Shouldn't the voice be personal identifiable information? Just as a picture is.

They could of course store a machine translated transcript instead (assuming no personally identifiable information is uttered (though almost always there is, the customer is expected to give a customer ID or something)).

Also, considering how bad computers are at understanding spoken words the usefulness would be debatable depending on context.

I would imagine most calls that people would want to record would contain PII anyway: account numbers, etc.