Source separation consists in processing multi-channel signals in order to extract from them (hypothetical) underlying elementary components (a typical application domain is the separation of sound sources using several microphones). Blind source separation (BSS) is the art of doing so by resorting only to statistical properties of the sources (mutual independence, non stationarity, sparseness, ...), i.e. without modeling propagation, transduction, etc. This talk presents the statistical ideas behind BSS and shows how simple source models, if properly exploited, are sufficient for "blind processing".