I'm trying to segment (or extract) different calls from a spectrogram. My current method, calculates a threshold value (variance) and then using RMS can detect voiced/unvoiced signals:
$$ Thres = (1 - c) * min(x) + c * max(x) $$
Where $c = 10$
This method works ok, for most signals. However, does not produce high results when using different data-sets. I have read this article (http://www.avisoft.com/classifying%20recordings.pdf) which uses Frequency contour cross-correlation in order to classify the different calls.
I'm wondering whether or not someone could provide an insight into this, or at least provide some research based material to which I can read and hopefully gain some kind of understanding.