This paper addresses the modeling of the source signal of a flute sound obtained by «removing» the contribution of the resonator. The resulting sound has then a more regular spectral behavior and can be modeled using signal models. The decomposition of the source signal into a deterministic and a stochastic part has been made using adaptive filtering. The deterministic part can then be modeled by non-linear synthesis models, the parameters of which are obtained using perceptive criteria. Linear filtering are used to model the stochastic part of the source signal.