Using lhotse cut simple to merge unsorted recordings.jsonl.gz and supervisions.jsonl.gz causes segment fault

[JinZr]

Adding the --force-eager option circumvents the issue, but maybe throwing a warning would be better.

Best
Jin

[JinZr]

Seems like this happens because Lhotse doesn't support <end-time> token in segments to be -1.

[pzelasko]

Adding the --force-eager option circumvents the issue, but maybe throwing a warning would be better.

Yes, a warning would be useful.. Would you mind creating a PR with it?

Seems like this happens because Lhotse doesn't support token in segments to be -1.

What tokens do you mean? The manifests have duration rather than end.

[JinZr]

Yes, I’ll look into that for a PR.

And for the second question, the load_kaldi_data_dir needs a segments file to exist for generating the supervision manifest. kaldi.py

In the segments file, each item follows the following format: speaker_name wavfilename start_time(in secs) end_time(in secs) channel-id(0 or 1), where the end_time token can be set to -1 to represent "reading to the end of the wavfile".

[pzelasko]

Hmm I didn’t know segments can have -1. A support for that would need to be added in load Kaldi data dir. can you also add this? Thanks!

[JinZr]

I think end_time holding -1 is a rare case extract-segments.cc, I'll make it a PR later, thanks for responding!

[kfmn]

Hi all,

I have the similar situation.
I need to create CutSet from my Kaldi data folder. And I would like to use my precomputed features via feats.scp
My data has segments file, so single recording may contain several supervisions.

I have imported kaldi data folder into {recordings, supervisions, features}.jsonl.gz with lhotse cut import
Then I wanted to obtain CutSet from recordings and supervisions only with lhotse cut simple. On this stage I also had Segmentation Fault caused by inconsistent sorting but --force-eager helped me to overcome this issue.
Then I made lhotse cut trim-to-supervisions to obtain single supervision per cut, and decomposed the CutSet back with lhotse decompose.

After decomposing I added features.jsonl.gz to make CutSet once again. Without --force-eager I have an exception like this:
AssertionError: Mismatched recording_id: Features.recording_id == 00541_inline1_landline, but Recording.id == '00541_inline1_mobile' which seems to be caused by inconsistent sorting too.
With --force-eager all goes well but supervisions in obtained CutSet turn to be empty! What do I do wrongly?

Could you please provide a minimal lhotse recipe on how to prepare CutSet in settings like my?

[pzelasko]

@kfmn you can use features.scp directly in lhotse kaldi import if you provide the frame shift (I think via -f option). You wouldn't need to decompose anything etc, it should just work. LMK how it went.

[kfmn]

@kfmn you can use features.scp directly in lhotse kaldi import if you provide the frame shift (I think via -f option). You wouldn't need to decompose anything etc, it should just work. LMK how it went.

I used --frame-shift (same as -f) option, so it doesn't solve my issue

[pzelasko]

In that case, can you use lhotse cut simple with all three of recordings, supervisions, and features? Then trim-to-supervisions will correctly trim the features as well.

[JinZr]

Update this issue with a pull request. #952 <#952 (comment)>

On Mon, 16 Jan 2023 at 22:17 Piotr Żelasko ***@***.***> wrote: In that case, can you use lhotse cut simple with all three of recordings, supervisions, and features? Then trim-to-supervisions will correctly trim the features as well. — Reply to this email directly, view it on GitHub <#925 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AOON42HKV5NRQF2VU5ZGUY3WSVKBHANCNFSM6AAAAAAS4HFQYE> . You are receiving this because you authored the thread.Message ID: ***@***.***>

-- Best Regards

[desh2608]

Closing via #952.