CN101193090A - Signal processing method and its device - Google Patents
Signal processing method and its device Download PDFInfo
- Publication number
- CN101193090A CN101193090A CNA2006101403202A CN200610140320A CN101193090A CN 101193090 A CN101193090 A CN 101193090A CN A2006101403202 A CNA2006101403202 A CN A2006101403202A CN 200610140320 A CN200610140320 A CN 200610140320A CN 101193090 A CN101193090 A CN 101193090A
- Authority
- CN
- China
- Prior art keywords
- evolution
- difference
- signal
- spectrum
- frequency
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Granted
Links
Images
Abstract
The invention discloses a signal processing method and a device thereof. The signal processing method comprises the steps that: the linear prediction analysis is applied to the signal in the time domain so as to obtain the predicted residual signal; the extraction of root processing is applied to the spectrum amplitude of the residual signal or spectrum amplitude difference data; the parameters used for the extraction of root processing are recorded and the unified processing and constant or vector quantization are applied to the spectrum amplitude or spectrum amplitude difference data after being extracted . The invention can effectively flatten the signal in a better way.
Description
Technical field
The present invention relates to a kind of signal processing method and device thereof.
Background technology
In the signal processing field, the processing of voice, audio signal generally will be through filtering or to the time-domain signal windowing, after carry out time-frequency conversion, transform to frequency domain and handle, so that encode or transmit.
Wherein, the transform domain auditory model of audio frequency is mainly considered two factors, one is the threshold of hearing threshold values, another is the frequency domain capture-effect, and using the purpose of threshold of hearing threshold values is to make the quantization error of each frequency be lower than the threshold of hearing threshold values, avoids being heard, the this point ratio is easier to realize, both can finish by the quantization step of adjusting each frequency at frequency domain, it is smooth also can to make that with the FIR filter weight frequency domain threshold of hearing threshold values flattens in time domain, and realizes in the frequency domain uniform quantization.The frequency domain capture-effect changes along with the fluctuation of signal frequency domain amplitude, only exceeding the part of covering curve just is quantized, that is to say that quantizing noise is lower than slightly covers curve and gets final product, but the mutual capture-effect of calculating between each Frequency point of frequency domain is more loaded down with trivial details thing, has very high processing complexity.
In the transform domain processing procedure, if the frequency coefficient quantization encoding adopts identical signal to noise ratio, then all quantization parameters has an identical quantification number of significant digit; If adopt the quantification treatment of rule, at first will solve big problem is the distribution problem of coefficient in transform domain, just consider to seek a useful quantitative method then, this is because the data ratio of relatively concentrating that distributes is easier to find easy quantization method, if the data distribution is bigger, just means that also the regularity of distribution is complicated and be difficult to simplify processing.
If frequency-region signal has bigger distribution, the quantification of frequency spectrum, coding acquire a certain degree of difficulty.If the effective range that can effectively compress frequency-region signal just can improve the efficient of follow-up coding link greatly.The prior art proposition adopts the method for flattened spectral response effectively to compress the effective range of frequency-region signal, such as at (the MPEG4 of Motion Picture Experts Group, Motion Picture Exlperts Group 4) comprised two standards in the T/F coding tools: one is the advanced speech (AAC that encodes frequently, Advance Audio Coding), another is the transform domain weighting vector quantization (TwinVQ, Transform-domain Weighted InterleaveVector Quantization) that interweaves.Wherein the TwinVQ standard mainly is to adopt the method compression frequency-region signal of planarization conversion coefficient, promotes the coding effect of low code check audio signal.TwinVQ is made of three major techniques: 1) utilize linear predictive coding (LPC, Linear Predictive Coding) coefficient comes modified model discrete cosine transform (MDCT) conversion coefficient is done planarization, that 2) utilizes that interframe predicts backward comes the MDCT coefficient to be done planarization, 3) the weighting vector quantization that interweaves.
The MDCT conversion coefficient planarization of TwinVQ comprises two links, a frequency coefficient that is transform domain MDCT coefficient divided by LPC, obtain predicting frequency-domain residual, another link is the average power that calculates each critical band, as the critical band envelope, a plurality of coefficients in each critical band are divided by its average power, and it is smooth that it is further flattened.Fig. 1 is the concrete process flow block diagram of TwinVQ, comprises step:
101, lpc analysis: calculate the auto-correlation coefficient of input signal data, go out filter coefficient by Lie Wen-Du Bin algorithm computation;
102, the LPC coefficient is to the conversion of (LSP) coefficient;
103, LSP vector quantization of coefficient;
104, quantize the LSP coefficient and transform, calculate the frequency domain value of LPC then to the LPC coefficient;
105, Frame is carried out the MDCT conversion:
106, critical band scale factor calculation;
107, critical band package (scale factor) quantizes;
108, energy normalized.
Above method is at frequency domain audio frequency to be carried out the processing of planarization, but has following technical problem: (1) realizes the planarization of frequency coefficient at frequency domain with linear forecast coding coefficient, and frequency domain is handled frame by frame, can cause the discontinuous of signal characteristic interframe switching instant; (2) do planarization with the critical band average power, calculate a large amount of divisions, make the foundation that grouping is divided with critical band simply, do not consider to utilize smooth, the difference characteristic of actual spectrum, the planarization effect is difficult to accomplish very desirable.
Summary of the invention
The invention provides the comparatively simple simultaneously signal processing method of a kind of raising signal planarization level for solving the problems of the technologies described above.
The present invention also provides a kind of raising signal planarization level comparatively simple simultaneously signal processing apparatus.
