This is more an ML than SP question, but there's overlap and I've worked with this problem so I'll comment.
This is a problem of information. Lengthening a signal = making something out of nothing. Cutting a signal = tossing contents out. When we're talking time lengths that are double and triple each other, that's excessive and cannot be resolved easily. Three approaches come to mind:
Time-invariant features. Something that completely collapses the time axis. An example is wavelet scattering. Simpler, DFT (FFT), but that's terrible features.
Multi-scale models. Mainly, multi-input networks, each input being different time length, then concatenating later.
Resampling: to same sampling rate. A common approach, and the optimal one if the procedure is perfect (unaliased). Upsampling can work close to ideal, unlike downsampling which can be severely lossy.
Resampling: to same length (but different SR). Works well if done right. Segments A and B can both be 36 samples, but B be twice longer in time. Then, feeding
(2, 36) to 1D CNN, which assumes uniformly strided input, may mess up features, more so if there's cross-channel operations along time. It is the norm for images, but images don't care about alignment along a spatial axis, unlike some temporal tasks. This is worth asking about separately, or just testing.
Many tricks can force lengths to be same, but nothing bypasses the problem of information: if lengths become same, some transformation has been done that necessarily changes information contents. This is true even in lossless compression, which preserves synthesis information (recovery) but not analysis information (stability, discriminability, invariance). Hence, if we take a network that "takes care of" lengths for us, e.g. by padding, a close inspection is due.
Another point, task-specific exceptions can be made, e.g. vastly different lengths are okay for telling apart different speakers, but not for predicting what both of them will say at some same instant.
Lastly, a large factor is feature-network synergy: different architectures react differently. For an RNN, left-padding is much preferred to right-padding, while a conv-net doesn't care, for example.