Digital encapsulation of audio signals

Abstract

Encoding and decoding systems are described for the provision of high quality digital representations of audio signals with particular attention to the correct perceptual rendering of fast transients at modest sample rates. This is achieved by optimising downsampling and upsampling filters to minimise the length of the impulse response while adequately attenuating alias products that have been found perceptually harmful.

Claims

1. An encoder for producing a digital audio signal from a signal representing an audio capture, the encoder having: a downsampler receiving the signal representing the audio capture at a first sample rate and downsampling the signal to provide the digital audio signal at a transmission sample rate which is half the first sample rate; an encapsulation filter providing attenuation of at least 45 dB at frequencies within 10 kHz of a first Nyquist frequency corresponding to the first sample rate; and a flattening filter having an amplitude response that is symmetrical about the transmission Nyquist frequency and that provides amplitude boost to compensate at least 50% of an amplitude droop of the encapsulation filter at frequencies between 80% and 100% of the transmission Nyquist frequency.

2. An encoder according to claim 1, wherein a combined response of the encapsulation filter and the flattening filter is flat over the frequency range 0-20 kHz.

3. An encoder according to claim 1, wherein a z-transform of the flattening filter contains only even powers of z.

4. A system for conveying the sound of an audio capture, the system comprising: an encoder, the encoder producing a digital audio signal from a signal representing the audio capture, the encoder having a downsampler receiving the signal representing the audio capture at a first sample rate and downsampling the signal to provide the digital audio signal at a transmission sample rate which is half the first sample rate, the encoder having an encapsulation filter providing attenuation of at least 45 dB at frequencies within 10 kHz of a first Nyquist frequency corresponding to the first sample rate and a flattening filter having an amplitude response that is symmetrical about the transmission Nyquist frequency and that provides amplitude boost to compensate at least 50% of an amplitude droop of the encapsulation filter at frequencies between 80% and 100% of the transmission Nyquist frequency; and a decoder, the decoder receiving the digital audio signal and producing a reconstructed signal.

5. A system according to claim 4, wherein a combined response of the encapsulation filter and the flattening filter is flat over the frequency range 0-20 kHz.

6. A system according to claim 4, wherein a z-transform of the flattening filter contains only even powers of z.

7. A system according to claim 4, wherein the flattening filter is an Infinite Impulse Response (IIR) filter having a pole in its z-transform, and wherein the decoder comprises a filter having a zero in its z-transform whose z-plane position coincides with that of the pole, the effect of which is thereby cancelled in the reconstructed signal.

8. A system according to claim 4, wherein an impulse response of the encoder and decoder in combination has a duration for its cumulative absolute response to rise from 1% to 95% of its final value not exceeding five sample periods of the transmission sample rate.

Description

BRIEF DESCRIPTION OF THE DRAWINGS

(1) Examples of the present invention will be described in detail with reference to the accompanying drawings, in which:

(2) FIG. 1 shows a known (continuous) ‘brickwall’ antialias filter response for use with 96 kHz sampling, and (dotted) an apodised filter response;

(3) FIGS. 2A and 2B show known impulse responses corresponding to linear phase filters having the frequency responses shown in FIG. 1;

(4) FIG. 3 shows a system for transmitting an audio signal at a reduced sample rate, with subsequent reconstruction to continuous time.

(5) FIG. 4 shows the response of a (½, 1, ½) reconstruction filter, normalised for unity gain at DC;

(6) FIG. 5A shows the frequency response of an unflattened downsampling filter.

(7) FIG. 5B shows the frequency response of a downsampling filter incorporating flattening;

(8) FIG. 6 shows the response of a reconstruction filter including upsampling to continuous time and a third-order correction for the passband droop of FIG. 5A;

(9) FIG. 7 shows the total system impulse response when the filters of FIG. 4 and FIG. 5B are combined with further upsampling to continuous time;

(10) FIG. 8 shows the spectrum of two commercial recordings having a strongly rising ultrasonic response.

(11) FIG. 9 shows the response of a flattening filter symmetrical about 48 kHz for use with the downsampling filter of FIG. 5B;

(12) FIG. 10 shows (lower curve) the response of the downsampling filter of FIG. 5A and (upper curve) the response after flattening using the symmetrical flattener of FIG. 9;

(13) FIG. 11 shows a linear B-spline sampling kernel;

(14) FIG. 12A illustrates impulse reconstruction at 88.2 kHz from 44.1 kHz infra-red encoded samples aligned with even samples of an original 88.2 kHz stream.

(15) FIG. 12B illustrates impulse reconstruction at 88.2 kHz from 44.1 kHz infra-red encoded samples aligned with odd samples of an original 88.2 kHz stream.

(16) FIG. 13A shows the response of a downsampling filter having zeroes to provide strong attenuation near 60 kHz;

(17) FIG. 13B shows the response of an upsampling filter having poles to cancel the effect on total response of the zeroes in the filter of FIG. 13A;

(18) FIG. 13C shows the end-to-end response from combining the responses of FIG. 13A, FIG. 13B and an assumed external droop; and,

(19) FIG. 14 shows the normalised cumulative impulse response of the filter shown in FIG. 5A plotted against time in sample periods.

DETAILED DESCRIPTION

(20) The present invention may be implemented in a number of different ways according to the system being used. The following describes some example implementations with reference to the figures.

