extraction; So the selection precision of this method can be very not high.In other words, between frame number that this method is selected and the optimum number bigger deviation is arranged with relatively large length; Not enough or exceedingly use short block will cause the unnecessary decline of pre-echo noise or code efficiency, the use of this straightforward procedure is based on considering of complexity to a great extent.

In MPEG-2 AAC standard, block length is definite according to perceptual entropy (perceptual entropy), perceptual entropy (PE) is defined as energy threshold, the function of the energy of quiet threshold values and frequency spectrum each several part, MPEG-2 AAC calculates and the corresponding PE value of various block lengths earlier, if the PE value of long piece is bigger, short block is used, otherwise, long piece is used, this method is based upon psychologic acoustics about on stable state/astable notion, it is based on the frequency domain solution of global information (all coefficient of frequencies), directly cause on the original time-domain signal of pre-echo of short duration and significant variation does not obtain enough emphasizing and utilizing, therefore, the elimination of pre-echo will be accompanied by the excessive descent of code efficiency; The performance cost ratio can not be very high.

As mentioned above, when using the audio coder of piece conversion, input signal is of short duration on the time domain and significant variation (hop signal) can cause the pre-echo noise when the forward masking timeliness can not cover the front end of PCM frame; In order to suppress or to eliminate this pre-echo, short piece should be used so that the ripple of hop signal front is limited in the forward masking timeliness, yet, thereby less block length causes the decline of frequency domain resolution to reduce code efficiency inevitably, to having the input term signal of hop speciality, between code efficiency and pre-echo elimination, there is an equilibrium problem.

The content of invention

The object of the present invention is to provide in the audio signal compression and predict the length selection method that carries out, under the prerequisite of eliminating the pre-echo noise, at utmost improve and optimize code efficiency based on adaptive threshold and typical sample.

A kind of length selection method based on adaptive threshold and typical sample prediction of the present invention is removed the pre-echo noise fully under the prerequisite that keeps code efficiency, it is characterized in that method is made up of following steps:

A) according to the configuration of the conversion of being adopted, input audio data is decomposed framing;

B) above-mentioned frame is further resolved into S isometric subframe, find out the peak value of PCM data absolute value on each subframe, in the peak value of each subframe, select those local maximum points;

C) predict the typical sample value that is positioned at d subframe place before the local maximum point with several subframe peak value pi of above-mentioned local maximum point front, calculate current local maximum point and the difference and the ratio of the typical sample value that doped;

D) according to the subframe peak value of the difference that is calculated and ratio, front and corresponding to a series of adaptive thresholds of optional block length, determine the best block length that is associated with this local maximum point;

E) according to the shared number percent in totalframes of the frame number with specific block length, adjust corresponding current threshold values;

F) selected or last local maximum point is reached up to the shortest block length to repeat aforesaid operations;

G) with block length that each local maximum point is associated in, select the block length of reckling as this frame.

Subframe peak value p wherein _iThe method of front d subframe place typical sample value of prediction:

$S\left(d\right)=\frac{1}{1+2\·\underset{j=1}{\overset{N}{\mathrm{\Σ}}}{2}^{-j}}(p_{i-d}+\underset{j=1}{\overset{N}{\mathrm{\Σ}}}[(p_{i-d-j}+p_{i-d+j})/2^j]$

Wherein: S (d) is sub-Frame peak value p _iThe forecast sample peak value that postpones d sub-Frame place forward;

p _kIt is the sample peak value of k sub-Frame;

N is half of the related sub-Frame number of computing.

