- We use the Bach Doodle Dataset to generate new four-bar phrases of music.
- User melodies are harmonised in the style of Bach Chorales by CoCoNet.
- We then train an LSTM-GAN to be able to generate new examples.
- Generating music with long-term structure such as repeated themes or chord progressions is still a challenge.
- The main idea is to develop an RNN architecture that effectively captures ideas from previous bars.
- For example, Lookback and Attention are methods that can be useful to create long-term structure.
- We can explore how the latent features affect the characteristics of the music produced.
- This could assist the composer in personalisation of the music.
- During the learning process, we could add a penalty term as well to the loss function that rewards various themes depending on the task.
- For Part 1, this could be in the form of penalising the presence of Parallel Fifths / Octaves.
- In Part 2, we could reward the model for repeating melodic themes or chord progressions.