Many current ICA-algorithms do not estimate the noise level from the data. The ability to estimate the noise is not due to the Bayesian analysis, however. It stems from the generative model used here. Similar models have been used for instance in [2,11]. The treatment in the latter is very similar to ours: a factorial ensemble was fitted to the posterior pdf of the source signals by minimising their Kullback-Leibler information. However, the posterior did not include other parameters than the sources and only the maximum of the posterior pdf was used. As argued in section 2.1, large density does not necessarily imply large mass, and therefore the EM-algorithm used in [11] does not necessarily find a probable model for the data, and can not, in any case, be used for comparing models with different structures.