Ah yes, voice. The ones I used were done from the network, I wonder how much would be required locally.Depends what you're doing. Comparison to sampled reference data can use huge datasets (eg. voice recognition). It's certainly possible Sony have 1GB of reference data for people tracking.