Applications in Intelligent Music Analysis - Intelligent Audio Analysis - page 224

[ 44 ]. The music data considered provide evidence that drum components generally
have short peaks of similar length. As opposed to this, harmonic components tend
to have longer peaks varying more in length.

11.1.1.2 Synthesis

Subsequent to the assignment of components to their 'bags' by automatic classifi-
cation, time signals for each class can be computed as follows [ 44 ]:Per class, the
magnitude spectrograms of the components belonging to that class are added, where
the magnitude spectrogram of a component is the dyadic product of its spectral and
gains vectors.the class spectrograms.Then, a column-wise IDFT is performed on As
alluded above,the phase values from the corresponding columns of the phase matrix
of the original signal are used in this step. Time signals are finally computed by win-
dowing the columns with the square root of the Hanning function using overlap-add.

11.1.2 Performance

For evaluation the set “20 Years on MTV” (1981-2000, Sony/BMG) was used. It
consists of 200 songs, from each of which one data instance of 15-30 s duration is
extracted. With the framework as described in Sect. 11.8 , spectrograms were com-
puted using the square root of the Hanning function with a window size of 60 ms and
50 % window overlap. For subsequent NMF application, 30 components were used.
Out of the 6000 resulting components 344 were manually selected by perceptual
quality. Music experts carried out the labelling attaching either the label “Drum”
(95 components) or “Harmonic” (249 components) to these. Evaluation with linear
kernel SVM of this data is carried out in ten-fold SCV after scaling features in the
range [− 1 , 1 ] . Different feature subsets are considered:

• The “complete” feature set contains all features described above and led to a WA

of 95.9 %.

• The “reduced” feature set as proposed in [ 44 ] includes standard deviation,10

MFCCs, noise-likeness, spectral centroid and roll-off for spectral vectors, and
average peak length, percussiveness, peak fluctuation and periodicity for gains
vectors. It led to a slightly improved WA of 96.2 %.

11.1.3 Summary

In this section a separation of music into drum-beat and the harmonic parts was shown.
Generally judging, the audible results are well usable in, e.g., DJ applications or music
remixing. There are, however, some cases which seem to pose difficulties to the