A new data set for audio segmentation with high accuracy: CMMSD


Henrik von Coler just released a new data set with annotations for monophonic audio excerpts [1]. Three segment classes have been manually annotated:

  • note (steady state, sustain)
  • transition (transient, release)
  • rest (silence, noise, reverberation)
overview of CMMSD pieces

Compositions selected for the CMMSD data set

The data set consists of 36 items, excerpts of commercial recordings of solo performances of classical music. It covers the instruments bass, basson, cello, flute, horn, oboe, trombone, trumpet, and violin, and spans several epochs. Although the problem of note-level segmentation is assumed to be solved for monophonic music, non-percussive instruments with weak transients, recordings with larger amounts of reverberation, and notes with large intra-note variations still pose a problem to concurrent approaches. This makes this data set a valuable complement to existing data sets.

distribution of composition and recording years

Distribution of year of composition (top) and year of recording (bottom) in the CMMSD

More details can be found in the accompanying paper [1] and on the web site for the data set  here.

[1] Coler, H. v.; Lerch, A.: CMMSD: A Data Set for Note-Level Segmentation of Monophonic Music (2014), Proc. of the AES 53rd International Conference on Semantic Audio, London, UK

Leave a Reply

Your email address will not be published. Required fields are marked *