The Problem With Music

The Problem with Music: Modeling Distance Distributions of Large Music Collection.

Talk at Dartmouth Computer Science Colloquium

Wednesday 21st Januray 2009, Silsby Room 028 (*note location change*), Dartmouth College
Talk Slides

Abstract: Recently, a number of piano recordings by different artists were found in a classical music catalog that exhibited a striking resemblance to each other. Could this resemblance be purely coincidental? We set about building a system that could answer this question and others in large recorded collections of music. The AudioDB system listens to polyphonic music recordings and encodes important perceptual information about them at fine time scales. The information extracted corresponds to traditional music-theoretic concepts such as, harmony, timbre, pitch, texture and rhythm yielding a high-dimensional representation; consisting of 300-1200 dimensions.
Our music databases have 10⁴-10⁷ recordings, each with thousands of vectors, so brute-force methods for similarity computation are not practical. Instead, we use locality-sensitive hashing (LSH) which searches in high dimensions with sub-linear time complexity. We propose a method for automatically estimating the radius threshold for efficient and accurate LSH retrieval. Our method employs statistical sampling of the background distance distribution and solving for the minimum distance distribution using order statistics.
Using these methods, we are able to quantify the “purely coincidental” resemblance in the piano recordings mentioned above, demonstrating that their similarity is not, in fact, coincidental. The newer recordings are altered copies of the older ones. Detecting fraudulent recordings with human hearing is difficult; even the music critics were fooled into highly commending these newer recordings. We conclude that scalable audio search systems, such as AudioDB, are required to address the emerging multimedia needs of the Internet, commercial music services and large multimedia archives.