Piotr Żelasko

`+` like we talked offline, we have to work around global_rank being set incorrectly before `trainer.predict()` is called - the dataloader has to be initialized with the correct global_rank (and...

I couldn't reproduce your issue with the following snippet: ```python from lhotse import * cuts = CutSet.from_file("libri-train-5.jsonl.gz") cuts = CutSet.mux( cuts, cuts.perturb_speed(0.9), cuts.perturb_speed(1.1), ) cuts = CutSet.mux( cuts, cuts.narrowband(codec="mulaw", restore_orig_sr=True),...

Which NeMo version is this and at which exact spot did you insert the mux with narrowband call? It'll help me nail down what's going on.

Your manifest seems to be wrong, you say: > The total length of the sound file: 48.92 s but the Recording manifest inside MonoCut says: > recording=Recording(id='shorts-20230702.wav_ln-25.0', ..., sampling_rate=16000, num_samples=79168,...

MonoCut seems OK but the Recording inside has incorrect metadata. You can check with `lhotse.qa.validate(cut.recording, load_data=True)`

Set the actual duration and num samples of the audio file in the Recording object. Cut is a slice, so the recording must always have full audio metadata.

There’s quick-path code optimization when Cut’s start is 0 and duration == Recording.duration that loads the full audio without slicing. Since your metadata is wrong it errors out. When start...

It’s not a bug in Lhotse. I don’t know how you created the Recording object but the issue is likely there somewhere.

Can you show the Nemo manifest line for that recording?

I think we were resolving an issue in NeMo for parsing manifests with offsets a while ago; can you try this with the latest NeMo version (or using the unreleased...