A kind of signal processing method comprises: signal is carried out linear prediction analysis in time domain, the residual signals that obtains predicting; Spectrum amplitude or spectrum amplitude variance data to described residual signals are carried out the evolution processing; Write down described evolution and handle the parameter of using, and spectrum amplitude behind the evolution or spectrum amplitude variance data are unified processing, constant or vector quantization.
A kind of signal processing apparatus comprises: linear prediction analysis unit, signal is carried out linear prediction analysis in time domain, the residual signals that obtains predicting; The evolution processing unit carries out the evolution processing to the spectrum amplitude or the spectrum amplitude variance data of described residual signals; Quantifying unit writes down described evolution and handles the parameter of using, and spectrum amplitude behind the evolution or spectrum amplitude variance data are unified processing, constant or vector quantization.
Above-mentioned signal processing method carries out linear prediction analysis and the amount of calculation that causes is big and the discontinuous problem of Frame switching instant with respect to prior art at frequency domain, because the mode that employing is carried out linear prediction analysis in time domain is carried out signal filtering and other processing, signal is to pass through continuously, does not need to transform to frequency domain earlier and the calculating carried out; Simultaneously, carry out the envelope characteristic that linear prediction analysis is equivalent to extract frequency-region signal, and eliminate its influence, can be to the preliminary planarization of signal; Again, employing is carried out the evolution processing to the spectrum amplitude of residual signals, by frequency spectrum difference ratio evolution is handled the disparity range of dwindling spectral line, can at utmost realize planarization, owing to keep evolution information in quantizing process, the data after the planarization can be recovered out at an easy rate again; Thereby do planarization with respect to prior art with the critical band average power simultaneously and will calculate a large amount of divisions, the all flattened spectral responses of present embodiment are handled and are all handled at log-domain, calculating of many complexity to simplify, solved the algorithm computation complexity issue, signal processing efficient improves greatly; The processing method of evolution is beneficial to follow-up coding and processing very much, improves the decoding quality.
Said signal processing device is owing to adopt linear prediction analysis unit that signal is carried out linear prediction analysis in time domain, and therefore signal is continuous when forecast analysis; Simultaneously, carry out the envelope characteristic that linear prediction analysis is equivalent to extract frequency-region signal, and eliminate its influence, can be to the preliminary planarization of signal; Because adopting the evolution processing unit that the bin magnitudes of described residual signals is carried out evolution handles, by frequency spectrum difference ratio evolution being handled the disparity range of dwindling spectral line, can at utmost realize planarization, owing to keep the evolution parameter in quantizing process, the data after the planarization can be recovered out at an easy rate again; Calculating to simplify of many complexity solved the algorithm computation complexity issue, and signal processing efficient improves greatly; Be beneficial to follow-up coding and processing, improve the decoding quality.
Description of drawings
Fig. 1 is the flow chart of prior art signal flattening method;
Fig. 2 is the flow chart of signal processing method execution mode of the present invention;
Fig. 3 carries out the frequency envelope figure that linear prediction obtains respectively by low-frequency range, high band in Fig. 2 flow process;
Fig. 4 carries out the spectrum envelope figure that linear prediction obtains to low-frequency range in Fig. 2 flow process;
Fig. 5 handles the reference curve figure that obtains to the low-frequency range frequency spectrum in Fig. 2 flow process;
Fig. 6 is the schematic diagram that reference curve figure among Fig. 5 is divided into groups;
Fig. 7 is the planarization design sketch to second grouping of reference curve among Fig. 6;
Fig. 8 is the planarization design sketch to the band evolution number of times of second grouping of reference curve among Fig. 6;
Fig. 9 is the structure chart of signal processing apparatus execution mode of the present invention.
Embodiment
The present invention carries out linear prediction analysis to signal in time domain, the residual signals that obtains predicting, and the spectrum amplitude of described residual signals or spectrum amplitude variance data are carried out evolution handle the planarization frequency spectrum.
Time domain linear prediction wherein is equivalent to extract the envelope characteristic of frequency-region signal, and eliminate its influence, its treatment effect directly perceived is to have howed than smooth before handling through the spectral shape of time domain linear prediction back signal, if can accomplish the complete planarization of frequency domain spectral shape, then the amount of information of signal drops to minimum at this moment, all information of signal just are included in and make in the smooth process of signal spectrum, if can be contained information extraction of these processing procedures and expression, then signal just can be encoded well.Based on this criterion, starting point of the present invention is to make in the planarization of frequency-region signal, by frequency spectrum difference ratio evolution is handled the disparity range of dwindling spectral line, keeps different information again, can recover out at an easy rate.Further, scope by signal spectrum reality is carried out the several times evolution to it, can be controlled at the spectral line value behind the evolution small range,, just can guarantee to recover original signal value by certain signal to noise ratio as long as the effective in other words quantization digit of quantified precision of the value of n power is opened in control rationally.
The invention provides a kind of signal processing method basic embodiment, described signal processing method comprises:
Signal is carried out linear prediction analysis in time domain, the residual signals that obtains predicting;
Spectrum amplitude or spectrum amplitude variance data to described residual signals are carried out the evolution processing;
Write down described evolution and handle the parameter of using, and spectrum amplitude behind the evolution or spectrum amplitude variance data are unified processing, constant or vector quantization.
