Samplers Audio Quality Comparison, Resampling/Aliasing Test

[Logga]

I tested a bunch of virtual sample players (hereinafter samplers) to see how well they perform at resampling and avoiding aliasing.
I used the latest available versions as of September 2023.

The test:
I loaded a 44.1k frequency sweep file into each sampler and rendered it at its base pitch, but with the DAW/sampler running at 48k.
For consistency the rendered files were downmixed to mono and normalized to -6dB peak, if needed.
Some samplers have quality settings. For those, there are two results: (1) with the default/playback/realtime setting and (2) with the highest quality setting.

This is the test audio file's spectrogram, which is also how the absolute ideal sampler output would look:



And here are the spectrograms for all the tested samplers:


Ableton Sampler - (1) Normal Interpolation & Ableton Simpler


Ableton Sampler - (2) Best Interpolation


ADSR Drum Machine


Apisonic Labs Speedrum


Bitwig Sampler


Decent Sampler


Decomposer Sitala


discoDSP Bliss - (1) No OS, Normal Interpolation


discoDSP Bliss - (2) No OS, Extreme Interpolation


FL Studio Sampler - (1) Content Resample


FL Studio Sampler - (2) 512-point sinc


NewSonicArts Vice


NI Kontakt - (1) Standard & NI Battery


NI Kontakt - (2) Perfect


OneSmallClue Poise


Plogue Sforzando


Reaper ReaSamplomatic5000 - (1) Sinc Interpolation 64pt


Reaper ReaSamplomatic5000 - (2) r8brain free


Renoise Redux


sfizz - (1) No OS, Polynomial


sfizz - (2) 8x OS, Sinc 72


Steinberg Halion & Cubase Sampler - (1) Standard


Steinberg Halion & Cubase Sampler - (2) Extreme


Studio One ImpactXT & SampleOne


TAL-Drum


TAL-Sampler - clean



All audio files and images can be downloaded from this direct link (20MB zip).


Some comments:
Many samplers perform poorly, with very audible aliasing. There's plenty of room for improvement.
Ideal performance needs more latency for effective filtering. But 'good enough' performance doesn't need much; Reaper's 64pt sinc resampling adds just 38 samples of latency, for example, which is arguably imperceptible (that's less than 1ms even at 44.1k).
It would be interesting to know the latency for all samplers and their various modes, but I couldn't find a reliable way to measure that.
Other kinds of tests, probably better and more in-depth, are possible, but I tried to keep it simple. Here are some older sampler tests with pitch shifting: https://www.discodsp.com/bliss/aliasing/
Feel free to post similar tests or to test samplers not listed here in the same way. And let me know if you find any mistakes (I don't use most of these samplers, so maybe I missed some settings).

[imrae]

Very interesting results, thanks for putting this together!

[Logga]

I want to add that using an even multiple samplerate can indeed help. Often that produces a cleaner result, with imaging in the ultrasonic range, but no audible aliasing. At least if it's played at the default pitch.
Here's our 44.1k test sample played at 88.2k in Speedrum:


But shifted by a semitone, it's not so good anymore:



Using higher sample rates can also reduce audible aliasing, because more artifacts are pushed up into the ultrasonic range.

But the best solution is to just have higher quality resampling in samplers and DAWs. I think a lot of us wouldn't mind a small increase in CPU usage or a tiny amount of latency.

[enroe]

Thank you for this test - and all your effort you put in here.

For all newbies, it should be mentioned: Even the worst picture here causes
aliasing, which in practice is in most cases not audible at all.

[fladd]

I would like to replicate this, but I am having already difficulties to replicate the initial correct spectrogram (i.e. just rendered out of REAPER as audio on a track). My line is not as defined as yours, but much more blurry, with frequencies close to the line being activated, too.



What am I doing wrong?

[Logga]

It's because you used the default (Hann) 'window function' for the spectrograms, whereas I used Kaiser.
Windows batch file for my SoX spectrograms:

Code: Select all

FOR %%A IN (%*) DO "c:\whatever\sox.exe" %%A -n remix 1 spectrogram -t "%%~nA" -x 506 -y 362 -z 120 -w Kaiser -o "%%~nA.png"

[audiojunkie]

I may be misunderstanding, but I don't believe "Clean" is TAL-Sampler's best mode. I believe the highest mode in Tal-Sampler is S1000. Clean is a subset of the variable-mode. It always uses the sample samplerate.

[Logga]

Yes, the S1000 mode seems better (it sounds better too).

[wangeroge]

Nice! Can you post a ranking from best to worst sampler that a five year old could understand?

[werp]

Nice! Can you post a ranking from best to worst sampler that a five year old could understand?

Votes this comment up, I have no idea what those images are saying.
Just gimmee a list of the best samplers in order.

[koalaboy]

What is 'best' defined as here ?

[BertKoor]

Well, I guess some of us have never studied a spectrogram. It has time on the horizontal axis and audio frequency content on the vertical axis.

So this is what an ideal sine frequency sweep looks like:

It is a single straight line, so the audio contains just one frequency at any given point in time.

A parallel line going up is harmonic distortion, of which this is a good example where a frequency of 4 kHz higher is quite strongly present after 2 seconds at 10 kHz:


Any aliasing will show as a line going down instead of up, of which it shows a little bit at the low frequencies (where it also bounces back upwards) but very strong at the high frequencies.

I leave it up to you to interpret the other images with the gathered knowledge

[DrGonzo]

Aliasing is the new analog warmth.

(logga - thanks for a truly awesome post)