The aim is to estimate the integrated volatility of a process observed discretely, in the setting of high frequency data, and when there is a microstructure noise. We use a kind of pre-averaging approach, which is rate-optimal when the noise is i.i.d., and may probably be even variance-optimal for a good choice of the kernel involved. However, the main innovative aspect is that it accommodates other types of noise, and in particular the case where the observations are rounded values of the underlying process plus an additive noise.