Revelation wrote on Thu, 08 June 2006 01:29 |
I am not as smart as you tech guys, but a former engineer from the BBC stated this in regards to recording at 88 and 96 as compared to 44.
"Aside form the changing storage requirements and data through rates, there aer plenty of other changes too.
Metering is significantly more accurate at higher sample rates, as is dynamic processing such as compressors and limiters.
High EQ is more 'analogue like' with higher rates too because the response curves on the HF side don't have to be curtailed due to the brick-wall response filters.
And then there is the anti-alias/reconstruction filtering which, being an octave higher, is significantly less critical, allowing relatively poor designs to sound rather good.
personally, I work at 24/96 most of the time....
hugh
-------------------- Technical Editor, Sound On Sound
Do you guys agree with this?
|
Hello Hue,
Let me elaborate a little on what I sated earlier:
I agree that 44.1KHz may be a bit tight at the high audio frequency range. And I also agree that having an "octave of margin" is helpful. At 88.2 or 96KHz we certainly have a lot of margin, a bit too much from a converter stand point, but I do not want to argue about 10-20KHz, where there are some benefits to a X2 rate - say you want to make a CD (44.1KHz), then 88.2KHz is a "natural" place to be from a down sample stand point, because it calls for a 2:1 rate reduction, which is a SYNCHRONUS sample rate, certainly easier to avoid the SRC problems associated with running asynchronous clocks. So while I would prefer 60-70KHz for the optimum sample rate, I see some wisdom in 88.2 or 96KHz (at least in terms of "backward compatibility to CD 44.1KH rate or 48KHz).
But having an octave or so margin does not call for another octave, then another and another... One MUST ACCEPT the fact that there is such a thing as OPTIMAL SAMPLE RATE. The industry salesmen trying to promote 192KHz, and now some are at 384KHz have no technical basis for it, not even many years AFTER the introduction of such a concept. They sure tried to explain it.
First by the wrong notion that more dots yield a more accurate wave (the analogy was pixels on a picture). That was based on using "street sense", totally contrary to reality.
Then, by claiming all sorts of stuff about narrow impulse response yielding better time location. That argument totally ignored the direct connection between impulse width and bandwidth of the system INCLUDING THE EAR.
Then came the argument about tight analog filters. That argument does not hold for the last dozen or so years because we over sample the AD's front end, and we up sample the DA's back end, so a 3 pole filter at some high frequency yields the needed performance.
Then came an argument about latency, which only addresses a small range of applications, and in fact is more about AD architecture. FIR based sigma delta has a relatively long delays, say 1-2msec at 44.1KHz. The proponents of low latency wished for 500usec which is 6 inches of acoustic distance (mic location, speaker location...)
I could go on and on. Needless to say I did take a technical stand against all the false arguments. First at the AES, chairing a tutorial about AD's, then with my paper "Sampling Theory". I will continue to do so, with a new paper about the trade offs between speed (sample rate) and accuracy (distortions and noise).
One octave up may be a "comfortable margin", or a slight over kill, but I can accept it as reasonable. But folks need to realize that going faster, not only increases the data size, not only requires a lot more signal processing, it also reduces the quality of AD and DA conversion. All things held equal, a 88.2KHz-96KHz is better then the sales and marketing driven 192KH.
The comments about a limiter or compressor are in order. I believe I was the first to point out that some of the complaints regarding “digital sound” are based on the fact that non linear processing and digital do not “go hand in hand” because the aliasing problems. I specifically stated that a non linear polynomial curve y=a0+a1*X+a2*X^2+a3*x^3… requires double the bandwidth for the X^2 term, a triple bandwidth for the X^3 term and so on. So clearly one must watch what they do when dealing with non linear transfer function in digital.
One way is to do some processing in analog (such as final compression, hard limiting, tube sound…). The other way is to ups sample way up, to the point where the high frequency due to non linearity is at very low amplitude. Only then can you “live with” the aliased energy.
Of course the need for a couple of processes is not enough to call for 10MHz sampling, or whatever it takes to accommodate say a hard limiter.
One more time: There is a big difference between 2 cases:
1.Converting say at 96KHz (48KHz bandwidth), then up-sampling to very high rate, such as 4fs, 8fs…1024fs. In this case, we are using a computational process, and the audio bandwidth stays the same.
2. Having an AD or DA convert at that higher rate such as 4fs, 8fs…1024fs. In this case the “audio” bandwidth is at near half the sample rate. Accommodating such a wide signal range is very costly in terms of performance. The reasons for it are both theoretical (tradeoff between bandwidth and accuracy for sigma delta) and practical (many analog circuits tradeoffs come to mind).
So clearly, if one insists on a digital solution to non linear processing, go for case 1 (the up sampling case), not case 2 which trades of nearly everything – distortions, noise, file size, processing requirement and more!
Regards
Dan Lavry
www.lavryengineering.com