SBERT-based Chord Progression Estimation from Lyrics Trained with Imbalanced Data

In this research, we developed a model that can estimate appropriate chord progression based on lyrics input. It outputs a sequence of chord that can be used to compose the corresponding lyrics input. By training the model with different datasets, it is also possible to estimate other musical components that are correlated with lyrics, for example rhythm pattern, instrument, tempo, and drum pattern. Using this set of musical components as a setup recommendation for composition can potentially automate the configuration process on AI-based composition tools. We sourced our training data from "Orpheus", a web-based automatic composition system, resulting in more than 6,000 paired data of lyrics and musical components chosen by users who published their songs in the platform. Lyrics are pre-processed into semantics embedding using Sentence-BERT before being fed as training data into the multi-layer perceptron model as a classifier to estimate chord progression. Evaluation of this model is done objectively with ROC and F1 score, and subjectively through a survey.