I don't think it's the same purpose. YouTube, TV and Movies offers enough speech samples and a lot of content is dubbed to other languages, and alot of this content already has the transcripts available.

They know who's calling, and the greeting was something like "Hello again". They are catching up at building a competitive database of persons and their preferences at the scale of FAANG. They're moving over from collecting info for their models to collecting info from their users for their agents. This is what they need to offer good agents.

But I might be wrong and it's just phoneme collection, as you speculate.