[Datasets] Improve Covost 2

[patrickvonplaten]

It's currently quite confusing to understand the manual data download instruction of Covost and not very user-friendly.

Currenty the user has to:

1. Go on Common Voice website
2. Find the correct dataset which is not mentioned in the error message
3. Download it
4. Untar it
5. Create a language id folder (why? this folder does not exist in the .tar downloaded file)
6. pass the folder containing the created language id folder

This PR improves this to:

1. Go on Common Voice website
2. Find the correct dataset which is mentioned in the error message
3. Download it
4. Untar it
5. pass the untared folder

Note: This PR is not at all time-critical

[lhoestq]

Cool thanks !

I removed some dummy data files leftovers. I also tried to remove the file you added in the .hypothesis directory but for some reason GitHub doesn't allow me to remove it in the browser.

Feel free to delete this file and merge :)

[shaikmoeed]

I am trying to use load_dataset with the French dataset(common voice corpus 1) which is downloaded from a common voice site and the target language is English (using colab)

Steps I have followed:

1. untar:
!tar xvzf fr.tar -C data_dir

2. load data:
load_dataset('covost2', 'fr_en', data_dir="/content/data_dir")

0 rows are loading as shown below:

Using custom data configuration fr_en-data_dir=%2Fcontent%2Fdata_dir
Reusing dataset covost2 (/root/.cache/huggingface/datasets/covost2/fr_en-data_dir=%2Fcontent%2Fdata_dir/1.0.0/bba950aae1ffa5a14b876b7e09c17b44de2c3cf60e7bd5d459640beffc78e35b)
100%
3/3 [00:00<00:00, 54.98it/s]
DatasetDict({
    train: Dataset({
        features: ['client_id', 'file', 'audio', 'sentence', 'translation', 'id'],
        num_rows: 0
    })
    validation: Dataset({
        features: ['client_id', 'file', 'audio', 'sentence', 'translation', 'id'],
        num_rows: 0
    })
    test: Dataset({
        features: ['client_id', 'file', 'audio', 'sentence', 'translation', 'id'],
        num_rows: 0
    })
})

Can you please provide a sample working example code to load the dataset?

[lhoestq]

Hi ! I think it only works with the subsets of Common Voice Corpus 4, not Common Voice Corpus 1