I am all for people knowing what they are talking about before they decide to talk but JB's posts were technical nonsense.
Elaborate.
Alrighty then.
The standards for CDs represent the absolute minimum for recording sound, not music. The samples are too small and the sample rate is too low. According to sampling theory, the low sample rate required a very steep, ‘brick wall,’ anti-image filter, which contributed to the bad sound of the first generation DACs.
B.S.
44,100 samples per second and 16 bits resolution yeilds a bandwidth of 22,050hz and noise level of -96db.
It is a mathematical fact that any error resulting from encoding at 44.1k and 16 bits must be either above 22.05khz or below -96db, Period!
This leaves plenty of room for full bandwidth music on a blank (silent) canvas.
You could go more but it is not necessary (even for audiophiles).
Your hearing might exceed the bandwidth of the 44khz sampling system which will be less than the nyquist frequency to limit the ringing of the ant-alias filter and could in theory exceed the dynamic range of the 16 bit system which would require a very high output system and very quiet listening environment.
This is simply a trivial fact and should not get in the way of understanding that 44khz and 16 bits has the potential to encode incredibly detailed music.
Fortunately the anti-alias filter is at the top of the audio range where your hearing is barely there so don't get too hung up on this detail.
A filter with rounder knee can also be used if you like.
At the time, microprocessors were too slow to be able to manipulate the sample data in real time and DAC designers did what they could to increase the clock rate of the DAC to reduce the severe requirements of the anti-image filter. This brand new idea was called 2X over-sampling. It involved doubling the clock rate and inserting a null sample in between each recorded sample. Today, that technique is sometimes called decimation. The problem is decimation halves the energy content of the signal and adds noise. I believe preserving the energy content is essential to accurately reconstruct music.
The added noise is above the passband and there is no consequence to the resulting level being -6db (or even -60db) unless your filter lacks the mathematical precision for accurate calculation of the signal.
The next attempt, called 4X oversampling, involved using 4 DACs that operate on successive quarters of the sample period. The outputs of the DACs are summed. The sample rate is quadrupled and the effective sample size is increased by two bits. However, instead of a vertical step between recorded samples, as there would be with no oversampling, there is a diagonal line between the two points. Instead of a square wave, the DAC outputs more of a triangular wave. In terms of energy content, a triangular wave does not approximate the energy in a sine wave as well as square wave does. Today we call it linear analog interpolation.
Actually a trangle wave approxamates a sine wave much better than a square wave.
The amplitude of a square waves harmonics is equal to the reciprocal of the harmonic number while the amplitude of a triangle waves harmonics is equal to the reciprocal of the harmonic number squared.
The triangle wave is much more pure.
This is not how four times oversampling systems were implemented although I'm shure there have been some implemented in this manner as it is a good idea that does not require a digital filter as the multiple DAC's diffused timing will cause high frequencies to cancel creating a filter.
Also since no formal digital filter is employed there are no samples generated with values between the steps of the 16 bit system which would require higher resolution DAC's and/or redithering to properly decode.
This approach could be further refined by using more DAC's with giving finer diffusion and adjusting the gain of each DAC in the sequence so that their resultant response mimics that of a desirable filter function (the gain of each of the DAC's in the timed series would be adjusted acording to the impulse response of a target LP filter).
Having all of the DAC's in the series at the same gain also works but the resultant filter has some ripple and the slope is limited, having more DAC's and tuning their gains would allow for even more refined filter responses (If you were trying to build a high-end DAC with early 80's technology this would be the way to do it).
Standard oversampling systems (2x, 4x, 8x....) oversample by inserting null samples between the existing samples then use a digital filter and then a DAC running at the oversampling sample rate.
Employing multiple DAC's can increase the effective number of bits (S/N) of a DAC system either by oversampling and dithering or by feeding the data to them in such a way that more steps are available or by oversampling and noise shaping however the above described multiple DAC method does not give an effective increase in the number of bits.
When microprocessors were fast enough to manipulate the sample data in real-time we got 8X oversampling and the digital filter. The technique, more properly called digital interpolation, attempts to recreate the shape of the signal in between each pair of samples from the original recording. To do so, a large number of samples preceding and following the sample period being interpolated are examined to determine the shape of the signal through the period in question. The accuracy of the interpolation depends on the accuracy of the samples and a similar shape of the signal leading up to and following the interpolated period. The effective sample size is increased by three bits.
Microprocessors were only used in ultra expesive models when they became available.
Digital filters are what was used since the first 2x oversampling DAC's were employed not introduced for the first time with 8x oversampling.
A digital filter is not a microprocessor but an aplication specific IC which is setup to perform a single task.
The effective resolution is increased 1.5 bits not 3.
So far, only the added samples are the result of calculations, the original samples are preserved, and the sample rate is increased by an integer power of two. The next generation of digital interpolation used non-integer sample-rate multipliers and was called UPsampling to differentiate it from all previous incarnations of OVERsampling. The only thing that is different is the non-integer multiplier and the fact that now every sample is the result of a calculation. The output is essentially a digitally synthesized version of the original recording.
Asynchronous sample rate conversion is a variation of upsampling. Instead of a fixed sample rate multiplier applied to the input sample clock, the sample rate multiplier is determined by continuously computing the difference between the input sample clock and another reference or output sample clock.
In my opinion, the evolution of CD playback has gone from bad to worse. Although upsampling improves the quality of steady-state sine waves, it doesn’t improve the quality of recorded music. That conjecture is proven by the renaissance of NOS DACs. If the newest generation of CD playback technology were as good as the promoters say it is, there would be no desire for music lovers to want to revisit the past.
Unfortunately, in their zeal to reject everything that was wrong with prior attempts to reproduce music, NOS proponents have thrown the baby out with the bath water. Yes, the brick-wall filter is bad and so is the digital filter, but a proper reconstruction filter, also called an anti-image filter, is an absolute requirement for any digital to analog converter. While 8X upsampling and extending the 16-bit sample to 24-bits with what is essentially noise doesn’t enhance the music, 24-bit DACs are superior to 16-bit DACs whether or not you use all the bits. Also, there’s nothing wrong with upping the sample rate provided you preserve the energy content of every sample.
The quality of a DAC can be quantified by measuring it's output signal and seeing how close it comes to perfectly representing the information in the recording (linear transfer function, linearity, noise).
The availability of fast DSP and accurate DAC's simply gives you more options in deciding how to decode the available information without the compromises that the early designers were forced to make.
Higher resolution DAC's (24 bit) allow you to use a lower level of dither (or none) when rounding the sample values generated in a digital filter.
Digital filter, Anti-alias/image filter and reconstruction filter are all the same thing you can choose all sorts of filter transfer functions depending upon how you prefer to make your compromises.
In the early days we saw CD players with 12 bit DAC's, no oversampling, no dither and wild deviations in low level linearity but these somehow did not get a mention.
2x, 4x and 8x oversampling Dacs are essentially the same animal except for a faster speed being introduced each time and possibly a more refined filter kernel allowing a more benign analog filter with each faster incarnation despite JB's misguided descriptions.