More than, carry out linear prediction analysis and the amount of calculation that causes is big and the discontinuous problem of Frame switching instant with respect to prior art at frequency domain, because the mode that employing is carried out linear prediction analysis in time domain is carried out signal filtering and other processing, signal is to pass through continuously, does not need to transform to frequency domain earlier and the calculating carried out; Simultaneously, carry out the envelope characteristic that linear prediction analysis is equivalent to extract frequency-region signal, and eliminate its influence, can be to the preliminary planarization of signal;
Again simultaneously, employing is carried out the evolution processing to the spectrum amplitude of residual signals, by frequency spectrum difference ratio evolution is handled the disparity range of dwindling spectral line, can at utmost realize planarization, owing to keep evolution information in quantizing process, the data after the planarization can be recovered out at an easy rate again;
Thereby do planarization with respect to prior art with the critical band average power simultaneously and will calculate a large amount of divisions, the all flattened spectral responses of present embodiment are handled and are all handled at log-domain, calculating of many complexity to simplify, solved the algorithm computation complexity issue, signal processing efficient improves greatly;
The processing method of evolution is beneficial to follow-up coding and processing very much, improves the decoding quality.
The present invention also provides a kind of signal processing apparatus basic embodiment, and described signal processing apparatus comprises: linear prediction analysis unit, signal is carried out linear prediction analysis in time domain, the residual signals that obtains predicting; The evolution processing unit carries out the evolution processing to the spectrum amplitude or the spectrum amplitude variance data of described residual signals; Quantifying unit writes down described evolution and handles the parameter of using, and spectrum amplitude behind the evolution or spectrum amplitude variance data are unified processing, constant or vector quantization.
Because adopt linear prediction analysis unit that signal is carried out linear prediction analysis in time domain, therefore signal is continuous when forecast analysis; Simultaneously, carry out the envelope characteristic that linear prediction analysis is equivalent to extract frequency-region signal, and eliminate its influence, can be to the preliminary planarization of signal; Because adopting the evolution processing unit that the bin magnitudes of described residual signals is carried out evolution handles, by frequency spectrum difference ratio evolution being handled the disparity range of dwindling spectral line, can at utmost realize planarization, owing to keep the evolution parameter in quantizing process, the data after the planarization can be recovered out at an easy rate again; Calculating to simplify of many complexity solved the algorithm computation complexity issue, and signal processing efficient improves greatly; Be beneficial to follow-up coding and processing, improve the decoding quality.
Below in conjunction with execution mode and accompanying drawing, describe the present invention.
Signal processing method embodiment of the present invention is made up of two parts: a part is linear prediction analysis, earlier signal is carried out lpc analysis in time domain, obtains predicted residual signal; Another part, just subsequent step carries out self adaptation to the spectrum amplitude of residual signals or spectrum amplitude variance data and opens the n power and handle, to the spectrum amplitude number behind the evolution or spectrum amplitude variance data according to unify to handle, the constant or the vector quantization of rule.Concrete steps are as follows:
To the audio signal sampling of input, sampling mode can be that per 256,512 or 1024 point of sample data are a frame, forms a plurality of Frames, handles frame by frame.
Here be every frame data to be obtained linear prediction analysis obtain transfer function, use this function that every frame data are carried out analysis filtered, obtain predicted residual signal.
Here, can carry out linear prediction respectively by low-frequency range, high band earlier and obtain frequency envelope figure.Consult Fig. 2, it is obviously poor near low-frequency range Fs/4 that spectrum envelope approaches effect as can be seen.
To the analysis filtered that gives a forecast respectively with low frequency, high-frequency signal after high pass, the low pass filter separation.With the low-frequency range is that example is introduced, and Fig. 3 has shown the effect of a frame of low-frequency range being carried out analysis filtered.
The low-frequency range frequency spectrum is carried out preliminary treatment: all peak points that at first count frequency spectrum, and find out wherein maximum and position thereof, peaked position is as the starting point of carrying out the log-domain linear compensation, starting point is a horizontal reference line with this maximum before, all peak points from the starting point to the low frequency cut-off points calculate its decay oblique line, all frequency spectrums are positioned under this oblique line, the decay oblique line is as far as possible near the spectrum peak point, horizontal reference line and decay oblique line are formed the reference line of low-frequency range frequency spectrum, and effect can be consulted Fig. 4.
Frequency spectrum grouping purpose is that the similar spectrum group of flatness is lumped together, and carries out planarization with identical or close parameter.The frequency spectrum group technology can be followed different principles and realize different dividing mode, provides two kinds of packet modes below: 1, the grouping of frequency spectrum division considers that the power power adjacent coefficient with 2 is one group, and from left to right carries out, and the size of each group is determined separately; 2, be to divide into groups in the center with local peaking's frequency spectrum, adjacent two groupings are divided with middle peak valley frequency spectrum.Determine two factors of the main consideration of group size, one is the width of the adjacent spectral coefficient of amplitude relatively flat, general tabular surface is wide more, and then to be distributed in the grouping of one's respective area just big more, and another factor is the restriction of code check, and grouping dimension is big more more at least for whole frame coding figure place.Fig. 5 is a grouping schematic diagram.
The optimization method of best packet further is provided here: (1) at first marks off all groupings by the minimum packets size; (2) calculate the associating mean square deviation of all adjacent packets; (3) find out minimum associating mean square deviation, if the size sum of these two groupings is less than the largest packet size, then merge this two groupings, otherwise search the size restrictions condition whether time little associating mean square deviation meets merging, transfer part if meet merging without any adjacent packets, then end of packet merge to be handled, and handles if there is grouping to merge, and then forwards for second step to restart the next round grouping and merge; (4) consider the code rate restriction, rule of thumb threshold value merges the less adjacent packets that is grouped into.
All will carry out flattened spectral response to each grouping and handle, 2 be that example is introduced the planarization process to divide into groups below: (1) finds out maximum spectrum value in the grouping, the difference of coding maximum spectrum value and reference line; (2) calculate log-domain (can adopt with 2 is that the logarithm at the end calculates) difference of all frequency spectrums and maximum spectrum place reference value; (3) if the log-domain difference of maximum spectrum amplitude and reference line less than
Then calculate and adjust to the deviant that this value needs, this deviant is added on the difference of the group intermediate frequency spectrum coefficient that calculates and maximum place reference line.
