I developed a genre fusing melody generator using machine learning to create personalized fusion music. The model blends melodies from various genres using Long Short-Term Memory (LSTM) networks, allowing users to adjust the influence of each genre. I began with the data preprocessing of MIDI files to extract music notes, followed by generating individual genre melodies with Recurrent Neural Networks (RNNs). These melodies are then fused using an LSTM network, guided by user-defined weights for each genre to personalize the composition. An interactive user interface (UI) allows users to select genres and set weights, providing visual and auditory feedback. This project demonstrates AI’s potential in personalized music generation, with the potential for improvement through the addition of other genres and training that moves beyond simple melodies to incorporating other musical elements such as tempo, time signature, and instrumentation.

Applying data preprocessing filters for length and user ratings, I collected melodies from the BachDoodle melody library to train a transformer model. This type of neural network, typically used for processing and generating natural language, was adapted to learn the context and patterns of melody sequences. In this context, the transformer model leverages its attention mechanism to understand and predict the relationships between notes in a sequence. By determining the probabilities of subsequent note occurrences and selecting those with the highest probabilities, the transformer model generates a unique melody. These melodies range from popular trained sequences, like nursery rhymes, to original compositions.