In August 2026, researchers reported that a model identified pianists in held-out recordings with 94.4% accuracy across 20 named players. Its input was a transcription of each performance, allowing it to learn patterns in the music.[1]
The team assembled 1,629 performances spanning 84 hours. Its best model was a residual neural network reading automatically transcribed piano rolls: notes represented by pitch, timing, duration and velocity. A separate model combined four networks examining melody, harmony, rhythm and dynamics; that version scored 91.3%. The analysis associated Bill Evans with descending seventh-chord arpeggios: chord notes played in a falling sequence.[1]
What a player does with the beat
A 2024 study helps explain the rhythmic clues. Researchers separated piano, bass and drums in 300 recordings by ten pianists, then measured 19 timing features. A random forest classifier averaged 59% correct identification across five validation folds, against 10% for balanced random guessing. This used a different dataset and method from the 2026 study.[2]
Three ways to describe rhythmic character
| Musical quality | What the analysis measures |
|---|---|
| Feel | How the pianist places notes relative to the accompanying bass and drums.Source 2 |
| Complexity | The density and unpredictability of rhythmic patterns.Source 2 |
| Swing | The balance between longer and shorter subdivisions of the beat.Source 2 |
Feel and complexity were particularly informative. Recognition could therefore depend on how a pianist moves around the pulse and organises activity within it, beyond the tune being played.[2]
A recording becomes something researchers can compare
That kind of analysis needs carefully prepared material. The separately published Jazz Trio Database assembled 44.5 hours of piano solos with bass and drums, together with annotations of note onsets, beats and downbeats. Its piano transcriptions also record notes in MIDI, a machine-readable musical format.[3]
Source separation gives researchers distinct instrumental parts to examine. The annotations let them study a pianist’s playing alongside the ensemble’s timing. This matters because “late” is a relationship: a note can arrive behind the surrounding pulse even when the recording itself contains no written score telling the analyst where it belongs.[3]
The database combined listening statistics and discographic metadata to select recordings before applying inclusion criteria. Its coverage therefore reflects a constructed research collection. A recogniser’s musical world depends on which performances its builders choose to include.[3]
Recognition opens a listening question
The 2026 test concerns a defined roster, rather than every pianist or artistic merit. Its authors offer a public explorer; the website currently requires a desktop or laptop.[1][4]
One useful listening exercise is to take two recordings you already know and follow only the pianist’s entrances for a minute. Then listen again for the spaces between phrases. This is an invitation to notice recurring choices, without treating a classifier’s label as the final interpretation of a performance.
Sources & further reading
- Machine learning of artistic fingerprints in jazz Nature Machine Intelligence · 17 Aug 2026 · Accessed 12 Oct 2026
- Rhythmic qualities of jazz improvisation predict performer identity and style in source-separated audio recordings Royal Society Open Science / University of Cambridge · Accessed 12 Oct 2026
- Jazz Trio Database: automated annotation of source-separated piano trio recordings Transactions of the International Society for Music Information Retrieval / University of Cambridge · Accessed 12 Oct 2026
- Learning Jazz Fingerprints Huw Cheston and collaborators · Accessed 12 Oct 2026