WO1994003988A2 - Systeme de traitement de signaux numeriques de superposition - Google Patents

Systeme de traitement de signaux numeriques de superposition Download PDF

Info

Publication number
WO1994003988A2
WO1994003988A2 PCT/GB1993/001644 GB9301644W WO9403988A2 WO 1994003988 A2 WO1994003988 A2 WO 1994003988A2 GB 9301644 W GB9301644 W GB 9301644W WO 9403988 A2 WO9403988 A2 WO 9403988A2
Authority
WO
WIPO (PCT)
Prior art keywords
signal
noise
dither
lower resolution
dither noise
Prior art date
Application number
PCT/GB1993/001644
Other languages
English (en)
Other versions
WO1994003988A3 (fr
Inventor
Michael Anthony Gerzon
Peter Graham Craven
Original Assignee
Michael Anthony Gerzon
Peter Graham Craven
Priority date (The priority date 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 date listed.)
Filing date
Publication date
Application filed by Michael Anthony Gerzon, Peter Graham Craven filed Critical Michael Anthony Gerzon
Publication of WO1994003988A2 publication Critical patent/WO1994003988A2/fr
Publication of WO1994003988A3 publication Critical patent/WO1994003988A3/fr

Links

Classifications

    • H—ELECTRICITY
    • H04—ELECTRIC COMMUNICATION TECHNIQUE
    • H04B—TRANSMISSION
    • H04B14/00—Transmission systems not characterised by the medium used for transmission
    • H04B14/02—Transmission systems not characterised by the medium used for transmission characterised by the use of pulse modulation
    • H04B14/04—Transmission systems not characterised by the medium used for transmission characterised by the use of pulse modulation using pulse code modulation
    • H04B14/046—Systems or methods for reducing noise or bandwidth
    • G—PHYSICS
    • G11—INFORMATION STORAGE
    • G11B—INFORMATION STORAGE BASED ON RELATIVE MOVEMENT BETWEEN RECORD CARRIER AND TRANSDUCER
    • G11B20/00—Signal processing not specific to the method of recording or reproducing; Circuits therefor
    • G11B20/10—Digital recording or reproducing
    • G11B20/10527—Audio or video recording; Data buffering arrangements

