In the context of growing digital media and new classification/indexing demands, the task of Automatic Instrument Recognition in the field of Music Information Retrieval (MIR) has increasing importance. Through the use of deep learning techniques, namely convolutional neural networks, and different automatic source separation algorithms, developed at the Fraunhofer Institut für Digitale Medientechnologie (IDMT), this Master thesis investigates this recognition task and how different pre-processing stages can improve its classification performance. Several experiments have been conducted in order to reproduce and improve upon the results of the reference system reported by Han et al. Two systems are proposed in this research: an improved system using harmonic/percussive separation and post-processing using class-wise thresholding, and a combined system using solo/accompaniment separation and transfer learning methods for the special use case of jazz solo recognition. To validate the obtained results, diverse tests have been performed with multiple music data sets, with different complexities and instrument selections. Download PDF
The documentation of the code created is found in this link.
An example of a real-time implementation using of the neural network can be seen in this video:
