El presente trabajo busca evaluar los modelos basados en Transformers en una tarea de clasificación, utilizando como insumo los espectrogramas de modulación calculados a partir de señales de electroencefalogramas, para este proyecto, se implementó un modelo soportado en la plataforma Tensorflow, al tiempo que se usó una librería para cálculos de los espectrogramas. El modelo fue probado con tres diferentes experimentos basados en recopilación de electroencefalogramas para actividades cerebrales diferentes, las cuales son: elicitación de emociones, cálculos aritméticos, estimulación con luces parpadeantes. El modelo arroja un rendimiento en la exactitud de 42,19%, 66,77% y 52,93% respectivamente. Aunque inicialmente el modelo no refleja grandes resultados, constituye una base para trabajos futuros.The use of electroencephalograms is broadly known in medicine—diagnosis and treatment processes are supported on this technique—diseases such as Alzheimer's and epilepsy also have a wide association with studies based on the analysis of signals from electroencephalograms. However, its use also covers other fields such as: psychology, neuromarketing, bionics, among others. Consequently, regardless of the context, the relevance of the study of techniques oriented to the analysis of electroencephalograms is remarkable, especially if they are supported in a computational environment. On the other hand, Spectral analysis is a starting point for signal processing, even more so when relating with non-periodic signals, such as those originated by electroencephalograms. Modulation spectrograms are a technique for calculating the second-order spectral components of a signal, which makes it possible to access a frequency analysis that would not be easy to perform with the calculation of a conventional spectrogram. The present work evaluates computational models based on Transformers—an evolution of recurrent neural networks—for different classification tasks, using modulation spectrograms calculated from EEG signals as input. For this study, a model supported by the Tensorflow platform was implemented, while a library was used for spectrogram calculations. The model was tested with three different experiments based on the collection of electroencephalograms for different brain activities, which are: elicitation of emotions, arithmetic calculations, stimulation with flashing lights. The model shows an accuracy performance of 42.19%, 66.77% and 52.93% respectively. Although initially the model does not reflect great results, it constitutes a basis for future work.