|
What makes a piece of music sound "sad"
or "rocking" and what accounts for the "groove"?
In existing approaches to these questions the semantic description
of songs is mostly given by catalog-oriented classification.
Unfortunately, this labelling according to predefined categories
is cumbersome and time-consuming. An automatic algorithm can
help here, even if no computer knows from the start what sounds
sad, rocking or groovy.
A musical signal can be analyzed by means of distinct measures
("descriptors"). Reasonable combination of these
low-level descriptors make it possible to derive semantically
meaningful characteristics like tempo  ,
rhythm, singing voice or song structure - all that in an automated
manner without any catalog.
The possibilities do not stop here. The melody recognition
system Query
by Humming transcribes monophonic tunes
and retrieves the respective artist and title. Another important
semantic feature is constituted by the harmonies of a song.
The automatic harmony recognition
finds the chords as well as the musical key of piece of music.
With the help of the above mentioned descriptors reoccurring
structural elements of a song can be identified .
In this way, the track can be split into its main parts (like
chorus, verse, bridge) and a representative snippet can automatically
be chosen for a pre-listening function (e.g. in music download
portals).
The result of all these analysis steps presents an objective
categorization of music - that does not always match the expectations
of a human listener. Especially music is purely a matter of
taste …
Nevertheless, these technologies can provide an automatic
genre classification or recommend similar sounding songs from
a huge database. This is accomplished by the software modules
GenreID and SoundsLike,
developed at Fraunhofer IDMT.
Retrieval and recommendation tools are not bound to audio
related applications. Many of the principles apply as well
to the image and video domain (see Photo
and Video Analysis).
|
