Aliasing and Oversampling Explained: Why Plugins Sound Harsh

What is aliasing in digital audio?

Aliasing is what happens when a digital system tries to hold a frequency higher than half its sample rate: the frequency cannot be stored, so it reappears as a lower, false frequency. That false tone is the alias, and it has no musical relationship to the note you played.

The rule behind it is the sampling theorem. A signal can be reconstructed exactly from its samples only if it contains nothing at or above half the sampling rate.^[1] That limit is the Nyquist frequency: 22.05 kHz in a 44.1 kHz project, 24 kHz at 48 kHz.

Energy above the limit does not vanish. Julius O. Smith's DSP text at Stanford's CCRMA describes it as folding: a frequency that sits a certain distance above Fs/2 lands the same distance below Fs/2, a mirror image around the Nyquist frequency.^[2] At 44.1 kHz, a 30 kHz component folds to 14.1 kHz, right in the middle of the range your ears care about.

Why recordings rarely alias but plugins do

Converters use an anti-aliasing filter at the input of the analog-to-digital stage,^[9] so a clean recording enters the DAW already band-limited. The trouble starts inside the box, when a plugin generates brand-new high frequencies that were never in the recording. For the basics of project sample rates, see sample rate and bit depth for producers.

Why do saturation and clipping plugins create aliasing?

Saturators, clippers, distortion and fast limiters are nonlinear: they bend the waveform, and bending a waveform adds harmonics above the original note. When those harmonics go past the Nyquist frequency, they fold back as aliases.

Hard clipping is the worst case. As a clipped waveform approaches a square wave it produces strong high-frequency harmonics, mostly odd-numbered ones.^[3] Take a 7 kHz tone, roughly the fizz on a hi-hat or the top of a bright lead, and clip it hard in a 44.1 kHz session. Here is where its odd harmonics end up:

None of those folded tones (9.1, 4.9, 18.9 or 11.2 kHz) is a multiple of 7 kHz, so they clash with the note instead of reinforcing it. On a low 808 the harmonics stay mostly below Nyquist and the effect is small. On hi-hats, vocals with air, and bright synth leads it adds up fast.

What does aliasing sound like?

Aliasing usually sounds like harsh, grainy or metallic noise in the top end that follows the loudness of the part but not its pitch. Once you know the symptoms, it is easy to spot on a solo track.

• Inharmonic ringing Faint tones that sit out of key with the note, most obvious on sustained high notes and pads pushed through distortion.
• Backward-moving tones Play a pitch bend or a rising sweep through a hard clipper. If you hear some tones moving down while the note goes up, those are aliases reflecting off the Nyquist limit.
• Gritty cymbals and sibilance Hi-hats and S sounds turn sandy after heavy saturation or a hot master clipper, even when the level meter looks fine.
• Mix that gets harsher as you add drive If the top end gets tiring faster than the added loudness explains, compare the same settings with oversampling on.

Aliasing is not always a flaw. Ableton describes its Redux effect as producing effects that range from harsh distortion and digital aliasing artifacts to 8-bit grit, on purpose.^[7] If that is the sound you want, see free bit crusher plugins and leave oversampling off.

How does oversampling reduce aliasing?

Oversampling temporarily raises the sample rate inside the plugin, so new harmonics have room to exist without folding, then removes them with a low-pass filter before converting back to your project rate. FabFilter's Pro-MB manual describes it as running the internal process at a sample rate two or four times higher than the host rate.^[6]

1. Upsample
   The plugin inserts extra samples between your originals, for example turning 48 kHz into 192 kHz at 4x. Nothing audible changes yet.
2. Process
   The saturation or clipping runs at the high rate. Harmonics up to the new Nyquist (96 kHz at 4x) are stored correctly instead of folding.
3. Filter
   A steep low-pass filter removes everything above your project's Nyquist frequency, including those ultrasonic harmonics.
4. Downsample
   The plugin drops back to 48 kHz. The harmonics that would have folded into the audible band are already gone.

When should you turn oversampling on?

Turn oversampling on for any plugin that adds harmonics or reacts very quickly to peaks, especially on bright sources. Leave it off on linear processors like reverb and delay, where it only burns CPU.

The manufacturers say as much. FabFilter's Saturn 2 help says High Quality mode drastically reduces aliasing by oversampling the saturation section, and that the Good setting (8x) is more than sufficient in most cases.^[4] For the Pro-L 2 limiter, FabFilter recommends 4x or possibly 8x for normal use, together with true peak limiting.^[5] The Pro-MB manual adds that compressors need it more with fast attack and release or high ratio settings.^[6]

Stock devices follow the same logic. In Ableton Live, Saturator has a Hi-Quality mode that reduces aliasing at a slight CPU cost, and Glue Compressor's Oversampling option processes at twice the sample rate to reduce aliasing and transient harshness.^[7] If you are shopping for a saturator, check for an oversampling switch in the best free saturation plugins before you commit.

• Always worth it Hard clippers, heavy distortion, master bus limiters, saturation on hi-hats, cymbals, vocals and bright leads.
• Often worth it Fast compressors on drums, tape and tube emulations driven hard, EQs boosting near the top of the spectrum.
• Rarely needed Saturation on sub bass and 808s, gentle bus glue, slow compressors.
• Leave off Reverb, delay, chorus, clean EQ cuts, and anything where aliasing is the intended effect.

What does oversampling cost in CPU and latency?

Oversampling multiplies the processing work and adds a small delay from its filters. FabFilter's Pro-MB manual states that the plugin uses more CPU with oversampling and that oversampling introduces a small latency on top of any lookahead.^[6]

The CPU cost climbs steeply at high factors. FabFilter warns that Saturn 2's Superb mode (32x) can raise CPU load quite drastically depending on the number of bands and distortion styles.^[4] For Pro-L 2 it calls 16x and 32x too taxing for real-time use on most systems and suggests them for offline rendering instead. It also notes that the linear-phase filtering can introduce a very slight pre-ring.^[5]

Latency matters most while you record. Your DAW compensates for plugin delay on playback, but a chain of oversampled plugins on a live input adds delay you can hear while singing or playing. The guide to latency and buffer size shows how those milliseconds stack up.

A practical routine

Track with oversampling off or at 2x. Mix at 2x to 4x on the plugins that need it. Before the final bounce, raise the factor on the master clipper or limiter and the heaviest saturators, since offline rendering does not care about real-time CPU limits.

How can you test a plugin for aliasing?

Feed the plugin a pure high sine tone or a sweep, drive it hard, and look at a spectrum analyzer for peaks that are not multiples of the input frequency. Those extra peaks are aliases.

1. Generate a test tone
   Use a test oscillator or a synth with a pure sine at around 5 to 8 kHz. Keep your monitor level low.
2. Add the plugin and drive it
   Push the drive or input until the plugin clearly distorts. Leave oversampling off for the first pass.
3. Watch the analyzer
   Place a spectrum analyzer after the plugin. Harmonics sit at exact multiples of the tone; aliases appear at odd in-between frequencies, often below the tone itself.
4. Sweep the pitch
   Slowly raise the tone. Real harmonics move up with it. Aliases move in unpredictable directions as they reflect off the Nyquist limit.
5. Compare oversampling settings
   Enable 2x, then 4x or 8x. The stray peaks should drop or disappear. Stop at the lowest setting that sounds clean on real material.

For precise measurements, DDMF's Plugindoctor offers harmonic analysis with a sine input of variable strength and frequency, plus a sweep function for fundamental and THD response.^[8] Running the whole project at 96 kHz also pushes Nyquist higher, with trade-offs covered in 48 kHz vs 96 kHz recording.