Contrastive Unsupervised Learning for Audio Fingerprinting

Metric Learning with Sequence-to-sequence Autoencoder for Content-based Music Identification

“…Recent researchers have utilized contrastive learning to enhance the effectiveness of audio fingerprinting. For instance, Z. Yu et al introduced a self-supervised methodology for audio fingerprinting based on contrastive learning [8]. The authors conducted evaluations using different processing units and achieved impressive performance with the VGG-16 network variant.…”

“…The content-based music identification problem cannot be modeled as a typical multi-class classification problem due to the regular release of new music. Therefore, in the literature [7,8], the common approach is to compute the pairwise similarity between query fingerprints and reference fingerprints for identifying the music performance from which the query is extracted. Metric learning is a field of study that addresses this problem.…”

“…Metric learning with triplet loss is one such method found in the literature [7,8,15]. This approach involves working with triplets of data points, which consist of an anchor data point, a positive data point, and a negative data point.…”

“…The performance of the music identification system is assessed using the hit rate metric as widely used in the literature [8,14]. The calculation detailed in Eq.…”

“…Various techniques have been explored for audio fingerprinting by researchers [1,[4][5][6][7][8], including hand-crafted rule-based algorithms, computer vision-based algorithms, and, more recently, data-driven algorithms. Data-driven algorithms incorporate machine learning techniques for fingerprint generation, with Convolutional Neural Networks (CNN) being the most prominent approach [7,8]. However, only limited research has been conducted considering a musical performance as a dynamic system.…”