Βαθιά νευρωνικά δίκτυα για τη σύνθεση μουσικών σημάτων και εφαρμογή στην ανάλυση ηχογραφήσεων
Deep neural networks for synthesizing music signals with an application on the analysis of audio recordings

View/ Open
Keywords
Artificial Intelligence ; Music genre classifiers ; AI music generation ; Python ; Music information retrieval ; Music tagging ; Domain adaptation ; Generative musicAbstract
In 2022, an event took place that marked our daily lives. This was the release of ChatGPT, which made artificial intelligence (AI) accessible and known to the public. AI is a fast-growing field, constantly offering new possibilities to make our lives easier. One of its uses is the synthesis of musical signals. This paper focuses on the use of Deep Neural Networks (DNN) for music signal synthesis and analysis of recordings. In particular, the application of MusicGen (Meta), a generative artificial intelligence model, for the generation of artificial music tracks is studied. The generated tracks are used as data for training a classifier, which undertakes the analysis and categorization of audio signals according to the musical genre.
A comparison is then made with real data from a dataset, which are classified using the same classifier.
The methodology provides a powerful tool for the analysis of sound recordings, especially in cases where real sound recording data is limited. The use of artificial data from MusicGen broadens the scope of applications such as improving detection models, categorizing large music files, and genre-based audio analysis in a variety of contexts.
This thesis investigates whether incorporating generative models for generating audio files into classifier training can be an effective source of data, enhancing the understanding and analysis of audio data in an ever-evolving field.


