Datasets:

covost2

like 14

great!
just my personal irrational request: let's use English-Turkish or English-Arabic config as an example?
inspired by https://twitter.com/mervenoyann/status/1623132574976450562?s=20&t=aTp86htYhDm68vI84OhTuA 💔

LGTM! Have we checked that the load_dataset snippet works? i.e. is datasets able to find the CV 4.0 data stored locally by itself? Or do we need to point it to the correct download dir? E.g. as we do for the TIMIT dataset: https://huggingface.co/datasets/timit_asr#dataset-summary

@polinaeterna - Brilliant idea! added your suggested change.
@sanchit-gandhi - It works the same way as TIMIT, I added more details to make it more explicit. We also convey the same in the load script as well: https://huggingface.co/datasets/covost2/blob/main/covost2.py#L89

Okay! there is a problem: https://github.com/Vaibhavs10/scratchpad/blob/main/CoVoST_datasets_download_error.ipynb

The CoVoST2 datasets expect the CV4 TSV whilst the CV4 download is in arrow format.
What do you think is the best path forward @polinaeterna here?

@reach-vb ah I believe one should download original data archives from Common Voice website in this implementation.

Just double checked this works. I've updated the wording in the README @polinaeterna . If you are okay with it then feel free to merge! :))