A learning-based method for automatic music transcription

Abstract

The primary objective of this project is to explore the capabilities of Deep Learning approaches for tackling the challenge of automatic music transcription. We will delve into various data preprocessing techniques to prepare the raw audio data for neural network models. This stage ensures the data is structured in a way that the models can effectively learn from it. Next, we will experiment with a selection of neural network architectures. Our focus will be on transcribing polyphonic music specifically produced by the piano instrument. By evaluating the performance of different models, we aim to contribute valuable insights into the effectiveness of deep learning for AMT tasks. In our approach, the musical's signal was treated as sequential data and a network model, LSTM (Long Short-Time Memory), was built for realizing audio-to-score conversion. Recordings of piano music in .wav file format is inputted into our algorithm, and the output is MIDI (Musical Instrument Digital Interface) file that can be easily converted into music score.