Syllables might require a certain duration to be recognizable by humans. Over part of that time, voiced syllables might have a pitch, or a clear periodicity in waveform. Pitch can be described as having a fundamental frequency or F0 (the reciprocal of the perceived repetition period). Note that there a lots of other answers ont this Q&A site on various ways to measure pitch, and the potential problems related to such.
Pitch can change over time. If you measure pitch at several points in time (say 9 points for example) over the duration where a syllable is voiced, you might find that the measured pitch in fact has changed (or not). With those pitch measurements over time one could interpolate a curve (spline or polynomial regression, etc.) thru those measured points. You might end up making a bunch of different looking curves from different snippets of audio. One could also publish nice looking graphs of those interpolated curves and call the shapes of those curves contours.
Feed a lot of those nice looking pitch/F0 frequency contours (plus a large amount of other hopefully related data) to a machine learning algorithm, and it might be able to infer some possibly interesting correlations between things.