基于强化学习引导预训练模型的情感音乐生成

Emotion-conditioned Music Generation Based on Reinforcement Learning-guided Pre-trained Model

生成具有特定情感的音乐是可控音乐生成的一个重要子任务。以往的监督学习方法需要依赖带有情感标注的音乐数据集,且存在训练目标与模型优化目标不一致的问题。本文提出了一种强化学习引导的情感音乐生成方法,使用训练好的符号音乐情感分类模型对生成的音乐进行打分,以此作为强化学习的反馈来优化基于GPT-2的自回归音乐生成模型。该方法突破了数据集标注的限制,能够在曲风流派和数据类型相似的无标注符号音乐数据集上训练模型进行情感音乐生成。客观和主观评价结果表明,本文提出的方法可以生成与指定情感类别相匹配的高质量音乐。

Generating music with specific emotions is an important subtask of controllable music generation.Previous supervised learning methods rely on emotion-annotated music datasets and suffer from the issue of inconsistent training objectives and model optimization goals.In this paper,we propose a reinforcement learning guided approach for emotion-based music generation,using a pre-trained symbolic music emotion classification model to score the generated music as feedback for optimizing the GPT-2-based autoregressive music generation model.This method overcomes the limitation of annotated datasets and enables training models for generating music with emotions on unlabeled symbolic music datasets with similar musical genres and data types.Objective and subjective evaluation results demonstrate that our proposed method can generate high-quality music matching the specified emotions.