We propose CLASP, a lightweight model which performs recommendations for cold-start users for which only textual user description is provided and no previous interactions are observed. Our architecture learns lightweight latent alignment functions that project textual preference representations into a pretrained collaborative filtering latent space. The CF component can be pretrained on a large set of dense interactions from warm-start users, whilst the alignment maybe trained on a distinct set of users for which textual descriptions are present, allowing an efficient use of heterogeneous information. Extensive experiments on Steam, MovieLens-1M, and Netflix demonstrate that CLASP consistently outperforms retrieval-oriented and encoder-only semantic recommendation baselines while remaining substantially more computationally efficient than LLM-based recommendation methods. Remarkably, this happens despite the fact that some of the baselines have access to item text descriptions, whilst CLASP doesn’t. Lastly, we also provide a dataset which allows a granular study of the variability in user profile generation from small subsets of interactions.