It is not theoretically necessary if the local unit has sufficient capacity, which it doesn't in most current implementations. Hence why most implementations of voice recognition in consumer hardware does send the clips off to remote servers for processing.

Judge the delay for yourself.