Applications in Intelligent Music Analysis - Intelligent Audio Analysis - page 218

[ 10 , 29 ] are presented. Then, the structure of music is analysed targeting chorus
sections [ 30 , 31 ]. Such extracted information is then used in the assessment of mood
in music [ 8 , 32 , 33 ]. Finally, in analogy to speech analysis, it is attempted to assess
traits of singers [ 34 - 36 ].
Other examples could have been chosen, such as the full transcription of music on
the note event level [ 37 ], the recognition of genre [ 38 , 39 ], music spotting in audio
streams [ 40 ], query by humming [ 41 , 42 ], or the recognition of vibrato singing [ 43 ],
and of course the querying per se [ 2 , 3 ], to name a few. However, the chosen examples
provide a good overview on core topics and allow for a general understanding of the
principles and methods involved. Each topic is addressed by selected exemplary data
with according test results to provide the reader with a feeling for obtainable state-
of-the-art performances under realistic conditions as were outlined in Sect. 11.1 .

11.1 Drum-Beat Separation

For the analysis of music, we will first see how the harmonic section and the drum
beat in Rock, Pop, or similar music can be separated. This was first shown in [ 21 ].
Non-negative Matrix Factorisation (NMF) is known for its suitability in BASS of
drums and melodic parts of music recordings [ 44 - 46 ]. An isolation of these parts can
serve as enhancement in manifold MIR tasks such as the ones to follow including
automatic onset, metre, tempo detection or key and chord labelling and even the
recognition of singer traits. Let us thus consider in this section the combination of an
NMF based blind music separation into several isolated audio tracks with subsequent
classification of obtained isolated NMF components to label them as either rhythmic
or melodic.
In [ 47 ] drum beat separation based on ICA was introduced. Opposed to this, in
[ 48 ] it is relied on NMF for separation of sources within transcription of polyphonic
music. Remarkable results were reported on piano music. Also, the work in [ 44 ]is
based on NMF. There, a feature extraction and subsequent classification is already
used. The authors report promising results for the separation of drum beats in popular
music. Such an approach was later proven beneficial for drum transcription [ 45 , 46 ]
and vocal separation [ 49 ].

11.1.1 Methodology

Let us first take a look at different cost functions and parameters. As we remember
from Sect. 11.8 , given a matrix V

n × m

∈ R and a constant r ∈ N , NMF computes two

≥ 0

n × r
≥

r × m
≥

matrices W ∈ R and H ∈ R , such that

0 0

≈ .

V W H (11.1)