Fusion of Music Styles Using LSTM Recurrent Neural Networks
This repo is the result of a collaboration between Jacob Sundstrom, Harsh Lal, Dave DeFillipo, and Nakul Tiruviluamala in a class entitled "Machine Learning and Music". We each expressed interest in extracting "musical features" from music and recombining them to form new, fused musical works. A paper was written and submitted.
The abstract is pasted below.
Appeal of a musical composition is almost exclusively subjective in that it is a combination of the tastes, preferences, and history of an individual's experiences. That is, it is perceived and judged qualitatively in a different way by different individuals. In this project we propose to build a deep learning system which could take n different samples of a jazz soloist - especially a variety of samples of specific 'styles' - and generate sound using current input as well as feedback and memory from the past samples. This generation can then be judged by a 'human' agent and the parameters of the neural network could be adjusted accordingly to generate a fusion music that is more closer and appealing to agent's expectations. Recurrent neural networks with Long Short Term Memory (LSTMs) in particular have shown promise as a module that can learn long songs sequences, and generate new compositions based on the song's harmonic structure and the feedback inherent in the network.
- Jacob Sundstrom, Department of Music, UCSD
- Harsh Lal, Computer Science and Engineering, UCSD