Music style translation aims to generate variations of existing pieces of music by altering the style-related characteristics of the original piece while content, such as the melody, remains unchanged. These alterations could involve timbre translation, re-harmonization, or music rearrangement. Previous studies have achieved promising results utilizing time-frequency and symbolic music representations. Music style translation on raw audio has also been investigated and applied to single-instrument pieces. Although processing raw audio is more challenging, it provides richer information about timbres, dynamics, and articulations. In this paper, we introduce Music-STAR, the first audio-based translation system that translates the existing instruments in a piece into a set of target instruments without using source separation. To conduct our experiments, we also present an audio dataset that contains two-track pieces performed by two instrument sets alongside their stems. We carry out subjective and objective evaluations to compare Music-STAR with a variety of baseline methods and demonstrate its superiority.

Direct link to video