Only a proposal:
The problem seems to be: having a clear definite point to determine whether you are in one vocal or another for example.
As you said you could define transition zones.
And then take a temporal "barycenter" of those zones as the discrete transition points so to base your calculation of durations.
My point is: You can get some inspiration from Fuzzy Logic, which helps a lot with the need to describe a quality / a feature that is not a discrete value. I think it is exactly the case for phonems.
As you highlighted with your varying frequency curves, and when looking at the IPA diagram, it is clear that the space between vocals is totally continuous, and not discrete:
Same for consonants, I don't have a diagram at hand, but when playing didjeridoo/yidaki, I spend my time exploring the nuances between all sorts of d from the back of the palate, to the d on the front of the palate, then approaching the different t and d we can do on the back of teeth, either at the base of the tooth/gengiva or at the top of the teeth.
Same for k and g. You have all the palette of tongue position, and strenght in the breath, that makes them sound slightly different.
And for ch, f, v, and h, the palette is continous also.
Like you have asked, the difficulty is determining the spectral/frequency features to define one phonem. You are asking for reference about that, I don't have any. Neither have I used Praat.
But you can start arbitrarily fixing them. And decide yourself: when this or that parameter has gone to far away from this value, we are no longer in this phoneme. When this parameter and this parameter is there: we are totally in this phonem.
That's the core of Fuzzy logic: only defining the points "We are full in it", "We are no longer in it".
Taking an ultra simplisitc example from the desciption of Fuzzy Logic with only one dimension / degree of liberty, (here it's temperature) you happen to have something like this:
- 0 : Quality is false
- 1 : Quality is true, fully present
It works really well for multi-dimension continous space: that's namely the strength of this technique indeed.
And I tend to think you could apply that to several parameters for defining phonems.
As said, it's only a proposal. I had worked on a simplistic system which was determining at which moment the sound was a vocal or a consonant. And it was mostly based only on these two qualities:
I was not using formant frequencies like you seem to do (that's what I interpreted from the graph).
Ultimately, if the description of each phonem you are interested in is too tedious, you could think about using Machine Learning.
You would have to provide examples of frequency analysis (not the sheer impulse amplitude data which is too precise, detailing too much, too much local) and the corresponding phonem. That could be a good candidate for what is called "Classification" in machine learning. But you would have to invest in learning some ML classification techniques.