This paper introduces a clustering-based unsupervised approach to the problem of drum transcription. The proposed method is based on a stack of multiple clustering and seg-mentation stages that progressively build up meaningful audio events, in a bottom-up fashion. At each level, the inherent redundancy of the repeating events guides the clustering of objects into more complex structures. Comparison with state-of-the-art approaches demonstrate the potential of the proposed approach, both in terms of efficiency and of ability to generalize.