Training a convolutional neural network for note onset detection on the clarinet

Abstract

Although computational models for note onset detection have improved drastically in the last decade, mainly due to the advances brought by the field of Deep Learning, such models have not been perfected yet. When dealing with specific data, like clarinet recordings, those models still produce a significant number of false positives and negatives. In this paper, we evaluate pre-trained onset detection models from the library madmom on a dataset composed of solo clarinet recordings, in particular, to investigate their performance on this kind of data. Moreover, we use the clarinet dataset to train the same neural network (CNN) employed in one of those models, to investigate whether training the model on this specific data leads to an improvement when dealing with clarinet recordings. The results obtained from the model trained strictly on clarinet data are considerably better than those from models trained on generic data.