Automatic Labeling of Unpitched Percussion Sounds

Herrera P, Dehamel A, Gouyon F. Automatic Labeling of Unpitched Percussion Sounds. In: AES Convention 114. Audio Engineering Society; 2003. Paper 5806. Available from: https://aes.org/publications/elibrary-page/?id=12599

@inproceedings{Herrera2003_12599,
  author = {Herrera, Perfecto and Dehamel, Amaury and Gouyon, Fabien},
  title = {{Automatic Labeling of Unpitched Percussion Sounds}},
  booktitle = {AES Convention 114},
  note = {Paper 5806},
  year = {2003},
  month = mar,
  publisher = {Audio Engineering Society},
  url = {https://aes.org/publications/elibrary-page/?id=12599}
}

TY  - CPAPER
TI  - Automatic Labeling of Unpitched Percussion Sounds
AU  - Herrera, Perfecto
AU  - Dehamel, Amaury
AU  - Gouyon, Fabien
T2  - AES Convention 114
M1  - Paper 5806
PY  - 2003
DA  - 2003/03/06
UR  - https://aes.org/publications/elibrary-page/?id=12599
PB  - Audio Engineering Society
LA  - en
AB  - We report a series of studies related to automatically labeling sounds from unpitched percussion instruments. Different databases have been set up in order to study relevant factors for labeling different sets of classes (up to 32 instruments, drum kit sounds, electronic sounds, manufacturer 'acoustic' signatures). Usual spectral features (i.e. centroid, skewness, etc.), Bark-band relative energies, and Mel-Frequency Cepstral coefficients, besides some additional original descriptors have been evaluated alongside different feature selection strategies. Classification techniques having different flavors and tradeoffs (k-NN, Kernel Density, Canonical Discriminant Analysis, Binary trees, etc.) have been also evaluated. It is shown that the feature set can be reduced by factors ranging from 2 to 3 without affecting performance, and that performance differences between classification techniques are in the order of 10% between the best (usually Kernel Density estimation) and the worst tested technique. For the most complex problem (classification of 32 sound categories) best hit rates were of up to 80%, whereas for drum kit problems containing 9 or less classes, performance increased to 88%, even using independent sample cross-validation. Very high success rates have been also achieved when labeling manufacturer-model, and when classifying drum machine sounds.
ER  -