Emotion recognition from electroencephalogram (EEG) signals is limited by scarce data, high inter-subject variability and noisy recordings. We propose CNN-BiGRU-Attention, a compact hybrid model (381,447 parameters) combining multi-scale convolutions, a bidirectional GRU and multi-head attention pooling, preceded by Z-score normalization, ANOVA F-score feature selection and PCA fitted on training data only. A grid search over 108 hyperparameter combinations on Dataset 1 (three classes: positive, negative, neutral) was run for six architectures under identical conditions, and the best model was compared with a reimplemented GANN baseline. On Dataset 1 (two participants, so within-subject evaluation), CNN-BiGRU-Attention reached 97.09 ± 0.52% accuracy in five-fold cross-validation (98.59% on a fixed 80:10:10 split), 4.06 and 4.36 percentage points above GANN, with a train–test gap (1.23%) comparable to GANN’s (1.52%). It had the highest mean accuracy and lowest fold-to-fold variation of the six architectures, the gain over GANN was statistically significant (corrected paired test), whereas differences from the other recurrent models were not. Because Dataset 1 supports only within-subject evaluation, subject-independent generalization was tested with leave-one-subject-out (LOSO) validation on Dataset 2 (20 participants, four valence–arousal classes): the proposed model reached 92.37% accuracy, 6.03 points above GANN, but below its 98.55% fixed-split accuracy on that dataset. An ablation study showed that the convolutional front end, feature selection and batch normalization contribute most. The model needs about 16 M MACs per inference, suggesting potential for real-time brain–computer interfaces; latency was assessed only on a laptop CPU and emulated resource-constrained profiles.