🎵Aria Model Generates Jazz Piano Etudes in 12 Pianists' Styles
Aria nails jazz pianist styles with 70% accuracy
TL;DR
Aria, a transformer model trained on piano MIDI, generates jazz piano etudes in the style of 12 pianists with 70% accuracy when conditioned. Real recordings are identified with 95% accuracy by a second classifier.
Aria, a transformer model pretrained on piano MIDI, has been fine-tuned to generate jazz piano etudes in the style of 12 pianists from the PiJAMA dataset. The model uses a gated cross-attention layer to condition on a learned embedding for each pianist, achieving 70% accuracy in attributing generated music to the intended pianist. This breakthrough could revolutionize jazz education and performance, allowing musicians to study and emulate the styles of legendary pianists. The model's perplexity on held-out music is 11.4125%, and it identifies real recordings with 95% accuracy.
Key Points
Aria uses a gated cross-attention layer to condition on a learned embedding for each of 12 pianists.
The model generates music attributed to the intended pianist 70% of the time when conditioned.
A second classifier identifies real recordings with 95% accuracy, validating the model's output.
The model's perplexity on held-out music is 11.4125%, nearly blind to the pianist's style.
Aria imitates some pianists better than others, with 96% accuracy for Hank Jones and Dick Hyman.
Why It Matters
If you're a jazz pianist or educator, Aria's ability to generate music in the style of 12 pianists with 70% accuracy changes how you study and teach. Real recordings are identified with 95% accuracy, ensuring the synthetic music is of high quality. This model could democratize access to jazz education and performance techniques.
Comments
Be the first to comment
Enjoyed this article?
Get it daily. 7am. Free. Reads in 5 minutes.