Testing the quality of re-sampling on the DT

Gino · September 2, 2019, 2:58pm

Hi Everyone,

I woke up this morning thinking about re-sampling on the DT, and how I’ve avoided it because you theoretically lose audio fidelity. Obviously you lose stereo, but what if mono is fine (or even preferable)?

The first thing I did was sample 33s of Pat Metheny from a YouTube video , and yes I could have started with something more pristine, but as I was testing re-sampling degrading fidelity, I felt it was a good sound source.

After I had the sample, I set up track 1 to play it, and re-sampled it internally, assigning that sample to track 2. I repeated this pattern, being careful to mute previous track until I got the last sample on track 8. The sample on track 8 had been re-sampled internally 7 times.

Here is the result as an A:B test - use your ears to guess which one is the 8th gen sample, and which one is the original:

I won’t reveal which is which until a few of you have had a chance to guess.

Best regards,

Gino

LyingDalai · September 2, 2019, 3:10pm

From 0:10 to 0:15 there’s a clear drop in the treble. Happens again on 0:19.
So I guess resampling doesn’t work so well with maintaining high frequencies, am I right?
Have you made sure there wasn’t any FX nor filter on your tracks?

I got to check and confirm this. Not that I resample 7 times that often anyway.

Anfim · September 2, 2019, 5:03pm

I’m rolling a generation loss/resampling project on the Octatrack right now using the same source material. Been going for almost 30 minutes (in 5-second iterations) and starting to hear noticeable artifacts but I’m surprised it took this long. Did it in 16 bit to match the DT. Quite enjoying this as a meditation, feeling all Alvin Lucier over here. Will post results in a few hours…

tnussb · September 2, 2019, 9:53pm

If there isn’t anything else involved beside some minimal gain differences it’s not really surprising that pure digital re-sampling (which isn’t real re-sampling at all, because there is no D/A+A/D involved) takes quite long until you hear artifacts.

If it would be possible to exactly match the gain staging it wouldn’t degrade at all …

Anfim · September 2, 2019, 5:09pm

Yeah that’s true - I guess the more surprising thing then is that @Gino experienced treble drop after just 8 resampling iterations on the DT

tnussb · September 2, 2019, 5:19pm

Well, the DT performs an auto-normalizing step after (re-)sampling. So you’ll need to take extra care about the gain staging, because you don’t hear that the gain is first reduced internally for processing and then added again during normalization. This way your real bit resolution may drop quickly.

@Gino: which channel have you used to resampled? Only one side or a summing of L+R?

Gino · September 2, 2019, 7:04pm

I used INT L+R for re-sampling, and initialized each track sound with [TRK] + [PLAY].
Then I set the sample in [SRC] and used P-locks on a trig with 1st to set note length, and envelope to INF

knobgoblin · September 2, 2019, 7:54pm

Had me thinking of this too

Anfim · September 2, 2019, 8:24pm

OK, this took a really long time - toward the 3 hour mark I got bored and started shortening the loop to speed things up. At the very end I start adding filters - it handled 1 per track OK but once I put filters on the second FX slot things fell apart very quickly. Gainstaging was tricky - I had the sample gain at +10 and each track LEV at 117, had to periodically tweak LEV to keep things in the right range, so these volume waves are probably the main culprit behind loss of quality. Might run this again overnight now that I have a better handle on gain.

Anyway, here’s a compilation of how the loss progressed - each time you hear a longer cut we’re jumping ahead 30 minutes in time (hundreds of generations). Total time elapsed here is just over 3 hours.

Gino · September 2, 2019, 8:27pm

Maybe I’m weird, but I really enjoyed listening to that!

J0n35y · September 2, 2019, 8:31pm

The three hour version coming up…

Gino · September 2, 2019, 8:38pm

Also in case anyone was curious, it’s the BRIGHTER of the 2 samples that was the 8th gen!

So there’s something that sweetens the highs / high-mids on the DT for each re-sample. Might just be the normalizing algorithm, but it was interesting to me. Not the result I was expecting.

The A:B swaps are about every 5 seconds.
The most interesting A:B swap for me is at 25s, as it’s very hard to hear then the one at 30s confirms the boost of high-mids.

Anfim · September 2, 2019, 8:34pm

It was fun to have on in the background and dip into once in a while! Love that you can automate this kind of nonsense on the OT.

That’s wild that the brighter one of yours is 8th gen, as you said that confirms there’s some sneaky master EQ happening there. Would be cool to hear more iterations if you have the patience to do it.