>>/9167/
> You would imagine they have catalogued the data and can pull out 100s of years of at least averagely famous artists songs in decent quality.
The higher the bitrate, the higher the operational costs and the training time becomes. As per some chart I found, for a 60 minute audio file these are the (guesstimation) uncompressed file sizes. Now imagine how many hours of audio you would have in those "100s of years" of songs. I really, really, really doubt that they would have anything even remotely close to "decent" quality, much less lossless. For instrumentals and backing tracks it would be much easier to train a model to learn since you can simplify it down to midi files (or similar) with basically just unicode representing the notes and timings for the instruments but vocals cannot be simplified like that, you'd need the actual audio recording.
I've talked about it before but every AI voice in songs has that speaking-into-a-fan texture that I am not certain is because of being compressed and reduced in bit rate to save on storage space and operational costs