I tried modelling what happens when typing and correcting an error, to find out what kind of improvement can be expected by increasing one's accuracy.

Assumptions:
- Let's assume x is the number of correct characters written per minute ("net cpm"), and p is the accuracy.
- For simplicity's sake error distribution is assumed to be uniformly distributed, so in addition to x correctly typed characters, x/p-x are mistyped per minute.
- One has to delete the wrong char (thus 1 backspace per error, assuming one notices the error after the first wrong char).
Subtotal (per minute) so far: correct chars + wrong chars + backspaces = x*(2/p-1).

- There's a small delay t for noticing an error and getting back into the flow after correcting it. How many characters would have been typed in that time if the error had not occured depends on the gross speed, which is the sum of the number of correct chars, the number of wrong chars, the number of backspaces (thus the subtotal above) plus the number of characters which could have been typed in the pause. The delay occurs for every error, thus the number of delays is x/p-x. The number of characters per delay is t*g/60, where g is the "gross cpm".

With these one can calculate the gross speed, which is: g = (x*(2/p - 1)) / (1 - x*t/60*(1/p - 1))

Result:
In the following the delay t is 0.2s. For speeds of 60 to 100 wpm, and accuracies from 95% to 100%, this is what it looks like (horizontal axis is net wpm, vertical axis is accuracy, color signifies achievable gross wpm).

Examples:
- Let's say someone types at (net) 64 wpm and 95% accuracy, the achievable gross speed is 75 wpm (at 100% accuracy). Maybe the goal is 98% accuracy, which would mean a net 70 wpm can be reached simply by increasing the accuracy (gross wpm stays at 75 wpm).

- I usually type at about 68 wpm and 97% accuracy. By improving my accuracy to 99% I'd have a net speed of 72.5 wpm (at 100% it'd be 75 wpm).

- More examples going from 97% to 99% accuracy:
60 wpm at 97% accuracy would be 63.8 wpm at 99% accuracy.
62 wpm at 97% accuracy would be 65.9 wpm at 99% accuracy.
64 wpm at 97% accuracy would be 68.1 wpm at 99% accuracy.
66 wpm at 97% accuracy would be 70.3 wpm at 99% accuracy.
68 wpm at 97% accuracy would be 72.5 wpm at 99% accuracy.
70 wpm at 97% accuracy would be 74.6 wpm at 99% accuracy.
72 wpm at 97% accuracy would be 76.8 wpm at 99% accuracy.
74 wpm at 97% accuracy would be 79.0 wpm at 99% accuracy.
76 wpm at 97% accuracy would be 81.2 wpm at 99% accuracy.
78 wpm at 97% accuracy would be 83.4 wpm at 99% accuracy.
80 wpm at 97% accuracy would be 85.6 wpm at 99% accuracy.

The pause has a fixed length in this model, which means that the higher the net wpm the more significant is the speed improvement when accuracy is increased. (assuming a longer pause would make the colored bands slant even more to the right and thus enhance this effect).

Anyway, I'm sure this model can be improved, but this should give a rough idea.

Edited to update assumptions, formula, picture and examples.