That depends on your covariance matrix in the observer part.
The lower this matrix, the faster the biases converge to the actual biases. On the other hand,
you will "trust" less in your kinematics model, so if your observation (lets say gravity) is not right (because linear acceleration), then
your bias is wrong estimated.
There is a trade off here. One of the "tricks" is to tune this convariance matrix in the observer part. If for correcting the pitch and the roll
is based mainly on the gravity vector. You can assign to that matrix a large value if the norm of the acceleration vector is pretty different from the
norm of the gravity vector. The way of how to do that (what function to pick up) is up to you.