Audio Processing Units

Designing and implementing the upcoming APU TrueGain plug-in has involved creating a large number of FIR filters with very tight frequency response requirements. For TrueGain in particular, there are thousands of individual filters which give the plug-in its unique capabilities. There are minimum and linear phase variations of many different sample rates and many different frequency-response contours.

The naive approach of packaging all of these filters would result in massive download and installation sizes. Because plug-ins are typically self-contained, this means all the filters need to be packaged for macOS, Windows, x86, x64, ARM, VST, AU, AAX, Standalone application. Even with some basic optimizations, this could easily result in GBs of downloads.

If you've ever wondered why a mix sounds different at low versus high playback volumes, you're bumping into a century-old psychoacoustic phenomenon: the Fletcher-Munson curves, also known as equal-loudness contours. These curves describe how our ears perceive loudness differently across frequencies, depending on sound pressure level (SPL). While they have evolved over time into more refined ISO standards, the original Fletcher-Munson data remains a seminal contribution to both audio engineering and psychoacoustics.

Audio Units can be tricky to work with as they are a bit notorious for holding on to old versions you thought you had already removed, failing to load anything at all, and providing little information suitable for diagnosing. No doubt millions of users have banged their head against this issue and there is little help to be found other than to keep rebooting, keep refreshing, crossing your fingers that it works.

So with all that said, here is a set of commands which have so far given me the best results:

Although the relatively long window durations of LUFS and RMS are useful for capturing the audio's dynamics accurately, the relatively slow response times can lead to artifacts in certain scenarios. For example, when the audio is first played, the initial loudness measurements will be very low. This is because the window is initially filled with silence, weighing the measurements down. This ultimately leads to a delay in the measurement response, which ultimately can prevent the initial transient from being compressed at all.

In my previous blog post about LUFS, we delved into some details of the LUFS algorithm. One of the key elements of the algorithm is the calculation of the sum of all amplitudes within the audio waveform. For the purpose of this discussion, we'll refer to this sum as the energy of the waveform.

When measuring LUFS, we typically focus on a continuous stream of measurements. This approach allows us to observe the loudness fluctuations over time, providing insights into the audio's dynamic range.

When working with audio, there often arises a fundamental need for an objective way to measure loudness. Historically, RMS (Root Mean Square) has served this role, providing a good first-order approximation. This style of measurement quickly became ubiquitous across audio tools of all sorts. Even though it has limitations, which we'll get into, it is a whole lot better than nothing.

Technology has advanced over time and more accurate forms of measurement have become feasible. One of the leading algorithms for this purpose today is LUFS. Its use is becoming increasingly common, especially given its adoption by many popular streaming platforms.