The study of the impact on the marine ecosystem of an offshore wind farm benefits from the knowledge of the underwater noise observed at a single turbine, as the wind speed varies. The calculation of the noise spectral average at a given wind speed requires many recordings, each acquired in a limited time interval: an extremely time-consuming process. This study investigated how to approach the spectral average using only very few noise recordings for each wind speed, leveraging supervised and unsupervised machine learning techniques. Three different prediction methods, based on mean and interpolation, principal component analysis (PCA), and nonnegative matrix factorization, in combination with four techniques for coefficient estimation as the wind varies, are tested. Prediction based on principal component analysis, combined with Gaussian process regression, outperforms other methods in all three case studies considered. The latter, in addition to the problem described above, include the prediction of the noise spectrum: at wind speeds where no noise recordings are available, and using a few recordings acquired at another (nominally identical) wind turbine.