Music informer as an efficient model for music generation

Abstract

Abstract Music Transformer has been extensively employed in music generation, however, the self-attention mechanism consumes significant memory due to its complexity. To address this issue, Music Informer is proposed, drawing inspiration from the widely-used Informer model in fields like time series forecasting, weather prediction, etc. Music Informer primarily utilizes the ProbSparse self-attention mechanism, relative local attention mechanism, and LSTM structure. Objective results indicate that Music Informer conserves 21.73%, 31.87%, and 41.33% of computational resources compared to Music Transformer, Performance RNN, and Multi-Track Music Transformer under identical experimental conditions. Additionally, music samples generated by Music Informer outperform those from the three baselines in the metrics of Pitch Class Entropy, Number of Pitch Classes, Pitch Entropy, Number of Pitches, Average Inter-Onset Interval. Furthermore, Music Informer demonstrates higher Overlap Area values in the features Number of Pitches, Pitch Class Entropy, Average Inter-Onset Interval, Groove Consistency, and Pitch Entropy, as well as lower Kullback–Leibler Divergence in Average Inter-Onset Interval and Groove Consistency, highlighting a strong alignment with the feature distribution of the real dataset for these specific characteristics. Subjective results suggest that listeners prefer music generated by Music Informer over the baselines due to its improved coherence and overall quality.