Existing deterministic+stochastic spectral models assume that the sounds are with low noise levels. The stochastic part of the sound is generally estimated by subtraction of the deterministic part: It is assumed to be the residual. Inevitable errors in the estimation of the parameters of the deterministic part result in errors – often worse – in the estimation of the stochastic part. We propose a new method that avoids these errors. Our method analyzes the stochastic part without any prior knowledge of the deterministic part. It relies on the study of the distribution of the amplitude values in successive short-time spectra. Computations of the statistical moments or the maximum likelihood lead to an estimation of the noise power density. Experimentations on synthetic or natural sounds show that this method is promising.