Adaptive Accompaniment with ReaLchords

2025-06-17

International Conference on Machine Learning

AI Summary

Existing music generation models lack real-time responsiveness, hindering synchronous improvisational ensemble with human performers. This paper proposes an online chord generation framework for real-time accompaniment. First, we design a multi-objective reward model balancing harmonic validity and rhythmic consistency. Second, we introduce a “future-aware” teacher-distillation reinforcement learning paradigm, enabling dynamic output optimization conditioned on both the current and predicted melodic segments. Third, we build upon a maximum-likelihood pre-trained sequence model and integrate a low-latency inference mechanism. Experiments demonstrate that our method significantly outperforms baselines in adaptability to unseen melodies, accompaniment fidelity, and end-to-end latency (<120 ms). In subjective listening evaluations, it achieves— for the first time—high-fidelity, low-latency human-AI real-time improvisational collaboration.

Technology Category

• Machine Learning: Large Multimodal Models (LMMs)
• Natural Language Processing: Generation
• Cognitive Modeling & Cognitive Systems: Computational Creativity

Application Category

• Economics, Online Markets and Human Computation: Economic ramifications for generative AI infrastructure and applications
• Search and Retrieval-Augmented AI: Web learning to rank, online learning, and counterfactual learning for ranking
• Semantics and Knowledge: Methods to enhance, augment, integrate or synergize semantic models such as knowledge graphs and LLMs

Abstract

Jamming requires coordination, anticipation, and collaborative creativity between musicians. Current generative models of music produce expressive output but are not able to generate in an emph{online} manner, meaning simultaneously with other musicians (human or otherwise). We propose ReaLchords, an online generative model for improvising chord accompaniment to user melody. We start with an online model pretrained by maximum likelihood, and use reinforcement learning to finetune the model for online use. The finetuning objective leverages both a novel reward model that provides feedback on both harmonic and temporal coherency between melody and chord, and a divergence term that implements a novel type of distillation from a teacher model that can see the future melody. Through quantitative experiments and listening tests, we demonstrate that the resulting model adapts well to unfamiliar input and produce fitting accompaniment. ReaLchords opens the door to live jamming, as well as simultaneous co-creation in other modalities.

Problem

• Online generative model for chord accompaniment
• Improving harmonic and temporal coherency with melody
• Enabling live jamming with adaptive music generation

Innovation

• Online generative model for chord accompaniment
• Reinforcement learning with harmonic-temporal reward
• Future-aware distillation from teacher model