Hand-v2 pilot · phrase forecasting
Cadence
Play a phrase. Generate its continuation. Inspect both hands.
Example 01
First 10 seconds: supplied RH. After 10 seconds: generated RH + LH.
This is RH-prompt forecasting, not live following. The model does not hear what you play next. Short endings and weak examples are retained — no cherry-picked replacement takes. Original clip 04 is featured first at Hugo's request; original IDs and recordings are unchanged.
What has — and hasn't — been verified
391 validation windows share identical future targets. These diagnostics support conditioning sensitivity; they do not establish musical quality. The empty-prefix variant is also an out-of-distribution perturbation.
All eight examples contain newly generated RH and LH; their MIDI and MP3 files passed structural checks. Seven ended with model EOS rather than the horizon cap. A note-score review found promising harmonic continuity as well as repetition, dense textures and imperfect hand separation. It was not an auditory listening study.
The reference is an actual withheld continuation, not a model baseline or the only musically valid future. Hand labels are predicted, and this fine-tuning split is not proven unseen by Aria pretraining.
Checkpoint SHA-256: 29e8e1b…7eca3b0
Play a right-hand phrase
Record up to 10 seconds of right-hand notes on the virtual keyboard, then press Generate to send that exact phrase to the trained checkpoint once. This is a single turn-taking request, not live accompaniment: the model never hears anything you play after you press Generate. There is no offline fallback — if the endpoint is unreachable, this page says so instead of showing a substitute result.
Computer keys — white: Z X C V B N M = C4–B4, Q W E R T Y U = C5–B5. Black: S D G H J = C#4 D#4 F#4 G#4 A#4, 2 3 5 6 7 = C#5 D#5 F#5 G#5 A#5. Labels are also printed on each key; pointer and touch play every key too. This is a plain browser oscillator preview, not the trained model.
Idle. Press Record, then play up to 10 seconds on the keyboard above.
Your generation
Fresh result from the live endpoint — not one of the eight fixed examples above, and not cherry-picked. No withheld reference exists for your own recording, so there is no Reference track here.
Live accompaniment — candidate
candidateB200-measured timingone offline take listened & acceptedAn experimental streaming session: play the right hand continuously and the candidate checkpoint schedules a left hand back over the same connection, acknowledged only once your own audio clock actually crosses each onset. This targets the interaction-v1 candidate documented in Candidate evaluation takes below — evaluated privately over SSH on a rented RTX before it reached this serving path. Experimental, one session at a time: the release serving this page (v0.0.4) answers both the legacy Generate flow above and this stream, but never both at once — while a Generate is running, Connect is refused with busy, and a live stream makes Generate wait. Timing and musical quality on this platform are measured, not promised; the status line below says exactly what the model did.
Computer keys drive this keyboard while a session is open (before Connect, and after End, they drive the recorder above): white Z X C V B N M = C4–B4, Q W E R T Y U = C5–B5; black S D G H J = C#4 D#4 F#4 G#4 A#4, 2 3 5 6 7 = C#5 D#5 F#5 G#5 A#5. Hold a key to sustain, Shift for an accent, several keys for a chord. Pointer and touch work at any time.
Idle. Press Connect to open one streaming session — nothing connects on page load.
While a session is open this keyboard owns the computer keys, so the recorder above cannot record what you play here; End hands them back. Both hands are voiced by the same piano-like synth on this page — your right hand a little brighter, the model's left hand a little softer and marked on the keys — so what you hear is the model's notes, not the instrument. Experimental — a different checkpoint from the fixed examples and the recorder. Timing was measured on this deployed B200 serving path over the real data channel on 2026-09-19 (a 134 s scripted right hand: 524 requests, 0 deadline misses, 0 missing acknowledgements, 35 ms round trip), after the same numbers on a rented RTX. Listening acceptance so far covers one offline re-rendered live take (see the gallery below); the live stream itself, as served here, has not been listened to yet.
Candidate evaluation takes
Loading…