Abstract
Style translation in symbolic music has traditionally been hindered by a lack of parallel datasets. This paper presents a supervised approach to music style translation by leveraging synthetic data generated from commercial accompaniment software. By training an attention-based encoder-decoder model on over 658,000 parallel musical pairs across 70 styles, the authors demonstrate that models trained on synthetic data can effectively generalize to real-world human performances while maintaining harmonic integrity.
Introduction
While style transfer is a mature field in computer vision, symbolic music faces a significant "parallel data" challenge. Supervised learning requires thousands of examples of the same musical content rendered in different styles—data that does not naturally exist in the real world. Previous attempts have relied on unsupervised methods like CycleGANs or Variational Autoencoders (VAEs), which often struggle to produce results as compelling as those seen in other domains. This research asks: can we manufacture the data we need?
Methodology
The core innovation lies in the data generation pipeline. Using Band-in-a-Box (BIAB), the researchers rendered 3,500 lead sheets into 70 distinct musical styles, creating a vast parallel corpus where the underlying harmonic content remains constant while the stylistic execution (rhythm, patterns, instrumentation) varies.
The architecture employs an RNN encoder-decoder with attention. Two primary input representations were tested:
- roll2seq: Processes a 2D piano roll via a CNN.
- seq2seq: Processes a MIDI-like sequence of event tokens, including
NoteOn,NoteOff, andTimeShift.
To enable multi-style translation within a single model, the decoder is conditioned on a target style ID ($z$). This allows one model to translate between any of the 70 learned styles.
Results and Key Insights
The model achieved high content preservation, with a chroma similarity of approximately 0.78 on synthetic tests, nearly matching reference arrangements. Crucially, the model generalized well to the Bodhidharma MIDI dataset—a collection of real-world recordings—despite being trained exclusively on synthetic data. This suggests that the structural rules of style learned from software are applicable to human performances.
A key finding was the importance of style diversity: models trained on 70 styles simultaneously were significantly more robust than those trained on only two styles. To evaluate the "groove," the authors utilized "Style Profiles" (2D histograms of timing and intervals), confirming that the outputs closely matched the target style's rhythmic and melodic characteristics.
Conclusion and Future Work
This work demonstrates that synthetic data is a viable solution for supervised music style translation. While the model occasionally struggles with complex syncopation and independent track cohesion, it provides a strong foundation for future "beat-aware" architectures and joint generation of multi-instrumental ensembles. Scaling this approach to full-length songs with structural transitions remains a primary challenge for future research.
Summary updated