(21) Axioms

(22) Most adult listeners are unable to hear isolated sinewaves above 20 kHz and it has hitherto often been assumed that this implies that frequency components of a signal above 20 kHz are also unimportant. Recent experience indicates that this assumption, though plausible by analogy with linear-system theory, is incorrect.

(23) Current understanding of human hearing is very incomplete. In order to make progress we have therefore relied on hypotheses that have been only partially or indirectly verified. The invention will thus be explained on the basis of the following hypotheses: The ear does not behave as a linear system As well as analysing tones in the frequency domain, the ear also analyses transients in the time domain. This may be the dominant mechanism in the ultrasonic region. “Ringing” of filters used for antialiasing and reconstruction is undesirable, even if in the high ultrasonic range 40 kHz-100 kHz. Aliasing of frequencies above 48 kHz to frequencies below 48 kHz is not catastrophic to sound quality, provided the aliased products do not fall within the conventionally audible range 0-20 kHz. A pre-ring is usually more of a problem than a post-ring, but both are bad. It seems best if the temporal extent of the total system impulse response can be minimised.

(24) Regarding the last of these points, the “total system” is intended to include the analogue-to-digital and digital-to-analogue converters, as well as the entire digital chain in between. Ideally, one might include the transducer responses too, but these are considered outside the scope of this document.

(25) Sampling and Aliasing

(26) A continuous time signal can be viewed as a limiting case of a sampled signal as the sample rate tends to infinity. At this point we are not concerned whether an original signal is analogue, and therefore presumably continuous in time, or whether it is digital, and therefore already sampled. When we talk about resampling, we mean sampling a notional continuous-time signal that is represented by the original samples.

(27) A frequency-domain description of sampling or resampling is that the original frequency components are present in the resampled signal, but are accompanied by multiple images analogous to the ‘sidebands’ that are created in amplitude modulation. Thus, an original 45 kHz tone creates an image at 51 kHz, if resampled at 96 kHz, the 51 kHz being the lower sideband of modulation by 96 kHz. It may be more intuitive to think of all frequencies as being ‘mirrored’ around the Nyquist frequency of 48 kHz; thus 51 kHz is the mirror image of 45 kHz, and equally an original 51 kHz tone will be mirrored down to 45 kHz in the resampled signal.

(28) If a transmission channel involves several resamplings at different rates, images of the original spectrum will accumulate and there is every possibility that an audio tone will be mirrored upward by one resampling and then down by a subsequent resampling, landing within the audible range but at a different frequency from the original. It is to prevent this that ‘correct’ communications practice teaches that antialias and reconstruction filters should be used at each stage so that all images are suppressed. If this is done, resamplings may be cascaded arbitrarily without build-up of artefacts, the limitation being merely that the frequency range is limited to that which can be handled by the lowest sample rate in the chain.

(29) However, we take the view that filters that would be considered correct in communications engineering are not audibly satisfactory, at least not at sample rates that are currently practical for mass distribution. We accept that aliasing may take place and are proposing to balance aliasing against ‘time-smear’ of transients due to the lengthening of the system's impulse response caused by filtering.

(30) Thus, unlike in traditional practice, aliasing is not completely removed and will build up on each resampling of the signal. Hence, multiple resamplings to arbitrary rates are not undertaken without penalty and it is best if the signal is always represented at a sample rate that is an integer multiple of the rate that will be used for distribution. For example, analogue-to-digital conversion at 192 kHz followed by distribution at 96 kHz is fine, and conversion at 384 kHz may be better still, depending on the wideband noise characteristics of the converter.

(31) Following distribution, the consumer's playback equipment also needs to be designed so as not to introduce long filter responses, and indeed the encoding and decoding specifications should preferably be designed together to give certainty of the total system response.

(32) Downsampling from 192 kHz for 96 kHz Distribution

(33) We consider the problem of taking a signal that has already been digitised at 192 kHz, downsampling the signal to 96 kHz for transmission and then upsampling back to 192 kHz on reception. It is understood that the principles described here apply to storage as well as transmission, and the word ‘transmission’ encompasses both storage and transmission.

(34) Referring to the system shown in FIG. 3, the input signal 1 at a sampling rate such as 192 kHz is passed to a downsampling filter 2 and thence to a decimator 3 to produce a signal 4 at a lower sampling rate such as 96 kHz. After passing through the transmission or storage device 5, the 96 kHz signal 6 is upsampled 7 and filtered 8 to furnish the partially reconstructed signal 9, at a sampling rate such as 192 kHz.

(35) The main focus of this document is the method of producing the partially reconstructed signal 9, but we also note that further reconstruction 10 is needed to furnish a continuous-time analogue signal 11. The object of the invention is to make the sound of signal 11 as close as possible to the sound of an analogue signal that was digitised to furnish the input signal 1. This does not necessarily imply that signal 9 should be as close as possible in an engineering sense to signal 1. Moreover, the further reconstruction 10 may have a frequency response droop which can, if desired, be allowed for in the design of the filters 2 and 8.

(36) FIG. 3 shows the filter 2 and downsampler 3 as separate entities but it will sometimes be more efficient to combine them, for example in a polyphase implementation. Similarly the upsampler 7 and filter 8 may not exist as separately identifiable functional units.