Definitions

  • the present invention relates to digital waveform encoding and decoding systems, and in particular to systems in which dither is used to counter the effect of low-level non-linearities on signals.
  • Such dither noise when of a "Gaussian" white noise form, degraded the signal-to-noise ratio of the 16 bit digital channel to not better than 92 dB.
  • the peak signal level used to calculate signal-to-noise ratio is a sharp limit, rather than the soft overload of analogue media, so that in practice, the signal-to-noise ratio of CD is often comparable to that of an analogue medium with around a 75 dB S/N (signal to noise ratio) .
  • the 16 bit medium has proved to inadequate.
  • the present inventors pointed out that the inherent performance of a 16 bit channel could be improved substantially by a combination of two strategies: 1) the use of noise shaping around the quantization/dither error process, so as to shape the spectrum of the noise error to make it subjectively less audible, and 2) the use of subtractive dither, which allowed an improvement of 6 dB (relative to Gaussian nonsubtractive dither) or 4.8 dB (relative to triangular nonsubtractive dither) for suitably designed reproducers, while retaining compatibility with existing nonsubtractive reproducers.
  • a synchronised subtractive dither system uses a dither that is dependent on the relative timing of a signal sample with the synchronizing clock, any time-shift of the signal would require a requantization of the signal, subtracting the "old" dither signal, and requantizing with the "new" dither signal, resulting in a 3 dB loss of S/N. Only two such retimings within a signal processing chain would cause a loss of all the S/N advantages of subtractive dither over nonsubtractive dither. The alternative was to find a way of "retiming" the synchronizing clock quickly whenever necessary, which further complicated the design of a synchronization standard.
  • a digital signal processing system comprising means for processing a signal of a first resolution to produce a signal of a second, different, resolution, the said means including a generator for dither noise,the system using the dither noise in processing the signal to ameliorate the effect of system non-linearities, characterised in that the generator for dither noise includes means for generating in functional dependence upon samples of the lower resolution signal a pseudo-random dither noise output for use in processing a current sample of said lower resolution signal.
  • the present invention uses a new approach to subtractive dither that requires no synchronizing clock signal, but which uses the signal samples themselves to compute the dither signal.
  • This has a first important advantage that no extra data need be transmitted or recovered to synchronize the subtractive dither process, which means that subtractive dither can be used with existing media and chips.
  • a second advantage is that retiming of the subtractive dither can be achieved in as few as 16 samples across an edit or signal retiming, thereby allowing such processing as time delays or "slipping" between simultaneous or consecutive tracks to be achieved without any loss of S/N.
  • the process of generating and regenerating dither from the dithered signal samples themselves means that other processes which conventionally do not cause loss of S/N in the digital domain, such as polarity inversion and/or the swapping of stereo channels, still do not cause a loss of S/N.
  • the dither generator may produce a dither output dependent on, eg, the value or polarity of the presently processed sample as well as previous samples of the lower resolution signal. It is much preferred however that for the current sample the output of the generator depends on previous samples only.
  • subtractive dither synchronized by a clock signal requires requantization and redithering whenever any of these signal manipulations occur, resulting in a 3 dB loss of S/N in the first "generation".
  • autodither generates a dither signal for each sample that is dependent only on a finite number (typically 16 or 24) of the previously quantized/dithered samples.
  • the autodither process itself requires standardization, and the standard must include a specification of the noise-shaping being applied.
  • autodithered signals may be sent via existing linear digital channels in a manner fully compatible with existing uses, although only subtractively dithered playback will recover the full available dynamic range, with nonsubtractively dithered playback being about 4.8 dB worse.
  • the autodither process is thus operationally robust, and does not cause unnecessary losses of S/N in ordinary studio operation.
  • the theoretically achievable subjective S/N of about 116 dB is far better than existing A/D converters, and so leaves adequate margins for loss of S/N due to subsequent signal manipulations.
  • 116 dB S/N allows fade- outs to near silence in the digital domain, and provides a reasonable safety margin for conservative recording levels to avoid peak overload in live recording of music with a wide dynamic range. It also allows compact disc recorders to be used for live recording applications, while obtaining the same dynamic range as currently achieved using 20 bit recording media.
  • a lowered noise floor also means that there is less need to push up the levels recorded on CDs so that peaks run to near the peak 0 dB level. Rather, it is possible to record inherently quiet music (e.g. clavichords or lutes) at their natural levels even if this means peaks at -30 dB or so.
  • the digital signal processing system may be implemented using some stages operating in the analogue domain. For example, as further described below, noise subtraction in a decoder might be carried out in the analogue domain.
  • the digital signal processing system further comprises filter means for applying noise-shaping to the dither noise to reduce the perceptibility of the dither noise in the reproduced signal.
  • the combination of autodither with noise-shaping makes it possible to increase the effective dynamic range of, e.g., 16 bit audio media to the equivalent of around 19.5- 20 bits. With a 20 bit system, such as those used for studio recordings, it makes possible a 24 bit performance.
  • the digital signal processing system may be a digital encoder for encoding a digital signal of a higher resolution to produce a digital signal of a lower resolution, in which case the encoder comprises a quantizer for rounding or truncating the higher resolution signal to produce the lower resolution signal, the output of the quantizer being connected to an input of the generator for dither noise, and an addition means for adding the dither noise output by the generator to the higher resolution signal input to the quantizer.
  • the digital signal processing system may be a digital decoder for decoding a lower resolution digital signal to reproduce a digital signal of higher resolution, the decoder including subtraction means for subtracting dither noise from the lower resolution signal, the dither noise generator receiving the lower resolution digital signal at its input and outputting dither noise to the subtraction means.
  • a transmission system for transmitting a higher resolution signal via a lower resolution digital channel comprises in combination an autodither encoder and an autodither decoder as defined above.
  • the lower resolution digital channel will typically be provided by a recording medium such as CD, DAT or other digital tape media, or may be a digital storage or memory medium such as a RAM or ROM or a hard disc, floppy disc or optical disc.
  • a recording medium such as CD, DAT or other digital tape media
  • a digital storage or memory medium such as a RAM or ROM or a hard disc, floppy disc or optical disc.
  • Figure 1 shows a schematic of nonsubtractive dither.
  • Figure 2 shows a schematic of subtractive dither.
  • Figure 3 shows the use of a transmitted clock signal to synchronize subtractive dither.
  • Figure 4 shows noise shaping around a dithered quantizer.
  • Figure 5 shows an alternative form of noise shaped dithered quantizer equivalent to figure 4.
  • Figure 6 shows the subtractive reconstruction of signals for noise-shaped dither.
  • Figure 7 shows the autodither encode process.
  • Figure 8 shows the autodither decode process.
  • Figure 9 shows autodither encoding with noise shaping
  • Figure 10 shows an alternative autodither encoding with noise shaping equivalent to figure 9.
  • Figure 11 shows autodither decoding with noise shaping.
  • Figure 12 shows the cross-fade splicing of autodithered signals.
  • Figure 13 shows the generation of pseudo-random dither noise with a triangular pdf by means of adding two uniform pdf's.