The log-domain difference table of adjusted grouping frequency spectrum is shown floating number form a (n) * (2^m (n)), a (n) is a mantissa, the scope of best value is 0.35355339～0.5, m (n) is a power exponent, span is 0～6, and 2^m (n) is the number of times of evolution, and m (n) is integer or decimal, a (n) is the evolution result after log-domain is transformed into actual value, and the symbol n with the back represents the serial number of processed frequency spectrum data at present frame herein.Make that Sqrt_factor (n)=2^m (n) is the number of times to the spectral magnitude evolution, all frequency spectrums in organizing are done evolution respectively to be handled, spectrum amplitude fluctuation range after the processing obviously diminishes, and is more approaching with the distance of reference line, and contrast effect can be referring to a following width of cloth figure.Can obviously make flattened spectral response in the group by this processing, make the data distribution be controlled at small range, be convenient to the processing of subsequent quantizatiion, coding link.
| Prediction residual | Prediction residual (db) | Mantissa | Index | Mantissa quantizes | The evolution number of times | The evolution result |
| 32.837418 | 5.037269 | 0.314829 | 4 | 0.312500 | 16 | 1.241858 |
| 45.882641 | 5.519877 | 0.344992 | 4 | 0.343750 | 16 | 1.2691 |
| 24.634417 | 4.622603 | 0.288913 | 4 | 0.281250 | 16 | 1.15869 |
| 45.028618 | 5.492770 | 0.343298 | 4 | 0.343750 | 16 | 1.2691 |
| 47.592545 | 5.572664 | 0.348291 | 4 | 0.343750 | 16 | 1.269051 |
Table one: the planarization result of partial frequency spectrum amplitude
For frequency spectrum more smooth, can select the evolution number of times adaptively during to the frequency spectrum processing in organizing, but evolution number of times parameter must keep and encode, limit in order to split degree number encoder rate, can be the scope (its corresponding spectrum amplitude differences is up to more than 50 times) that is limited in 0～4, precision can be loosened to 0.5 or littler precision.The change curve of the actual evolution of frequency spectrum time number of times is shown in following figure, and the data that the black broken line is represented are m (n).Under low code check situation, limit number of coded bits possibly to index, can consider that the evolution number of times in the grouping is limited in one or certain several values, so also can correspondingly enlarge the scope of mantissa.
Several different methods is arranged, and can be the vector quantization of standard, also can handle with the arithmetic coding mode.A kind of step of cataloged procedure is as follows: calculate the difference that m (n) does consecutive value (1), obtains difference sequence m ' (n), and m (0) does not do the adjacent data Difference Calculation, and do the Difference Calculation of front and back frame, difference sequence is done the vector quantization of some dimensions; (2) the 3rd to the 5th b4 (n) that forms 3 number of significant digit behind the decimal point of binary representation pressed in mantissa, the 6th to the 7th b2 (n) that forms 2 number of significant digit, b4 (n) and b2 (n) are done the difference the same with m (n) to be calculated, obtain b4 ' (n) and b2 ' (n), and then do the vector quantization of some dimensions (to be analyzed determine) equally, satisfy under the situation of signal to noise ratio precision at b4 (n), can not use b4 (n).
More than, carry out linear prediction analysis and the amount of calculation that causes is big and the discontinuous problem of Frame switching instant with respect to prior art at frequency domain, because the mode that employing is carried out linear prediction analysis in time domain in step 202 is carried out signal filtering and other processing, signal is to pass through continuously, does not need to transform to frequency domain earlier and the calculating carried out; Simultaneously, carry out the envelope characteristic that linear prediction analysis is equivalent to extract frequency-region signal, and eliminate its influence, can be to the preliminary planarization of signal;
Again simultaneously, in carry out step 204, adopt the evolution processing is represented-also promptly carried out to the bin magnitudes formation logarithm of residual signals, by frequency spectrum difference ratio evolution being handled the disparity range of dwindling spectral line, can at utmost realize planarization, owing to keep evolution information in quantizing process, the data after the planarization can be recovered out at an easy rate again;
Thereby do planarization with respect to prior art with the critical band average power simultaneously and will calculate a large amount of divisions, the all flattened spectral responses of present embodiment are handled and are all handled at logarithm, calculating of many complexity to simplify, solved the algorithm computation complexity issue, signal processing efficient improves greatly;
Owing in step 207, calculate the difference of frequency spectrum and maximum place reference line, and in step 208, be expressed as relocatable, be beneficial to very much follow-up coding and processing, improve the decoding quality.
The control method of quantization encoding bit number is provided in addition, here:
The bit number of control quantization encoding is the control requirement of adjusting code rate for adapting to, adapts to the performance change demand of different transmission channels.Will realize that in the present embodiment control quantizes number of coded bits, can start with from several links that one is to regulate frequency spectrum grouping size, another is control mantissa part quantified precision and coding figure place.
Frequency spectrum grouping size is as follows to the influence process of coding figure place: to when grouping is big, diversity ratio in the group between each frequency spectrum is bigger, with the maximum spectrum in the grouping is reference, then the less spectral line of amplitude will reach the same degree of planarization, then need bigger evolution number of times, under the situation of restriction code check, can consider the highest evolution number of times of restriction, actual needs evolution number of times is the unified threshold value that is restricted to during greater than a certain threshold value, the mantissa part that the exponential form of these spectral lines (log-domain is represented) is represented can not encoded yet, replace with a random value during decoding, so just can reduce overall coding figure place less than thresholding; If grouping is regulated to less direction, the figure place of then encoding can increase.
