Sound source separation for musical pedagogy

Abstract

Harmonics hopes to support musical pedagogy, offering a concrete product with which those interested in learning to play an instrument can practice. We trained a model to identify and isolate the singular tracks of a song through TensorFlow and tools to make the separation of auditory sources and produce genuine sheet music, based on a musical transcription algorithm (specifically for pianos, basses, drums, and voice) that beginners can visualize, edit, and download (in .PDF and .MIDI formats), adjusting at their own pace. Three methods of source separation were considered, under the following restrictions: Use a single song as an input file, which it was moderately complex (composed of a set of between three and six instruments), and that the number of samples -songs composed by relevant instruments and tracks of each standalone instrument - suitable for model training, would be extremely scarce.