The Dragon is almost always on, except when I'm proofreading. Because there's a (user) adjustable pause before speaking a command, sometimes this pause, when added to the necessary processing time, takes longer than it would to key an abbreviation, or even type it in full, so I'll use the keyboard. Also, the quality of the audio governs whether I use Dragon. If it's of poor quality, and I have to play and replay to get the correct sense, the Dragon would be a hindrance. I do what is called "shadow speaking" and repeat what I hear into the Dragon, so it only has to recognise my voice ... plus I also have to punctuate on the fly. But poor quality audio I have to listen to more than once and sometimes speak a single word at a time. Whilst the processing time increases with a longer utterance, it's not linear, so the processing time for a 15 second utterence is not three times longer than that of a 5 second utterance. I've sometimes spoken for nearly a minute without stopping, and the Dragon recognised it all with no trouble.
However if (as the job I'm currently doing) it is one person speaking clearly at reasonable speed, I can quite literally step on the gas, and push the replay speed up by 10 or 12% above "normal" and do it in better than real time. Say I can transcribe 12 seconds of speech in 9-10 seconds.
Really whether I use Dragon, keyboard abbreviations or "vanilla typing" (typing everything in full) is a second-by-second decision making process.
Yes, the Dragon does love to eat up the RAM (I've got 8 GB) but my current version of Dragon (11.1) straight out of the box is brilliant, and hasn't slowed things down noticeably.
Joe