Add frequency change metrics?

[gthompson]

Q: Some signals such as hybrids and lp-rockfalls change frequency (high to low for hybrids, low to high for lp-rockfalls). So can we think of a good way to measure the frequency change during the signal itself (and not during the noise). Perhaps scan the signal in 3-s sliding window and record the highest and lowest number of zero crossings per second, and their times, and then these frequencies, the frequency change and the time between low f and high f (-ve time for hybrid, +ve for lp-rockfall) could be features?

A: She noted a spectrogram already includes this information, so maybe we should just use a spectrogram as a feature? The problem, as Alexis had mentioned, is that waveforms are vastly different in length, and AAA needs a fixed length feature vector. Nevertheless, I could compute what I suggested above from the spectrogram, similar to where I compute other frequency metrics from the spectrogram.