Neural Harmony: Advancing Polyphonic Music Generation and Genre Classification through LSTM-Based Networks

Creating complex multi-track music compositions from simple vocal melodies is a significant challenge, particularly for those with limited musical expertise. Our research addresses this by utilizing Google Magenta’s onsets and frames system combined with LSTM architecture to process vocal inputs. This involves refining raw vocals through tempo control, normalization, and noise removal, transforming them into musical notes. Advanced string algorithms, such as Longest Repeated Subsequence (LRS) and Longest Common Subsequence (LCS), facilitate effective melody extraction. The result is genre-specific, polyphonic compositions presented as MIDI files, which are efficient for storage and manipulation. This approach enables accurate music genre prediction and the creation of sophisticated compositions from basic vocal inputs. By integrating machine learning with music theory, our research democratizes music creation, allowing musicians of all skill levels to generate elaborate compositions, thus paving the way for more inclusive and innovative music production.