Transfer Learning with Deep Neural Embeddings for Music Classification Tasks

Mateusz Modrzejewski, Piotr Szachewicz, Przemyslaw Rokita

Abstract: In this paper we present an approach for transfer learning with deep neural embeddings applied to a selection of music information retrieval (MIR) classification tasks with several datasets. The tasks include genre recognition, speech/music distinguishing, predominant instrument recognition and performer identification. We propose the usage of pre-trained $$L^3$$ neural networks for feature extraction and apply several supervised classification algorithms to the obtained feature representations in order to compare their performance. The deep neural embedding representations are compared with traditional, hand-crafted features and are shown to outperform the baselines.