Machine Learning to Identify Neural Correlates of Music and Emotions - Guide to Brain-Computer Music Interfacing - page 89

Fig. 5.1 Taxonomy of
feature types used to describe
neurological datasets

5.4 Measuring Music

Musical properties and structure may be described by a number of models or
methodologies. However, many of these methods rely on speci c descriptions of

musical pieces, which are often grounded in speci c musical styles or cultures. For

example, descriptions of chord structures in a piece of music stem from European
musical history (Christensen 2002 ) and may not apply equally well to music from
other cultural backgrounds.
An alternative view of music is to treat it as a complex time-varying set of
sounds. From this view, music is merely a label one may apply to a set of complex
sounds with speci

c structural properties. Thus, one may take acoustic properties of
a recording of a piece of music and use them as alternate descriptors of the music.
The advantage of such an approach is that it allows one to describe all sounds in
the same manner. Thus, music from any cultural background, genre, or style may all
be described in the same manner and via the same framework of features. In
addition, non-musical sounds such as speech, environmental noise, animal cries,
etc., may also be described under the same framework.
A very large number of feature types may be extracted from a piece of sound
(Mitrovic et al. 2010 ). These may be broadly grouped into six types: temporal-
domain features, frequency-domain features, cepstral features, modulation fre-
quency-domain features, eigen-domain features, and phase space features.
Temporal- and frequency-domain audio features are analogous to EEG features.
Cepstral-domain features are heavily used in speech analysis (Liu and Wan 2001 )
and attempt to capture timbre and pitch information by taking frequency-smoothed
representations of the log magnitude spectrum of the signal. Modulation frequency
features capture low-frequency modulation information; sounds induce different
hearing sensations in human hearing (e.g. rhythm) (Tzanetakis and Cook 2002 ).
Eigen-domain features describe long-term information in the audio signal, such as
statistical markers of noise versus structured sound (Mitrovic et al. 2010 ). Finally,
phase space features attempt to capture nonlinear properties of the auditory signal,
such as turbulence introduced by the vocal tract (Kokkinos and Maragos 2005 ).