The tolerance criterion of wherein related hop input, above-mentioned peak value is determined as follows with the difference and the ratio of the typical sample value that is doped:

$D\left(d\right)=p_i-S\left(d\right)$

$=p_i-\frac{1}{1+2\·\underset{j=1}{\overset{N}{\mathrm{\Σ}}}{2}^{-j}}(p_{i-d}+\underset{j=1}{\overset{N}{\mathrm{\Σ}}}[(p_{i-d-j}+p_{i-d+j})/2^j]$

$R\left(d\right)=p_i/S\left(d\right)$

$=p_i\·(1+2\underset{j=1}{\overset{N}{\mathrm{\Σ}}}{2}^{-j})/(p_{i-d}+\underset{j=1}{\overset{N}{\mathrm{\Σ}}}[(p_{i-d-j}+p_{i-d+j})/2^j]$

A kind of avoiding excessively used the method long than short block, and " local maximum masking analytic approach " is presented below:

A) find out before the current local maximum point first and possess the local maximum peak dot of backward masking pre-echo noise, the pre-echo noise is here caused by current local maximum point;

B) if before the peak dot of finding out, there is the subframe that peak value is enough little with backward masking ability, be the shortest block length with the elimination pre-echo, otherwise, use longer piece to improve code efficiency.

Wherein the threshold values method of adjustment is made of following steps:

A) calculate the number percent that the frame number that adopts specific block length accounts for totalframes;

B) threshold values of correspondence is increased or reduce a step-length, to control corresponding number percent, this step-length and above-mentioned number percent are proportional, and adjusted threshold values is that next incoming frame is used.

Fast definite method of each the subframe peak value that wherein relates to is made of following step:

A) to the absolute value { x of a frame PCM signal ₀, x ₁..., x _LCarry out sub sampling, at the sub sampling version Find out each subframe peak point in (M is the sub sampling factor);

B) be in (2M-1)-sample neighborhood at center with each sub sampling peak point, finding out the maximum as original PCM frame { x ₀, x ₁..., x _LThe subframe peak value.

Description of drawings

Fig. 1 is based in the codec of MDCT, the synthetic caused ripple of hop signal.

Fig. 2 is the length selection method flow process that MD (ATRAC) scrambler uses.

Fig. 3 is the block diagram of low complex degree MPEG-2 AAC scrambler.

Fig. 4 is an operation process involved in the present invention.

Fig. 5 is a block length selection algorithm block diagram of the present invention.

Specific implementation

The technical scheme that realizes the object of the invention is to predict the length selection method that carries out based on adaptive threshold and typical sample in a kind of audio signal compression, under the prerequisite that keeps code efficiency, remove the pre-echo noise fully, it is characterized in that method is made up of following steps:

A) according to the configuration of the conversion of being adopted, input audio data is decomposed framing;

B) above-mentioned frame is further resolved into S isometric subframe, find out the peak value of PCM data absolute value on each subframe, in the peak value of each subframe, select those local maximum points;

C) with several subframe peak value p of above-mentioned local maximum point front _iPredict the typical sample value that is positioned at d subframe place before the local maximum point, calculate current local maximum point and the difference and the ratio of the typical sample value that doped;

D) according to the subframe peak value of the difference that is calculated and ratio, front and corresponding to a series of adaptive thresholds of optional block length, determine the best block length that is associated with this local maximum point;

E) according to the shared number percent in totalframes of the frame number with specific block length, adjust corresponding current threshold values;

F) selected or last local maximum point is reached up to the shortest block length to repeat aforesaid operations;

G) with block length that each local maximum point is associated in, select the block length of reckling as this frame.

The subframe peak value p that is further characterized in that of the present invention _iThe method of front d subframe place typical sample value of prediction:

$S\left(d\right)=\frac{1}{1+2\·\underset{j=1}{\overset{N}{\mathrm{\Σ}}}{2}^{-j}}(p_{i-d}+\underset{j=1}{\overset{N}{\mathrm{\Σ}}}[(p_{i-d-j}+p_{i-d+j})/2^j]$

Wherein S (d) is sub-Frame peak value p _iThe forecast sample peak value that postpones d sub-Frame place forward.

p _kIt is the sample peak value of k sub-Frame.

N is half of the related sub-Frame number of computing.

