Detecting Level and How It Sounds

Alert readers will recall I’ve been reviewing the fundamentals of compressors over the past few months. This month, I’m going to conclude this review by considering the nature of the level detector itself, and how it affects the sound quality of the compressor. If only it were simple.


The level detector is a circuit that observes the incoming audio signal and derives from it a DC control voltage that will, after manipulation by the Ratio, Makeup Gain, Attack and Release controls, regulate the gain of the voltage-controlled amplifier that is at the heart of the compressor. Exactly how that detector derives such a DC control voltage is an important determinant of how the compressor will regulate gain and how it will sound.

At the heart of such concerns are the concepts of peak, average and RMS levels and their detection. These days, digital compressors often give us a choice between these modalities.

“Peak” detection derives a control voltage from the highest signal peaks encountered. The detector senses the highest audio level and determines the control voltage from that, usually slowly decaying until another peak is encountered. Such detection concentrates on preventing overloads or clipping, by emphasizing quickly the maximum signal levels encountered.

“Average” detection is derived from an average of the changing signal level over some time period, usually between 100 milliseconds and 3 seconds. Such detection of amplitude is fairly slow (depending on the signal itself) and gain regulation is comparatively restrained.

“RMS” (Root Mean Square, get it?) detection is derived from the relative power variations of the signal. Power changes as the square of the amplitude, so RMS detection is more volatile over time, for a given time constant. It is fairly well correlated to how we humans hear, but it is also fairly complex to compute in real time, so it is not used as much as it might be, particularly in low-cost compressors.

Loudness, as I’ve noted in earlier columns, is a subjective sensation, not a physical value. Our perception of loudness varies as a function of amplitude, frequency, frequency bandwidth and time. It is a complex sensation. It defies direct physical measurement.

With compressors we are attempting to manipulate relative loudness and the behavior of the level detector can have a significant effect on such perception of loudness. In the case of peak detection, a squashing of loudness range occurs for any program with regularly occurring peaks, while for programs with only one peak there is little reduction in loudness except for the few moments following that single peak.

What should be clear from this is that peak detection will have a distinct set of behaviors and effect on loudness, which may not be appropriate, but will be distinctive. At the same time, such peak detection does head off overloads and clipping pretty definitively.

API 527 Compressor In the case of both average and RMS detection systems, we are actually taking an average of the level. Often, we integrate that average over time, so that recent events have a greater effect on gain regulation than do not-so-recent events.

But, in any case, such detection is of an average level, NOT a maximum level. This means overshoots remain possible (governed, of course, by the Attack and Release controls).

The difference between average and RMS is that average values respond to relative amplitude while RMS values respond to the relative power of those amplitudes. That power curve is exponential, so we obtain a different range and nature of behaviors.

Which is the correct one? Neither. They sound different—you gotta use your ears to decide what “sounds best” for you in any given application. Dang!


One final complication to all of this is the “crest factor,” which describes the variable relationship between peak and RMS values. Compression and limiting both may change the crest factor of a signal, depending on how their time constants are set. At the same time, crest factor is often surprisingly large. For random noise the crest factor is slightly less than 12 dB, meaning that the measured peak level of a pink noise signal will be almost 12 dB greater than the RMS measured level. For some voice signals or certain types of percussive music, a crest factor of 20 dB is reasonable to expect.

Crest factor is not only audible, but it is also a basic determinant of audio character, which is to say that a program with a large crest factor has a distinctly different sound character and sonic and musical meaning than a program with a small crest factor. When we squash the crest factor with a fast-acting limiter, we change something fundamental about that sonic “meaning” of the signal. Not necessarily good.

Compressors and limiters are complex devices. Their action is quite audible in ways that are often unexpected, and sometimes quite hard to describe. We need them in our work, but we often don’t understand all the sonic implications of what they are doing.

The behavior of the level detector in each given compressor is a big part of this. And beyond the parameters I have already described, designers can add all sorts of things to manipulate the detector. Unfortunately, such designs and the capability of any given level detector are usually not revealed (except in marketing terms such as “hyperacoustic temporal sensing”), so we are limited to guessing at what they are doing, which can be annoying and difficult.

Such confusions require that we use our ears, which is a good thing, because that’s all that our end-users have.

Thanks for listening. And thanks to Eddy Bogh Brixen (Audio Metering, Broadcast Publishing) for some really useful information for this article.

Dave Moulton is trying to limit his weight gain, while making up his portfolio gain and limiting his equity attenuation. You can complain to him about anything, except your portfolio, at his Web site

Dave Moulton