AccoMontage-3 is a symbolic accompaniment arrangement system based on a novel methodology of sequential style transfer with prior. It first leverages content-style disentanglement and transfer to address local clip-level constraints in harmony and instrumentation. Subsequently, it resorts to global style planning (prior) to guarantee a coherent long-term structure. Our research roadmap begins with AccoMontage, a piano arranger that applies chord-texture disentanglement and a primitive, rule-based style planner to maintain a long-term piano texture structure. In this work, we introduce Q&A-XL, a multi-track orchestrator with a more generic latent style prior model, which characterizes the global structure of orchestration style. From end to end, the complete system is named AccoMontage-3, which is capable of generating full-band, multi-track accompaniment for whole pieces of music, with cohesive multi-track arrangement and coherent long-term structure. This demo page showcases orchestration samples by Q&A-XL, as well as full-band arrangement samples by the complete AccoMontage-3 system.
Please refer to our paper, GitHub repo, and Colab tutorial for more technical details. This page is developed with the html-midi-player library and is best to be viewed on a Google Chrome browser.
Pop Music Orchestration
Q&A-XL is an orchestror module and here we apply it to pop music orchestration, i.e., to rearrange the accompaniment part of a pop song from piano solo to full-band, multi-track music. The piano arrangements (conditional inputs) are selected from POP909 dataset.
1/6
2/6
3/6
4/6
5/6
6/6
Full-Band Accompaniment Arrangement
AccoMontage-3 is capable of full-band accompaniment arrangement given an input lead sheet (i.e., a lead melody with chord progression). It first applies AcoMontage to generate a piano sketch and then orchestrates the piano with Q&A-XL. Such a modular pipeline aligns with musicians' creative reasoning in real composition scenarios. The lead sheets (conditional inputs) are selected from Nottingham dataset.
1/7
2/7
3/7
4/7
5/7
6/7
7/7
Summary
To wrap up, we contribute AccoMontage-3, a self-supervised, modular system for full-band accompaniment arrangement. The main novelty lies in our proposed sequential style transfer with prior, which is a generic methodology for conditional generation with sequential semantic conditioning. In our case, the prior model is an encoder-decoder framework connecting an upstream piano arranger (AccoMontage) with a downstream orchestrator (Q&A), which guarantees both global context planning and history context dependency. Demo samples show that our entire system generates faithful, creative, and natural full-band arrangements for whole pieces of music.
For more technical details, refer to our paper, GitHub repo, and Colab tutorial.