Category:

Guide

Duration: 1:22:04

19 views

0

A key problem facing us in the 21st century is information retrieval and management - how to retrieve, process, and store the information one seeks from the huge and ever-growing mass of available data, including multimedia. Music, from mp3s to ring tones to digitized scores, is one of the most popular categories of multimedia. Music collections are typically indexed by such features as title, composer, and performer. However, people often wish to perform tasks that require knowledge of their musical content, such as the melody ("What is the name of that song that goes like this -whistled melody-?") and selection of individual sound sources ("Could you make the flute part a bit louder in this recording?"). To extend the range of ways in which audio recordings can be accessed and manipulated, researchers must create systems that can access and manipulate perceptually relevant structures in the audio signal. This requires advances in areas such as audio source separation (picking out a single instrument), and higher-level structure identification (labeling a sung melody as "Hey Jude"). Systems able to reliably access such features represent a fundamental improvement in our ability to access and manipulate audio data, opening up new applications. In this talk, Bryan Pardo will discuss and demonstrate two interlinked research systems addressing these issue: the VocalSearch system for query-by-humming and the ASE system for automatic separation of sound sources in a stereo mix to isolate individual instruments.