Recognising Classical Works in Historical Recordings

In collections of recordings of classical music, it is normal to find multiple performances, usually by different artists, of the same pieces of music. While there may be differences in many dimensions of musical similarity, such as timbre, pitch or structural detail, the underlying musical content is essentially and recognizably the same. The degree of divergence is generally less than that found between ‘cover songs’ in the domain of popular music, and much less than in typical performances of jazz standards. MIR methods, based around variants of the chroma representation, can be useful in tasks such as work identification especially where disco/bibliographical metadata is absent or incomplete as well as for access, curation and management of collections. We describe some initial experiments in work-recognition on a test-collection comprising c. 2000 digital transfers of historical recordings, and show that the use of NNLS chroma, a new, musically-informed chroma feature, dramatically improves recognition.