Thanks @shucongzhang! I think @mravanelli, @Adel-Moumen and @pplantinga could have a look (or even review) :-)
One last thing to do @shucongzhang, as it's a new recipe, you'd need to add a recipe testing case (see the tests/ folder). You'd need to add a csv file into a folder with a line equivalent to the ones we have for the LibriSpeech test csv. You may also want to write somewhere that the full dataset is 50k hours+ so prettttttty heavy to store :p

Hi all,

I ran the PR, and the scripts are all working! (nice)

I only have concern regarding the splits. Indeed, the current state of the PR forces the user to use the dev/test sets of Libriheavy. Unfortunately, those splits contains audio all located in the large split. If you are only using the small/medium (usually because of storage bottleneck) sets, you can't use this PR. What I would suggest in this case, is:

1. do not force the dev nor test set from libriheavy in the prepare.py
2. Document this behaviour in the README and make the user is aware that he can alternatively use the dev/test of LibriSpeech directly by following our recipes to prepare the data and change the location of valid_csv and test_csv.

Other than that, it's all good to me. I made one change by adding the possibility to modify SpeechBrain audio backend instead of using soundfile directly. We should also add this note in the README saying that we are using Soundfile for performances reasons.