Getting great accuracy requires practice, good equipment, and good enunciation. What I look for after the initial training is whether a person can read about 50 words of text and have it come out with at most 2-3 errors -- around 95% accuracy. If this isn't happening, particularly after correcting the text and re-reading it, then something needs to be changed. Homonyms don't concern me at this point, and incorrect proper names I often ignore.
First, find a paragraph or two of your text that is about 100 words -- that is enough for a fair test, should take about 1 minute to dictate, and it is easy to calculate percentages. 2 errors in about 100 words gives you about 98% accuracy, 5 errors about 95% accuracy, 20 errors gives 80% accuracy. Read the paragraph once. Select the entire paragraph and click the "Playback" button. Read the text silently as it is played back and decide whether each misrecognition was caused by the computer or yourself. Fix the errors you feel were made by the computer using the correction commands. Reread the paragraph and check your accuracy.
If you see roughly 80% accuracy or poorer, even after correcting and rereading the text, there are major problems either with your dictation or with your equipment. Generally this occurs with very poor microphones, incorrect connections, or a gross misunderstanding of how the software works. Get help or start experimenting. If you ignored advice on sound cards, microphones, etc.; it may be time to pay attention -- we aren't just trying to sell you stuff you don't need. Record your sound, play it back and see if it is clear. If not, figure out why not.
If close to 90% accuracy, then it is worth listening to what you say. See if you can figure out why the software puts out the words that it does. 90%-95% accuracy is common with people who do not enunciate their language.
The change may need to come from the person or, it may need to come from the equipment. Occasionally it needs to come from the environment if background noise is interfering with recognition. Sometimes a new enrollment is useful. How can you tell what needs to change?
One can verify the equipment with another person. If one person gets excellent results but another gets poor results, the odds are good that it is the person that needs to change. (Yes, I'd love for the computer to adapt to the person and in a few years I expect this to happen.) That is why I like to test a few minutes of dictation on a machine before a client tries the software -- I know what to expect, and if I don't see it there is no reason for a novice to waste time and get frustrated.
Sometimes, there are intermittent problems that are nasty to diagnose. Loose wiring in the microphone, driver software that causes a 1/2 second of sound to be missed every 3 minutes, environmental sounds that go unnoticed, etc.
After using the product a while, expectations go up. Many users get to around 98% accuracy and then level off. Some get to 99%, and a few even higher. The more narrow your vocabulary, the higher the eventual accuracy. But as a practical matter, once you hit about 98% accuracy, improvements will come slowly. You may get slight improvements with better equipment, but there will be few who will be able to point you to better equipment. You may get slight improvements as you improve your dictation style -- but your style must be good to be getting 98% in the first place. I typically suggest aiming to automate repetitive documentation rather than work on improving the accuracy at this point.