This paper presents a statistical model aiming at quantitatively evaluate the spectral similarity between two sounds. The measurement of similarity is a central issue in the field of Music Information Retrieval as several popular applications rely on comparisons between sound objects as for instance musical sequence seeking in a big database or automatic transcription. To take musicological considerations into account, the measure is intended to be invariant to pitch shifting and to amplitude scaling. The main idea of the method is to compare a target spectrum to a reference spectrum using the reference to drive a statistical model, the target being an outcome of the model. The likelihood of the target spectrum is then derived in order to measure the similarity between both spectra. To be able to compare sounds of unequal intensity and pitch, the reference spectrum is made tunable in term of transposition and rescaling. Transposition and scaling parameters maximizing the likelihood are selected and the values are kept to compute the similarity measure. Thanks to a joint model, the measure is then made symmetrical. The measure is used to assess the similarity between two simple sounds (i.e. single isolated notes). Experimental results illustrate the usefulness of the approach: Applications of the method to classification and multipitch estimation are presented.