Mantissa part has been limited in a less fixed range, such as 0.35355339～0.5, in this interval, roughly become evenly to distribute, be controlled at the error precision that this interval interior quantification levels just can directly be controlled the decoding frequency spectrum, also can adjust the coding figure place simultaneously.If exponential part is limited in the limited several values chooses, then the distribution of mantissa part may be bigger, and the regularity of distribution also can change to normal distribution from even distribution, and so Dui Ying subsequent quantizatiion, encoding process also will be made corresponding change,
Consider to adjust the quantization digit of frequency spectrum grouping size and mantissa part simultaneously, just can realize flexible control code rate.Adjust whole frequency spectrum grouping size and make the mantissa part of some spectral lines not be encoded fully, can think coarse adjustment; The quantification levels and the quantization error of control mantissa part are comparatively careful control, can think fine tuning.
Handle the parameter of using for the evolution that quantification is write down,, need encode as evolution number of times or group mode.Coding to the frequency spectrum group mode: (1) frequency spectrum grouping size has 4,8,16,32 4 kind of situation, encodes or indicates with two; (2) several coding modes can combination in any, that is to say that a grouping can be adjacent with the grouping of any pattern; (3) original position of any corresponding frequency spectrum that divides into groups is lower than its grouping decision by all frequencies, that is to say that the size addition that begins all groupings from low frequency promptly is the frequency spectrum original position of current group.
Consulting Fig. 9, is the structure chart of signal processing apparatus embodiment of the present invention.Described signal processing apparatus comprises sampling unit 910, linear prediction analysis unit 920, evolution processing unit 930 and quantifying unit 940.
910 pairs of samples of signal of described sampling unit form a plurality of Frames, and are input to linear prediction analysis unit 920.920 pairs of signals of described linear prediction analysis unit carry out linear prediction analysis in time domain, the residual signals that obtains predicting.The spectrum amplitude of 930 pairs of described residual signals of described evolution processing unit or spectrum amplitude variance data are carried out the evolution processing, comprise and asking counting unit 931 and evolution unit 932.Described asking asked logarithm to the spectral magnitude or the spectrum amplitude variance data of 931 pairs of all residual signals of counting unit, and the logarithm that obtains the predicted residual signal spectrum envelope is represented.Described evolution unit 932 is used to obtain the reference curve of all frequency amplitudes, and calculate difference between all frequencies and the corresponding frequency amplitude maximum place reference curve value, carry out evolution and handle, perhaps calculate the peaked difference of all frequency spectrums and spectral magnitude and carry out the evolution processing.Comprise grouped element 9321 in the described evolution unit 932.
Described grouped element 9321 is used to carry out frequency spectrum self adaptation grouping to be divided and handles, and is one group with a power power adjacent coefficient of 2, and from left to right carries out that the size of each group is definite separately that is:; Or be to divide into groups in the center with local peaking's frequency spectrum, adjacent two groupings are divided with middle peak valley frequency spectrum.
The described evolution of described quantifying unit 940 records is handled the parameter of using, and the amplitude data behind the evolution is unified processing, constant or vector quantization.Specifically be that described log-domain difference table is shown floating number form a (n) * (2^m (n)), wherein a (n) is a mantissa, and m (n) is a power exponent.Wherein, comprise difference computational unit 941 and mantissa's processing unit 942.Difference computational unit 941 is used for m (n) is done the difference calculating of consecutive value, obtains difference sequence m ' (n), the Difference Calculation of frame before and after wherein m (0) only does; Difference sequence m ' (n) is done the vector quantization of some dimensions; B4 (n) and b2 (n) are done the difference the same with m (n) calculate, obtain b4 ' (n) and b2 ' (n), and then do the vector quantization of some dimensions equally, satisfy under the situation of signal to noise ratio precision at b4 (n), do not use b4 (n).Described mantissa processing unit 942 be used for to mantissa by the decimal point of binary representation after the 3rd to the 5th b4 (n) that forms 3 number of significant digit, the 6th to the 7th b2 (n) that forms 2 number of significant digit.
With the audio signal is example, and during running, sampling unit 910 per 256,512 or 1024 sampled datas are a frame, handle frame by frame.Linear prediction unit is then carried out the time domain linear forecast analysis to the Frame of input, obtains linear prediction transfer function H (Z), with H (Z) every frame data is carried out analysis filtered, obtains predicted residual signal e (n).930 couples of predicted residual signal e of evolution processing unit (n) do the FFT conversion, calculate the amplitude A (n) of all frequencies.With asking the amplitude A (n) to 931 pairs of all frequencies of counting unit to ask logarithm, the logarithm that obtains the predicted residual signal spectrum envelope is represented Alg (n) then; The reference curve of entire spectrum is obtained in evolution unit 932, and 9321 pairs of frequency spectrum groupings of grouped element are adopted in the back.Every group of data after the grouping are input to quantifying unit 940.Maximum spectrum value MaxAlg in the grouping is found out in evolution unit 932, calculates the difference DiffAlg between all frequency spectrums and the reference line value RefAlg of maximum place or the maximum spectrum value MaxAlg, according to the distance of maximum and reference line difference is adjusted.Difference computational unit 941 in the quantifying unit 940 is done the difference of consecutive value and is calculated to m (n), obtain difference sequence m ' (n), the Difference Calculation of frame before and after wherein m (0) only does; Difference sequence m ' (n) is done the vector quantization of some dimensions; B4 (n) and b2 (n) are done the difference the same with m (n) calculate, obtain b4 ' (n) and b2 ' (n), and then do the vector quantization of some dimensions equally, satisfy under the situation of signal to noise ratio precision at b4 (n), do not use b4 (n); Adopt mantissa's processing unit 942 that DiffAlg (n) is expressed as relocatable DiffAlg (n)=a (n) * (2^m (n)) at last, a (n) is a mantissa, and the scope of best value is 0.35355339～0.5, and m (n) is a power exponent, and span is 0～6.
