Feel the Music

A playlist is a list of titles to be played like a musical program. A user interacting with a database might
ask for the automated generation of such a list. As Pachet et al. (2000) point out the creation of such a list
has to be taken seriously, since “[t]he craft of music programming is precisely to build coherent sequences,
rather than just select individual titles”. A first step towards coherence is to set certain criteria the songs have
to fulfill in order to be grouped into one playlist. The user could be asked to provide a seed song for the
playlist and the computer would try to find tracks from the database which have similar descriptor values.
Pachet et al. (2000) go further and look at an even more advanced way of playlist generation capturing the
two contradictory aspects of repetition and surprise. Listeners have a desire for both, as they state, since
constant repetition of already known songs will cause boredom, but permanent surprise by unknown songs
will probably cause stress. In their experiments Pachet et al. (2000) use a hand edited database containing,
among others, attributes like type of melody or music setup. We can see a correspondence here to tonal and
timbral complexity, that encourages the utilization of complexity descriptors for playlist generation.
An alternative way of playlist generation, which gives more control to the user, is that of using user
specified high-level concepts as for example party-music or music for workout. Inside the SIMAC project
methods were explored to arrive at such a functionality. A playlist could then be easily compiled by selecting
tracks with the according label. The bottleneck here is the labelling of the tracks, which might be a lot of work
in a big collection. Since the labels are personalized and may only have validity for the user who invented
them, there is no way to obtain them from a centralized meta-data service. Instead the user can try to train
the system to automatically classify his tracks and to assign the personalized labels [e. g. citePlugIn]. For
this process semantic descriptors are needed that help in distinguishing whether a track should be assigned a
certain label or not. It depends of course very much on the nature of the label to identify descriptors that are
significant for this distinction. In any case, the complexity descriptors certainly have a potential to be useful
here, as can be seen from the examples at the beginning of this section.

We have seen big changes that came with the spreading of perceptual audio coding technology in combination
with fast network connections during the past decade. The next revolution should be one that helps us to
manage this incredible amount of digital content. The call is for intelligent devices and services that can
actively assist us in our needs and interests in relaxation, entertainment, and culture whenever and wherever
we want. By providing access to semantic aspects of musical audio signals without the need for manual
annotation we are giving another spin to the wheel of innovation. This dissertation will hopefully help to
bring our vision a bit closer to reality, by providing means for a multi-faceted content description in a multifaceted
field of research.

For song retrieval there are different possibilities in a music database. The most obvious one is the direct
specification of parameters by the user. Since the complexity descriptors consist of only one value per track,
they can be used very easily in queries. The user can specify constraints only for those facets he is interested
in and narrow down the set of results. This way it is very straightforward to find music that, for example,
does not change much in loudness level over time, or contains sophisticated chord patterns.
A second way of querying is the so called query-by-example approach. The user presents one or several
songs to the database and wants to find similar ones. So, as explained for the visualization using similarity
measures, here the complexity descriptors can easily be integrated into the computation again. The weighting
and/or the tolerance for the different descriptors could be specified by the user directly, extracted from the
provided example, or taken from a pre-computed user profile. Such a user profile would be established by
monitoring the user’s listening habits (i. e. songs he/she has in his/her collection; songs he/she listens to very
frequently, etc.)