This actually doesn't seem to be the case. Take a look at Google Translate's offline voice recognition AND translation - it's really amazing, considering it's all happening on your device.

I forget where it was, but they published something about training a very small very fast neural network that could fit comfortably in the phone's memory. Tricky tricky. :D

Plus the only way to train these things at scale is to upload the recordings once you have some usage.

Worse for battery life than firing up the radio?