i would build a combination of different methods and then introduce a priority list how they are combined, i.e. 1. look for quick changes in power of the whole signal, if you dont find that so easily, raise the influence of 2. quick changes in timbre/ power of the HF content, if you still cant find anything, 3. raise the influence of found changes of the analysed pitch ... every paraaameter which doesnt work out of the box should be made available as GUI sliders for the user / later use ...

My understanding is that spectral flux is the best way to go. This gets into details: https://www.badlogicgames.com/wordpress/?p=161