Upsampling for Mastering

Having problems using the greatest Visual Mastering software of the century? Use this area of the Forum to post your technical questions to Earle and Paavo regarding Har-Bal or ask questions regarding how to work on a certain area of the software? Post away!
Post Reply
adamlloyd83
Posts: 29
Joined: Tue Nov 01, 2005 9:36 am
Location: Los Angeles, CA

Upsampling for Mastering

Post by adamlloyd83 »

Hey Guys, been a while, I have a quick question.

These days I find myself upsampling to 96k for mastering, for various reasons. Is there any reason why Har-Bal should be used before upsampling, and not after? I thought I read something about this on the forum a long, long time ago but can't find it now.

Still love the program after all these years :)

peace
HarBal
Site Admin
Posts: 761
Joined: Mon Apr 19, 2004 8:18 pm
Contact:

Re: Upsampling for Mastering

Post by HarBal »

Hi Adam,

To quickly answer your question, use Har-Bal on the lowest sampling rate version. Why? Because this will give the best frequency selectivity in Har-Bal (ie. more spectrum control) due to the FIR length being fixed. Another point I'd make is that if your target format is 44.1kHz then you should be using 88.2kHz and not 96kHz because conversion from 44.1 to 96 is much harder than 44.1 to 88.2 so your much less likely to get degradation (aliasing distortion mostly) from the conversion than with 44.1 to 96. If your target is 48kHz then use 96KHz.

I must admit that I'm slightly curious about your reasoning for up-sampling. I seem a number of explanations for why up-sampling is the bees knees but nothing that has really convinced me of its worth. For the most part, scientifically speaking, such claims have no justification.

I think the most prominent reason I've heard of recently is to avoid the misnamed "digital over". Truth is if it isn't outside the numerical range of the format it is not an over - period. No distortion occurs. Clipping distortion can only occur in reconstruction (ie. conversion from digital to analog and one sampling rate to another) and then only if the processing code / hardware / whatever has been poorly designed. Any design of any merit will have designed into the system sufficient headroom to reconstruct without clipping and those that haven't have got faulty designs.

I've only ever had one experience of hardware not behaving correctly and causing clipping because of the phantom digital over and that's a 15 year old ARCAM CD player given to me by a friend cos it wasn't working. The reconstruction distortion is obvious because any modern (loudness war era) recordings sound very harsh. I verified the clipping with a test CD of full scale square wave played back and viewing the signal on an oscilloscope. All the gibs effect ringing peaks were clipped by the digital reconstruction filter. I did the same test on my now retired Yamaha CDX-750 (which is of the same age) and it passed in flying colours. No clipping.

Digital overs causing a harsh sound is an indication of hardware/firmware/software design fault. It isn't an over, it only becomes one in the badly designed piece of kit, which is why the name is completely wrong and draws attention away from the real cause of the problem.

Actually, in a funny sort of irony the whole digital overs reasoning for upsampling is introducing an environment where clipping can occur in the recording / mixing / mastering process. If you don't follow I'll explain. For a so called digital over to turn into clipping it requires reconstruction / interpolation and interpolation occurs in sample rate conversion so your doing the very thing you shouldn't to avoid the problem. I'm guessing the reason people do it so much is because they use plugins that internally up-sample to 96kHz or something similar and thereby introduce clipping when doing that conversion. I've never understood why plugin designers think upsampling is a great thing to do and here's a perfect reason why it isn't!

Sorry for the rant. I got a bit carried away by it all. In summary, my advice would be to stick to the one sample rate if at all possible, but if you must use a plugin that forces a sample rate change to an internal rate, whatever that may be, let say 96kHz, then you'll need to up sample to that rate first before processing with that plugin to avoid the misnamed digital over.

cheers,


Paavo.
adamlloyd83
Posts: 29
Joined: Tue Nov 01, 2005 9:36 am
Location: Los Angeles, CA

Re: Upsampling for Mastering

Post by adamlloyd83 »

Hey Paavo.

Actually, my reasons for upsampling don't really have anything to do with inter-sample clips/digital overs, which I've found can be present regardless of sampling rate and especially when trying to push something loud. I use the free SSL meter (x-ism I think it's called) to keep tabs on those.

I simply upsample because the recordings seem to sound bigger and more open when I master them at higher sampling rates. I don't really know why. Well, I sort of do, but I just try to use my ears more than try to intellectually justify it. But I'm not talking about plugins that upsample internally, I've never had very great results with that from my experience. I use an izotope resampler and master in Pro Tools in a session at the new sample rate. I seem to get less artifacts from limiters, and again, the recordings sound bigger and seem to have more depth than when I've tried mastering at 44.1k...

