Machine learning to analyze and classify pieces of classical music

The work of analyzing music from its theory and technique can reach complex levels in some cases. With the help of artificial intelligence tools, it is possible to reach more accurate conclusions, free of human biases.

Based on this thesis, a research team from the Laboratory of Digital and Cognitive Music of the Swiss Federal Institute of Technology of Lausanne (EPFL) developed an unsupervised machine learning model, trained to listen to and categorize more than 13,000 pieces of Western classical music.

Sound analysis based on your musical modes

In particular, the aforementioned tool developed in the EPFL can identify the presence of major and minor modes in these sound records.

This analysis criterion, more typical of music theory, could involve concepts not dominated by those who are less familiar with this topic. However, intuitively, it is more common to develop a perspective that allows us, at least to differentiate bleak, sad or tense sounds, which would correspond to a smaller scale; as well as that corresponding to its counterpart, the larger scale, which usually evokes feelings of happiness or strength.

Throughout history, there have been periods in which several other modes besides major and minor were used, as well as, in other cases, it is difficult to find a clear separation between modes.

Encompassing the Late Renaissance, Baroque, Classical, Early Romantic and Romantic musical periods, developed from the 15th to the 19th century, this tool was created to understand and visualize these differences over time.

There are cases like that of music that emerged after the end of the classical period, in which the distinction of these modes is more diffuse. For example, Franz Liszt's music, corresponding to the late romantic period, given its complexity, failed to be addressed under this analysis criterion by the researchers of this team.

Uns supervisory AI analysis

In this case, the experiment began from the base of letting the computer analyze the data autonomously, without human intervention.

In itself, this machine learning system is much more difficult to implement than a supervised one, but as a modality, it has the advantage of being able to eliminate the possibility that the analysis contemplates tags issued by humans or that ends under the influence of a particular bias.

In fact, the 13,000 tracks that make up the analyzed database only provide sound information. Audios don't even contain metadata so that the analysis is only under the designated approach.

As outlined by the research team, the categorization results obtained with acceptable and credible from a cognitive point of view.

This initiative arises from a group of monomaniacal scientists who, in the future, intend to implement this same initiative, but now focused on jazz.

The details of these results, plus the conclusions obtained after the development of this research, were collected by the journal Humanities and Social Sciences Communications. Also, the same educational institution involved in this project published an article with more information about this project.

Post a Comment