Let’s Jam: A Machine Learning Framework For Sample Pairing and Filtration Using Melodic and Tonal Movement Classification

Abstract

The search engines of music sampling services are often inadequate when addressing samples that tonally and sonically pair well with each other. Using search filters and other functions within these services, one is able to find samples that are within the same key as the desired sample, and with the same sonic qualities. Current technology is either hyperspecific with the sample selections, or focuses too much on the instrumentation of the sample alone. Using a machine learning model (MLM), we plan to create a sample-sorting system that categorizes samples, not on the instrumentation or sonic similarity, but on the tonal qualities, like the chord structure the sample would fall on when assigned chords based on common practice chord assignments (I, II, III, and so on). This would allow the model to adequately categorize the samples by their movement (by breaking up the given sample into chord progressions such as I-IV-V). This delivers further classification advantages (such as filtering samples by what chords they would fall under) and allows sample libraries to provide sample pairings that are not sonically identical, but still fall under the same movement.  Samples with similar chord structures are far more likely to sound harmonious without sounding exactly the same. This would lead to more diversity and efficiency with sample pairing and selection, while also allowing the user to exercise more tonal specificity when searching, improving workflow.