zehuiw/piano_music_transcription

Piano music transcription using Convolutional Neural Networks

This is the final project of Stanford CS230: Deep learning

Introduction

This project tries to solve piano music transcription problem using convolutional neural networks (CNN). The input to our system is piano music in .WAV format. We then apply constant Q transform (CQT) to the input so that the audio file is divided into multiple frames. Each frame has a ground truth label, thus we can use our neural networks to predict the pitch within each frame. The output is then processed to a MIDI (Musical Instrument Digital Interface) file, which can be directly converted to music scores.

Dataset

The dataset we use for this project is MAPS (MIDI Aligned Piano Sounds) dataset. Here is the link to the dataset.

Source Code

• Preprocessing and evaluation scripts: forked from this repo
• model/train.py defines models and training process
• model/train_load.py loads a previous checkpoint
• script/npy2midi.py coverts output numpy arrays to MIDI files.
• script/run.sh integrates data preprocessing, model prediction and MIDI conversion together.

Prediction Result

Here is a Youtube video that presents the prediction result.