Singer diarization: application to ethnomusicological recordings

Abstract

A music audio document can be structured automatically by many ways according to the final objective. In the context of a project on indexing ethno-musicological au­dio documents, we asked ourselves the questions: who is singing and when. By analogy with speaker diarization which consists in detecting who is speaking and when, we called the fact of detecting changes of singers, singer diarization. Figure 1 illustrates the task. The ground truth consists of a manual annotation in singing turns, and eventual entry/exit of instruments. In the context of the ANR DIADEMS1 project (Descrip­tion, Indexing, Access to ethno-musicological and Sound Documents) on indexing ethno-musicological audio doc­ uments, singer diarization automatically appeared to be essential. In this paper, we present our developed singer diarization system which is applied on ethno­ musicological recordings.