研究人员开发AI模型学习爵士钢琴家演奏风格

研究人员利用基于Aria变压器架构的AI模型，通过在MIDI数据上集成门控跨注意力层，成功使机器学习了12位爵士钢琴家的独特演奏风格。1该模型在生成音乐时能够以70%的准确率保持目标钢琴家的风格特征，相比无条件生成的37%实现了显著提升。1

这项研究采用了PiJAMA数据集中的自动转录商业录音进行训练，研究对象涵盖了Tatum、Garner、Peterson和Dick Hyman等知名爵士钢琴家。1模型在某些演奏家的风格模仿上表现出色，对Hank Jones和Dick Hyman的风格准确率达到96%，但对Cedar Walton的模仿准确率仅为29%。1此外，合成专用分类器在保留音乐上的准确率达到98.8%，进一步验证了该条件化生成方法的有效性。1该研究将在2026年于阿布扎比举行的ISMIR会议上呈现。1

Researchers have developed an AI model based on the Aria transformer architecture that learns to replicate the distinctive playing styles of twelve jazz pianists by applying gated cross-attention layers to MIDI data.1 The model demonstrates significant success in style transfer, with pianists' classifiers correctly identifying the intended performer in conditionally generated music at a 70% accuracy rate, substantially outperforming the 37% baseline achieved through unconditional generation.1

The study, presented at the ISMIR 2026 conference in Abu Dhabi, trained on automatically transcribed commercial recordings from the PiJAMA dataset and examined artists including Art Tatum, Erroll Garner, Oscar Peterson, and Dick Hyman.1 The Aria architecture incorporates sixteen transformer layers with cross-attention blocks inserted into the final eight layers to enable style conditioning.1 Results reveal considerable variation in learning effectiveness across pianists: the model achieved 96% accuracy in capturing the styles of Hank Jones and Dick Hyman, while performance for Cedar Walton reached only 29%.1 A specialized classifier trained on preserved musical features achieved 98.8% accuracy in distinguishing the synthesized performances.1