This dissertation addresses the analysis of performance disparities between subgroups of data within speech recognition models. More specifically, particular focus will be put on how the performance of subgroups evolves during the training of the models themselves. Having considered previous studies that have highlighted the presence of subgroups and discriminatory biases in speech recognition models, this work will also focus on understanding how these disparities are created and propagated during training. For instance, populations characterized by parameters such as gender, accent, speech rate, or age may undergo either a decline or enhancement in the model's performance as it develops. In conclusion, the detailed examination of the evolution of these disparities across diverse datasets and speech recognition models will be conducted to acquire an in-depth understanding of their propagation.