  • Figure 14 shows the use of a first noise shaping characteristic on dither and a second different characteristic on quantizer error.
  • Figure 15 shows a schematic of noise-shaped dithered encoding with adaptive noise shaping around the quantizer.
  • Figure 16 shows a scheme for ensuring that autodither decoding works correctly when signal polarity is inverted.
  • Figure 7 shows a first example of an encoder embodying the present invention.
  • a high resolution digital signal is input on line 1 via an addition node 2 to a quantizer 3.
  • a low resolution signal is output on the line 4.
  • the low resolution signal is fed to a buffer memory 5.
  • M successive samples in the buffer memory 5 are output in parallel to address a look-up table 6.
  • the look-up table 6 then outputs a value which is added to the incoming high resolution signal at the node 2 as dither noise.
  • Figure 8 shows a complementary autodither decoder structure.
  • An incoming low resolution digital signal on line 7 is input to a level reconstruction circuit 8, of the type well known in the art.
  • the samples from the line 7 are used to address a look-up table 10 via a buffer memory 9 to generate autodither noise in the same manner as described for the encoder.
  • the contents of the look-up table 9 are matched to those of the look-up table 6 in the encoder.
  • the output from the look-up table 9 is combined with the output of the level reconstruction circuit 8 at a subtractive node 11 so that the dither noise is subtracted from the output signal.
  • the invention may conveniently be implemented using digital signal processing algorithms on general-purpose digital signal processing (DSP) chips, such as those of the Motorola DSP 56000 family or those of the Texas Instruments TMS320 family, in which a digital signal representing a waveform at one resolution is input to the DSP chip and the required digital signal output is output from the chip.
  • DSP digital signal processing
  • the individual signal-processing blocks described in this description will, in such convenient implementations, be realized as signal processing sub-algorithms of the digital signal processing algorithm programmed on the DSP chip.
  • the signal processing block referred to as a quantizer may consist of an algorithm rounding or truncating a digital signal word to a digital word having fewer bits at a lower resolution.
  • the memory on board the DSP chip may be used, but where larger amounts of memory are required, as for example in look-up tables, an external RAM or ROM chip or chips may be provided addressed by the DSP chip.
  • external RAM may be provided by a high speed static RAM such as the Integrated Device Technology IDT 71256 when 32 K of 8 bit memory is required.
  • static RAM such as the Integrated Device Technology IDT 71256 when 32 K of 8 bit memory is required.
  • Dither is the process of adding a random (or pseudo-random) noise signal before a quantizer, as shown in Figure 1, so as to eliminate the effect of the "staircase" transfer-function non-linearity of the quantizer. Dither achieves this by replacing a single input signal level which is quantized by a range of such signal levels determined statistically by the added noise signal, so that the transfer-function non-linearity is averaged over the possible range of input signal levels created by adding the noise.
  • Figure 5 shows an alternative form of noise-shaping of a dithered quantizer in which the noise-shaping of the quantizer and the dither is separated. It is easily shown [1] that Figures 4 and 5 have exactly equivalent performance, and are equivalent to one another.
  • Figure 5 reveals that it is possible, as shown schematically in Figure 14 to use a different filter, say H'(z "1 ), around the quantizer than for shaping the dither noise spectrum, making the quantization noise error have a different spectrum from that of the added noise.
  • the final noise error spectrum from the subtractive dither process with the dither noise- shaped by H(z " ) is thus shaped by H' (z -1) .
  • Thi.s means that i.f a subj.ectl.vely better noise-shaping filter H'(z * ) is determined after the decoding noise-shaping filter has already been standardized as H(z -1) , the subtractively dithered decoded results wi.ll incorporate the improved noise-shaping.
  • This noise- shaping will give about 18 dB improvement in S/N subjectively as compared to no use of noise-shaping.
  • the use of a ninth order noise-shaping filter H(z ' ) as described in references [3] or [2] is preferred.
  • the noise heard by the user of a subtractively dithered decoder will always be statistically uncorrelated with the input signal if the dither noise (before noise-shaping filtering) has independent samples with a rectangular probability distribution function (pdf) with a peak-to-peak level of a least significant bit (lsb) step size, it can be shown [1,4] that this does not eliminate audible modulation noise effects for nonsubtractive listeners. It is therefore strongly preferred that the dither noise used, before noise-shaping, should have a triangular pdf with peak-to- peak amplitude of 2 lsb's, as described in refs.
  • any other pdf comprising this triangular probability distribution convoluted with any other pdf will also avoid variations in the mean square noise level heard by nonsubtractive listeners.
  • Such a pdf may be acheived by adding to a triangular pdf dither a second statistically independent dither signal with an arbitrary pdf.
  • a quadratic B-spline pdf with peak-to-peak amplitude of 3 lsb's may be used.
  • Such a quadratic B- spline pdf dither may be formed by adding three statistically independent rectangular pdf dithers with peak levels ⁇ % lsb.
  • the form of the noise and its generation has to be standardized in a subtractively dithered system in order that the dither signal reconstructed at the decoding stage should exactly match that used at the encoding state.
  • a dither noise with Gaussian statistics at a level of about 4.8 dB or more above the rms (root mean square) level of rectangular pdf noise with 1 lsb peak-to-peak level will also give satisfactory results both for subtractive and nonsubtractive decoding.
  • our preferred option of triangular pdf dither described above will minimize the noise energy heard with nonsubtractive decoding.
  • the autodither process of the presently described embodiments instead generates dither from the least significant bits of the quantized digital data stream in the last M samples, where M is a number that is typically chosen equal to 14 or 16 or 24.
  • M is a number that is typically chosen equal to 14 or 16 or 24.
  • the encode process of turning a high resolution digital signal (which may have a much larger number of bits than the final number of say 16) into a quantized data stream is shown schematically in Figure 7.
  • the look-up table converts these into one of 2 values of dither noise.
  • the mapping from the 2 states of the state variables n the buffer to the noise signal should appear as random or "chaotic" as possible, and may be chosen by random or Monte-Carlo methods, or by any known method of
  • the generated dither samples may have any probability distribution function that is suitable for subtractive dither, but for our purposes a triangular probability distribution function with peak levels ⁇ 1 lsb is preferred, since such dither also performs optimally without subtractive decoding as shown by Lipshitz et al
  • a look-up table with 2 entries is expensive to implement, and for triangular dither may alternatively be o implemented as two look-up tables each having 2 entries, producing rectangular probability distribution function
  • 2 states to a product of two state variables with 2 values may be any deterministic mapping, which may be implemented by means of any one-to-one function followed by a separation of the 16 binary digits into two subsets (e.g. even and odd samples after the mapping) .
  • the mapping from the 2 states to the rectangular dither should generate each of 2 values of rectangular dither quantized to 8 bits once only. Also, in order that the pseudo-random generation process be unchanged by signal polarity inversion, the mapping should be such that if a given set of 8 digits goes to a dither value d, then the digits of opposite polarity go to -d; this assumes as is preferred that the mapping from 2 to
  • 2 8 x 28 state variable is preserved under polarity inversions, as is the case when this mapping is done by selecting 2 subsets of the 16 digits in the buffer. (We assume here that polarity inversion negates all digits - i.e. that analog zero is halfway between quantization levels) .
