End-to-end music transcription using a neural network

Abstract

We present various neural network models that learn to produce music transcriptions directly from audio signals. Instead of employing commonplace processing steps, such as frequency transform front-ends or temporal smoothing, we show that a properly trained neural network can learn such steps on its own while being trained to perform note detection. We demonstrate two models that use raw audio waveforms as input and produce either a probabilistic piano roll output or text in music notation format that can be directly rendered into a score.