This paper reports on the participation of the Information Management System Research Group of the University of Padua to the Music Identi cation task of the MusiCLEF Laboratory in CLEF 2011. The system under evaluation is FALCON, an open-source engine for content-based cover song identi cation written in Java that applies classic techniques derived from textual Information Retrieval to music identi cation. The obtained results show how such approach yields satisfying results using little computational resources.
Automatic identi cation of music documents has become an essential component
of many popular web services. Audio ngerprinting techniques are widely used
in order to identify copies of a recording, which can di er from the original
because of data compression, A/D conversion or environmental noises [
Audio ngerprint approaches aim at the identi cation of a particular performance/recording rather than work : even alternate takes of a composition are considered di erent items instead of being regarded as di erent instances of the same piece. This assumption imposes signi cant restrictions to the range of possible di erences between copies of the same recording, allowing audio ngerprinting techniques to be particularly e cient.
Aside from audio ngerprinting techniques, past research on music identi
cation has focused, for obvious commercial reasons, on popular music; the research
eld is commonly referred to as cover song identi cation, a more general name
for this research eld being version identi cation [
The experience gained in the development of text search engines shows that
in most cases simple and e cient techniques can be generally employed in various
retrieval tasks, with results that are often comparable with more complex and
less e cient approaches. Following this idea, we developed a music identi cation
engine named FALCON2. The fact that our software is implemented on top of
Apache Lucene3, an open source text search engine, substantiates the claim that
the problem of music identi cation can be modelled as a more general retrieval
task. In this paper we study its adaptability to classical music, making use of the
dataset prepared for the Music Identi cation task of the MusiCLEF 2011
benchmarking activity. The collection is constituted by circa 7 thousand recordings for
which the MusiCLEF Laboratory organizers provided precomputed descriptors.
Testing FALCON on this test collection has allowed us to investigate the
scalability of our proposed approach, thus extending previous investigations carried
out on popular music [
Our approach is based on a two-level bag-of-features hierarchy; the input
descriptors (chroma features) are transformed into hashes which are subsequently
grouped into a \set of hash sets" representation for each recording. The
cardinality of set intersection is adopted as a similarity measure. The methodology,
described in detail in [
v S(Q; D) = jQutuj Y max n X min q2Q d2D t2q\d hf(t; d) hf(t; q) o
; jdj jqj (1) 2 http://ims.dei.unipd.it/falcon 3 http://lucene.apache.org/ where q and d denote two segments of a query and a collection recording, respectively of length (number of hashes) jqj; jdj; similarly, jQj; jDj denote the number of distinct segments for Q; D and hf(t; d) denotes the frequency of the hash t in the segment d. Formula 1 can be interpreted as follows: the rst step (inner summation) computes local similarity between segments as the (normalized) number of terms they have in common; the second step aggregates the contributions of all the query segments, by computing the geometric mean of the best local similarities.
The procedure is repeated multiple times according to the most probable keys detected in the audio analysis step; for each recording in the collection, the highest similarity value is preserved. 3
3.1
Test Collection and Setup
The test collection comprises 6679 recordings of classical music works, which add
up to more than 572 hours of music. Of these recordings, 2,671 are associated
to works that are represented at least twice in the data base, forming 945 cover
sets4. The audio descriptors were extracted using the MIRToolbox package [
All experiments were repeated twice, one time using a key- nding algorithm
to preserve independency to transposition, and one time without such strategy.
The other parameters were set according to the values reported in [
Identi cation Accuracy and Computational Load In FALCON, the trade-o between accuracy and speed of retrieval privileges the latter, as the architecture was designed with large collections in mind. It is therefore important to measure the computational resources required by the software. All experiments were run on a machine with a 3.4 GHz dual-core processor (4 logical cores), 24 GB of RAM and a 7200 RPM hard disk.
Table 1 shows the results of our experimentation. As can be seen, the software can index about 3.5 hours of music per second and is able to perform a typical query in under 3 seconds. The average query time is di erent because the keynding algorithm repeats a query multiple times (in this case 3) in parallel; each thread is completely independent of the others. 4 The term \cover set" is commonly used to de ne a set of recordings of the same piece of music. no transposition with transposition total indexing time average query time MAP MRR 163s 1.79s 0.6754 0.7617 166s
This paper reports our contribution to the Music Identi cation task of the MusiCLEF Laboratory in CLEF2011. Our experimental results show how a simple approach based on text retrieval techniques can yield satisfying results using little computational resources.
These results suggest that the methodology implemented in FALCON can be adopted as the rst of a two-step methodology. The identi cation performed by FALCON would aim at providing candidate versions of the query recording at the top k rank positions. Then identi cation could be re ned by means of more sophisticated but also more resource consuming music alignment techniques, that when exploited in isolation and on the entire collection are hardly scalable.