  • the subtractive dither decoding process used to recover a higher-resolution digital signal with more than (say) 16 bits from the encoded signal is shown schematically in Figure 8.
  • This uses the same buffer memory and look-up table arrangement as in the encoder, but now operating on the last M samples of the input digital data stream, and the pseudo-random dither signal produced is now subtracted from the digital signal recovered from the data stream by reconstructing the quantization levels.
  • This operates exactly as the subtractive dither scheme shown in Figure 3, but with the specialized scheme for generating and recovering the dither signal using the lsb's of the digital data stream.
  • the digital signal may be quantized to 16 bits with digits 0 (most significant bit or msb) to digit 15 (least significant bit or lsb) , and the rectangular dither signal operates at digits 16 to 23, so that the resulting triangular dither obtained by adding two rectangular dithers operates at digits 15 to 23.
  • the resulting reconstructed signal is of 24 bit length.
  • the encoding process will similarly operate on input digital signals expressed to well over 16 bit accuracy, e.g. 24 bits, even if the inherent noise in that signal is well above the 24 bit digital quantization noise floor.
  • the digital signal is derived from an oversampling analog-to-digital converter (ADC) or as a result of earlier digital signal processing, the results ideally should not be "rounded off” below 24 bits before being encoded by the autodither encoding of Figure 6.
  • ADC analog-to-digital converter
  • any build up of perceived quantization noise or distortion is thereby minimized.
  • the invention is not confined to being implemented purely in the digital domain.
  • the level reconstruction blocks in various implementations of the invention may be realised as digital to analogue converters (DAC's) and the subtracted dither noise may be realised in the analogue domain or may also be converted to the analogue domain by a second DAC, so that the subtraction of dither from the reconstructed signal may be implemented in the analogue domain.
  • DAC digital to analogue converters
  • any non-linear mapping approximating the required pseudo-random and polarity-inversion properties may be used, including methods of deriving such a mapping by means of a length L pseudo-random logic sequence generator, of the kind well known in the mathematical and signal processing literature for generating pseudo-random number sequences.
  • Such use of extra information from each sample will generally increase the number of states for a given number of samples in the buffer, and so possibly complicate the look up table or logic, but can be used to provide a higher approximation to chaotic or random behaviour of the dither noise.
  • Noise-shaping can be added to autodither systems in exactly the same manner as already described in connection with Figures 4 to 6 for the general subtractively dithered case.
  • Figure 9 shows the use of noise-shaping with autodither encoding analogous to that shown in Figure 4. The only difference is that here, the unshaped dither noise is derived from the digits of the previous samples of the encoded quantized signal via a buffer memory and look up table (or a pseudo-random logic means) as described above. If it is desired that the noise shaping of the quantizer be different, the feedback filter in Figure 9 can be replaced by another filter H' (z " ) provided that the dither noise is filtered by the filter of Eq. (2) before being fed to the addition node, again as described earlier in connection with Figure 4.
  • Figure 10 is an alternative means, equivalent to that of Figure 9, of encoding autodithered signals with noise- shaping. This is the version of Figure 5 using the buffer and look up table (or a pseudo-random logic means) operating from the digits of the previous sample outputs of the quantizer to derive the dither noise. Again, the feedback filter H may be replaced by an alternative H' to alter the quantization noise spectrum.
  • Signals quantized and encoded by any of these autodither methods described in connection with Figures 9 and 10 may be decoded as shown in Figure 11.
  • noise-shaped dither noise added in the encoding process is removed again in the decoding process, leaving only noise-shaped quantization noise which, before shaping, has rectangular pdf of ⁇ % lsb with statistically independent samples that are also uncorrelated with the input signal to the encoder.
  • noise-shaping itself can give a perceptual improvement of around 18 dB as compared to a system with spectrally flat noise, using the published noise-shaping filters described in refs. [2] or [3]. This leads to a perceptual improvement of around 24 dB as compared with properly dithered conventional quantization using nonsubtractive Gaussian noise - i.e. a subjective improvement in performance equivalent to 4 additional bits of resolution.
  • the same kind of autodither technique can be used for other kinds of waveform coding using subtractive dither and noise-shaping.
  • it can be used for quantizing the subbands of low-bit-rate subband encoding systems, such as aptX-100 or MUSICAM with subtractive dither (and possible noise- shaping) in each subband that has a signal present, without the need for any synchronizing clock.
  • the use of rectangular pdf dither, rather than triangular may give better compatibility with nonsubtractive decoders, since lower noise energy may here be more important than modulation noise effects.
  • Autodither can also be used with companding systems of the NICAM type.
  • This method may also be used for subtractively dithered video signals, such as described by Roberts [5], and again some degree of noise-shaping can improve perceptual results (although the best noise-shaping characteristic depends on viewing distance - [6]) .
  • rectangular pdf dither may be preferable to triangular for maximum compatibility with nonsubtractive decoding.
  • both the buffer memory and the noise-shaping be two-dimensional.
  • the buffer memory should store the lsb's of data about recent previous image samples in both (say) the horizontal and vertical image directions, and that the noise-shaping filter should be of the form H(z-, ,z 2 ), where z. ⁇ is a one sample delay in (say) the horizontal di Irecti.on and z 2 -1 i.s a one-sample delay i.n the verti•cal direction, and where H has no constant term.
  • H has no constant term.
  • Such erroneous samples will have a 50% chance of having the wrong lsb's, which will affect the dither recovery.
  • the effect of such errors on the recovered dither before noise shaping will be of limited duration. typically 16 samples or whatever the length of the buffer memory used is, after the occurrence of an error.
  • the original subtracted dither noise energy N will be increased to 5N (where N is the noise that would have been heard in the absence of error, 2N is the noise energy that is no longer being subtracted, and 2N is the new noise that erroneously is being subtracted) .
  • This noise error energy after the sample error is noise shaped, and so should not be particularly serious as compared with the erroneous sample itself.
  • the erroneous sample itself multiplies the noise-shaped nonsubtractive noise signal by a one-sample duration delta function impulse response. This is more serious than it might seem, since the noise-shaping itself increases the unweighted power of the noise, which can be as much as 38 dB higher than the perceptually weighted noise power, meaning that a sharp click may be heard for an isolated sample error, especially with low wanted signals.
  • Occasional small "clicks" may be acceptable with a medium such as CD where the error rate is low, and the problem can be mitigated by using an interpolation procedure itself incorporating noise shaping or spectral weighting in playback.
  • a prime advantage of autodither is that it recovers quickly in the presence of splices between signals, typically taking the duration of the buffer to recover the correct dither signal for subsequent samples.
  • the original signals may be used before and after the cross-fade, so that no redithering or requantization is necessary.
  • the original signals are autodither decoded to a higher word length (e.g. 24 bits) , cross-faded at that word length and then re-coded. This re-encoding will give a noise level of 2N during the cross fade (depending on the precise cross-fade law) .
  • Figure 12 shows a schematic of the basics of cross-fade splicing for autodithered signals.
  • the transition at the end of the cross-fade also needs careful handling, since a sudden switch at the end of the transition from the noise-shaped autodithered signal to the original post-fade autodithered signal can cause a modulation sideband "click".
