Also here's the graph of the f0 estimate from the TWM. It's still somewhat flawed (see the jump down at frame 37), and it required using unusual parameters (Kaiser-Bessel beta 2.2 instead of the 1.95 recommended by the study, and only 6 harmonics instead of 11) to avoid instabilities even in relatively trivial scenarios. Actually I think I've finally figured out what's been wrong with the TWM the whole time - I forgot to convert the frequency in bins to the frequency in hertz. I think this was originally intentional, because only realized much later that the error formula is neither linear nor even relative. I haven't test the fix yet, however I'd imagine it should finally solve the problems I've had with TWM.

This graph specifically shows the estimated fundamental frequency for each 256-point frame of an E4 /e/ phoneme.