Context and motivation: Conventional Machine Learning (ML) algorithms and Natural Language Processing (NLP) techniques have effectively supported the automatic software requirements classification. The continuous evolution of Deep Learning (DL) has allowed the creation of complex architectures including pre-trained models. The emergence of pre-trained models for Natural Language Processing (NLP) provides promising results in several types of text classification. Question/problem: Most ML/DL approaches on requirements classification show a lack of analysis for requirements written in the Spanish language. Moreover, there has not been much research on pre-trained language models, like fastText and BETO (Bert for the Spanish language), neither in the validation of the generalization of the models. Principal ideas/results: We aim to investigate the classification performance and generalization of fastText and BETO classifiers in comparison with other ML/DL algorithms. The findings show that Shallow ML algorithms outperformed fastText and BETO when training and testing in the same dataset, but BETO outperformed other classifiers on prediction performance in a dataset with different origins. Contribution: Our evaluation provides a quantitative analysis of the classification performance of fastTest and BETO in comparison with ML/DL algorithms, the external validity of trained models on another Spanish dataset, and the translation of the PROMISE NFR dataset in Spanish.