More than, owing to adopt 920 pairs of signals of linear prediction analysis unit to carry out linear prediction analysis in time domain, therefore signal is continuous when forecast analysis; Simultaneously, carry out the envelope characteristic that linear prediction analysis is equivalent to extract frequency-region signal, and eliminate its influence, can be to the preliminary planarization of signal; Because adopting the bin magnitudes of 930 pairs of described residual signals of evolution processing unit to carry out evolution handles, by frequency spectrum difference ratio evolution being handled the disparity range of dwindling spectral line, can at utmost realize planarization, owing to keep evolution information in quantizing process, the data after the planarization can be recovered out at an easy rate again; Calculating to simplify of many complexity solved the algorithm computation complexity issue, and signal processing efficient improves greatly; Be beneficial to follow-up coding and processing, improve the decoding quality.
More than a kind of signal processing method provided by the present invention and device thereof are described in detail, used specific case herein principle of the present invention and execution mode are set forth, the explanation of above embodiment just is used for helping to understand method of the present invention and core concept thereof; Simultaneously, for one of ordinary skill in the art, according to thought of the present invention, the part that all can change in specific embodiments and applications, in sum, this description should not be construed as limitation of the present invention.
Claims (20)
1. a signal processing method is characterized in that, comprising:
Signal is carried out linear prediction analysis in time domain, the residual signals that obtains predicting;
Spectrum amplitude or spectrum amplitude variance data to described residual signals are carried out the evolution processing;
Write down described evolution and handle the parameter of using, and spectrum amplitude behind the evolution or spectrum amplitude variance data are carried out reprocessing, constant or vector quantization.
2. signal processing method according to claim 1 is characterized in that, described signal is carried out comprising before the linear prediction analysis in time domain: to sample of signal, form a plurality of Frames.
3. signal processing method according to claim 2, it is characterized in that, describedly carry out linear prediction analysis and obtain predicted residual signal and be meant: every frame data are carried out linear prediction analysis obtain transfer function, use this function that signal is carried out time domain analysis filtering, obtain predicted residual signal.
4. signal processing method according to claim 1 is characterized in that, the spectrum amplitude of described residual signals adopts fast fourier transform to obtain.
5. signal processing method according to claim 1 is characterized in that, described spectrum amplitude or spectrum amplitude variance data to residual signals carried out evolution and handled and comprise:
Spectral magnitude or spectrum amplitude variance data to residual signals are asked logarithm, and the logarithm that obtains the predicted residual signal spectrum envelope is represented;
Obtain the reference curve of spectral magnitude;
Calculate the difference between all frequency spectrums and the corresponding spectral magnitude maximum place reference curve value and carry out evolution and handle, perhaps calculate the peaked difference of all frequency spectrums and spectral magnitude and carry out the evolution processing.
6. signal processing method according to claim 5, it is characterized in that, also further comprise: the log-domain difference of maximum spectrum amplitude and reference curve less than the set point situation under, calculate and adjust to the deviant that this set point needs, this deviant is added on the difference of the frequency amplitude that calculates and described corresponding maximum place reference line.
7. signal processing method according to claim 6 is characterized in that, described log-domain is to be to take the logarithm at the end with 2, and the optimum value of described set point is
8. signal processing method according to claim 5, it is characterized in that, also further comprise: described log-domain difference table is shown floating number form a (n) * (2^m (n)), wherein a (n) is a mantissa, m (n) is a power exponent, and 2^m (n) is the number of times of evolution, and m (n) is integer or decimal, a (n) is the evolution result after log-domain is transformed into actual value, and n represents the serial number of processed frequency spectrum data at present frame.
9. signal processing method according to claim 8 is characterized in that, described amplitude data behind the evolution or spectrum amplitude variance data is quantized to comprise:
M (n) is done the difference of consecutive value and calculate, obtain difference sequence m ' (n), m (0) does not do the adjacent data Difference Calculation, and the Difference Calculation of frame before and after doing; Difference sequence is done the vector quantization of some dimensions;
Mantissa is pressed the 3rd to the 5th b4 (n) that forms 3 number of significant digit behind the decimal point of binary representation, the 6th to the 7th b2 (n) that forms 2 number of significant digit, b4 (n) and b2 (n) are done the difference the same with m (n) to be calculated, obtain b4 ' (n) and b2 ' (n), and then do the vector quantization of some dimensions equally, satisfy under the situation of signal to noise ratio precision at b4 (n), do not use b4 (n).
10. signal processing method according to claim 8 is characterized in that, the scope of the best value of described mantissa is 0.35355339～0.5, and described power exponent optimum valuing range is 0～6.
11. signal processing method according to claim 5 is characterized in that, described spectral magnitude or spectrum amplitude variance data to all residual signals asks logarithm to comprise before: to the linear prediction of carrying out of residual signals, and carry out forecast analysis filtering.
12. signal processing method according to claim 5 is characterized in that, the described reference curve of obtaining all spectral magnitudes comprises:
Count all peak points of frequency spectrum, and find out wherein maximum and position thereof, with peaked position as the starting point of carrying out the log-domain linear compensation, starting point is a horizontal reference line with this maximum before, all peak points from the starting point to the low frequency cut-off points calculate its decay oblique line, all frequency spectrums are positioned under this oblique line, and the decay oblique line is as far as possible near the spectrum peak point, and horizontal reference line and decay oblique line are formed the reference curve of described frequency spectrum.
