Speech recognition software is stymied by word combinations that sound alike (homophones), says columnist David Pogue.