Detectors of fake speech are prone to poor performance when presented with data unseen during training. Unseen data can include not only new methods for generating fake speech and various transformations applied to fake speech, but also fake speech in new languages. This study investigates two methods for improving generalization capability of detectors that use a self supervised model as a front-end. The first method uses an encoder-decoder architecture to learn representations robust to different types of fake speech. The second method uses a subspace-based decomposition and learns common information present in various types of fake speech. The proposed methods are evaluated on the standard ASVspoof 2021 dataset, as well as on a multilingual speech dataset of Indic languages. Experiments reveal that state-of-the-art detectors struggle with speech in unseen languages, and the proposed generalization methods help in reducing the performance gap.