We propose a semiparametric framework to disentangle the autocovariance structure of high-frequency equity returns. The observed price is modeled as the sum of an efficient component, evolving as a nonparametric continuous-time Itˆo-semimartingale, and a market microstructure component governed by a discrete-time moving-average process. Our estimation strategy is based on a quasi-likelihood approach that employs a misspecified moving-average model selected via information criteria. We establish a pointwise central limit theorem for the resulting autocovariance estimator, as well as uniform consistency over a broad class of models allowing for arbitrary noise magnitude and flexible dependence structures. Simulation results demonstrate that our estimator outperforms leading nonparametric alternatives, particularly in regimes with small noise magnitude, and highlight that the performance of all estimators is critically driven by the level of noise. Applying our methodology, alongside existing nonparametric approaches, to S&P 1500 constituents from 1996 to 2024, we find that noise magnitude is persistently small throughout the sample. This empirical finding has important implications for the measurement of noise magnitude, noise persistence, and volatility.