While there is a multitude of music information retrieval algorithms that have distance functions as their core procedure, comparing the similarity between recordings is a costly procedure. At the same, the recent growth of digital music repositories makes necessary the development of novel time- and memory-efficient algorithms to deal with music data. One particularly interesting idea on the literature is transforming the music data into reduced representations, improving the memory usage and reducing the time necessary to assess the similarity. However, these techniques usually add other issues, such as an expensive preprocessing or a reduced retrieval performance. In this paper, we propose a novel method to summarize a recording in small snippets based on its self-similarity information. Besides, we present a simple way to compare other recordings to these summaries. We demonstrate, in the scenario of cover song identification, that our method is more than one order of magnitude faster than state-of-the-art adversaries, at the same time that the retrieval performance is not affected significantly. Additionally, our method is incremental, which allows the easy and fast update of the database when a new song needs to be inserted into the retrieval system.