ISMIR 2026 · Late-Breaking Demo
This demo page presents audio samples generated by PianoKontext, a latent flow matching model that can perform classical piano and chamber music. Despite the limited training data (MAESTRO and MusicNet, each around 30 hours), Classifier-Free Guidance on context tokens enables the model to be faithful to compositions. The demo page also includes an unsupervised bridge baseline model (CFG Bridge) that performs style transfer (from "deadpan" to "expressive").
Comparative rendering on held-out MAESTRO scores (solo piano music).All samples are resynthesized with the Music2Latent audio codec. Deadpan audio files are synthesized from MIDI scores using the YDP Grand Piano soundfont from the FreePats project. Please adjust your volume settings as some of the audio samples might be loud.
Comparative rendering on held-out MusicNet scores (multi-instrumental chamber music). All samples are resynthesized with the Music2Latent audio codec. Deadpan audio files are synthesized from MIDI scores using the Fluid R3 Mono soundfont. Please adjust your volume settings as some of the audio samples might be loud.
PianoKontrol can generate excerpts of variable duration predefined by a user.