Aria fine-tuned to learn the styles of twelve jazz pianists
What happened
On October 6, Hacker News featured research on learning jazz pianists' styles through cross-attention conditioning. Starting from Aria, a 16-layer Transformer pretrained on piano MIDI, the researchers inserted gated cross-attention modules into the last eight layers, reading a learned embedding for each pianist, and fine-tuned on solos by twelve jazz pianists from the PiJAMA dataset. The share of generated clips a classifier judged to be the target pianist rose from 37% under unconditional generation to 70%; a synthetic classifier trained only on generated music reached 95% accuracy identifying real recordings.
Written by AI from the coverage · updated 59 minutes ago
Coverage
Follow the reports to see the story from different sides.
- Hacker News front pageLearning Jazz Pianist Style with Cross-Attention Conditioning
研究者基于钢琴 MIDI 预训练的 16 层 Transformer 模型 Aria，在最后八层插入门控交叉注意力模块，读取每位钢琴家的学习嵌入向量，用 PiJAMA 数据集中十二位爵士钢琴家的独奏进行微调。生成片段被分类器判定为目标钢琴家的比例从无条件的 37% 提升到 70%，仅用生成音乐训练的合成分类器识别真实录音的准确率达 95%。
Heat over time
Not enough continuous observations to draw a trend yet.