V2N
V2N · To appear at ISMIR 2026

Video to Notes

Piano, transcribed by sight.

The first complete visual piano transcription system.
It reads the keyboard and writes full MIDI from silent video alone.

Demo

Watch it transcribe, note by note.

A real recording plays on top. Below, V2N's prediction lines up against the ground truth. Pick a performance, and switch between the real audio and V2N's transcription.

Video in, MIDI out
V2N output, replayed in sync
V2N prediction filled ground truth outline velocity soft to loud
video keeps playing, sound switches

Filled bars are V2N, outlines are the ground truth. Where they line up, the model got it right, across onset, offset, key hold, and velocity. Nothing runs live. These are saved predictions, so you can check them against the truth and against the sound.

Why video

Audio cannot tell when a key lifts.

Press the sustain pedal and a note keeps sounding long after the key lifts. Audio cannot tell when the key lifted, only when the sound fades. Video sees the key itself.

Key mechanics. Key Hold, read from video, spans key press to key release. Audio Active, read from audio, extends past it through the sustain pedal.
From the paper. Key Hold (video) runs from key press to key release. Audio Active (audio) reaches further, held up by the sustain pedal, so the two agree only when no pedal is down. V2N reads the key edge that audio never sees, and the same view carries velocity.

Physical key release

On video, you can see whether a key is down or up. So V2N catches the exact moment it lifts, even when the pedal hides that from audio.

Velocity, from sight

The first video-only system to read note velocity, how hard each key is struck, from the motion alone rather than the sound.

When audio fails

Video still works when the sound is silent, noisy, reverberant, or mixed with other instruments.

How it works

One backbone, four heads.

Video in, MIDI out, one second at a time.
head 01

Onset

The frame a key goes down.

head 02

Offset

The frame it physically lifts.

head 03

Key hold

The full span it stays pressed.

head 04

Velocity

How hard it was struck.

V2N architecture. A modified ResNet-18 video module feeds a shared backbone of three Conformer convolution blocks, then four BiLSTM heads for onset, offset, key hold, and velocity.
The full architecture. One second of grayscale video, 25 frames, runs through the S2S[1] video module and three Conformer blocks into the four heads, then offset-guided decoding writes the MIDI.
Results

State of the art, and the first with velocity.

89.5 F1
Physical key releases
PianoVAM +Off · 50 ms
82.8 F1
Note velocity, a VPT first
PianoVAM +Vel · 50 ms
94.7 F1
Onset, on par with audio-visual
PianoVAM · 50 ms
0.04 RTF
28.2 M parameters
RTX 5080 · bfloat16

Read the columns as cumulative. Onset counts a note when its pitch and onset land within the tolerance. +Off also requires the offset to match, +Vel also requires velocity within 10 percent, and +Off+Vel requires all four at once.

F1 at 50 ms on PianoVAM, R3s, and R3x, under our strict offset convention. V2N leads every column, p<0.01 against the runner-up. Parenthesized scores mark a degenerate baseline head, constant velocity for all prior systems and near-zero-duration offsets for Li et al. The full 50 and 100 ms tables are in the paper.
MethodOnset+Off+Vel+Off+Vel
Trained on PianoVAM, tested on PianoVAM
S2S [1]60.121.8(39.4)(14.3)
V2R [2]86.555.4(58.5)(36.8)
Li et al. [3]94.2(21.0)(63.6)(12.8)
PPAN [4]84.945.9(57.0)(29.7)
V2N94.789.582.878.3
Trained on R3s+R3x, tested on R3s
S2S44.017.2(24.9)(9.8)
V2R38.715.8(21.6)(8.8)
PPAN40.714.9(23.4)(8.7)
V2N78.869.557.951.2
Trained on R3s+R3x, tested on R3x
S2S46.614.7(33.5)(10.3)
V2R46.414.9(33.7)(10.3)
PPAN47.713.9(33.7)(9.3)
V2N73.368.559.756.6
What makes it work

What each part adds.

Multi-task heads and multi-frame context. We add each one at a time and measure the gain on PianoVAM, F1 at 50 ms.

Multi-task heads

Predicting every attribute at once lets the heads teach each other. Velocity climbs from a constant-value proxy to real dynamics, and dedicated offset supervision is what unlocks physical key release.

Heads trainedOnset+Off+Vel
Key hold92.881.4(62.3)
+ velocity92.785.780.6
+ onset94.386.782.8
+ offset94.789.582.8

Multi-frame context

Seeing more of the performance at once compounds. Supervising every frame, modeling the frames as a sequence, and widening the window each add accuracy, and the long context is decisive on the harder R3 repertoire.

ModelOnset+Off+Vel
Center-frame loss 0.2 s90.578.577.4
+ multi-frame loss 0.2 s91.883.078.5
+ sequence model 0.2 s94.388.780.5
+ longer context 1 s94.789.582.8

Each row adds one component over the row above. Bold marks the best score in each column.

Open and reproducible

Code, checkpoint, dataset.

@inproceedings{kim2026v2n,
  title={Multi-Task Multi-Frame Visual Piano Transcription},
  author={Yonghyun Kim and Hoyeol Sohn and Juhan Nam and Alexander Lerch},
  booktitle={Proceedings of the 27th International Society
             for Music Information Retrieval Conference (ISMIR)},
  year={2026},
  address={Abu Dhabi, UAE}
}

Complete refers to the full set of MIDI note attributes recovered from video alone: onset, offset, key hold, and velocity. It is not full symbolic notation, so fingering, articulation, and other expressive markings are out of scope. Prior visual systems focus on onset and leave offset accuracy far behind, and V2N is the first video-only system to report note-level velocity.