A Scalable Peer-to-Peer System for Music Information Retrieval

Currently, a large percentage of Internet traffic consists of music files, typically stored in MP3 compressed audio format, shared and exchanged over peer-to-peer (P2P) networks. Searching for music is performed by specifying keywords and naïve string-matching techniques. In the past years, the emerging research area of music information retrieval (MIR) has produced a variety of new ways of looking at the problem of music searching. Such MIR techniques can significantly enhance the ways users search for music over P2P networks. For that to happen, there are two main challenges that need to be addressed: scalability to large collections and number of peers; and a richer set of search semantics that can support MIR, especially when the retrieval is content-based. In this article, we describe a scalable P2P system that uses rendezvous points (RPs) for music metadata registration and query resolution. The system supports attribute-value search semantics as well as content-based retrieval. The performance of the system has been evaluated in large-scale usage scenarios using "real," automatically calculated musical content descriptors.

One could argue that both the ideas of MIR and P2P became familiar to the general public with Napster (www.napster.com). Although crude both in terms of search capabilities and P2P performance, Napster provided for the first time an example of sharing vast amounts of musical data over large ad-hoc networks. Despite this early connection between MIR and P2P, there has not been much progress in combining these two areas. Although better P2P paradigms have been proposed, searching for music is currently still performed using traditional keyword-based text searching. Although a variety of novel ways of searching and retrieving music, especially in audio format, have been proposed, they have not found their way into P2P networks and remain largely academic exercises.

There are many advantages to P2P networks, such as distributed computing and storage power, fault-tolerance, and reliability. Owing to copyright restrictions, however, major recording labels have been reluctant to follow this paradigm, but the emergence of audio fingerprinting technology (Haitsma and Kalker 2002) is likely going to change this attitude. Although the intellectual property issues behind P2P networks are complicated, we believe that the use of techniques such as finger-printing will allow digital music distribution while protecting copyrights. Although the main focus of the system is the use of P2P networks in MIR, the proposed architecture can be adapted for the requirements of fingerprinting. One of the greatest potential benefits of P2P networks is the ability to harness the collaborative efforts of users to provide semantic, subjective, and community-based tags to describe musical content. We believe this aspect of P2P networks provides a unique opportunity for changing the way music is produced and distributed.

Centralized P2P networks such as Napster are not robust and may be vulnerable to Denial-of Service attacks, because the central server forms the system's single point of failure. Such a system does not scale well as registration and query load increase. Distributed P2P systems, such as Gnutella (www.gnutella.org) and KaZaA (www.kazaa .com), are more robust, but because peers do not explicitly register their shared files, a query may have to be broadcast throughout the network to get resolved. The potentially large number of messages involved limits the system's scalability and performance. Distributed Hash Table (DHT)-based systems, such as those described in Stoica et al. (2001) and Rowstron and Druschel (2001), achieve good scalability by deploying a structured overlay P2P network that supports efficient content location. However, the basic set of applications built on top of DHT only supports exact file name lookup and does not allow the rich search semantics desired for MIR.

In this article, we describe a robust, scalable P2P system that provides flexible search semantics based on attribute-value (AV) pairs and supports automatic extraction of musical features and content-based similarity retrieval. The system is shown to perform well under realistic loads consisting of features automatically extracted from a large database of audio recordings. The main contributions of this article...