There's nothing wrong with that process using 800 MB of real memory.
On three OS X machines I surveyed: 5 MB of shared, 825 MB private and 865 MB of real memory was typical. The virtual memory ranged between 3.0 and 4.0 GB.
What's important is the overall memory pressure on the system. Here you can see that speech recognition does have one of the larger footprints, but the system was in no way over committed with respect to RAM.
Since the process doesn't even start until you dictate, killing it won't really do much if you continue to use that feature. Since it is run on a per-user basis, logging out and back it will release the memory it uses and you can continue killing it as well, but there'd need to be other issues for this allocation to cause problems on a system.