13. signal processing method according to claim 5, it is characterized in that, before the difference or each frequency spectrum and the peaked difference of frequency amplitude that calculate between each frequency spectrum and the spectral magnitude maximum place reference curve value, comprising: carry out the grouping of frequency spectrum self adaptation and divide processing, that is:
Power power adjacent coefficient with 2 is one group, and from left to right carries out, and the size of each group is determined separately; Or be to divide into groups in the center with local peaking's frequency spectrum, adjacent two groupings are divided with middle peak valley frequency spectrum; And, described group size is considered two factors, and one is the width of the adjacent spectral coefficient of amplitude relatively flat, and general tabular surface is wide more, and then to be distributed in the grouping of one's respective area just big more, another factor is the restriction of code check, and grouping dimension is big more more at least for whole frame coding figure place.
14. signal processing method according to claim 13 is characterized in that, described self adaptation grouping is divided to handle and is comprised:
Mark off all groupings by the minimum packets size;
Calculate the associating mean square deviation of all adjacent packets;
Find out minimum associating mean square deviation, if the size sum of these two groupings is less than the largest packet size, then merge this two groupings, otherwise search the size restrictions condition whether time little associating mean square deviation meets merging, transfer part if meet merging without any adjacent packets, then end of packet merge to be handled, and handles if there is grouping to merge, and then forwards for second step to restart the next round grouping and merge;
Merge the less adjacent packets that is grouped into according to code rate restriction and experience threshold value;
After dividing into groups, calculate difference or interior all frequency spectrums of grouping and the interior peaked difference of frequency amplitude of this grouping between the interior all frequency spectrums of grouping and this grouping spectral magnitude maximum place reference curve value.
15. signal processing method according to claim 1 is characterized in that, described evolution is handled the number of times that the parameter of using is meant evolution.
16. a signal processing apparatus is characterized in that, comprising:
Linear prediction analysis unit is carried out linear prediction analysis to signal in time domain, the residual signals that obtains predicting;
The evolution processing unit is used for the spectrum amplitude or the spectrum amplitude variance data of described residual signals are carried out the evolution processing;
Quantifying unit is used to write down described evolution and handles the parameter of using, and amplitude behind the evolution or spectrum amplitude variance data are unified processing, constant or vector quantization.
17. signal processing apparatus according to claim 16 is characterized in that, further comprises sampling unit, and sample of signal is formed a plurality of Frames, and is input to linear prediction analysis unit.
18. signal processing apparatus according to claim 16 is characterized in that, described evolution processing unit comprises:
Ask counting unit, be used for the frequency amplitude of all residual signals is asked logarithm, the logarithm that obtains the predicted residual signal spectrum envelope is represented;
The evolution unit is used to obtain the reference curve of all spectral magnitudes, and calculates all frequency spectrums and corresponding difference of dividing into groups between the intermediate frequency spectrum amplitude maximum place reference curve value, carries out evolution and handles, and perhaps the difference value between spectral magnitude is carried out evolution and handles.
19. signal processing apparatus according to claim 18 is characterized in that, described evolution unit comprises grouped element, be used to carry out the grouping of frequency spectrum self adaptation and divide processing, that is: be one group with a power power adjacent coefficient of 2, and from left to right carry out that the size of each group is determined separately; Or be to divide into groups in the center with local peaking's frequency spectrum, adjacent two groupings are divided with middle peak valley frequency spectrum.
20. signal processing apparatus according to claim 16, it is characterized in that, described quantification treatment unit is further used for described log-domain difference table is shown floating number form a (n) * (2^m (n)), wherein a (n) is a mantissa, and m (n) is a power exponent, and 2^m (n) is the number of times of evolution, m (n) is integer or decimal, a (n) is the evolution result after log-domain is transformed into actual value, and n represents the serial number of processed frequency spectrum data at present frame, and comprises:
Difference computational unit is used for m (n) is done the difference calculating of consecutive value, obtains difference sequence m ' (n), the Difference Calculation of frame before and after wherein m (0) only does; Difference sequence m ' (n) is done the vector quantization of some dimensions; B4 (n) and b2 (n) are done the difference the same with m (n) calculate, obtain b4 ' (n) and b2 ' (n), and then do the vector quantization of some dimensions equally, satisfy under the situation of signal to noise ratio precision at b4 (n), do not use b4 (n);
Mantissa's processing unit, be used for to mantissa by the decimal point of binary representation after the 3rd to the 5th b4 (n) that forms 3 number of significant digit, the 6th to the 7th b2 (n) that forms 2 number of significant digit.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN2006101403202A CN101193090B (en) | 2006-11-27 | 2006-11-27 | Signal processing method and its device |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN2006101403202A CN101193090B (en) | 2006-11-27 | 2006-11-27 | Signal processing method and its device |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN101193090A true CN101193090A (en) | 2008-06-04 |
| CN101193090B CN101193090B (en) | 2011-12-28 |
Family
ID=39487843