The tolerance criterion of hop input involved in the present invention, above-mentioned peak value is determined as follows with the difference and the ratio of the typical sample value that is doped:

$D\left(d\right)=p_i-S\left(d\right)$

$=p_i-\frac{1}{1+2\·\underset{j=1}{\overset{N}{\mathrm{\Σ}}}{2}^{-j}}(p_{i-d}+\underset{j=1}{\overset{N}{\mathrm{\Σ}}}[(p_{i-d-j}+p_{i-d+j})/2^j]$

$R\left(d\right)=p_i/S\left(d\right)$

$=p_i\·(1+2\underset{j=1}{\overset{N}{\mathrm{\Σ}}}{2}^{-j})/(p_{i-d}+\underset{j=1}{\overset{N}{\mathrm{\Σ}}}[(p_{i-d-j}+p_{i-d+j})/2^j]$

A kind of method of avoiding excessively using than short block length of the block length that the present invention is the shortest, " local maximum masking analytic approach ", be presented below: find out before the current local maximum point first and possess the local maximum peak dot of backward masking pre-echo noise, the pre-echo noise is here caused by current local maximum point; If before the peak dot of finding out, there is the subframe that peak value is enough little with backward masking ability, be the shortest block length with the elimination pre-echo, otherwise, use longer piece to improve code efficiency.

Threshold values method of adjustment of the present invention is made of following steps:

A) calculate the number percent that the frame number that adopts specific block length accounts for totalframes;

B) threshold values of correspondence is increased or reduce a step-length, to control corresponding number percent, this step-length and above-mentioned number percent are proportional, and adjusted threshold values is that next incoming frame is used.

Fast definite method of each the subframe peak value that the present invention relates to is made of following step:

A) to the absolute value { x of a frame PCM signal ₀, x ₁... .., x _LCarry out sub sampling, at the sub sampling version Find out each subframe peak point in (M is the sub sampling factor);

B) be in (2M-1)-sample neighborhood at center with each sub sampling peak point, finding out the maximum as original PCM frame { x ₀, x ₁...., x _LThe subframe peak value.

As a kind of scheme of eliminating pre-echo, block length was determined before carrying out conversion usually, and the target that block length is selected is: under the prerequisite of eliminating the pre-echo noise, improve block length as far as possible, reduce operand, to optimize code efficiency, operand is another important index of code efficiency.The present invention is based upon on the theory of psychologic acoustics about forward direction and backward masking timeliness, the so-called PCM Frame time that is divided into is gone up isometric subframe, the duration of each subframe is about half (1.5ms) of forward masking timeliness, the peak dot of PCM data absolute value is found out on each subframe, identifies those local maximums from these peak points.

Several subframe peak values of a local maximum point front are used to the typical sample value that prediction is positioned at a current local maximum point front d subframe place, the difference of current local maximum point and predicted value and ratio are used as the criterion that detects the hop signal, " local maximum masking analytic approach " proposed by the invention is used to the psychologic acoustics theory to avoid the excessive use of short block, thereby optimize the rate-distortion performance of codec, a sub sampling mechanism is used to reduce the needed operand of process of seeking the subframe peak value, and other means comprise: confirm whether there is the subframe that has than low peak between frame front end and forward masking timeliness front end.

When application was of the present invention, staged operation was performed: according to the configuration of the conversion of being adopted, input audio data is decomposed framing, frame further is broken down into subframe, and then, compare operation is performed to determine the peak dot of each subframe.If a subframe peak dot is a local maximum, the typical sample value at its forward delay d subframe place is by linear prediction.The difference of local maximum and forecast sample value and ratio are calculated, if difference that calculates and ratio are all greater than separately thresholding, judge and have the hop signal, confirm to possess the local maximum peak dot of backward masking pre-echo, if between frame front end and forward masking timeliness front end, there is the subframe that peak value is enough little, use than short block to eliminate the pre-echo noise, otherwise, use longer piece to improve code efficiency, repeating above-mentioned block length selects computing selected or last local maximum point is reached up to the shortest block length, with block length that each local maximum point is associated in, select the block length of reckling as this frame, the process of looking for each subframe peak value can be simplified: search sub sampling version subframe is is also evaluated and tested the neighborhood of sub sampling peak value, and the process of aforesaid operations as shown in Figure 4.

