Automatic singer identification in polyphonic music.

Abstract

Trained human listeners show a remarkable ability to identify singers from their voices alone even in new contexts. The identity of a singer can provide a good deal of information about a particular song, and thus systems that can identify a singer's voice can be important for browsing and searching through large databases of multimedia content. Music, however, comprises a very rich signal class that is often difficult to analyze due to the prevalence of complex mixtures of sound sources. In some areas of audio content analysis, this difficulty is addressed without performing source separation by using monophonic (single-source) recordings. Practical systems for musical content analysis, however, must be able to handle polyphonic (multi-source) recordings, which are by far the norm. In this work, we develop a set of methods for singer identification in polyphonic music and present a detailed performance evaluation of these methods. In particular, we seek to determine whether an approach that separates the voice signal from the polyphonic mixture yields superior performance over one that does not. We begin by developing and evaluating a set of basic but extensible methods for singer identification in monophonic music. These methods are then extended through the introduction of PESCE, a system for voice detection and separation in musical mixtures. With PESCE used for pre-processing, we evaluate the performance of our singer identification methods under a large variety of conditions. In particular, we examine performance for features that employ voice separation, temporal voice location only, and neither separation nor location. These evaluations include a basic database of piano-accompanied songs, a database of vocal music with orchestra, and a database containing a variety of popular music. Contrary to our initial hypothesis, we find that there is no uniform performance improvement when using voice separation. Notably, however, knowing when the singing voice appears within the musical mixture improves performance substantially. In general, identification performance is found to degrade for types of music with greater interference from accompanying instruments. Additionally, performance decreases when the system attempts to generalize to novel recordings by a given singer.