In this work, automatic segmentation is done using different original representations of music, corresponding to rhythm, chroma and timbre, and by calculating a shortest path through the selfsimilarity calculated from each time/feature representation. By varying the cost of inserting new segments, shorter segments, corresponding to grouping, or longer, corresponding to form, can be recognized. Each segmentation scale quality is analyzed through the use of the mean silhouette value. This permits automatic segmentation on different time scales and it gives indication on the inherent segment sizes in the music analyzed. Different methods are employed to verify the quality of the inherent segment sizes, by comparing them to the literature (grouping, chunks), by comparing them among themselves, and by measuring the strength of the inherent segment sizes.