Sample Rate Conversion Versus D/A - A/D Re-Recording

[ADWolpert]

Hey, Guys --

I has always thought that sample rate conversion was something to be avoided -- that, if I had to take a 96k mix and reduce it to 44k, it was considered better to just re-record it -- taking the 96K output of a D/A, bringing the mix into the analog realm, and then redigitizing it at 44k. Is this considered true these days?

I also have heard that it's better to record at 88k because the sample rate conversion down to 44k is simpler and more graceful than it would be going from 96k to 44k.

Anway, I'm interested in the advice of the many experts here. Is it better to do SRC than to bring a mix back into the analog realm for resampling at 44k?

Assuming decent software and converters (I'd be willing to invest in a dedicated sample rate conversion program, like RBrain Pro, and would be using Lynx Aurora converters), which method does less harm?

TIA,

Alan

[ORC]

SRC's we use in mastering today are quite good. Use whatever sample rate you wish. I recieve most pop, rock, etc.... at 24/48 or on analog tape.

[Alexey Lukin]

The process of D/A and A/D involves digital SRC at least twice, because all D/A and A/D nowadays are oversampled. So, a decent digital SRC will have less affect on your signal than D/A+A/D chain, and this is a way to go unless you prefer some specific coloration introduced by your D/A+A/D chain.

[ADWolpert]

Thanks fore the great responses, guys.

So, what consititutes "decent SRC" these days? I have Wavelab 7 -- is that a good choice? Or should I get a dedicated program like RBrain Pro?

Separate question: Is it still fair to say that if I want to convert down to 44k, it's better to record at 88k, rather than 96, because the simpler math will give me better sound quality?

Thanks again,

Alan

[Alexey Lukin]

1. Yes, Wavelab 7 must be good. More details here: SRC Comparisons

2. It does not matter when SRC is good. Most people still do it "for safety".

[Trakworx]

Originally Posted by ADWolpert

Separate question: Is it still fair to say that if I want to convert down to 44k, it's better to record at 88k, rather than 96, because the simpler math will give me better sound quality?

It is my understanding that it doesn't matter because SRC works by up-sampling to an extremely high SR that is a multiple of all the common SRs, then down-sampling to the target SR. So the conversion is done using whole numbers either way. But good quality SRC must be used, as the "SRC Comparisons" page clearly shows.

I'm open to someone correcting me if I am wrong about any of this.

[nms]

It's a bit of a grey area. You could assume conversion between 44 to 96 will be as clean and efficient, but there's an amount of speculation there and all SRC's are not created equally. I prefer 88khz for audio. I just thought of a test I could try tomorrow. SRC a piece of music from 44 to 88 and back then do the same with another copy between 44 and 96k. Then use Diffmaker to null the results against the original and see what comes out on top. Would this work Alexey? I could see how it could go either way.

Oh, and Izotope 64 bit SRC is the best around IMO. Fine work by Alexey.. I use it religiously.

[Alexey Lukin]

A null test may not work for SRC, because SRCs (esp. with fractional ratios) tend to introduce subsample delay in the signal. It is not an artifact, it is just like if you've digitized the same audio signal 10 µs later. If your null-test app is able to align signals with subsample delays (like RX2's Azimuth correction tool), you'd be able to cancel out 2 waveforms to a good degree. Otherwise they may not null, but should sound identical.

[Trakworx]

Originally Posted by Alexey Lukin

...they may not null, but should sound identical.

Bold added. To me this is all that matters. If I can't hear the difference in a blind listening test, then for all intents and purposes there is no difference. Thanks!

P.S. I haven't done this listening test yet, but I will soon and report back.

[nms]

The problem with that can be that maybe you don't catch any difference on one test but with other content or with a bunch of samples or stems resampled up and stacked then resampled down at the end the difference is more obvious. I like the better safe than sorry route so that I don't even need to wonder about it. Having a great SRC makes a difference here though. I'll try the test with Diffmaker as soon as I'm in my studio here shortly as it adjusts for subsample shift.

[Petrus]

Get iZotope cheaply with Sample Manager or Wave Editor (something like $70).

Sox is also very extremely good, and it is free (but command line only).

[nickmoon]

I've really enjoyed Weiss Saracon SRC (software version). If you want to spend less but still good batch procession Wavelab is killer.
I don't think twice about SRC. Good math, fast computers.

[Audiop]

Originally Posted by Petrus

Get iZotope cheaply with Sample Manager or Wave Editor (something like $70).

Sox is also very extremely good, and it is free (but command line only).

Agree. Used SoX on PC but now on Mac I use Sample manager. I think both are basically equally good. I did some blind tests of SM and could not hear a difference between original and down+up-sampled version. Several other listened to my files as well and could not detect a difference.


