Differential Melody Generation Based on Time Series Prediction

Long-term melody generation may encounter the challenges such as inadequate melodic variation, resulting in monotony or unreasonable melodic variation. In this work, we introduce the time series prediction and propose a method of Music-FED to generate more creative and harmonic melodies. The proposed approach adopts first-order difference to describe the melodic relative motion, and designs a temporal music representation that makes the model more easily aware of the temporal hierarchy of notes. It then learns the distribution of melody motion variation with time series prediction-based model in a non-autoregressive manner. The objective and subjective evaluations demonstrate that the proposed Music-FED can generate pop music melody with high harmony and rich contents to a certain extent.