Adding trim_to_supervisions to Cut; new options to include acoustic context when using trim_to_supervisions

[pzelasko]

No description provided.

[pzelasko]

It seems that with some effort, I got it working with a slow-ish, but mostly reasonable processing time. An example of what it does, using one of GigaSpeech recordings that are very long:

60 seconds of the recording after the first supervision starts (there are many supervisions):

Creating 20 second sub-cuts, as many as there are supervisions, each one centered on its corresponding supervision:

And the next two cuts:

Note that this due to keep_overlapping=False, the supervisions that didn't fully fit in the cut were rejected. This makes sure we don't learn bogus alignments in alignment-free training, but still use that speech as the acoustic context. Maybe we can figure out a way to split supervisions based on alignment in these cases in the future, but likely not before the tutorial.

[pzelasko]

... the next step is to actually train snowfall models on some GigaSpeech subset -- using "vanilla" supervisions with padding first, and then "contextual" supervisions. Hopefully it's going to show some improvement...

[danpovey]

Cool!

…

On Wed, Jul 7, 2021 at 12:04 PM Piotr Żelasko ***@***.***> wrote: ... the next step is to actually train snowfall models on some GigaSpeech subset -- using "vanilla" supervisions with padding first, and then "contextual" supervisions. Hopefully it's going to show some improvement... — You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub <#331 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AAZFLO2VSHIQKD6SSFA56XLTWPG57ANCNFSM475FT6TA> .