Tavim Blog Open Tavim

Tavim · how it works

A piano with no samples

Why a sampled piano sounds mechanical, why it can never be your piano, and the piano we built instead.

We wanted two things from a piano. It should sound like a real piano being played. And it should sound like yours.

Why a sampled piano sounds mechanical

A sampled piano is a real instrument, recorded one note at a time. Each key is struck alone, in silence, and left to ring.

But in music no note is ever alone. The hammer meets a string that is still moving from the last time it was struck. The strings beside it are already ringing in sympathy. The pedal is half down. A repeated note interrupts itself.

The sampler has none of that. It has the lone note. So it has to cut: stop one recording, start the next, and join them. However good the recordings, the ear hears the joins. That is the mechanical feeling, and no amount of extra samples removes it, because the number of situations a note can be in has no end.

A SAMPLER: TWO RECORDINGS, CUT AND JOINED the first note is cut off here A PIANO: THE STRING IS STILL MOVING the second strike builds on the first Each sample was recorded from silence. In music the same key is struck while its string, and its neighbours, are still ringing. A sampler can only stop one recording and start the next.
The same key struck twice.

Why it can never be your piano

A sampler is one instrument, recorded once, in one room. It can be a magnificent piano. It cannot be the upright in your living room.

What nobody had built

The alternative is a piano that learned from real performances, where every note is heard in its true situation. It needs something that did not exist: many different real pianos, hour after hour, each with an exact account of every note and how it was played. One research project had shown the idea could work. Its sound stayed far from the recordings.

We had the missing piece. Tavim takes a recording apart into the score and the performance, so we ran it over a very large collection of real piano recordings and trained a model on the result. It has no samples to trigger. It learned every note inside real music.

How it plays

Give it the notes, the performance, and 15 seconds of any piano. It plays that performance on that piano.

The test

Sixty recordings the model had never heard. Each system received the same notes, and the result was compared with the real recording.

TavimA sampled grandA modelled (DDSP) pianoAn earlier neural prototype
Notes reproduced, of 10096879574
Sounds closest to its own piano, among 609 in 101 in 103 in 105 in 10
Tone error, dB (lower is better)2.46.35.34.4

Ask each to play a recording back. One of them sounds like the piano in that recording.

Your piano, or another

Play your take back on your own instrument, without the phone and the room. Or move the same performance to a bright Steinway, a Yamaha concert grand, a dark Baldwin.

Try it on your own recording.Free during the beta.
Open Tavim