Some analysis of the existing genre classifications has been done. The authors of [3] have thoroughly examined the genre taxonomies used by different official music sources, including:
"-Record Company catalogues: Universal, Sony Music, EMI, BMG
-Record shops and megastores: Virgin Megastore, Tower Records, Fnac...
-Music charts: Billboard, Top 50, Cashbox...
-Musical websites and online record shops: Amazon, All Music, SonicNet, Mzz, Listen, Netbeat...
-Specialized press and books
-Specialized web radios." [2]
The results of this study showed first that most taxonomies in use are album-oriented. In fact, the music industry tends to produce and sell albums, more than single titles. It follows that most genre taxonomies are based on albums. However, you can always find, in an album, titles of different genres. Hence, these taxonomies are not appropriate for describing the genre of a music title. Moreover, the study showed that there is no consensus in the naming of classifications. The authors of [3] have compared 3 Internet genre taxonomies:
"-allmusic.com (531 genres)
-amazon.com (719 genres)
Only 70 words turned out to be common to the three taxonomies.
Also, the genres in each taxonomy turned out to have different definitions. Rock and pop for example did not denote the same set of songs across the three taxonomies. Furthermore, and as we saw a little bit in the Introduction, it has been shown from this study that within a given taxonomy, genres do not bear fixed semantics. In other words, we can have genres based on a period (baroque, 70s etc..) or based on country (Italian music), or on language (French variety) or on a dance type (Waltz), or on artist (Bach-like). These categories lead to confusion and certainly introduce redundancy, since many of them can overlap: We can have Chopin's waltzes for example classified as Polish music, Waltz music and Romantic music all at the same time, or ABBA's music classified as 70's music, pop music and even Swedish music, all at the same time.
You can verify this last fact by yourself. For example, if you go to www.timaltbaum.com (last accessed on Oct.30, 2005) you see that ABBA's "Dancing Queen" is classified as 70's music, whereas if you go to www.mp3.com (last accessed on Oct.30, 2005) it is classified as Pop/Rock. Also, if you go to www.seacoastonline.com (last accessed on Oct.30, 2005) you notice that Jim & Karrie Band is classified as 70's, 80's, 90's and Dance Rock all at the same time! This redundancy introduced by these different taxonomies is not a good description scheme for automatic genre recognition systems, although it is very good for humans. Humans have no trouble switching between taxonomies and can deal very well with redundancies.
Finally, it has been shown from this study that labels used for genre classification carry a lot of information, most of which is extrinsic, i.e., does not depend on internal properties of the music piece.
People analyzed existing genre classifications and, as you will see in the next section, they planned different approaches to solving the automatic genre recognition problem.
Other people (who would have loved to benefit from automatic recognition) did manual genre recognition. The genre taxonomy for Microsoft's MSN Music Search Engine was manually built. "The manual labeling of a few hundred-thousand songs for Microsoft MSN required musicologists to be brought as full time employees and took about 30 man-years" [2].
Another attempt to manually classify songs into genres was done by the authors of [2], in their CUIDADO project. They combined genre with other attributes (country, instrumentation, artist) in order to limit the number of genres that can arise. For example, they will not have genres of the form "French-Rock", "English-Rock", "Italian-Rock", rather, the genre will be "Rock" and the country attribute will be used. However, they gave up this effort for two main reasons:
1- There are many subtle distinctions between very close genres. They are so subtle that people tend to disagree on the classification. The example they gave is the following: Rock-California differs from Rock-FM only by the fact that Rock-California songs would have a more dominant acoustic guitar sound.
2- As with any art, music evolves with time. During this evolution, new genres arise which are made up of a combination of older genres and some genres split giving rise to new sub-genres (as we saw in the Introduction). The difficulty is not in building the taxonomy for these new constantly evolving genres but rather in maintaining it.
As a conclusion, one can say that manual classification reflects first the time-consuming task of classification, then the complexity of such a task. So manual input cannot be used to classify millions of music pieces. It can be used however to evaluate the performance of automatic genre recognition algorithms, which are addressed next.
HOME WHAT ? WHY ? HOW? WHO ? Class 2005 Projects .Go to Top