The mismatch between split and split-lazy, specifically counting from 0 or 1

[yfyeung]

I note that lhotse split will generate splits that counting from 1, while lhotse split-lazy will generate splits that counting from 0.
In some icefall recipes like gigaspeech, we initially use lhotse split and then change to use lhotse split-lazy. However, we didn't consider this mismatch, resulting in some bugs.

[yfyeung]

For lhotse split:

lhotse/lhotse/utils.py

Line 336 in 567ba29

def split_sequence(

For lhotse split-lazy:

lhotse/lhotse/utils.py

Line 277 in 567ba29

def split_manifest_lazy(

[pzelasko]

Sorry for that, I think initially I wanted to be compatible with Kaldi's convention of 1-based splits for data directories, and later I didn't remember that when introducing the lazy thing. We should fix that. I probably prefer to adopt 0-based counting everywhere, seems more consistent with the rest of the and library and Python in general. WDYT @yfyeung @csukuangfj @desh2608 @danpovey?

[yfyeung]

Thanks. I agree with adhering to a 0-based counting system.

Best regards.

[desh2608]

Submitting jobs on an SGE cluster requires indices starting at 1, which I suppose is why 1-based indexing is used in Kaldi. I would suggest adding a "start_index" option to those functions, which defaults to 0.

[desh2608]

The above is also why we added the num_digits option in the lazy version. In the original implementation, I think padding was by default, but this made it hard to use the split manifests in array job submissions.

[pzelasko]

Thanks, should be fixed now.