(37) Downsampling uses decimation, in this case discarding alternate samples from the 192 kHz signal, while upsampling uses padding, in this case inserting a zero sample between each consecutive pair of 96 kHz samples and also multiplying by 2 in order to maintain the same response to low frequencies. On downsampling, frequencies above the ‘foldover’ frequency of 48 kHz will be mirrored to corresponding images below the foldover frequency. On upsampling, frequencies below the foldover frequency will be mirrored to corresponding frequencies above the foldover frequency. Thus, upsampling and downsampling create upward aliased products and downward aliased products, which can be controlled by an upsampling filter prior to decimation and a downsampling filter following the padding. The upsampling and downsampling filters are specified at the original sampling frequency of 192 kHz.

(38) If the aliased products are ignored, the total response is the combination of the responses of the upsampling and downsampling filters. In the time domain, this combination is a convolution.

(39) We have found that good results are obtained by designing upsampling and downsampling filters such that the total response is that of a Finite Impulse Response (FIR) filter of minimal length. In the z-transform domain, zeroes can be introduced into each of these filters to suppress undesirable responses. In particular, it is likely that each filter will have one or more transfer function zeroes near z=−1 in order to suppress signals near the Nyquist frequency of 96 kHz. In downsampling without filtering, such signals would alias to audio frequencies, including frequencies below 10 kHz where the ear is most sensitive. Conversely, if upsampling is performed by padding without filtering, large low frequency signal content will create large image energy near 96 kHz which, whether or not of audible consequence, may place unacceptable demands on the slew-rate capabilities of subsequent electronics, and possibly also burn out loudspeaker tweeters.

(40) FIR filters whose zeroes are all close to the Nyquist will not, by themselves, cause overshoot or ringing: the impulse response will be unipolar and reasonably compact. However a (1+z.sup.−1) factor implemented at 192 kHz introduces a frequency response droop of 0.47 dB at 20 kHz. This would be considered only marginally acceptable in professional digital audio equipment, and if we need several such factors, say five or more, the passband droop and resulting dulling of the sound certainly becomes unacceptable. Accordingly, a correction or “flattening” filter is needed, as will be discussed shortly.

(41) Upsampling from 96 kHz for Playback

(42) It is usual for reconstruction to a continuous-time signal to be performed using a sequence of ‘2×’ stages. I.e., the sampling rate is typically doubled at each stage and a conversion from digital to analogue is performed when the sampling rate has reached 384 kHz or higher. We shall concentrate firstly on the first and most critical stage: that of upsampling from 96 kHz to 192 kHz.

(43) At the heart of this upsampling is an operation, conceptual or physical, of zero-padding the stream of 96 kHz samples to produce the 192 kHz stream. That is, we generate a 192 kHz signal whose samples are alternately a sample from the 96 kHz signal and zero.

(44) Zero-padding creates upward aliased products having the same amplitude as the frequencies that were aliased. In the current context, these products are all above 48 kHz and one might assume that they will be inaudible. However the signal will generally have high amplitudes at low audio frequencies, which implies high-level alias products at frequencies near 96 kHz. As already noted, these alias products need to be controlled in order to not to impose excessive slew-rate demands on subsequent electronics and risk the burn-out of loudspeaker tweeters. The purpose of an upsampling or reconstruction filter is to provide this control, and it will be seen that strong attenuation near 96 kHz is the prime requirement.

(45) The simplest reconstruction filter that we consider satisfactory for 96 kHz to 192 kHz reconstruction is a 3-tap FIR filter having taps (½, 1, ½) implemented at the 192 kHz rate. Its normalised response is shown in FIG. 4. This filter has two z-plane zeroes at z=−1, corresponding to the Nyquist frequency of 96 kHz. These zeroes provide attenuation near 96 kHz which may or may not be sufficient so further near-Nyquist zeroes may be required. The (½, 1, ½) filter also introduces a droop of 0.95 dB at 20 kHz, or 1.13 dB if operated at 176.4 kHz, which will need to be corrected.

(46) Passband Flattening

(47) Since the system includes a downsampler, correction to flatten a frequency response that droops towards the top of the conventional 0-20 kHz audio range could be provided either at the original sample rate or the downsampled rate, but to provide the shortest end-to-end impulse response on the upsampled output the flattening should be performed at the higher sample rate, such as 192 kHz. This still leaves choice about where the correction is performed: a. The encoder (downsampler) and decoder (upsampler) each incorporates a correction for its own droop b. The encoder provides correction for itself and for the decoder c. The decoder provides correction for itself and for the encoder d. Arbitrary distribution of correction between encoder and decoder.

(48) Option (a) may be convenient in practice since the resulting downsampled stream will have a flat frequency response and can be played without a special decoder, However the resulting combined of “end-to-end” impulse response of encoder and decoder is then likely to be longer than when a single corrector corrector is designed for the total droop.

(49) Options (b) and (c) may provide the same end-to-end impulse response, and so may option (d) if a single corrector to the total response is generated, factorised ad the factors distributed. However although the end-to-end responses may be the same, putting the flattening filter in the encoder prior to downsampling generally increases downward aliasing in the encoder, and listening tests have tended to favour putting the flattening filter in the decoder after upsampling, even though upward aliases are thereby intensified.

