Percussion recognition

[mike...@gmail.com]

Hello,

I have a very specific set of drums i'd like to recognize in realtime. I have about a hundred samples of each one, some of them hit very quickly (like 64th notes), and I wanted to train some models so I can use madmom to stream them out as recognized midi (mapped to notes). 

Does anyone have example code for training a model, or once a model is trained, converting it into a format readable by madmom's model libraries? Is there similar code to train it on the recognized onset times for those samples as well?

Thanks in advance

[Sebastian Böck]

Hi,

the specialities of your drums (hitting very quickly) shouldn't be a problem. You may ask Richard Vogl who has a working drum transcription (IIRC it is an implementation of the Recurrent Neural Networks for Drum Transcription paper). Most parts are already ported to madmom, and there is a drum_transcription branch with corresponding PR. As you stated correctly, it is mostly about format conversion (e.g. re-ordering dimensions), so the conversion process is straight forward. As far as I know there is not too much code released, mostly because it depends on which framework you use for training. You can find some example code in the madmom paper, however.

HTH