Power vs Decibel Averaging

Power and decibel (dB) averaging refer to what type of data goes into magnitude averages. Decibel averaging, sometimes called arithmetic averaging, is a simple average of decibel magnitudes at each frequency. Spatial power averaging is the average of squared linear magnitudes at each frequency with the result converted to decibels. Each has its own strengths and potential weaknesses to keep in mind.

Power averaging would be the typical, and in many cases the required choice for applications such as background noise surveys or evaluating the average power spectrum of sound across a wide area for any other reason. It tends to give more weight to the loudest sounds and when used for single channel signal analysis, where the focus is more on the sound being analyzed than the response of a system reproducing the sound, it can produce a result that “looks like it sounds.” Decibel averaging produces an averaged result where all magnitudes are equally weighted (unless you intentionally give some data more weight, which we will get to in a moment). You might say that it tends to give you more of a “consensus” view for all measurement positions than power averaging.

When evaluating sound system frequency response, power averaging works best if all measurements being averaged are approximately equal in level. Its natural bias toward the highest magnitudes means that if one of measurement in average comes in at a significantly higher level than the others, it will tend to dominate the result and could significantly change the shape of the averaged curve. In a decibel average, the higher-level measurement would simply move the entire averaged curve higher on the graph without affecting its overall shape more than any other contributor.

A common problem with simple decibel averaging is that it gives as much weight to the nulls in comb filters as it does to the lobes. The nulls, being much deeper than the lobes are tall (on a logarithmic magnitude scale), can produce dips in the averaged response that look like a cause for concern but may be largely inaudible to human listeners at any single location – our ears are generally more sensitive to boosts than cuts and the bandwidth of nulls is perceptually much narrower than the lobes, which also tends to make them less audible.

In this case, the natural bias of power averaging toward the highest magnitudes can be helpful as long as the overall levels of all measurements contributing to the average are very similar. This is mainly a concern when averaging spectrum measurement data, where the right answer depends on the purpose of the average and sound level calibration may be a complicating factor in adjusting input levels.

Next: Normalized Power Averaging