#1203 Update compute_supervisions_frame_mask() to skip negative end indices

[ArthLeu]

When flipping mask digits to 1.0, negative indices should never be used.

[pzelasko]

Right. I triggered the tests, if they're green, LGTM.

[ArthLeu]

It wants some newlines. Can this automatically apply the changes or do I need to manually change them?

[desh2608]

You would need to run black formatting. See development instructions here: https://github.com/lhotse-speech/lhotse#development-installation

[pzelasko]

Please follow these instructions, pre-commit will auto-apply relevant changes, but you'll need to commit them yourself.

https://github.com/lhotse-speech/lhotse#development-installation

[ArthLeu]

Sorry I was busy with something else. I upgraded to Lhotse 1.18, and seems like the Cut.supervision_frame_masks() now uses relative time to determine mask region.

So this means when we need to use_alignment_if_exists, the alignment timestamps should also be relative? I tried absolute alignment timestamps, which worked with 1.17, and does not work with 1.18.

Want to confirm this before making any further change.

Probably related to #1193

[desh2608]

Sorry I was busy with something else. I upgraded to Lhotse 1.18, and seems like the Cut.supervision_frame_masks() now uses relative time to determine mask region.

So this means when we need to use_alignment_if_exists, the alignment timestamps should also be relative? I tried absolute alignment timestamps, which worked with 1.17, and does not work with 1.18.

Want to confirm this before making any further change.

Probably related to #1193

That PR is a bug fix, not a change in behavior. Alignments were always supposed to have absolute start times (w.r.t. the recording), whereas supervisions are supposed to be relative (w.r.t the cut).