People's speech data processing?

[TParcollet]

Hello NeMo (I'm from the SpeechBrain community :p)

May I ask two things about the NeMO ASRSET:

1. How did you select the 12k hours of speech from people's speech? Do you have a script somewhere, or the manifest of the selected files and rational.
2. Where did you get the data from? The current HuggingFace repo only contains 6k hours, and training a model on this set results in really shitty performance (I mean, even just training on voxpopuli would give better WERs).

Thanks!
Pinging @erastorgueva-nv maybe :-)