(50) As for the design of the correction filter, the skilled person will be aware that in the case of a linear-phase droop, a linear-phase correction filter can be obtained by expanding the reciprocal of the z-transform of the droop as a power series in the neighbourhood of z=1. This total response can thereby be made maximally flat to any desired order by adjusting the order of the power-series expansion. In the present context however a minimum-phase correction filter is preferred in order to avoid pre-responses. To this end, the droop is first convolved with its own time reverse to produce a symmetrical filter and above procedure applied. This will result in a linear-phase corrector which provides twice the correction, in decibel terms, needed for the original droop. The linear-phase corrector is then factorised into quadratic and linear polynomials in z, half of the factors being minimum-phase and half being maximum-phase. The minimum-phase factors are selected and combined and normalised to unity DC gain to provide the final correction filter. This methodology was illustrated in section 3.6 of the above-mentioned 2004 paper by Craven, building on the work of Wilkinson (Wilkinson, R. H., “High-fidelity finite-impulse-response filters with optimal stopbands”. IEE Proc-G Vol. 120, no. 2, pp. 264-272: 1991 April).

(51) The effect of the correction filter is not only to flatten the passband but also to increase the near-Nyquist response of the encoder in case (b) or of the decoder in case (c), or potentially both in case (d), the increase probably requiring the introduction of further zeroes near z=−1 in order to achieve a desired near-Nyquist attenuation specification. The further zeroes will require an increase in the strength of the correction filer. Thus, the zeroes that attenuate near Nyquist and passband correction filter need to be adjusted together until a satisfactory result is obtained.

(52) Total System Response

(53) If fed with a zero-padded 96 kHz signal, the output of a 3-tap reconstruction filter having taps (½, 1, ½) implemented at the 192 kHz rate is a 192 kHz stream in which each even-numbered sample has the same value as its corresponding 96 kHz sample and each odd-numbered sample has a value equal to the average of its two neighbouring even-numbered samples. If now multistage reconstruction to continuous time similarly uses a 3-tap (½, 1, ½) reconstruction filter at each stage, the result will be equivalent to linear interpolation between consecutive 96 kHz samples.

(54) In the frequency domain, the response of such a multistage reconstruction is the square of a sinc function:

(55) ${\left(\sin c\left(\frac{f}{96\mathrm{kHz}}\right)\right)}^2$

(56) where f is frequency and

(57) $\sin c\left(x\right)=\frac{\sin \left(x\right)}{x}.$

(58) The passband droop may be approximated by a quadratic in f:

(59) $1-\frac{{{}^2\left(f/96\mathrm{kHz}\right)}^2}{3}\approx 1-3.29f/96{\mathrm{kHz}}^2,$

(60) which implies a response of −1.34 dB at 20 kHz if reconstructing from 96 kHz, or −1.61 dB at 20 kHz if reconstructing from 88.2 kHz.

(61) Reconstructed thus, the slew rate of the continuous-time signal is never greater than that implied by the 96 kHz samples on the basis of linear interpolation. Nevertheless, it will have small discontinuities of gradient. Viewed on a sufficiently small time scale, this is not possible electrically, let alone acoustically. It is outside our scope to consider the analogue processing in detail, but we note that an impulse response that is everywhere positive must, unless it is a Dirac delta function, have some frequency response droop. We prefer not to require the use of an analogue ‘peaking’ filter to produce a flat overall response since the shortest overall impulse response is likely to be obtained if all passband correction is applied at a single point. We therefore prefer that the digital passband flattening should have some allowance for analogue droop.

(62) Nevertheless, the more droop that is corrected, the less compact is the upsampling filter. In the filters presented here we have therefore compensated for the sinc(⋅).sup.2 droop for assumed multistage reconstruction from a 192 kHz stream to continuous time, with a further margin to allow for a small droop, amounting to 0.162 dB at 20 kHz, in subsequent analogue processing. This margin would allow for an analogue system having a strictly nonnegative impulse response of rectangular shape and extent 5 μs, or alternatively a Gaussian-like response with standard deviation approximately 3 μs.

(63) FIG. 5A shows the response of a 6-tap downsampling filter designed according to these principles having a near-Nyquist attenuation of 72 dB and z-transform response:
0.0633+0.2321z.sup.−1+0.3434z.sup.−2+0.2544z.sup.−3+0.0934z.sup.−4+0.0134z.sup.−5

(64) If paired with the previously discussed 3-tap upsampling filter having response (½+z−1+½z−2), we find that a 4-tap correction filter:
4.3132−5.3770z.sup.−1+2.4788z.sup.−2−0.4151z.sup.−3

(65) will correct the total droop from the downsampling filter and the 3-tap upsampling filter, to provide an end-to-end response flat within 0.1 dB at 20 kHz, including the effect of analogue droop as discussed above. If this correction filter is folded with the downsampling filter, the combined encoding filter has z-transform:

(66) $0.27289+\frac{0.66093}{z}+\frac{0.39002}{z^2}-\frac{0.20014}{z^3}-\frac{0.20992}{z^4}+\frac{0.04329}{z^5}+\frac{0.05411}{z^6}-\frac{0.00563}{z^7}-\frac{0.00555}{z^8}$

(67) and the response shown in FIG. 5B, which rises above 20 kHz in order to pre-correct the droop from the subsequent upsampling and reconstruction.

(68) Alternatively, the correction can be folded with the upsampling filter (½+z.sup.−1+½z.sup.−2) whose response is shown in FIG. 4 to produce a decoding filter having the response shown in FIG. 6 and the z-transform:
2.1566−0.5319z.sup.−1+0.7076z.sup.−2−1.6566z.sup.−3+0.0319z.sup.−4−0.2076z.sup.−5

(69) In this case it is the decoder that has a rising response, to correct the droop from the 6-tap encoding filter having the response of FIG. 5A. Listening tests have indicated that this 9-tap downsampling filter has a distinct superiority relative to longer filters and we have deduced that shorter filters are preferable generally.