I spoke with the dude that programmed the Izotope algorithm a while ago and I believe he said that with most modern resampling software, the samples are multiplied to a value that is divisible by both the source rate and the destination rate, so the whole 88.2 > 44.1 being "better math" is actually not true...

Could you explain this part a little more for me? >>> "...use Har-Bal on the lowest sampling rate version. Why? Because this will give the best frequency selectivity in Har-Bal (ie. more spectrum control) due to the FIR length being fixed." (especially the last part, spectrum control and FIR length) I mean, how is it different from if I'm working with a song that was originally recorded at 96k, and then I use Har-Bal? I feel like when I look at the upsampled version in HB, compared to the version at 44.1, I see more harmonic detail and can make really subtle, detailed tweaks (this part could be my imagination, I don't know). I've found over the years, as I've trained my ears, that the less drastically I use HB, the better the result (but that HB is still an extremely useful tool for me, for tightening and focusing the tone of the track).

Thanks for your time :)
HarBal
Site Admin
Posts: 761
Joined: Mon Apr 19, 2004 8:18 pm
Contact:

Re: Upsampling for Mastering

Post by HarBal »

adamlloyd83 wrote:I simply upsample because the recordings seem to sound bigger and more open when I master them at higher sampling rates. I don't really know why. Well, I sort of do, but I just try to use my ears more than try to intellectually justify it. But I'm not talking about plugins that upsample internally, I've never had very great results with that from my experience. I use an izotope resampler and master in Pro Tools in a session at the new sample rate. I seem to get less artifacts from limiters, and again, the recordings sound bigger and seem to have more depth than when I've tried mastering at 44.1k...

Yes, I can imagine that certain limiter algorithms when pushed hard will sound cleaner at higher sampling rates so I do believe there may be justification for it for that reason, if you exclude the insanity of "loudness wars". At the levels I like recordings being mastered at I doubt you'd hear the difference, but the last time anybody mastered at that level was at least 10 years ago. Here's me hoping things might change!
adamlloyd83 wrote:I spoke with the dude that programmed the Izotope algorithm a while ago and I believe he said that with most modern resampling software, the samples are multiplied to a value that is divisible by both the source rate and the destination rate, so the whole 88.2 > 44.1 being "better math" is actually not true...

I'd say that is definitely not true. The ratio between 88.2 and 96kHz puts much more demand on the interpolation filtering than 88.2 and 44.1kHz. It boils down to the transition band on the filter spec being much narrower than for the 88.2 to 44.1 case. Many of the converters out their have been designed for real time use with low CPU demand in mind and because of that the filter design is often compromised. Maybe the Izotrope algorithm uses direct evaluation of the reconstruction formula so it may not apply to that one, but generally, when considering the converters out there,it does matter, because they are using the upsample by X decimate by Y approach. If you care to look at this site, you can see that many commercial converters actually perform quite poorly:

http://src.infinitewave.ca/

adamlloyd83 wrote:Could you explain this part a little more for me? >>> "...use Har-Bal on the lowest sampling rate version. Why? Because this will give the best frequency selectivity in Har-Bal (ie. more spectrum control) due to the FIR length being fixed." (especially the last part, spectrum control and FIR length) I mean, how is it different from if I'm working with a song that was originally recorded at 96k, and then I use Har-Bal? I feel like when I look at the upsampled version in HB, compared to the version at 44.1, I see more harmonic detail and can make really subtle, detailed tweaks (this part could be my imagination, I don't know). I've found over the years, as I've trained my ears, that the less drastically I use HB, the better the result (but that HB is still an extremely useful tool for me, for tightening and focusing the tone of the track).


The frequency selectivity of an FIR filter (ie. the inverse bandwidth of the narrowest notch or peak it can realize) is determined by the filter length in samples and the sampling rate. Its basically proportional to the filter length divided by the sampling rate. Hence, because the length is fixed, if you increase the sampling rate the selectivity is reduced.

I'm not sure why you can do better at high rates. Maybe the plots look less confusing? In any case, the place where the loss in selectivity comes into play is at the low frequency end of the spectrum. Above a couple of hundred Hz there's more than enough selectivity to deal with this spectrum with equal authority. It's only below 200Hz that the difference in selectivity will show up.

regards,


Paavo.
Post Reply