|Home||Papers||Reports||Projects||Code Fragments||Dissertations||Presentations||Posters||Proposals||Lectures given||Course notes|
Werner Van Belle1 - email@example.com, firstname.lastname@example.org
Abstract : BpmDj is a tool to mix music on your phone. It uses beatgraphs to visualize the music in 2 dimensionsd. By sliding such images over each other, the user can align and mix his music. The software also helps the DJ to navigate his music collection by calculating the nearest neighbors towards any song in its database. The talk focuses on the scientific aspects of BpmDj. It starts off with fine tunings we made to an autodifference-based tempo detector. Then we explain how rhythmical patterns are extracted and normalised, which provides us with a series of content vectors in a >1000 dimensional space. To store these points we created a novel data structure that allows us to find the k nearest neighbors by accessing less than 30% of the data, and with 20% less storage than needed by vector approximation files. The last part of the talk focuses on a fast converging distance metric learning. It learns the weights of the various dimensions by using their cross correlation function. Thereby it learns a true distance metric (as opposed to a distance function).
bpmdj, tempo detection, nearest neighbors, distance metric learning, rhythm detection
Reference: Werner Van Belle; BpmDj = Tempo detection, rhythm extraction, nearest neighbors, bit-slicing & distance metric learning; Presented at Phonetisches Laboratorium; Universität Zürich; Switzerland; March 2014