Velocity Prediction for MIDI Notes with Deep Learning
| Autor: | KUO, CHIEN-SHENG, 郭建陞 |
|---|---|
| Rok vydání: | 2019 |
| Druh dokumentu: | 學位論文 ; thesis |
| Popis: | 107 When converting a MIDI (Musical Instrument Digital Interface) song (file) from a score, arranging the velocity of each note to exhibit velocity ups and downs is usually an important task that makes the song sound more musical and emotional. However, providing note velocity by hand requires a lot of human time, and it is also difficult to express delicate velocity changes manually. This thesis uses deep learning to construct neural network models for the prediction of note velocity. We use the MIDI songs recorded by Piano-e-Competition winners as our training dataset. The melody (a set of notes without velocity information) and the corresponding velocity of the melody are extracted from the songs, and are converted as the input and output of the neural network. That is, the model learns the velocity of the notes from the melody of the songs. We propose three neural-network architectures for training note velocity, including Convolutional Neural Network (CNN) and Autoencoder, using convolutional layers to capture the feature of melody for the prediction of note velocity. We experimented with several input and output formats and studied the training results. The experimental results showed that the note velocity predicted by our model is better than a model with either no or random velocity changes. |
| Databáze: | Networked Digital Library of Theses & Dissertations |
| Externí odkaz: |
|