(70) Of greater significance however is the total response when the downsampler, upsampler and assumed analogue response are combined. FIG. 7 shows the impulse response from the downsampler, a multi-stage upsampler as proposed above and an analogue system having a rectangular impulse response of width 5 μs. With no threshold applied, the total extent of the response is 13 samples or 67.7 μs, but with a threshold of −40 dB or 1% of the maximum, the absolute value of the response exceeds the threshold only in a region of extent 49.5 μs, i.e. 9.5 samples at the 192 kHz rate or 4.75 samples at the transmission sample rate of 96 kHz. Similarly, with a threshold of −20 dB or 10% of the maximum, the absolute value of the response exceeds the threshold only in a region of extent 32.2 μs, i.e. 6.2 samples at the 192 kHz rate or 3.1 samples at the transmission sample rate of 96 kHz. Thus, it is safe to say that the temporal extent of this filter does not exceed 4 sample periods of the transmission sample rate. When other criteria are tightened, the impulse response may need to be somewhat longer, but in nearly all reasonable cases it is possible to achieve an impulse response of length not exceeding 6 sample periods at the transmission sample rate.

(71) An encoder and decoder combination incorporating the downsampling and upsampling filters described above and with the total system response shown in FIG. 7 has been found to produce audibly good results on available 192 kHz recordings. Indeed the decoded signal has sometimes sounded better than conventional playback of the 192 kHz stream without downsampling, a result that could be attributed to the attenuation by the downsampling filter of any ringing near 96 kHz already present in the 192 kHz stream.

(72) Alias Trading Based on Noise Spectrum Analysis

(73) Much commercial source material has a noise floor that rises in the ultrasonic region because of the behaviour of analogue-to-digital converters and noise shapers. For example, the spectrum of a commercially available 176.4 kHz transcription of the Dave Brubeck Quartet's “Take 5”, shown as the upper trace in FIG. 8, reveals a noise floor that increases by 42 dB between 33 kHz and 55 kHz, these frequencies being equidistant from the foldover frequency of 44.1 kHz when downsampled. If there were no filtering before decimation, the resulting 88.2 kHz stream would have noise at 33 kHz composed almost entirely of noise aliased from 55 kHz and would thereby have a spectral density some 42 dB higher than in the 175.4 kHz presentation of the recording.

(74) The downsampling filter of FIG. 5B, if operated at 176.4 kHz instead of 192 kHz, would provides gain of +2.3 dB and −6.7 dB at 33 kHz and 55 kHz respectively, a difference of 9 dB. Downsampling “Take 5” with this filter, components aliased from 55 kHz would still dominate original 33 kHz components by 33 dB. The alternative downsampling filter of FIG. 5A provides 16.8 dB discrimination between these two frequencies, resulting in aliased components 25 dB higher than the original components. For this is a somewhat exceptional case, filters (to be described) having still larger discrimination might be preferable; nevertheless the filter of FIG. 5A has been found satisfactory in many cases, and to provide better audible results than the filter of FIG. 5B. Thus placing the correction filter in the decoder, as in option (c) discussed earlier, seems preferable to placing it in the encoder, option (b).

(75) The above discussion has concentrated on downward aliased signal components, but it should be noted that putting the correction filter in the decoder will have the effect of boosting upward aliased components. It is a matter of trading downward aliasing against upward aliasing, and for downsampling from 192 kHz to 96 kHz, or from 176.4 kHz to 88.2 kHz it seems audibly better to reduce downward aliasing even if upward aliasing thereby increased.

(76) There is no established criterion for how much aliased components should be reduced relative to original components, but a criterion may be derived based on balancing phase distortion in the audio band against total noise. We assume that the total response should be minimum-phase in order to avoid pre-responses. The flattening filter is always designed to give an total amplitude response flat to fourth order but Bode's phase-shift theorems tell us that when ultrasonic attenuation is introduced, phase distortion is inevitable in a minimum-phase system. When the phase response is expanded as a series in frequency, only odd powers are present. The linear term is irrelevant since it is equivalent to a time delay, hence the cubic term is dominant. If now additional attenuation δg decibels is introduced over a frequency interval δf centred on frequency f, we can deduce from Bode's theorems that the resulting addition to the cubic term in the phase response will be proportional to δg.Math.δf/f.sup.4. From the inverse fourth power dependence on f we can deduce that for lowest total noise consistent with a given phase distortion and a given end-to-end frequency response, the upward and downward aliasing should be balanced so that the ratio of the original noise power to the aliased noise power is equal to the inverse fourth power of the ratio of the two frequencies involved.

(77) In the case of downsampling to 96 kHz, this criterion implies that the noise spectral density at 36 kHz that results from original 60 kHz noise should be 8.9 dB below the noise spectral density at 36 kHz in the original 192 kHz sampled signal. Also, at the foldover frequency of 48 kHz, the spectrum of the noise after filtering by the downsampling filter should optimally have a slope of −12 dB/8ve. It follows that the slope of the downsampling filter of FIG. 5A is not sufficient in the case of “Take 5” according to this criterion, and a downsampling filter with a steeper slope near 48 kHz is indicated if this criterion is considered relevant. “Take 5” is somewhat exceptional but the spectrum of “Brothers in Arms” by “Dire Straits”, also shown in FIG. 8, also has a high slope near the foldover frequency.