  • An alternative, and less preferred, approach to splices is to continue autodither decoding the original post-splice signal and to re-encode it using noise-shaped autodither based on the new lsbs in the buffer memory. This strategy is less preferred since it continues to cause a generation loss (3 dB in the first generation) even well after the splice is over.
  • noise-shaped autodither is, by design, backward compatible with playback without subtractive dither, giving only a 4.8 dB weighted noise penalty for the case of triangular dither noise
  • forward compatibility i.e. of the response of an autodithered decoder to a signal that is not autodither encoded.
  • the autodither decoder will subtract a spurious noise-shaped dither signal from the original signal, thereby adding noise.
  • this added noise is at a weighted level much lower than the noise found in current nonsubtractively dithered systems, and will not in any case increase the perceived noise level by more than 3 dB even in the worst case.
  • autodither should prove to be practically compatible with most older material, the added noise being in any case well below the weighted noise level added by the imperfections in most existing DAC's (digital-to-analog converters) .
  • the potential for degradation exists, and in the case of existing signals that are not properly dithered, the reconstructed dither noise may not be random due to systematic patterns in the lsb's, especially at low signal levels, so that the subtracted "dither noise" may have undesirable perceptual properties.
  • an autodither "flag" in the digital data stream to indicate when the autodither should be subtracted.
  • this flag should be placed in different digital media is still under study.
  • the flag could be placed in the sub-code data stream, and similarly room can be found for the flag in the AES/EBU or MADI data stream.
  • the data rate for this flag need not be high, since no synchronization is involved, and a temporary error in the flag of a few milliseconds is relatively unimportant.
  • One flag per CD frame should prove adequate. This flag does not have to be used by an autodither decoder, as just noted, but is preferably used to minimise the potential degradation.
  • an autodither decoder determines from a comparison of the spectral statistics with and without autodither decoding, whether the signal is in fact autodither encoded, and to subtract the noise-shaped dither signal when it is.
  • Other equivalent methods of determining the presence of autodither encoding from signal statistics exist, such as measuring the cross-correlation of the regenerated noise- shaped dither noise and the signal, and comparing this to the autocorrelation of the shaped dither noise - the two should be similar over a sufficiently long measurement time interval. In the Fourier domain, this is equivalent to the spectrum of the noise-shaped regenerated dither signal being the same as its cross spectrum with the signal. The equality will be easiest to determine at frequencies at which the signal's spectral energy is smallest.
  • a conceptually simple form of "look up table” block is simply to use a look up table in which the history h of the LSB's of previous samples stored in the buffer memory is used as an address in a look-up table memory, at which is stored a value of pseudo-random dither corresponding to that history h of LSB's.
  • the "look up table” block is this case comprises a memory addressed by h, and achieves a mapping f : h -> f(h) from possible histories h to dither signal values f(h) .
  • the function f mapping histories h of the LSB's to dither signal values is termed the "scrambling function".
  • f may be chosen using a random number generator with uniform pdf (such as is available in the libraries of many computer computation programs and languages) and storing successive random numbers at addresses corresponding to successive values of the history h regarded as a digital number.
  • uniform pdf such as is available in the libraries of many computer computation programs and languages
  • f may be chosen using a random number generator with uniform pdf (such as is available in the libraries of many computer computation programs and languages) and storing successive random numbers at addresses corresponding to successive values of the history h regarded as a digital number.
  • uniform pdf dither noise one may store the sum of two successive random numbers at addresses corresponding to successive values of the history h regarded as a digital number. It has been found that in most cases tried that this "Monte Carlo" method of constructing a look-up table gives a high degree of pseudo-randomness in the dither.
  • a similar congruence method can be applied to generate a pseudo-random dither from the word h representing the least significant digits of the previous N samples.
  • d d, + d 2 .
  • d, and d 2 can be formed by selecting a subset of N, ⁇ N out of the N digits formed by the polynomial congruence relationships.
  • triangular pdf dither may be generated as in Figure 13, using a different look up table or polynomial congruence relations to generate each of the two uniform pdf dither signal that are substantially statistically indpendent of each other, and then adding the resulting uniform or rectangular pdf dithers.
  • the feedback in the noise shaping loop generally has many more state variables, so that the overall feedback loop has a much larger number of state variables, meaning that the likely repetition length of the pseudo-random sequence will generally be much larger and much more likely not to have obviously non-random short term patterns.
  • the transmitted signal is compared to zero; if non-zero, it forces the state of a flip-flop according to whether it is positive or negative, and this causes the dither noise signal as obtained previously (e.g. with reference to fig. 7) to be multiplied by either +1 or -1. (We assume here that the dither noise takes values symmetrically distributed about zero.)
  • the corresponding decoder will likewise multiply the dither noise by +1 or - 1, according to the sign of the most recent non-zero signal word. It is evident that polarity inversion in the transmission chain will thus cause the reconstituted dither also to be inverted. It will be noted that:
  • the invention is preferably used with pseudo-random dither noise having statistical properties known substantially to eliminate a low-level non-linear distortion produced by a quantizer truncation or rounding operation. It is known [7] [8] that a necessary and sufficient condition for the elimination of non-linear distortion by a dither signal, both for non-subtractive and subtractive reproduction is that the dither noise should have a pdf that, for each sample is the convolution of any pdf function with a uniform pdf having a peak-to- peak argument range substantially equal to one quantized step size or one least significant digit of the lower resolution signal.
  • Dither signals with such convoluted pdf's may be produced by adding a first pseudo-random dither signal with any statistics to a second statistically independent pseudo-random dither signal with uniform pdf the second dither signal's peak-to-peak range of values being substantially equal to one quantization step size or one least significant digit of said lower resolution signal.
  • the second dither signal may easily be produced by the linear or polynomial congruence method described above.
  • the non-subtractive reproduction has no modulation noise then this may be achieved using dither with appropriate statistics as described in references [7] [8] [9].
  • the dither noise should have a pdf that, for each sample, is the convolution of any pdf function with a triangular pdf having a peak-to- peak argument range substantially equal to twice the quantization step size or twice the least significant digit of the lower resolution signal, where the second dither signal may have statistically independent samples or be of the form known as high-pass dither [7] [8] [9] or be any other triangular dither whose cross-correlation between successive samples is time-invariant.
  • Dither signals with such convoluted pdf's may be produced by adding a first pseudo-random dither signal with any statistics to a second statistically independent pseudo-random dither signal with triangular pdf having a peak-to-peak argument range substantially equal to twice the quantization step size or twice the least significant digit of the lower resolution signal.
  • the second dither signal may easily be produce by adding to statistically independent signals with uniform pdf's derived by the linear or polynomial congruence method.

