alves@calvin.usc.edu (William Alves) (04/09/91)
In article <1991Apr5.172338.11258@vaxa.strath.ac.uk> cnbs30@vaxa.strath.ac.uk writes: > >How much of this is just marketing gimmicks? 4 x's versus 8 x's >oversampling, can we hear the difference? Oversampling is described >in some Hifi magazines as simply inserting extra samples between >adjacent samples in order to increase the resolution? I always >thought that oversampling was to avoid having to use expensive >analogue filters at the input (anti aliasing) and outputs?? Oversampling is, to oversimplify, "inserting extra samples," and it is also a way to avoid having to use steep roll-offs analog filters on the outputs. First of all, output filters are used because a digital sampling system produces artifacts known as "images," or high-frequency reflections of the input signal at multiples of the Nyquist frequency. Low-pass analog filters are the most straight-forward way to get rid these images (hence "anti- imaging filters). Oversampling is not used because the analog filters are more expensive, but because their roll-offs, even though they're very steep, will cut well into the audio spectrum. Not only will this effect the frequency response, but, as it turns out, more importantly, the phase response. Early CD players suffered from severe phase distortion in the high fre- quencies because of these filters. This was the probably culprit in early audiophile complaints of "graininess" or "harshness" in the high-end. Because digital-to-analog converters are generally n times faster than a comparably-priced analog-to-digital converter where n is the number of bits, Philips Corporation came up with a way of, in effect, shifting the sampling rate up 4 or 8 times upon playback. This isn't exactly interpo- lating 3 samples in between every two because that would cause secondary aliasing and (depending on the interpolation algorithm) more noise. Nevertheless, it's convenient to think about it that way. These extra samples do not add any information that wasn't recorded, of course, but they do shift the unwanted "images" four or eight times further up into the audio spectrum. There they can be easily filtered out by an analog filter with a much shallower roll-off. For more info, see: Wayne Schott, "Philips Oversampling System for Compact Disc Decoding," Audio, April 1984, 32-35. Ken C. Pohlmann, Principles of Digital Audio, (Indianapolis, IN: Howard W. Sams & Co., 1985). There's a second edition out now. Can one tell the difference between 4x and 8x oversampling? I couldn't on the CD-players I tested. Differences can be the result of a lot of things. The bottom line is trust your ears and not the numbers. Electronics manufacturers found out long ago that consumers respond over- whelmingly to big numbers, even when totally meaningless. Things that can- not be quantified on a spec sheet, on the other hand, get the short end. My personal pet peeve is the poor quality of analog output amps in CD players and synths which often totally obliterate the advantages of high-tech digital filters and such. Bill Alves
curt@cynic.wimsey.bc.ca (Curt Sampson) (04/12/91)
In article <31771@usc> alves@calvin.usc.edu (William Alves) writes: > Oversampling is not used because the analog filters are more expensive, > but because their roll-offs, even though they're very steep, will cut > well into the audio spectrum. Not only will this effect the frequency > response, but, as it turns out, more importantly, the phase response. Actually, the roll-offs, being very steep, stay out of the audio spectrum. The problem here is solely due to phase shift. > ...Philips Corporation came up with a way of, in effect, shifting the > sampling rate up 4 or 8 times upon playback. This isn't exactly interpo- > lating 3 samples in between every two because that would cause secondary > aliasing and (depending on the interpolation algorithm) more noise. > Nevertheless, it's convenient to think about it that way. > > These extra samples do not add any information that wasn't recorded, of > course, but they do shift the unwanted "images" four or eight times further > up into the audio spectrum. There they can be easily filtered out by an > analog filter with a much shallower roll-off. This is not completely correct. When the sample is converted from 44.1 to, say, 176.4 KHz, no interpolation or other changing of the data is done. The data is simply duplicated. If you started with the following sequence: 3 7 11 15 it would be turned into this when 4x oversampled: 3 3 3 3 7 7 7 7 11 11 11 11 15 15 15 15 Then a digital filter is applied, which is essentially a way of "smoothing" the samples. The "edges" you see (because it is essentially a sqare wave before filtering) are what cause the high frequences to be present. cjs -- | "It is actually a feature of UUCP that the map of curt@cynic.uucp | all systems in the network is not known anywhere." curt@cynic.wimsey.bc.ca | --Berkeley Mail Reference Manual (Kurt Schoens)
ogata@leviathan.cs.umd.edu (Jefferson Ogata) (04/13/91)
In article <1991Apr12.060657.19964@cynic.wimsey.bc.ca> curt@cynic.wimsey.bc.ca (Curt Sampson) writes: |> |> This is not completely correct. When the sample is converted from |> 44.1 to, say, 176.4 KHz, no interpolation or other changing of the |> data is done. The data is simply duplicated. If you started with the |> following sequence: |> |> 3 7 11 15 |> |> it would be turned into this when 4x oversampled: |> |> 3 3 3 3 7 7 7 7 11 11 11 11 15 15 15 15 |> |> Then a digital filter is applied, which is essentially a way of |> "smoothing" the samples. The "edges" you see (because it is |> essentially a sqare wave before filtering) are what cause the high |> frequences to be present. A digital filter will do the same thing as interpolation in this context. It may not be linear interpolation but it amounts to the same effect as just straight interpolation with an appropriate algorithm. You still aren't saying anything about the analog Nyquist filter that has to come after the DAC. Effectively oversampling allows this analog filter to have lower attenuation because the high frequency components of the digital steps have reduced amplitude from the effect of interpolation (or filtering). |> cjs |> -- |> | "It is actually a feature of UUCP that the map of |> curt@cynic.uucp | all systems in the network is not known anywhere." |> curt@cynic.wimsey.bc.ca | --Berkeley Mail Reference Manual (Kurt Schoens) -- Jefferson Ogata ogata@cs.umd.edu University Of Maryland Department of Computer Science