Content-Driven Automatic Equalization For Audio

Abstract

We can find equalizers in most of the modern day music players, and these equalizers usually include presets that the user can select, according to the genre of the song that is going to be played, or create their personal ones. With this thesis we intend to create a way to minimize the user interaction in this process, using machine learning techniques to categorize the audio inputs (songs) in one of ten categories (genres). This will be achieved using neural networks, to which we will input the mel-spectrogram of the audio files. This feature has been selected due to its high content of information about the audio and about parameters of it related to the genre, such as tempo, harmony or frequency bands related to certain instruments. These audio files will be as-signed to genres, as they conform one of the main categorization standards used for music and are heavily related to the frequency components, and, therefore, the equalization. After the audio files are assigned to one of the categories we will perform an equalization on them, through the use of a filterbank and stored presets for each of the genres. Then the bands will be united again to form an equalized version of the audio. We will measure the results of this process in two main parts: music genre classification system and equalization. In the first part we will focus on measuring the accuracy that the classifier achieved with the dataset, previously split between training and validation, with representations of different values and ways of evaluate it. Finally, we will explore the results of the equalization, first in analytic terms, with different representations of the songs before and after the process, and secondly with a survey conducted to evaluate the subjective quality of them.