Standard error correction codes (ECC) assume good data compression, thus expecting uniform apriori distribution of data. This assumption does not allow the exploitation of actual non-uniform priors, which may exist, to improve the threshold at which ECC decoding fails. This work presents a new scheme that builds a probabilistic model for data and uses this model to enhance ECC decoding performance. The approach is flexible because training is done online and does not assume any specific data type or structure but only the existence of some temporal correlation between codewords. This scheme can be helpful for a large class of systems, such as wireless sensors and autonomous platforms. The method was tested via simulation using standard ECC from the IEEE802.11 standard and in experiments in which an autonomous platform transmitted data to the base station. The results show significant improvement in decoding performance. Additionally, the article explains the nature of the performance gain.