Here i'd love to have an opinion about having this kind of recipe into SB from @Adel-Moumen @pplantinga and @mravanelli . It's quite uncommon to have such a big part of the code devoted to preparing a dataset. I could use an official review by someone as well.

I think this sort of recipe is sorely needed for open-source research. NeMo has a similar recipe that is not open-sourced for their ASRset. I'm wondering if this recipe could be developed further to involve a more sophisticated sample filtering by automatically transcribing via multiple ASR systems and taking the samples with low rates of transcription differences -- from what I understand this is a comon technique for large-scale ASR systems these days.

As for the recipe itself, it looks like it repeats a lot of the dataset preparation for datasets we already have. Is there any way we can re-use some of the scripts already available?

Agreed @pplantinga . Let me answer for the recipe part. All the scripts are different as the csv rows and steps / filtering are not the same. There is also some file copying involved. I cannot reuse existing data prep. As you can see in the PR there is also a new TextNormaliser class that I use. The ASR recipe PR will be much easier to review and merge ...

Hey, what should we do about this PR? If I am not mistaken, at some point you were thinking of closing this PR, right?

Yes, but not now.