Why peak-frequency is not a useful measure

Below you can save and/or play an artificial echolocation call by clicking on the play button.

testpulse   Download sound
testpulse   Play sound

If you look carefully at the oscillogram (in Batsound for example) you will notice a strange peak right at the start of the pulse. The pulse itself (sampling frequency: 44.1kHz expanded 10x) sweeps from 69 to 25kHz in 9.3ms. There is no doubt that the loudest part of the pulse is the peak at the start, 0.5 ms after pulse onset. The frequency of the loudest part of the pulse is 64 kHz.

You may now be surprised to find out that the peak frequency of this pulse is 33kHz and not 64kHz! Just try it yourself and measure the peak frequency by means of a power spectrum and you will see.

This simple exercise shows that peak frequency tells us little about the loudest part of the call, as this clearly is 64kHz and not 33kHz. It rather tells us which frequency was present for most of the time. This is of course the frequency of the QCF-part.

You may now think that the term peak frequency is therefore the same as the frequency of the QCF-part. However, even this is incorrect. Just put your cursor exactly in the middle of the most horizontal part of the QCF-part. You will now measure 25kHz and not 33kHz! Why is this? The reason for this is that a power spectrum takes the entire pulse into account, so the more prominent the initial downward sweep, the higher peak frequency will become. Peak frequency is therefore a measure which is highly biased by the frequency-time course of the pulse. Neither does it tell you which frequency is the loudest, nor does it give any information about the frequency-time course. As a measure it is highly biased by (and therefore correlated with) other signal parameters. This is not good. In statistical analyses, descriptive parameters must be uncorrelated. This is why we decided not to use peak frequency in our analyses and tables. We do use minimum or maximum frequency. These you can simply measure by putting your cursor exactly in the middle (loudest part) of the pulse in the spectrographic display. The Matlab scripts available on this site allow one to measure the frequency-time course and also the loudest (instantaneous-) frequency of the pulse, plus the frequency of the QCF-part, which are physically uncorrelated parameters.


Last updated 22.5.2009