This paper, following a series of papers on flow classification, proposes a very practical method for online network flows recognition and classification. It presents an approach based on non-supervised statistical machine learning that could realistically be implemented in real time in high-speed networks with only very limited information. The flow recognition is based only the packet size and directions of the first 5 to 7 packets of an applicative flow. In recognition of the simplicity and the practicality of the approach, this work received the Best Paper Award in the CoNext 2006 conference. RAE_import_type : Conference contribution RAE_uoa_type : Computer Science and Informatics