Landscapes

  • Engineering & Computer Science (AREA)
  • Signal Processing (AREA)
  • Multimedia (AREA)
  • Computer Networks & Wireless Communication (AREA)
  • Analogue/Digital Conversion (AREA)
  • Compression, Expansion, Code Conversion, And Decoders (AREA)

Abstract

L'invention se rapporte à des systèmes de traitement de signaux numériques conçus pour convertir un signal numérique d'une première résolution en un signal numérique d'une seconde résolution différente et comprenant un générateur de bruit de superposition 'dither noise'. Le bruit de superposition est utilisé de façon à masquer ou à réduire l'effet des non-linéarités du système. Le générateur génère le bruit de superposition que l'on utilise dans le traitement d'un échantillon d'un signal à faible résolution en fonction d'échantillons antérieurs du signal de faible résolution. Par exemple, le générateur de bruit de superposition se présente sous la forme d'une table de consultation adressée par les bits de poids faible de, par ex., 16 échantillons antérieurs. Les bits de poids faible de ces échantillons sont stockés dans une mémoire tampon. Dans un codeur numérique qui code un signal numérique à haute résolution pour produire un signal numérique à faible résolution, la sortie du bruit de superposition du générateur est ajoutée au signal de faible résolution à l'entrée d'un quantificateur. Dans un décodeur numérique complémentaire, la sortie d'un générateur de bruit de superposition correspondant est soustraite du signal numérique à faible résolution. Le codeur et le décodeur peuvent être combinés dans un système de transmission qui émet un signal à haute résolution par l'intermédiaire d'un canal numérique à faible résolution produit, par ex., par un support d'information numérique tel qu'un disque compact, une bande audionumérique ou autre format de bande numérique.
PCT/GB1993/001644 1992-08-05 1993-08-04 Systeme de traitement de signaux numeriques de superposition WO1994003988A2 (fr)