/Peter

[Trakworx]

Originally Posted by nms

I like the better safe than sorry route so that I don't even need to wonder about it.

I hear you, but do you use plug-ins? Because another thing I read on the internets, besides conflicting info about SRC, is conflicting info about various plug-ins (and converters) performing better at 96 or 48k than at 88.2 or 44.1k. So I have doubts as to which is safer.

Originally Posted by nms

I'll try the test with Diffmaker as soon as I'm in my studio here shortly as it adjusts for subsample shift.

I'd be interested in what you found out. I did a quick test using PT9 Tweak Head SRC. Converted a 44.1 sample to 96 and back to 44.1, then did the same to 88.2 and back. Nulled each against the original and analyzed with PT's Gain AS plug-in. I know this test didn't account for subsample shift, but the 96k nulled to -72.6 dBRMS and the 88.2 nulled to -72.8 dBrms, a 0.2dB difference between 96 and 88.2 SRC. Pretty small difference, but I'm not super-confident that my test measured what I wanted to measure. When I normalized the nulls and listened, the vast majority of the null was at 20kHz (and above? my RTA stops at 20k) and I could barely hear it, so I learned that I'm aging. Anyway, just reporting that, and wondering if anyone has any comment. Thanks,

J~

[GoldMember]

Originally Posted by ADWolpert

Hey, Guys --

I has always thought that sample rate conversion was something to be avoided -- that, if I had to take a 96k mix and reduce it to 44k, it was considered better to just re-record it -- taking the 96K output of a D/A, bringing the mix into the analog realm, and then redigitizing it at 44k. Is this considered true these days?

I also have heard that it's better to record at 88k because the sample rate conversion down to 44k is simpler and more graceful than it would be going from 96k to 44k.

Anway, I'm interested in the advice of the many experts here. Is it better to do SRC than to bring a mix back into the analog realm for resampling at 44k?

Assuming decent software and converters (I'd be willing to invest in a dedicated sample rate conversion program, like RBrain Pro, and would be using Lynx Aurora converters), which method does less harm?

TIA,

Alan

Protools bounce-to-disk is done in realtime...
& playback engine sounds very nice...
so... it`s much better to use a hardware SRC.
the: "What I Hear Is What I Get" approach.

i have a Drawmer M-Clock DMS-1 with 4-SRC.
but here is a review of the DMS-5 with 2x SRC.
Drawmer A2D2 DMS-3Dual Output A/D Converter & M-Clock Plus Review in Mix magazine product review of the Drawmer A2D2 DMS-3Dual Output A/D Converter & M-Clock Plus

i also like Bias Peak,
i dont like the others.

96khz--->DA--->AD-->44.1khz
add`s some noise, THD, & jitter.

SRC is too clean, some people like to add Dither.
A Good Digital SRC would keep more details than DA-->AD method, but some people like the loss of detail.

recording @ 96khz vs. 88.2khz is nonsense,
48khz with a good clock does sound better than 192khz or 96 or 88 with a bad clock.
but 96khz with a good clock...
192khz needs a perfect clock, unless you have Gr!mm Audio cc1 or Antelope Trinity with Atomic Clock, forget about 192khz.
but most converters sound better at 96khz because the aliasing-filter,
so its an: Anti-aliasing vs. Clock issue.
Plugins do sound better at 96khz..."Bias, has a weird issue"
decent ITB mastering, from a 44.1khz or 48khz file is done
in real time, with 2x different hardware SRC boxes.

1x for oversample to 96khz & Master@96khz, the other is to Downsample to 44.1khz...
and hear the final product@44.1khz, while working@96khz..

Mastering OTB, is also done oversampling the 44.1/48khz file to 96khz,
because most converters have an anti-aliasing filter that sound better@96khz.
unless they have an SRC built-in.
thats why recording, Mix & Master@96khz then SRC to 44.1khz sounds better than recording, mixing & master@44.1khz.

[nms]

Originally Posted by Trakworx

another thing I read on the internets, besides conflicting info about SRC, is conflicting info about various plug-ins (and converters) performing better at 96 or 48k than at 88.2 or 44.1k.

The only plugin I'm aware of that would have an audible preference for 96 vs 88 is Nebula because from what I gather it operates at 96k internally. When working at more double the standard sample rate your upper band limit filters and filters on your software have their ceiling raised enough that they can be inaudible, so personally I really don't believe in any difference worth mentioning when it comes to 88 vs 96. In contrast 48k will have an advantage over 44k because at that point the filters and aliasing are very much in the audible range still and can benefit from being pushed up another 4khz.

As for the SRC stuff.. Did lots of tests yesterday and found some great and conclusive stuff that I'm going to put into a thread here today. If attempting nulls for something like this you really do need software like Diffmaker which has capacity for aligning on a subsample level. Without that the results will have a margin for error.