The only data transformation which wouldn’t be trivial to reverse is encryption. But you still need to trust yourself, the machine, the network, and everything in between to not leak the key or the data.

If you “obscure” the data enough, then it won’t be useful. There’s no solution to your problem as far as I can tell.

Is there a way to obscure speech recording so there is no way to play it and get something intelligible, but still keep it useful for machine learning?

Depends how you define “obscure”, “useful”, etc.

Maybe feed it directly to your machine then trash it? The info will still be vulnerable when in transit or in memory though.