Evaluating and Comparing Deep Learning Methods with Vision Transformers for Musical Key Assessment

A composition's major key is an essential component that helps with transposition and arrangement, providing information about the work's harmonic structure, tonal center, and chord progressions. Furthermore, precise key estimate finds use in both academic and industrial contexts, such as in computerized music recommendation systems and computerized music reproduction. By utilizing their successes in various domains, the current publication provides a thorough comparison between the conventional deep learning structures and the upcoming visualization converters. They estimate the outcomes by analyzing six distinct deep learning methods on a specific subset of gathered GTZAN dataset. Our findings show that Dense-Net, a traditional deep neural network architecture, outperforms visual transformers with an impressive 93.64 % accuracy; yet, we examine each deep learning method in more detail to clarify its temporal characteristics. In particular, the SWIN and vision transformers perform better in temporal metrics than the Dense-Net design, but they show a slight decrease in overall effectiveness (1.82 % and 2.39 %, respectively). Our results are noteworthy because they offer a practical instrument for musical key valuation, an area where precise and effective computational methods are essential. Through analyzing the advantages and disadvantages of image transformers and artificial neural networks, we might potentially obtain important data for real-world applications like automatic music transcription and product recommendation systems. Our research opens up new avenues for future advances and motivates further study in this field