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
GB9216659.4 1992-08-05
GB929216659A GB9216659D0 (en) 1992-08-05 1992-08-05 Subtractively dithered digital waveform coding system

Publications (2)

Publication Number Publication Date
WO1994003988A2 true WO1994003988A2 (fr) 1994-02-17
WO1994003988A3 WO1994003988A3 (fr) 1994-03-31

Family

ID=10719889

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/GB1993/001644 WO1994003988A2 (fr) 1992-08-05 1993-08-04 Systeme de traitement de signaux numeriques de superposition

Country Status (2)

Country Link
GB (1) GB9216659D0 (fr)
WO (1) WO1994003988A2 (fr)

Cited By (15)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
WO1995018523A1 (fr) * 1993-12-23 1995-07-06 Philips Electronics N.V. Procede et appareil de codage de sons numeriques codes en bits multiples par vibration adaptative soustractive, par insertion de bits de canaux enterres et par filtrage, et appareil de codage et de decodage de mise en oeuvre de ce procede
WO2009016337A1 (fr) * 2007-07-31 2009-02-05 Wolfson Microelectronics Plc Circuit de réduction de longueur de mot
WO2010079166A1 (fr) * 2009-01-06 2010-07-15 Skype Limited Codage vocal
US8160132B2 (en) 2008-02-15 2012-04-17 Microsoft Corporation Reducing key picture popping effects in video
EP2529331A1 (fr) * 2010-01-29 2012-12-05 Hewlett Packard Development Company, L.P. Charge utile d'essai parallèle
US8396706B2 (en) 2009-01-06 2013-03-12 Skype Speech coding
US8750390B2 (en) 2008-01-10 2014-06-10 Microsoft Corporation Filtering and dithering as pre-processing before encoding
US9450601B1 (en) 2015-04-02 2016-09-20 Microsoft Technology Licensing, Llc Continuous rounding of differing bit lengths
US10026411B2 (en) 2009-01-06 2018-07-17 Skype Speech encoding utilizing independent manipulation of signal and noise spectrum
US10027963B2 (en) 2013-11-12 2018-07-17 Dolby Laboratories Licensing Corporation Pre-dithering in high dynamic range video coding
US10250905B2 (en) 2008-08-25 2019-04-02 Microsoft Technology Licensing, Llc Conversion operations in scalable video encoding and decoding
US10306227B2 (en) 2008-06-03 2019-05-28 Microsoft Technology Licensing, Llc Adaptive quantization for enhancement layer video coding
US10395664B2 (en) 2016-01-26 2019-08-27 Dolby Laboratories Licensing Corporation Adaptive Quantization
US10602146B2 (en) 2006-05-05 2020-03-24 Microsoft Technology Licensing, Llc Flexible Quantization
EP4072020A1 (fr) * 2012-04-20 2022-10-12 Analog Devices International Unlimited Company Système et procédé de convertisseur analogique-numérique

Families Citing this family (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
GB2466671B (en) 2009-01-06 2013-03-27 Skype Speech encoding

Citations (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US4377821A (en) * 1981-09-24 1983-03-22 Bell Telephone Laboratories, Incorporated Arrangement for providing a flickerless ordered dither image for a video display
EP0427524A1 (fr) * 1989-11-09 1991-05-15 The Grass Valley Group, Inc. Dispositif pour arrondir adaptivement le signal vidéo

Patent Citations (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US4377821A (en) * 1981-09-24 1983-03-22 Bell Telephone Laboratories, Incorporated Arrangement for providing a flickerless ordered dither image for a video display
EP0427524A1 (fr) * 1989-11-09 1991-05-15 The Grass Valley Group, Inc. Dispositif pour arrondir adaptivement le signal vidéo

Non-Patent Citations (4)

* Cited by examiner, † Cited by third party
Title
INTERNATIONAL CONFERENCE ON COMMUNICATIONS OF THE INSTITUTE OF ELECTRICAL AND ELECTRONICS ENGINEERS June 1975, SAN FRANCISCO pages 27.26 - 27.30 J.W. MARK 'A dithered adaptative predictive coding scheme for compression of analogue sources' *
JOURNAL OF THE AUDIO ENGINEERING SOCIETY vol. 40, no. 5, May 1992, pages 355 - 375 , XP000275656 S.P LIPSHITZ ET AL. 'Quantization and Dither : A theoretical survey' cited in the application *
PREPRINT 2822 OF THE 87TH AUDIO ENGINEERING SOCIETY CONVENTION October 1989, NEW YORK, US M.A. GERZON , P.G. CRAVEN 'Optimal noise shaping and dither of digital signals' cited in the application *
PREPRINT 3334 OF THE 92ND AUDIO ENGINEERING SOCIETY March 1992, VIENNA J.R. STUART, R.J. WILSON 'A search for efficient dither for DSP applications' *

