Abstract:

Recent unsupervised feature learning methods have shown great promise as a way of extracting features from high dimensional data, such as image or audio. In this paper, we apply deep belief networks to musical data and evaluate the learned feature representations on classification-based polyphonic piano transcription. We also suggest a way of training classifiers jointly for multiple notes to improve training speed and classification performance. Our method is evaluated on three public piano datasets. The results show that the learned features outperform the baseline features, and also our method gives significantly better frame-level accuracy than other state-of-the-art music transcription methods.