Only for a large dictionary ... so mostly for natural language web searching. For simply voice control the dataset is going to be tiny.Depends what you're doing. Comparison to sampled reference data can use huge datasets (eg. voice recognition).
I think in the end this is the only reasonable suggestion of why so much memory is needed. For the hugely limited corner case of people who want to do a voice based web search right after dropping out of a game you throw a couple of GB of precious memory under the bus for the common use?
They should be able to get everything done with 512 MB, 1 GB is luxury ... 3.5 GB is retarded.