A technique for the assimilation of spectral wave observations in wave models is presented and tested. The method uses the concept of spectral partitioning to project the entire wave spectrum onto a few essential mean parameters. Model and observed partition parameters are assimilated using an optimal interpolation (OI) technique. After data reduction, obtained by the partitioning, the cost of the assimilation is negligible compared to the cost of the model run itself. Therefore the optimal interpolation of partitions (OI-P) method is a very attractive assimilation technique for operational wave forecasting. The paper focuses on the assimilation of pitch-and-roll buoy spectra in a North Sea version of the WAM wave model. Treatment of the (non-fully two-dimensional) buoy spectra is discussed. Appropriate choices for the OI weight functions are made. The problem of correlating wave partitions in different spectra is addressed, which is essential for obtaining a robust and efficient system. In order to assess the influence of spectral wave observations on the analysis of the sea state, the method is compared to a second scheme, optimal interpolation of integral parameters (OI-I), which can only be used to assimilate observations of significant wave height and mean wave period. First, tests with synthetic data are described, which illustrate advantages of the partitioning method over the OI-I scheme. Also, the inherent limitations of OI are shown in both methods. Experiments with buoy observations for actual North Sea conditions show the benefits of the system, especially when several wave systems are present at the same time.¿ 1997 American Geophysical Union