[WIP] spokewoz recipe

[oplatek]

Hi @desh2608,

I created a recipe for https://spokenwoz.github.io/SpokenWOZ-github.io/

I am just testing it out. The download and prepare functions seem to work, and I prepared RecordingSet and SupervisionSet` manifests.

However, I was surprised that the following code returns MultiCut[s].

dev_recs = RecordingSet.from_file(manifest_dir / "spokenwoz_recordings_dev.jsonl.gz")
dev_sups = SupervisionSet.from_file(manifest_dir / "spokenwoz_supervisions_dev.jsonl.gz")
cuts_dev = CutSet.from_manifests(dev_recs, dev_sups)

Is MultiCut a good option for SpokenWOZ, where the system and user side of task-oriented dialogue are separated into two channels?

Personally, I do NOT want to tie together "system" and "user" turns together.
I just want to be able to traverse the whole conversations, which is possible thanks to SupervisionSegment naming where id is f"conversationID-{turn_number:03d}".

The practical aspect is that cuts_dev[0].plot_audio() takes ages for multicut of 183s. It took 10+min, before I restarted the jupyter kernel.

Thank you for any tips.

[desh2608]

I am not familiar with this dataset, but I suppose the recordings contain 2 channels, which is why the resulting cut is a MultiCut. If the SupervisionSegments contain the appropriate channel information, you would eventually get MonoCut objects when you run cut.trim_to_supervisions(keep_overlapping=False, keep_all_channels=False). You can then work with these resulting MonoCuts instead of the original MultiCut.

[desh2608]

Remember to also add this to docs/corpus.rst

[pzelasko]

In addition to what Desh stated: plot/play_audio does not actually support multi-channel data (yet). The reason it was slow for you is because matplotlib plot received a 2-d array (num_channels, num_samples) but it expected a 1-d array (num_samples,) (this is not the right way to use 2-d input in matplotlib). We should fix that at some point.