(78) Flattening the Downsampled Signal

(79) As discussed, aliasing considerations often suggest that that the downsampling filter be not flattened, flattening being postponed to a subsequent upsampler. The transmitted signal will thereby not have a flat frequency response, which may be a disadvantage for interoperability with legacy equipment that does not flatten.

(80) A way to avoid the disadvantage without affecting the alias property of the downsampler is to flatten using a filter with a response such as shown in FIG. 9 that is symmetrical about the transmission Nyquist frequency, i.e. half the transmission sample frequency. The transmission Nyquist frequency is 48 kHz if downsampling from 192 kHz to 96 kHz, giving the unflattened and flattened downsampling responses are shown in FIG. 10.

(81) The reason that the disadvantage is avoided is that the ‘legacy flattener’ is a symmetrical filter that treats each frequency and its alias image equally. The two frequencies are boosted or cut in the same ratio so the ratio of upward to downward aliasing in a subsequent decimation is not affected.

(82) The response shown in FIG. 9 is in fact the response of the filter:

(83) $\frac{1.660575124}{1+0.6108508622z^{-2}+0.04972426151z^{-4}}$

(84) which is minimum-phase all-pole and contains only even powers of z. Filtering with this filter prior to decimation-by-2 is equivalent to filtering the decimated stream using the all-pole filter:

(85) $\frac{1.660575124}{1+0.6108508622z^{-1}+0.04972426151z^{-2}}$

(86) which is a process that can be reversed in a decoder, for example by applying a corresponding inverse filter:
0.6022009998(1+0.6108508622z.sup.−1+0.04972426151z.sup.−2)

(87) to the received decimated signal prior to upsampling. Thus, z-plane poles in the encoding filter are cancelled by zeroes in the decoder. In the time domain, any ringing caused by the legacy flattener in the encoder is quenched by the corresponding ‘legacy unflattening’ in the decoder, and this is one of the ways in which the total impulse response of the combination of encoder and decoder is more compact than that of the encoder alone.

(88) After upsampling, a decoder can apply a psychoacoustically optimal flattener at the higher sample rate, just as if there were no legacy flattener. It is thus completely transparent that that the decimated signal has been flattened and then unflattened again.

(89) The ‘legacy unflattener’ can alternatively be implemented after upsampling, using:
0.6022009998(1+0.6108508622z.sup.−2+0.04972426151z.sup.−4)

(90) at the higher sampling rate. As this is an FIR filter, it may well be convenient to merge it with the upsampling filter and the end-to-end flattener. In this case the legacy unflattener may not be a separately identifiable functional unit. Thus, for both the legacy flattener and the legacy unflattener there is the option of implementation at the transmission sample rate or at the higher sample rate, in the latter case using a filter whose response is symmetrical about the transmission Nyquist frequency. In this document these two implementation methods are considered equivalent and a reference to just one of them may be taken to include the other. Moreover if implemented at the higher rate the flattener or unflattener may be merged with other filtering, though its presence may be deduced if the z-transform of, respectively, the total decimation filtering or the total reconstruction filtering has z-transform factors that contain powers of z.sup.n only where n is the decimation or interpolation ratio.

(91) It is not required that the legacy flattener be all-pole: it could be FIR or a general IIR filter provided its response is symmetrical about the transmission Nyquist frequency. For example the FIR filter:
1.444183138−0.5512608378z.sup.−1+0.1190498978z.sup.−2−0.01197219763z.sup.−3

(92) could be applied after decimation in an encoder and its inverse prior to upsampling in a decoder, this third-order FIR filter being similarly effective to the second-order all-pole filter of FIG. 9 in flattening the transmitted signal. In this case the decoder would have poles that cancel zeroes in the encoder. This FIR flattener could alternatively be implemented prior to decimation using:
1.444183138−0.5512608378z.sup.−2+0.1190498978z.sup.−4−0.01197219763z.sup.−6

(93) and in this form it could be merged with the downsampling filter and so not be identifiable as a separate functional unit.

(94) While the legacy flattener has here been explained in the context of a 2:1 downsampling, the same principles apply in the case of an n:1 downsampling, where the legacy flattening and unflattening may be performed at the transmission sample rate using a general minimum-phase filter and its inverse, or it may be performed at the higher sample rate using a filter containing powers of z.sup.n only. In both cases the legacy flattener has a decibel response that is symmetrical about the transmission Nyquist.

(95) Having noted that an invertible symmetrical filter applied at the original sample rate makes no difference to the alias characteristics of the filtering and that its effect can be reversed completely in a decoder, it follows that in comparing the suitability of one candidate downsampling filter with another, symmetrical differences in the decibel response are irrelevant. Hence we decompose the decibel response dB(f) of a given filter into a symmetric component:

(96) $\frac{\mathrm{dB}\left(f\right)+\mathrm{dB}\left({\mathrm{fs}}_{\mathrm{trans}}-f\right)}{2}$

(97) and an asymmetric component:

(98) $\frac{\mathrm{dB}\left(f\right)-\mathrm{dB}\left({\mathrm{fs}}_{\mathrm{trans}}-f\right)}{2}$

(99) where f is frequency, fs.sub.trans is the transmission sampling frequency, and a comparing between two downsampling filters we concentrate on the asymmetric component, leaving the symmetric component to be adjusted if necessary in a decoder. The asymmetric component is, in fact, half of the alias rejection:
alias rejection=dB(f)−dB(fs.sub.trans−f)

(100) Infra-Red Coding

(101) We refer to the paper by Dragotti P. L., Vetterli M. and Blu T.: “Sampling Moments and Reconstructing Signals of Finite Rate of Innovation: Shannon Meets Strang-Fix”, IEEE Transactions on Signal Processing, Vol. 55, No. 5, May 2007. Section III A of this paper considers a signal consisting of a stream of Dirac pulses having arbitrary locations and amplitudes, and the question is asked of what sampling kernels can be used so that the locations and amplitudes of the Dirac pulses may be deduced unambiguously from a uniformly sampled representation of the signal.

