Here is the paper we submitted to ICMC.
Here are a few examples from the project:
- Original recording of Escola - This is how one of the unprocessed rhythms sounds.
- Quantized recording of Escola - Note the subtle differences once the rhythm is quantized (i.e. the microtimings are all set to zero). For example, try focusing on the sound of the shaker.
- Escola microtimed using k-nearest neighbors (KNN) - This example shows how KNN has learned to microtime Escola. This sounds less mechanical than the quantized version.
- Escola microtimed using kernel ridge regression (KRR) - This example shows how KRR has learned to microtime Escola. This also sounds less mechanical than the quantized version.
- Original recording of Olodum - This is how one of the unprocessed rhythms sounds.
- Quantized recording of Olodum - Note the subtle differences once the rhythm is quantized.
- "Escolodum" cross-synthesis example (KRR) - This example used KNN trained on the Escola data to apply Escola-style microtiming to the quantized data from the Olodum sequence. In other words, the Olodum "score" as played in the style of Escola.
Source Code:
- ourutils.zip contains the code that we wrote along with the GPR code.
- You will also need the Midi Toolbox.
- The independent component analysis (ICA) dimensionality reduction method can also be used instead of principal component analysis (PCA).
- The Max patches we use to play these drum sequences, plus 7 megabytes of drum samples (all from findsounds.com or from Hydrogen drum kits) that we used.
 |
To Matt Wright's Web Page |
|
To Edgar Berdahl's Web Page |