The related block length decision component of MPEG-2 AAC scrambler is selected as implementation platform of the present invention, the framework of low complex degree MPEG-2 AAC scrambler as shown in Figure 3, an input audio signal is sampled with 44.1kHz, sampled signal is divided framing, every frame is formed (about 23.22ms) by 1024 samples, psychoacoustic model utilizes human auditory system's occlusion to remove imperceptible content from input signal frame, simultaneously, signal frame is buffered, then, the block length of buffered signal frame is determined.Then, carry out MDCT, subsequently, the quantizing noise of MDCT frequency spectrum is shaped by transient state, and intensity coupling module utilizes the insensitivity of high band minor matters information between a pair of sound channel with enhancing rate-distortion performance; Master/slave decomposition (Middle/Side) utilizes the coding of " the ears masking level constrains " characteristic with control noise and transient state/voice signal.At last, pretreated data are quantized and encode, and index value and minor matters information is packaged advances bit stream.

MPEG-2 AAC relates to two kinds of possible block lengths, long piece (2048 point) and short block (256 point), and realization details of the present invention such as following steps are described:

See also shown in Figure 5:

Step 801 begins to carry out this algorithm;

Step 802 is decomposed framing (1024 point) with input audio data;

Step 803 further resolves into 16 subframes (64 point) to incoming frame, calculates the peak value of PCM data absolute value on each subframe, finds the subframe with maximum absolute peak, and establishing its sequence number is b ^*

Step 804 is with sequence number b ^*Assignment is given b;

Step 805 judges that whether b sub-frame peak be greater than the sample around it.If then carry out 806, if otherwise carry out 812;

Step 806 is predicted the typical sample value at its 4 subframe places of forward delay relatively with 3 sub-frame peaks of subframe b front.The difference and the ratio of the typical sample value of calculating local maximum point and being doped;

$S\left(4\right)=\frac{1}{1+2\·\underset{j=1}{\overset{1}{\mathrm{\Σ}}}{2}^{-j}}(p_{i-4}+\underset{j=1}{\overset{1}{\mathrm{\Σ}}}[(p_{i-4-j}+p_{i-4+j})/2^j]$

$D\left(4\right)=p_i-S\left(4\right)$

$=p_i-\frac{1}{1+2\·\underset{j=1}{\overset{1}{\mathrm{\Σ}}}{2}^{-j}}(p_{i-4}+\underset{j=1}{\overset{1}{\mathrm{\Σ}}}[(p_{i-4-j}+p_{i-4+j})/2$

$R\left(4\right)=p_i/S\left(4\right)$

$=p_i\·(1+2\underset{j=1}{\overset{1}{\mathrm{\Σ}}}{2}^{-j})/(p_{i-4}+\underset{j=1}{\overset{1}{\mathrm{\Σ}}}[(p_{i-4-j}+p_{i-4+j})/2^j]$

Wherein: D (4) and R (4) are respectively local maximum points and the difference and the ratio of the typical sample value that is doped.

Judge that in step 807 with the difference of forecast sample value with than whether all greater than the threshold value that sets, there is the hop signal in peak value if then show this frame then, execution in step 808, if otherwise execution in step 812;

Step 808 is carried out the analysis of peak value backward masking;

Step 809 judges whether at the frame front end and shelters to have the subframe that peak value is enough little between the 2.5ms place, peak dot front, if then present frame is selected short piece (step 810), finishes this algorithm execution in step 811, if otherwise execution in step 812;

Step 812 is changed to present frame b=b+1 with next subframe;

Whether step 813 judges b less than 2, if otherwise enter step 805 cycle criterion again, be reached if then show last local maximum point, present frame is selected long piece (step 814), finish this algorithm and carry out (step 815).

The present invention has eliminated the pre-echo noise that is caused by the hop signal fully, and has obtained very high coding efficiency.Needed operand is very little.