(102) We consider that this question may be relevant to the reproduction of audio, in that many natural environmental sounds such as twigs snapping are impulsive and it is by no means clear that a Fourier representation is appropriate for this type of signal. The linear B-spline kernel shown in FIG. 11 is the simplest polynomial kernel that will enable unambiguous reconstruction of the location and amplitude of a Dirac pulse. We have given the name “infra-red coding” to a downsampling specification based these ideas.

(103) In downsampling, we start with a signal that is already sampled but the conceptual model is that this is a continuous time signal, in which the original samples are presented a sequence of Dirac pulses. The continuous time signal is convolved with a kernel and resampled at the rate of the downsampled signal. Referring to FIG. 11, the resampling instants are the integers 0, 1, 2, 3 etc while the original signal is presented, on a finer grid. Assuming that the original samples and resampling instants are aligned, then the continuous time convolution with the linear B-spline followed by resampling is equivalent to a discrete-time convolution with the following sequences prior to decimation: (1, 2, 1)/4 for decimation by 2 (1, 2, 3, 2, 1)/9 for decimation by 3 (1, 2, 3, 4, 3, 2, 1)/16 for decimation by 4 . . . (1, 2, 3, 4, 5, 6, 7, 8, 7, 6, 5, 4, 3, 2, 1)/64 for decimation by 8.

(104) These sequences are merely samplings at the original sampling rate of the B-spline kernel. Since the kernel has a temporal extent of two sample periods at the downsampled rate, in all cases the downsampling filter will have a temporal extent not exceeding two sample periods at the downsampled rate.

(105) Thus for decimation by 2 the downsampling filter would have z-transform (¼+½z.sup.−1+¼z.sup.−2). We have found that very satisfactory results can be obtained using this filter for downsampling in combination with the same filter, suitably scaled in amplitude, for upsampling, with also a suitable flattener, which can be placed after upsampling, or merged with the upsampler. For downsampling from 176.4 kHz to 88.2 kHz the combined downsampling and upsampling droop of 2.25 dB @ 20 kHz can be reduced to 0.12 dB using a short flattener such as:
2.1451346747−1.4364916731z.sup.−1+0.2913569984z.sup.−2 at 176.4 kHz.

(106) The total upsampling and downsampling response is then FIR with just 7 taps, hence a total temporal extent of six sample periods at the 176.4 sample rate or three sample periods at the downsampled rate. This is the shortest total filter response known to us that is often audibly satisfactory and maintains a flat response over 0-20 kHz.

(107) The infra-red prescription does not provide the strong rejection of downward aliasing considered desirable for signals with a strongly rising noise spectrum but there are many commercial recordings whose ultrasonic noise spectra are more nearly flat or are falling. With a downsampling ratio of 2:1 the slope of an infra-red downsampling filter is −9.5 dB/8ve at the downsampled Nyquist frequency; with a ratio of 4:1 it is −11.4 dB/8ve and in the limiting case of downsampling from continuous time it is −12 dB/8ve. This compares with a slope of −22.7 dB/8ve for the downsampling filter of FIG. 5A and for this type of source material the infra-red encoding specification may not be suitable.

(108) An encoder for routine professional use should ideally attempt to determine the ultrasonic noise spectrum of material presented for encoding, for example by measuring the ultrasonic spectrum during a quiet passage, and thereby make an informed choice of the optimal downsampling and upsampling filter pair to reconstruct that particular recording. The choice then should be communicated as metadata to the corresponding decoder, which can then select the appropriate upsampling filter.

(109) The above discussion has concentrated substantially on downsampling from a ‘4×’ sampling rate such as 192 kHz or 176.4 kHz to a ‘2×’ sampling rate such as 96 kHz or 88.2 kHz, but of commercial importance also is downsampling from a 4× or a 2× sampling rate to a 1× sampling rate such as 48 kHz or 44.1 kHz. In fact the same ‘infra-red’ coefficients ¼+½z.sup.−1+¼z.sup.−2 as discussed above for use at higher sampling rates have also been found to provide audibly good results when downsampling from 88.2 kHz to 44.1 kHz. This is perhaps surprising as one might have expected that the ear would require greater rejection of downward aliased images of original frequencies at this lower sample rate, but repeated listening tests have confirmed that this does not seem to be the case. The same filter can be used for upsampling, combined with or followed by a flattener. At this lower sample rate, a flattener with more taps is needed, for example the filter:
4.0185−5.9764z.sup.−1+4.6929z.sup.−2−2.4077z.sup.−3+0.8436z.sup.−4−0.1971z.sup.−5+0.0279z.sup.−6−0.0018z.sup.−7

(110) running at 88.2 kHz, flattens the total response of downsampler and the upsampler to within 0.2 dB at 20 kHz and has found to be audibly satisfactory.

