Emotion-Guided-Piano-Accompaniment-Generation

This is the codes repository for the project "Emotion-guided Piano Accompaniment Generation".

Baseline models are from Accomontage Repo.

Training and validation set is derived from POP909 dataset and emotion-guided music generation test set is a part of Nottingham Dataset.

Harmonization module is updated (2023/2/2) The paper is accepted at International Joint Conference on Neural Networks 2023(IJCNN2023).

Paper Link

https://arxiv.org/abs/2307.04015.

Run Demos

1. Click here to download the data folder and put them in ./data.

2. Put the melodies (MIDI Format) in the ./original_midi;

3. Run demo.py;

4. Wait a while and the accompaniment will be saved in the ./generate_midi.

5. For more emotional flow guided generation samples, please refer to https://soundcloud.com/ko9isjyplxrb/sets/demos-of-emotion-guided-generated-accompaniment.

Train the model

1. Click here to download the processed dataset and put them in ./data.

2. Run train.py

3. Log files will be automatically generated to the root directory, please use tensorboard related commands to monitor the training process.

Q&A

1. Q: Sometimes the generated chords dont fully harmonize the melody. How come?

   A: The randomness and controllabilty of LSTM-based harmonizer isnt satisfactory. Use music theory for the harmonization instead. You can rectify the solve() function from ./utils/Accompaniment_Generator.py for your personal usage.