Claims (6)

1, a kind of length selection method based on adaptive threshold and typical sample prediction is removed the pre-echo noise fully under the prerequisite that keeps code efficiency, it is characterized in that method is made up of following steps:

A) according to the configuration of the conversion of being adopted, input audio data is decomposed framing;

B) above-mentioned frame is further resolved into S isometric subframe, find out the peak value of PCM data absolute value on each subframe, in the peak value of each subframe, select those local maximum points;

C) with several subframe peak value p of above-mentioned local maximum point front _iPredict the typical sample value that is positioned at d subframe place before the local maximum point, calculate current local maximum point and the difference and the ratio of the typical sample value that doped;

D) according to the subframe peak value of the difference that is calculated and ratio, front and corresponding to a series of adaptive thresholds of optional block length, determine the best block length that is associated with this local maximum point;

E) according to the shared number percent in totalframes of the frame number with specific block length, adjust corresponding current threshold values;

F) selected or last local maximum point is reached up to the shortest block length to repeat aforesaid operations;

G) with block length that each local maximum point is associated in, select the block length of reckling as this frame.

2, the length selection method based on adaptive threshold and typical sample prediction according to claim 1 is characterized in that subframe peak value p _iThe method of front d subframe place typical sample value of prediction:

$S\left(d\right)=\frac{1}{1+2\·\underset{j=1}{\overset{N}{\mathrm{\Σ}}}{2}^{-j}}(p_{i-d}+\underset{j=1}{\overset{N}{\mathrm{\Σ}}}[(p_{i-d-j}+p_{i-d+j})/2^j\left]\right)$

Wherein: S (d) is sub-Frame peak value p _iThe forecast sample peak value that postpones d sub-Frame place forward;

p _kIt is the sample peak value of k sub-Frame;

N is half of the related sub-Frame number of computing.

3, the length selection method based on the prediction of adaptive threshold and typical sample according to claim 1 is characterized in that the tolerance criterion of related hop input, and above-mentioned peak value is determined as follows with the difference and the ratio of the typical sample value that is doped:

$D\left(d\right)=p_i-S\left(d\right)$

$=p_i-\frac{1}{1+2\·\underset{j=1}{\overset{N}{\mathrm{\Σ}}}{2}^{-j}}{p}_{i-d}+\underset{j=1}{\overset{N}{\mathrm{\Σ}}}[(p_{i-d-j}+p_{i-d+j})/2^j]$

$R\left(d\right)=p_i/S\left(d\right)$

$=p_i\·(1+2\underset{j=1}{\overset{N}{\mathrm{\Σ}}}{2}^{-j})/(p_{i-d}+\underset{j=1}{\overset{N}{\mathrm{\Σ}}}[(p_{i-d-j}+p_{i-d+j})/2^j].$

4, the length selection method based on adaptive threshold and typical sample prediction according to claim 1 is characterized in that a kind of method of avoiding excessively using than short block length, and " local maximum masking analytic approach " is presented below:

A) find out before the current local maximum point first and possess the local maximum peak dot of backward masking pre-echo noise, the pre-echo noise is here caused by current local maximum point;

B) if before the peak dot of finding out, there is the subframe that peak value is enough little with backward masking ability, be the shortest block length with the elimination pre-echo, otherwise, use longer piece to improve code efficiency.

5, the length selection method based on adaptive threshold and typical sample prediction according to claim 1 is characterized in that the threshold values method of adjustment is made of following steps:

A) calculate the number percent that the frame number that adopts specific block length accounts for totalframes;

B) threshold values of correspondence is increased or reduce a step-length, to control corresponding number percent, this step-length and above-mentioned number percent are proportional, and adjusted threshold values is that next incoming frame is used.

6, the length selection method based on the prediction of adaptive threshold and typical sample according to claim 1, what it is characterized in that each subframe peak value of relating to determines that fast method is made of following step:

A) to the absolute value { x of a frame PCM signal ₀, x ₁..., x _LCarry out sub sampling, at the sub sampling version Find out each subframe peak point in (M is the sub sampling factor);

B) be in (2M-1)-sample neighborhood at center with each sub sampling peak point, finding out the maximum as original PCM frame { x ₀, x ₁..., x _LThe subframe peak value.

Images