This paper deals with the problem of Query by Example Music Retrieval (QEMR). Retrieving music pieces that are "similar" to a musical query is crucial when exploring very big music databases. The term "similarity" in this paper is equivalent, for instance, to the rules permitting a human subject to build a list of songs to listen to. While the Query by Example Image Retrieval is becoming a mature domain, the QEMR is still in his infancy. This paper proposes a set of similarity measures aiming at expressing aspects of music similarity. The similarity is based on the distance between statistical distributions of the audio spectrum and it can be applied to the raw audio data with no format restriction. A QEMR algorithm relying on the presented similarity measures is evaluated on a dataset containing more than 4000 music pieces om seven musical genres. The results are encouraging both for subjective and non-subjece judgments