(111) A flattener and unflattener pair can be provided as was described previously to allow compatibility with 44.1 kHz reproducing equipment. To provide a maximally flat response with a droop not exceeding 0.5 dB at 20 kHz, a nine-tap all-pole flattener implemented at 44.1 kHz is theoretically required:

(112) $\frac{1.2305}{1+0.2489z^{-1}-\text{⁠}0.0231z^{-2}+\text{⁠}0.0058z^{-3}-0.0015z^{-4}+0.0003z^{-5}-\text{⁠}0.0001z^{-6}+0.8166{10}^{-5}{z}^{-7}-0.7262{10}^{-6}{z}^{-8}+0.3151{10}^{-7}{z}^{-9}}$

(113) though some of the later terms of the denominator here given could be deleted with minimal introduction of passband ripple. Either way, the expression here given can be inverted to provide a corresponding FIR unflattener. A high-resolution decoder would typically unflatten at 44.1 kHz, upsample to 88.2 kHz and then flatten using an optimally-designed flattener at 88.2 kHz such as the 7th order FIR flattener given above. In this case, the impulse response of the encoder and high-resolution decoder together has 12 nonzero taps, whereas the the encoder alone has an impulse response that continues longer, albeit at lower levels such as −40 dB to −60 dB.

(114) One or both of the flattening and unflattening filters presented here for operation at the 44.1 kHz rate could be transformed as indicated previously to provide the same functionality when operated at 88.2 kHz or a higher rate, if this is more convenient.

(115) Reconstruction as described above to continuous time from a 44.1 kHz infra-red coding of an impulse presented as a single sample at time t=0 within an 88.2 kHz stream is illustrated in FIGS. 12A and 12B. In FIG. 12A the reconstruction is from 44.1 kHz samples, shown as diamonds, coincident in time with even samples of the 88.2 kHz stream, whereas in FIG. 12B the reconstruction is from 44.1 kHz samples, shown as circles, coincident with odd samples of the 88.2 kHz stream points. The horizontal axes is time t in units of 88 kHz sample periods and the vertical axes shows amplitude raised to the power 0.21, which provides visibility of small responses but also may have some plausibility according to neurophysiological models of human hearing which suggest that for short impulses, peripheral intensity is proportional to amplitude raised to the power 0.21. The 44.1 kHz representations have been derived using the infra-red method as described above including flattening for compatibility with legacy equipment, while the two high-resolution reconstructions similarly use a legacy unflattener followed by infra-red reconstruction and a flattener implemented at 88.2 kHz.

(116) It will be noted that the 44 kHz stream shows a time response that continues long after the high resolution reconstruction of the impulse has ceased, thus demonstrating the effectiveness of the pole-zero cancellation in providing an end-to-end response that is more compact than the response of the encoder alone.

(117) FIGS. 12A and 12B also illustrate that the concept of an ‘impulse response’ needs to be defined more clearly when decimation is involved. In the case of decimation-by-2 the result is different for an impulse presented on an odd sample from that on an even sample. In this document we use the term ‘impulse response’ to refer to the average of the responses obtained in these two cases.

(118) It will be appreciated that infra-red coding as described provides two z-plane zeroes at the sampling frequency of the downsampled signal, and in the case of a downsampling ratio greater than 2, at all multiples of that frequency. This may be considered the defining feature of infra-red coding.

(119) Suppression of Downward Aliasing

(120) As noted, when encoding an item such as ‘take 5”, see FIG. 8, it may be desirable that the downsampling filter provide strong attenuation at frequencies such as 55 kHz where the noise spectrum peaks. It would be natural to think of placing one or more z-plane zeroes to suppress energy near this frequency. To do so would however increase the total length of the end-to-end impulse response: firstly because each complex zero requires a further two taps on the downsampling filter, and secondly because a zero near 55 kHz adds significantly to the total droop so a longer flattening filter will likely also be required.

(121) With one caveat, the increase in length can be avoided using pole-zero cancellation: the complex zero in the encoder's filter is cancelled by a pole in the decoder. In one embodiment, a downsampling filter incorporating three such zeroes is paired with an upsampling filter having three corresponding poles. The resulting downsampling and upsampling filter responses are shown in FIG. 13A and FIG. 13B and the end-to-end response from combining these two filters with an assumed external droop is shown in FIG. 13C. For consistency with other graphs, these plots assume a sampling rate of 196 kHz so the maximum attenuation is near 60 kHz rather than 55 kHz.

(122) The caveat here is that although downward aliasing has been suppressed, upward aliasing has been increased. For use on tracks such as ‘Take 5’, the increased upward-aliased noise is well covered by the steeply-rising original noise. However signal components near 33 kHz would also result in much larger aliases near 55 kHz. It is thus arguably misleading simply to present an end-to-end frequency response that ignores aliased components; nevertheless it appears that the ear is relatively tolerant to the upward aliases provided the boost applied to the alias is not excessive.

(123) The heavy boost of 38 dB at 57 kHz shown in FIG. 13B may seem at first unwise, but if a legacy flattener is used as described above then the decoder will incorporate a legacy unflattener which will compensate most of this boost, so the decoder as a whole will not exhibit the boost.

CONCLUDING REMARKS

(124) It is to be noted that some of the decoding responses described in this document have features that would normally be absent from reconstruction filters. These features include a response that is rising rather than falling at the half-Nyquist frequency of 44 .kkHz or 48 kHz, and a z-transform having one or more factors that are functions of even powers of z only, and thereby have individual responses that are symmetrical about the half-Nyquist frequency.