Feature Extraction?

Music with a duration of 5 minutes is usually represented by 13 million values. These values describe the physical properties of the acoustical waves, which we hear. When analyzing this data it is necessary to remove the irrelevant parts and emphasize the important features. The extraction of these features from the raw data is the most critical part in the process of creating a content-based organization in a music collection. If it were possible to extract one single feature that directly indicates which genre a piece of music belongs to, everything else would be trivial.

Good features should be intuitively meaningful, based on psychoacoustic findings, and robust towards variations which are insignificant to our hearing sensation. Furthermore, they should lead to an organization of the music collection that makes sense and not be too expensive to compute.

It is necessary to consider computational aspects because the raw data of even small music collections easily consumes several gigabytes of storage. A detailed analysis of all this information and all its possible meanings would be computationally prohibitive. It thus is necessary to reduce the amount of information to what is relevant in respect to the overall goal, which is to organize music according to its genre. These genres are not clearly defined and different people might assign the same piece of music to different genres. However, there are some attributes of the raw data, which definitely do not determine the genre. For example, removing the first second of a piece of music does not change its genre, but the raw data compared bit wise will be completely different. Generally the duration of a piece of music is not relevant. Neither does a particular melody define a genre. The same melody can be interpreted in different genre styles just as different melodies might be members of the same genre. Likewise, the number of instruments involved plays a minor role in defining the genre.

One of the attributes that is rather typical for a genre is its rhythm which is why this thesis primarily focuses on the dynamics of music, and in particular on the fluctuation strength of the specific loudness per critical-band.