Alignments in supervision segments

We should start supporting alignments in Lhotse. I think that the following scheme is expressive enough:

class AlignmentItem(NamedTuple):
  symbol: str
  start: Seconds
  duration: Seconds

SupervisionSegment(
  id='...',
  start=...,
  duration=...,
  text='...',
  ...
  alignment={
    'type': 'word',
    'items': [
      ['word', 0.0, 0.5],
      ['alignment', 0.5, 0.7],
      ...
    ]
  }
)

The types would be entirely user-dependent, e.g. it could be “phones”, “subwords”, “letters”, etc. I’m not sure if any other fields are needed besides start and duration…

[pzelasko]

Maybe we should mimic Kaldi’s approach to yield consistent setup with other approaches – anyway, it’s your call.

About the times – good question. I suggest being consistent with the reference point of supervision segment’s start, i.e. by default create them to reference the start of the recording, but when we create cuts / call with_offset or trim / etc. we should also modify AlignmentItem’s appropriately. In cuts, supervisions reference point is always the start of the cut (and they can have a negative start to indicate that it started before the cut).

[desh2608]

I believe Kaldi creates segments based on punctuation boundaries (at least in the AMI recipe). I already have an option in the Lhotse recipe for specifying max-pause between words to merge them into one supervision segment, so we could just use that.

Are the AlignmentItem start times w.r.t. the SupervisionSegment start?