Automatic Identification of Samples in Hip-Hop Music via Deep Metric Learning and an Artificial Dataset

Add to your list(s) Download to your calendar using vCal

• Huw Cheston (University of Cambridge)
• Tuesday 19 November 2024, 17:00-18:00
• CMS computer room, Faculty of Music (11 West Road, Cambridge, CB3 9DP).

If you have a question about this talk, please contact jbf43.

Abstract

Sampling, the practice of reusing recorded music or sounds from another source in a new work, is common in popular music genres like hip-hop and rap. Numerous services have emerged that allow users to identify connections between samples and the songs that incorporate them, with the goal of enhancing music recommendation. Designing a system that can perform the same task automatically is challenging, however, as samples are commonly altered with audio effects like pitch shifting or filtering, and may only be several seconds long. Progress on this task has also been blocked due to the availability of training data. Here we show that a convolutional neural network trained on an artificial dataset can identify real-world samples in commercial hip-hop music. We extract vocal, harmonic, and percussive elements from several databases of non-commercial music recordings using audio source separation, and train the model to fingerprint a subset of these elements in transformed versions of the original audio. We optimize the model using a joint classification and metric learning loss and show that it achieves 13\% greater precision on real-world instances of sampling than a fingerprinting system using acoustic landmarks, and that it can recognize samples that have been both pitch shifted and time stretched. We also show that, for half of the commercial music recordings we tested, our model is capable of locating the position of a sample to within five seconds. More broadly, our results demonstrate how machine listening models can perform audio retrieval tasks previously reserved for experts.

Biography

Huw Cheston is a PhD student at the Centre for Music and Science, University of Cambridge, focussing on music information retrieval. His PhD research uses large-scale quantitative and computational methods to investigate performance style in improvised music, drawing from audio signal processing, machine learning, data science, and corpus analysis. He is also interested in developing reusable software, models, and datasets that can be deployed by researchers across a broad variety of audio-related domains. His research has been published in journals including Royal Society Open Science, Transactions of the International Society of Music Information Retrieval, and Music Perception. The work Huw will be presenting at this seminar derives from research completed as an intern in Spotify’s Audio Intelligence laboratory during Summer 2024.

Zoom link

https://zoom.us/j/99433440421?pwd=ZWxCQXFZclRtbjNXa0s2K1Q2REVPZz09 (Meeting ID: 994 3344 0421; Passcode: 714277)

This talk is part of the CMS seminar series in the Faculty of Music series.

This talk is included in these lists:

• CMS computer room, Faculty of Music (11 West Road, Cambridge, CB3 9DP)
• CMS seminar series in the Faculty of Music

Note that ex-directory lists are not shown.