An automatic music generation and evaluation method based on transfer learning.

Area of Science:

• Artificial Intelligence
• Music Information Retrieval
• Deep Learning

Background:

• Deep learning models, particularly pre-trained models in Natural Language Processing (NLP), have advanced significantly.
• Direct application of NLP pre-trained models to music generation is hindered by representational differences between text and music symbols.

Purpose of the Study:

• To adapt advanced NLP pre-trained models for symbolic music melody generation.
• To develop a novel, objective music evaluation method for assessing generated melodies.

Main Methods:

• A text-like representation for music melody capturing pitch, rhythm, and pauses.
• The MT-GPT-2 (music textual GPT-2) model, utilizing generative pre-training-2 (GPT-2) and transfer learning.
• A symbolic music evaluation method (MEM) combining statistical, music theory, and signal processing approaches.

Main Results:

• The MT-GPT-2 model successfully generates symbolic music melodies.
• The proposed MEM provides a more objective evaluation compared to manual methods.
• Melodies generated by MT-GPT-2 demonstrate greater similarity to real music when compared against LSTM, Leak-GAN, and Music SketchNet models.

Conclusions:

• The proposed text-like representation and MT-GPT-2 model effectively enable the use of NLP pre-trained models for symbolic music generation.
• The MEM evaluation method validates the superior performance of the MT-GPT-2 model in generating realistic music melodies.