bandhand.io PL EN early access I want this for my releaseJoin beta

beta · rehearsal lab

Your rehearsals remember more than you do.

Bands overproduce ideas and underproduce finished songs. A riff from a February rehearsal disappears inside a two hour recording nobody will ever play again. Rehearsal Lab searches those recordings, pulls out the stretches where you were actually playing, works out which of them are the same song from different rehearsals, and lets you rearrange them like puzzle pieces.

working, built and tested on our own material planned, and labelled honestly as a plan below

// the problem

You record every rehearsal. You never go back to them.

Two hours from a recorder or an interface is a file with five minutes somewhere inside it that made the evening worth it. The rest is tuning, talking and three attempts at the same entry. Listening back costs as much time as the rehearsal itself, so nobody does it. Ideas are not lost because they were weak, but because no one ever heard them a second time.

// what already works

Eight rehearsals, 514 minutes, 66 seconds of computing.

That is how long it took to process one band's entire archive on an ordinary laptop, with no graphics card and no machine learning model running locally. The tool split those recordings into 112 stretches of playing, separated them from the breaks and the tuning, and scored each one on whether it looks like a song rather than a warm-up.

Splitting into stretches of playing

Silence and energy-drop detection splits a two hour rehearsal into the parts where someone is actually playing. Breaks, tuning and chatter fall away on their own.

yes

Recognising the same song

Fragments from different rehearsals group into songs by tonal profile, meaning by which notes actually occur in them. On the test material five songs turned out to have been played more than once, one of them across four separate rehearsals.

yes

Structure and pulse

For every fragment: tempo, section boundaries, recurring motifs and an energy curve. Cuts can be snapped to bar positions so they do not land in the middle of a phrase.

yes

The best fragments feed an overview: how much playing each recording contains, which songs keep coming back, and which take came out best.

Rehearsal overview: headline numbers, songs with their takes, and a table of fragments

// arranging

You move sections around and hear straight away whether they fit.

Every fragment is cut into sections. You drag them onto a timeline, reorder them, repeat a motif, throw out what does not work. You can name them yourself and rate them in stars, because that is the only rating here that says anything about the music.

Joint preview: the playhead runs straight through the transition.

Joint preview

You do not have to sit through a whole section to reach the transition. One button plays only the tail of one fragment and the head of the next. It is the single most useful thing in the tool.

yes

Fragments from other rehearsals

A verse from March, a chorus from April. If the same song was played several times, every take sits next to the others and they can be mixed.

yes

Trimming without losing anything

You shorten a fragment with handles, and the trimmed part does not vanish, it only dims. You can bring it back at any moment, because the original recording is never touched.

yes

// the test

We tested it against decisions the band had already made.

The band this was built for has spent years pulling fragments out of rehearsals by hand and keeping them in a separate folder. That is a ready-made control set: we know what a human judged worth keeping. We ran the tool on the raw recordings and compared.

Five of the seven hand-picked fragments were found by the tool on its own, and three of those it ranked in the top three of their rehearsal. The two it missed did not come from the analysed rehearsals at all, but from live recordings. This is not "the boundaries look reasonable", it is "it would have pointed at what you picked yourselves".

// from the recording

Bass and guitar written down from the rehearsal, before anyone sits down to transcribe.

The same recording that is already on your drive gives you a tab: note order and frets, laid out section by section. It does not replace your ear and does not pretend to know your timing - it shows what you played, so you stop rewinding the same riff for the twentieth time.

Bass straight from the mix

No stem separation needed. Down low the bass sits more than ten decibels above everything else, so a filter and pitch detection are enough. Across four rehearsal takes the confidence came out between 0.85 and 0.87.

Guitar after separation

Solos need their own track: on the raw mix one percent of the audio is readable, on a separated guitar it is ninety-six. So guitar is computed from cut sections, not whole songs.

Repeats fold up

Four hits on the same fret become one entry, 5x4. A verse that ran to eighty-nine lines fits in nineteen, and you can see the phrase instead of a wall of dashes.

// practice

Click a section and you get only what plays in it.

Lyrics, bass and guitar sit in three tabs above the player. Pick a section on the timeline and all three narrow down to it - and if the guitar is silent there, the tab says so instead of showing you somebody else's part.

Loop and tempo

Loop the chosen section with one click and take the tempo down to half speed without changing pitch. Learning a solo off the recording no longer means scrubbing back and forth.

Notes on the timeline

A note is saved where the player stood and stays on the timeline as a marker. Hover to see who wrote what, click to land on that exact spot in the recording.

Timing stays yours

A rehearsal take does not hold a grid - note onsets fit no tempo between 72 and 120 beats per minute. So the tab gives you notes and frets, and you fill in the rhythm by ear.

// honestly

What it does not do, and will not do.

It does not score hit potential

We tried. A measure of "how close is this fragment to your finished songs" failed its own control test: four mastered tracks by the same band differed from each other as much as the candidates differed from them. There is no shared profile, so the number never reached the interface.

It does not name sections for you

Automatic labels like "verse" and "chorus" come from models trained on finished productions. On a jam without a click track they are wrong often enough not to be worth showing. You type the names yourselves.

It does not generate music and it is not a DAW

Every output file is made by cutting and joining what you actually played. No mixing, no effects, no quantisation, no parts generated by a model.

// archaeology

A ten year archive is not a problem, it is material.

The second band we want to try this on has roughly a decade of recordings: multi-hour rehearsals, first jams, successive versions of the same compositions, the takes that ended up on a record, and plenty that never came out. All of it sits in the cloud, some on old file accounts such as box.net, some in Dropbox, in a folder structure nobody remembers any more.

For the tool that is not an obstacle, only scale. An hour of material costs a dozen or so seconds of computing, so thousands of fragments are a night's work rather than a project. Connecting Dropbox or any other cloud drive is a simple integration, because we read the files one by one and change nothing in them.

A song's history on a timeline

First jam, later rehearsals, demo, album version. You can see when the riff appeared, when the chorus was written, and what survived from the very first run-through.

planned

Forgotten ideas

Songs you have not played in years, and fragments that never got a second chance. Not another batch of new riffs, but a reminder of the ones you already have.

planned

What to finish first

Instead of a list of everything: which songs are closest to done, by your own judgement rather than a number produced by a model.

planned

There is a trap here worth naming up front. Released songs have more recordings in the archive because the band played them more often. An algorithm handed the whole archive will learn to detect "this was played many times" and will get almost everything right, except that this is circular: the song was played many times because the band decided to develop it. For the test to mean anything, the prediction has to come from a single earliest recording, with no information about how many other takes exist.

// what's next

Rehearsal mode.

The laptop sits in the room and everyone has an instrument in their hands. Big section buttons, loop the chorus with one click, readable lyrics and tab on screen, mute the guitar if separate tracks exist. Alongside it a home workbench: loop a fragment, slow it down without changing pitch, write out the tab, record your own part as a reminder. All so that at the next rehearsal the band develops the music instead of trying to remember what they played a month ago.

Beta. We are looking for bands with an archive.

If you record your rehearsals and suspect there is an album somewhere in those files, leave your address. We will get back to you personally. We are most interested in bands with years of archive, because that is where you can tell whether this really works.