actually, the 'machine' is trained with a voice model. so, when you're saying that you'll let a machine decide, in this particular case you're really letting google decide (and expecting them to have baked in an appropriate pronunciation of gif).

here's what we've got at nuance: GIF or JIF -> http://bit.ly/1938drD.

the google translate tts models are older than their android models (particularly network tts), and it seem as though it still treats it similarly -> http://bit.ly/12tpCol