Function to drop alignments from cut

[desh2608]

This is just a helper function to drop alignments if they are not required. I have been using it in my multi-talker dataset class to reduce size of the returned batch. Might be useful until such time as we decide to store alignments in a separate archive, like features.

[pzelasko]

Might be useful until such time as we decide to store alignments in a separate archive, like features.

This should be possible to do with Shar using sth like:

# Writing
from lhotse.shar import JsonlShardWriter

cuts = CutSet.from_file(...)
with JsonlShardWriter("alis.jsonl.gz", shard_size=None) as w:
  for cut in cuts:
    ali = get_ali(cut)
    if ali is not None:
      w.write({"cut_id": cut.id, "word_alignment": [item.serialize() for item in ali["word"]]})
    else:
      w.write_placeholder(cut.id)

# Reading
cuts = CutSet.from_shar({"cuts": "path/to/cuts.jsonl.gz", "word_alignment": "alis.jsonl.gz"})
for cut in cuts:
  ali = cut.word_alignment

It's a bit raw, because currently we can't attach the alignment to the supervision with Shar, it's instead attached to the cut. Either:

• we'd support attaching shar fields to supervision
• we'd rewrite Lhotse a bit to support alignments being a property on cut rather than supervision (which would probably cause problems for users so naturally I don't like it; also not sure if it makes sense in the overall design)

[pzelasko]

Anyway merging, this is useful :)