We present a general audio coder based on a structural decomposition : the signal is expanded into three features : its harmonic part, the transients and the remaining part (referred as the noise). The rst two of these layers can be very eciently encoded in a wellchosen basis. The noise is by construction modelized as a gaussian (colored) random noise. Furthermore, this decomposition allows a good time-frequency psycoacoustic modeling, as it dircetly provides us with the tonal and nontonal part of the signal.