[Bug] The total number of supervisions decreased after trimming to supervision groups.

[MartinKocour]

Hi, I am trying to trim the cuts to supervision groups by calling:

cuts = cuts.trim_to_supervision_groups(max_pause=2, num_jobs=num_jobs)

While I am getting this error:

AssertionError: The total number of supervisions decreased after trimming to supervision groups.
1104 != 1105
This is likely a bug. Please report it here: https://github.com/lhotse-speech/lhotse/issues/new, and provide the following information:
original cut: MonoCut(...)

(full error message at the end of this issue)

I suspect that the real reason behind this error is wrong conversion from seconds -> num_samples due to float precision on these lines:

Cut.truncate(offset, duration):
    new_start = max(
        add_durations(self.start, offset, sampling_rate=self.sampling_rate), 0
    )
    until = add_durations(
        offset,
        duration if duration is not None else self.duration,
        sampling_rate=self.sampling_rate,
    )
    new_duration = add_durations(until, -offset, sampling_rate=self.sampling_rate)

Which affects the selection of supervisions on these lines:

  if _supervisions_index is None:
      criterion = overlaps if keep_excessive_supervisions else overspans
      new_time_span = TimeSpan(start=0, end=new_duration)
      new_supervisions = (
          segment.with_offset(-offset) for segment in self.supervisions
      )
      supervisions = [
          segment
          for segment in new_supervisions
          if criterion(new_time_span, segment)
      ]

Full error message:

error.log

[pzelasko]

Makes sense... could you make a PR with the fix? Also, can you run lhotse.validate() on the input cut and see if it finds anything wrong with it?

[MartinKocour]

Hi Piotr, I've just checked lhotse.validate returns None, which I assume is OK.