Ummm...aliasing? What be this critter?

bitflipper


Aliasing is what happens when there isn't sufficient data to properly reconstruct the analog signal, causing the mathematical model to produce the wrong answers. The result is frequencies being generated that were not part of the original audio, and that are particularly unpleasant because they are usually not harmonically related to the original audio either.

The Shannon/Nyquist sampling theorem only works correctly when the sample rate is at least double the highest frequency in the audio. When we record, it is imperative that we remove all frequencies above that limit, which the interface does using a combination of analog and digital filters.

Problems can occur when high frequencies are subsequently generated within the DAW itself. Some processes cause harmonic distortion, with compressors being the #1 culprit. In the analog world this isn't a problem, since very high frequencies are naturally dissipated, but in the digital domain any frequencies generated that fall above the Nyquist frequency are not going to be properly reconstructed back into analog. Wrong frequencies will be created.

One way to mitigate this effect is to use oversampling inside DSP effects. Converting to a higher sample rate raises the Nyquist frequency, which means more of the harmonics fall into the "legal" range and won't be misinterpreted. After the plugin has done its thing, it can filter high frequencies out and downsample back to the project's sample rate.

Except in extreme cases, in practice we often can't hear aliasing. The aliased frequencies are just too quiet to be heard. But they are sometimes subliminally perceivable, causing us to sense there's something not quite right without being able to put our finger on why. To some extent it accounts for why one compressor might be preferred over another.

Although it's often hard to actually hear aliasing, it's fairly easy to measure and observe it using visual aids such as SPAN. This is where some folks get carried away over on KVR. They think if you can measure aliasing it's the same as hearing it, and therefore any plugin that shows measurable aliasing must be inferior. They don't take into account whether the measured errors are actually audible (or might become audible under certain circumstances).

At the moment there is some discussion over on KVR regarding IKM's LA-2A emulation. Somebody measured some aliasing, and a long debate ensued about how much aliasing there was, how much is acceptable, how to properly test for aliasing, blah blah blah. The consensus came down to "it sounds OK to me".

Eventually, all vendors will take the approach of Voxengo and others, which is to offer multiple oversampling options. A plugin can tell if it's in playback versus render mode, so it's easy to offer one sample rate for mixing and a higher rate for exporting/bouncing. That way, you get the CPU efficiency of a lower rate during playback and the precision of a higher rate when it comes time to commit your audio into permanence.

Finally, aliasing can still occur in the playback device despite all our efforts to keep it out of the signal chain during production. A cheap player may cause aliasing, as can very low-bitrate MP3s. Nothing we can do about that except to offer higher-quality MP3s (or FLAC) and hope that the people who listen to our music aren't doing so on iPods or ten-dollar players from the office supply store.

Actually, it would be nice if the oversampling options were at the DAW level.