Terascale music mining
This week, Seattle is hosting SC/05, the yearly conference on high performance computing, networking and storage. A new initiative this year at SC/05 is HPC Analytics. The goal of this initiative is to highlight "rigorous and sophisticated methods of data analysis and visualization used in high performance computing by showcasing powerful analytics applications solving complex, real-world problems."
One of the HPC Analytics finalist is a submission by Dr. Stephen
Downie from the University of Illinois at Urbana-Champaign called Terascale Music Mining. The
project's mission is the 'creation of secure, accessible, terascale
collections of music materials in a variety of audio, symbolic and
metadata forms. These collections, coupled with a set of standardized
experimental tasks and standardized evaluation metrics, will allow
members of the international MIR/MDL research community to participate
in the newly created, TREC-like, Music Information Retrieval Evaluation
eXchange (MIREX) contests.'
Dr. Downie's submission includes some work that Kris West and I did over the summer as part of the MIREX 2005 artist identification submission. For MIREX 2005, we created a system that analyzes a set of MP3 audio files and attempt to identify the performing artist of the audio based only upon the audio. The system is implemented in Music-to-Knowledge (M2K) an open-source system for developing large-scale music analysis systems. The system is a good example of the kind of problem that can only be solved with large or distributed computing systems. Our M2K submission took about 10 hours to analyze 1500 audio files. If we wanted to scale the same algorithms up to an iTunes-sized music collection, we'd need about a year of CPU processing time to perform the analysis. Two things seem pretty clear to me. Music analysis is going to be a big part of helping people find relevant music (i.e music that they want to listen to), and companies that want to do this sort of analysis are going to need some pretty big and fast computers.