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN2006101403202A Expired - Fee Related CN101193090B (en) | 2006-11-27 | 2006-11-27 | Signal processing method and its device |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN101193090B (en) |
Cited By (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN102396232A (en) * | 2009-04-24 | 2012-03-28 | 索尼公司 | Image-processing device and method |
| WO2015154397A1 (en) * | 2014-04-08 | 2015-10-15 | 华为技术有限公司 | Noise signal processing and generation method, encoder/decoder and encoding/decoding system |
| CN106034274A (en) * | 2015-03-13 | 2016-10-19 | 深圳市艾思脉电子股份有限公司 | 3D sound device based on sound field wave synthesis and synthetic method |
| CN106847296A (en) * | 2012-07-12 | 2017-06-13 | 诺基亚技术有限公司 | Vector quantization |
| CN110992739A (en) * | 2019-12-26 | 2020-04-10 | 上海乂学教育科技有限公司 | Student on-line dictation system |
| CN114025085A (en) * | 2021-10-29 | 2022-02-08 | 浙江大华技术股份有限公司 | Method, device and equipment for switching working modes of camera equipment and storage medium |
Families Citing this family (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN105988972B (en) * | 2015-02-03 | 2018-12-07 | 上海澜至半导体有限公司 | Realize Fast Fourier Transform (FFT)/inverse fast Fourier transform method and circuit |
Family Cites Families (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US4122432A (en) * | 1971-03-30 | 1978-10-24 | Fried. Krupp Gesellschaft Mit Beschrankter Haftung Atlas Elektronik Bremen | Device for identifying noise sources |
| CN1204721C (en) * | 2002-05-23 | 2005-06-01 | 华为技术有限公司 | Fast time delay optimization method of time domain equalizing training in discrete multisound system |
| CN1162838C (en) * | 2002-07-12 | 2004-08-18 | 清华大学 | Speech intensifying-characteristic weighing-logrithmic spectrum addition method for anti-noise speech recognization |
-
2006
- 2006-11-27 CN CN2006101403202A patent/CN101193090B/en not_active Expired - Fee Related
Cited By (11)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN102396232A (en) * | 2009-04-24 | 2012-03-28 | 索尼公司 | Image-processing device and method |
| CN106847296A (en) * | 2012-07-12 | 2017-06-13 | 诺基亚技术有限公司 | Vector quantization |
| CN106847296B (en) * | 2012-07-12 | 2021-01-22 | 诺基亚技术有限公司 | Vector quantization |
| WO2015154397A1 (en) * | 2014-04-08 | 2015-10-15 | 华为技术有限公司 | Noise signal processing and generation method, encoder/decoder and encoding/decoding system |
| US9728195B2 (en) | 2014-04-08 | 2017-08-08 | Huawei Technologies Co., Ltd. | Noise signal processing method, noise signal generation method, encoder, decoder, and encoding and decoding system |
| US10134406B2 (en) | 2014-04-08 | 2018-11-20 | Huawei Technologies Co., Ltd. | Noise signal processing method, noise signal generation method, encoder, decoder, and encoding and decoding system |
| US10734003B2 (en) | 2014-04-08 | 2020-08-04 | Huawei Technologies Co., Ltd. | Noise signal processing method, noise signal generation method, encoder, decoder, and encoding and decoding system |
| CN106034274A (en) * | 2015-03-13 | 2016-10-19 | 深圳市艾思脉电子股份有限公司 | 3D sound device based on sound field wave synthesis and synthetic method |
| CN110992739A (en) * | 2019-12-26 | 2020-04-10 | 上海乂学教育科技有限公司 | Student on-line dictation system |
| CN110992739B (en) * | 2019-12-26 | 2021-06-01 | 上海松鼠课堂人工智能科技有限公司 | Student on-line dictation system |
| CN114025085A (en) * | 2021-10-29 | 2022-02-08 | 浙江大华技术股份有限公司 | Method, device and equipment for switching working modes of camera equipment and storage medium |
Also Published As
| Publication number | Publication date |
|---|---|
| CN101193090B (en) | 2011-12-28 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN101193090B (en) | Signal processing method and its device | |
| JP7158452B2 (en) | Method and apparatus for generating a mixed spatial/coefficient domain representation of an HOA signal from a coefficient domain representation of the HOA signal | |
| CN101622662B (en) | Encoding device and encoding method | |
| EP2267698B1 (en) | Entropy coding by adapting coding between level and run-length/level modes. | |
| CN1866355B (en) | Audio coding apparatus and method, and audio decoding apparatus and method | |
| US9009036B2 (en) | Methods and systems for bit allocation and partitioning in gain-shape vector quantization for audio coding | |
| US8447597B2 (en) | Audio encoding device, audio decoding device, audio encoding method, and audio decoding method | |
| CN103733257B (en) | Audio-encoding method and apparatus, audio-decoding method and apparatus, and multimedia device employing same | |
| KR101995694B1 (en) | Device and method for execution of huffman coding | |
| CN1787383B (en) | Methods and apparatuses for transforming, adaptively encoding, inversely transforming and adaptively decoding an audio signal | |
| CN100459436C (en) | Bit distributing method in audio-frequency coding | |
| CN105144288A (en) | Advanced quantizer | |
| CN101800050B (en) | Audio fine scalable coding method and system based on perception self-adaption bit allocation | |
| EP0868031A1 (en) | Signal coding method and apparatus | |
| CN105957533A (en) | Speech compression method, speech decompression method, audio encoder, and audio decoder | |
| US20040230425A1 (en) | Rate control for coding audio frames | |
| CN101308657B (en) | Code stream synthesizing method based on advanced audio coder | |
| CN105122358A (en) | Apparatus and method for processing an encoded signal and encoder and method for generating an encoded signal | |
| CN101377926B (en) | Audio encoding method capable of quickening quantification circulation program | |
| KR100640833B1 (en) | Method for encording digital audio | |
| KR102658702B1 (en) | Method and apparatus for generating from a coefficient domain representation of hoa signals a mixed spatial/coefficient domain representation of said hoa signals | |
| Wang et al. | A new bit-allocation algorithm for AAC encoder based on linear prediction |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| C06 | Publication | ||
| PB01 | Publication | ||
| C10 | Entry into substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| C14 | Grant of patent or utility model | ||
| GR01 | Patent grant | ||
| CF01 | Termination of patent right due to non-payment of annual fee |
Granted publication date: 20111228 Termination date: 20161127 |
|
| CF01 | Termination of patent right due to non-payment of annual fee |