Your Name and Title: Jane Gilvin, Research, Archives & Data Information Specialist 3
Library, School, or Organization Name: NPR Research, Archives & Data Strategy team
Co-Presenter Name(s): N/A
Area of the World from Which You Will Present: United States
Language in Which You Will Present: English
Target Audience(s): Music or audio librarians & archivists, anyone interested in training algorithms, interest in automation and AI to increase productivity.
Short Session Description (one line): I will demonstrate the integration of an algorithm in a music database that assesses the likelihood of a music track being either vocal or instrumental. The discussion will be fairly non-technical as my role is product manager, not developer.
Full Session Description (as long as you would like): The NPR Research, Archives & Data Strategy (RAD) team has provided a music library for NPR journalists since the 1970s for production purposes. In preparation for moving to a new building, the NPR RAD team began building a digital database of music in anticipation of a digital, file based production workflow as well as lessening the need for physical storage. The effort included up to 4 librarians ripping and cataloging CDs for a year and a half. Since user demand was mostly for instrumental music, a large portion of the work was dedicated to verifying if a track was instrumental or vocal. When the move was completed, most work adding new tracks to the database was stopped because of the detailed and time consuming work required. The Essentia open source library & tools provided a pre-trained classifier model in determining vocal or instrumental probability. The accuracy of the pre-trained model did not meet our standards, but with training on NPR’s collection, we were able to achieve 85% accuracy. Our team was able to resume adding music, which is now born digital tracks, to the database since the most time consuming work had been alleviated. I will demonstrate the Orpheus database and the algorithm integration during this session.
Websites / URLs Associated with Your Session: N/A
Replies