This study investigates the use of contrastive learning to improve speech emotion recognition. Our method captures temporal features at various resolutions, ensuring the model generates consistent predictions with these features during contrastive learning. We further enhance our approach by integrating contrastive learning directly into the supervised training phase, rather than merely using it for pretraining. This integration allows for delicate adjustments to the neural network weights, steering the model past local optima and enhancing classification performance. Our results indicate that this integrated strategy outperforms current leading techniques.