We propose a new data mining method based on second-order hidden Markov models (HMM2) that implements a background model coupled with dedicated a posteriori decoding algorithms to extract DNA heterogeneities. An unsupervised training and a state splitting algorithm specify a HMM2 that observe fixed length sequences (k-mer and k-d-k mer) rather than nucleotides. The training process does not require any a priori knowledge. We tested this data mining method on the Actinomycete genomes (Streptomyces and Mycobacterium) and found many sequences that appear to be parts of the binding sites for transcriptional factors.