"This American Life" dataset recipe

[flyingleafe]

From here:

"This dataset consists of transcripts for 663 podcasts from the This American Life radio program from 1995 to 2020, covering 637 hours of audio (57.7 minutes per conversation) and an average of 18 unique speakers per conversation.

We hope that this dataset can serve as a new benchmark for the difficult tasks of speech transcription, speaker diarization, and dialog modeling on long, open-domain, multi-speaker conversations."

The website has been updated since the publishing of the transcripts, so I wrote a simple URL scrapper which works with the new website.

I also duplicated the archive from Kaggle to IPFS, so that it can be downloaded automatically without logging in with Kaggle account.

[flyingleafe]

@pzelasko ^ weird that every single CI action is currently successful, but nevertheless Github says some are not...

[pzelasko]

Looks good! Don't worry about codecov, it's just for information purposes, doesn't need to pass.

Before we merge, can you add an entry in docs/corpus.rst? Thanks!

[pzelasko]

Oh, you also need to import the recipe in lhotse/recipes/__init__.py

[flyingleafe]

@pzelasko done

[pzelasko]

Thanks!