CutSet.trim_to_supervisions changes the start of all Supervisions to 0

I am trying my luck on the csj corpus. It has a structure of multiple supervision segments corresponding to one long recording. I used the CutSet.trim_to_supervisions function so that one cut has exactly one supervision and one recording.

So far, I managed to compute fbanks for this cutset, but I noticed that the start of all supervisions have been changed to 0 while the start of the cut object itself is inherited correctly from the supervision. This was a problem when I ran validate_manifest.py on my cut_set.

Am I doing something wrong? I tried to find the implementation of trim_to_supervisions, but somehow in cut.py the _trim_to_supervisions_single and CutSet.trim_to_supervisions seem to point to each other for implementation.

code:

cut_set = CutSet.from_manifests(
    recordings=manifest["recordings"],
    supervisions=manifest["supervisions"]
)

cut_set = cut_set.trim_to_supervisions(keep_overlapping=False)

cut_set before trimming:

  {
    "id": "S00M0079-0-0",
    "start": 0,
    "duration": 592.936375,
    "channel": 0,
    "supervisions": [
      {
        "id": "S00M0079_0001",
        "recording_id": "S00M0079",
        "start": 0.621,
        "duration": 5.763,
        "channel": 0,
        "text": "テーマ は です ね あまり 考える 時間 が なく 私 の 私 生活 と いう こと で",
        "language": "Japanese",
        "speaker": "S00M0079",
        "custom": {
      },
      {
        "id": "S00M0079_0002",
        "recording_id": "S00M0079",
        "start": 7.597,
        "duration": 2.665,
        "channel": 0,
        "text": "普段 何 を し てる か と いう の を 話そ う と 思い ます",
        "language": "Japanese",
        "speaker": "S00M0079",
      },
      ...
  }

cut_set after trimming:

  {
    "id": "c9c8596d-032b-4618-8e7e-67785a561d4a",
    "start": 0.621,
    "duration": 5.763,
    "channel": 0,
    "supervisions": [
      {
        "id": "S00M0079_0001",
        "recording_id": "S00M0079",
        "start": 0.0,
        "duration": 5.763,
        "channel": 0,
        "text": "テーマ は です ね あまり 考える 時間 が なく 私 の 私 生活 と いう こと で",
        "language": "Japanese",
        "speaker": "S00M0079",
      }
    ],
    "recording": {
      "id": "S00M0079",
      "sources": [
        {
          "type": "file",
          "channels": [
            0
          ],
          "source": "/mnt/minami_data_server/xx/corpus/csj/WAV/noncore/S00M0079.wav"
        }
      ],
      "sampling_rate": 16000,
      "num_samples": 9486982,
      "duration": 592.936375
    },
    "type": "MonoCut"
  },
  {
    "id": "498e385b-15c0-4138-9eda-44d2db26f6d7",
    "start": 7.597,
    "duration": 2.665,
    "channel": 0,
    "supervisions": [
      {
        "id": "S00M0079_0002",
        "recording_id": "S00M0079",
        "start": 0.0,
        "duration": 2.665,
        "channel": 0,
        "text": "普段 何 を し てる か と いう の を 話そ う と 思い ます",
        "language": "Japanese",
        "speaker": "S00M0079",
      }
    ],
    "recording": {
      "id": "S00M0079",
      "sources": [
        {
          "type": "file",
          "channels": [
            0
          ],
          "source": "/mnt/minami_data_server/xx/corpus/csj/WAV/noncore/S00M0079.wav"
        }
      ],
      "sampling_rate": 16000,
      "num_samples": 9486982,
      "duration": 592.936375
    },
    "type": "MonoCut"
  },

[pzelasko]

For ASR it might be good to keep the check for ends, and also you don’t want to have negative supervision start time – these things are allowed in Lhotse because they indicate that a cut is smaller than a supervision. It makes sense to have data like this for some non-ASR tasks but not for ASR.

For feature extraction, the supervision information is ignored. You will need it to be sane at a later stage in a torch Dataset object that converts metadata into supervision tensors.

[teowenshen]

Great! Thanks for your confirmations. I will close this issue now.