Music Source Separation and Automatic Transcription

Automatic music transcription — is a critical problem in the field of music information retrieval. The existing solutions in this field mainly consider the transcription of only classical music or even only music played on the piano. This is largely due to the lack of a dataset containing transcriptions of modern music. The main task of this work is to propose the automatic music transcription algorithm which would perform best on modern polyphonic multi-instrument music. To solve this problem, a suitable dataset was found and supplemented. In this paper, a new model is presented which is designed for music transcription using modern neural network architectures. Since modern music has a clear rhythmic structure, the work also suggests using rhythmic data to improve onsets prediction quality. An important aspect of this work is to study the influence of the input data format on the performance of the algorithm. Comparing the proposed algorithm results with existing approaches on the same data, it can be concluded that the proposed solution shows a significant increase in the quality on multi-instrumental data both on standard metrics and on a trained metric, which better correlates with how a person perceives transcriptions.