The one-hot 0/1 encoding method is the most popularized encoding method of class labels for classification tasks. Despite its simplicity and popularity, it comes with limitations and weaknesses, like failing to capture the inherent uncertainty in data labels, and making classifiers more prone to overfitting. In this paper, these shortcomings are tackled with a framework for learning soft label embeddings. Two variants are proposed: first, a learnable general-class embedding which aims to capture information regarding inter-class similarities, and second, a neural architecture which can be added to any neural classifier and aims to learn inter-instance similarities. The inherent uncertainty in data labels is thus somewhat alleviated, allowing the network to focus on incorrectly classified samples, instead of difficult but correctly classified ones. The experimental study on multiple classification benchmarks of increasing difficulty, using neural networks of varying depth and width, show that the proposed method leads to better classification accuracy, highlighting its ability to generalize to unseen samples. •One-hot label encoding leads to classifier overfitting soft labels alleviate problem.•Learnable label embedding framework captures similarities between samples and classes.•Experiments on network of various depths and multiple datasets showcase effectiveness.