Oral Session 6 – Similarity
Chair: Roger Dannenberg
ON RHYTHM AND GENERAL MUSIC SIMILARITY
Tim Pohle, Dominik Schnitzer, Markus Schedl, Peter Knees and Gerhard Widmer
Abstract: The contribution of this paper is threefold:
First, we propose modifications to Fluctuation Patterns . The resulting descriptors are evaluated in the task of rhythm similarity computation on the “Ballroom Dancers” collection.Second, we show that by combining these rhythmic descriptors with a timbral component, results for rhythm similarity computation are improved beyond the level obtained when using the rhythm descriptor component alone.Third, we present one “unified” algorithm with fixed parameter set. This algorithm is evaluated on three different music collections. We conclude from these evaluations that the computed similarities reflect relevant aspects both of rhythm similarity and of general music similarity. The performance can be improved by tuning parameters of the “unified” algorithm to the specific task (rhythm similarity / general music similarity) and the specific collection, respectively.
- B&O recommender used OFAI
- Nice results
GROUPING RECORDED MUSIC BY STRUCTURAL SIMILARITY
Juan Pablo Bello
Abstract: This paper introduces a method for the organization of recorded music according to structural similarity. It uses the Normalized Compression Distance (NCD) to measure the pairwise similarity between songs, represented using beat-synchronous self-similarity matrices. The approach is evaluated on its ability to cluster a collection into groups of performances of the same musical work. Tests are aimed at finding the combination of system parameters that improve clustering, and at highlighting the benefits and shortcomings of the proposed method. Results show that structural similarities can be well characterized by this approach, given consistency in beat tracking and overall song structure.
- Normalized Compression Distance (NCD) a universal distance metric.
- Experimental setup – all classical music
A FILTER-AND-REFINE INDEXING METHOD FOR FAST SIMILARITY SEARCH IN MILLIONS OF MUSIC TRACKS
Dominik Schnitzer, Arthur Flexer, Gerhard Widmer
ABSTRACT We present a filter-and-refine method to speed up acous- tic audio similarity queries which use the Kullback-Leibler divergence as similarity measure. The proposed method rescales the divergence and uses a modified FastMap  implementation to accelerate nearest-neighbor queries. The search for similar music pieces is accelerated by a fac- tor of 10−30 compared to a linear scan but still offers high recall values (relative to a linear scan) of 95 − 99%. We show how the proposed method can be used to query several million songs for their acoustic neighbors very fast while producing almost the same results that a linear scan over the whole database would return. We present a work- ing prototype implementation which is able to process sim- ilarity queries on a 2.5 million songs collection in about half a second on a standard CPU.
Notes: Gaussian similarity features can be expensive.