Cited By (19)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
WO1995018523A1 (fr) * 1993-12-23 1995-07-06 Philips Electronics N.V. Procede et appareil de codage de sons numeriques codes en bits multiples par vibration adaptative soustractive, par insertion de bits de canaux enterres et par filtrage, et appareil de codage et de decodage de mise en oeuvre de ce procede
US10602146B2 (en) 2006-05-05 2020-03-24 Microsoft Technology Licensing, Llc Flexible Quantization
WO2009016337A1 (fr) * 2007-07-31 2009-02-05 Wolfson Microelectronics Plc Circuit de réduction de longueur de mot
US8750390B2 (en) 2008-01-10 2014-06-10 Microsoft Corporation Filtering and dithering as pre-processing before encoding
US8160132B2 (en) 2008-02-15 2012-04-17 Microsoft Corporation Reducing key picture popping effects in video
US10306227B2 (en) 2008-06-03 2019-05-28 Microsoft Technology Licensing, Llc Adaptive quantization for enhancement layer video coding
US10250905B2 (en) 2008-08-25 2019-04-02 Microsoft Technology Licensing, Llc Conversion operations in scalable video encoding and decoding
US10026411B2 (en) 2009-01-06 2018-07-17 Skype Speech encoding utilizing independent manipulation of signal and noise spectrum
US8655653B2 (en) 2009-01-06 2014-02-18 Skype Speech coding by quantizing with random-noise signal
EP2905776A1 (fr) * 2009-01-06 2015-08-12 Skype Codage vocal
US8396706B2 (en) 2009-01-06 2013-03-12 Skype Speech coding
WO2010079166A1 (fr) * 2009-01-06 2010-07-15 Skype Limited Codage vocal
US8797193B2 (en) 2010-01-29 2014-08-05 Hewlett-Packard Development Company, L.P. Parallel test payload
EP2529331A4 (fr) * 2010-01-29 2014-01-01 Hewlett Packard Development Co Charge utile d'essai parallèle
EP2529331A1 (fr) * 2010-01-29 2012-12-05 Hewlett Packard Development Company, L.P. Charge utile d'essai parallèle
EP4072020A1 (fr) * 2012-04-20 2022-10-12 Analog Devices International Unlimited Company Système et procédé de convertisseur analogique-numérique
US10027963B2 (en) 2013-11-12 2018-07-17 Dolby Laboratories Licensing Corporation Pre-dithering in high dynamic range video coding
US9450601B1 (en) 2015-04-02 2016-09-20 Microsoft Technology Licensing, Llc Continuous rounding of differing bit lengths
US10395664B2 (en) 2016-01-26 2019-08-27 Dolby Laboratories Licensing Corporation Adaptive Quantization

Also Published As

Publication number Publication date
GB9216659D0 (en) 1992-09-16
WO1994003988A3 (fr) 1994-03-31

Similar Documents

Publication Publication Date Title
US5864311A (en) Systems for enhancing frequency bandwidth
KR100684051B1 (ko) 비트스트림신호의데이터처리
EP0826274B1 (fr) Procede de codage sans pertes de donnees en forme d'onde
ES2251742T3 (es) Compresion y expansion de datos de una señal de audio.
WO1994003988A2 (fr) Systeme de traitement de signaux numeriques de superposition
JP3272438B2 (ja) 信号処理システムおよび処理方法
WO1984004859A1 (fr) Procede et dispositif de transmission d'un signal numerique
WO1994018762A1 (fr) Transmission de mots contenant des donnees numeriques representant une forme d'onde de signal
KR100800373B1 (ko) 디지털 입력신호의 워드 길이를 축소하는 방법 및 장치와, 그 디지털 입력신호를 회복시키는 방법 및 장치
Craven et al. Compatible improvement of 16-bit systems using subtractive dither
EP0983636B1 (fr) Prediction de donnees dans un systeme de transmission
JPH02501180A (ja) 音響放送信号符号化復号化システム
JP2004233901A (ja) オーディオ信号処理装置及びオーディオ信号処理方法
JP2004164845A (ja) 符号情報のコピー防止方法
JP2004164844A (ja) オーディオ信号のデコード方法及びオーディオ再生用記録媒体
JP2003263832A (ja) 符号化情報のエンコード方法及びデコード装置

Legal Events

Date Code Title Description
AK Designated states

Kind code of ref document: A2

Designated state(s): JP US

AL Designated countries for regional patents

Kind code of ref document: A2

Designated state(s): AT BE CH DE DK ES FR GB GR IE IT LU MC NL PT SE

AK Designated states

Kind code of ref document: A3

Designated state(s): JP US

AL Designated countries for regional patents

Kind code of ref document: A3

Designated state(s): AT BE CH DE DK ES FR GB GR IE IT LU MC NL PT SE

121 Ep: the epo has been informed by wipo that ep was designated in this application
DFPE Request for preliminary examination filed prior to expiration of 19th month from priority date (pct application filed before 20040101)
122 Ep: pct application non-entry in european phase