Every restoration is a claim about the past. To take noise out of a recording is to say which part of the waveform was the performance and which part was the room, the tape or the cable. To declip a peak is to say what the signal would have been had the converter not run out of headroom. To match a master to a reference is to say what the music should weigh against everything else that will play beside it. Restoration is representation under uncertainty, and the honest version shows what it changed.
The problem has been redefined
For years, source separation meant pulling stems apart: vocals here, drums there. Recent research has widened that into something closer to restoration. In May 2025, Zang, Dai, Plumbley and Kong defined Music Source Restoration: recovering unprocessed instrument stems from mixes that have been through the real path of production and distribution. The degradations they model are the ones every engineer knows: spectral filtering, dynamic range compression, harmonic distortion, reverb and lossy codecs. Their RawStems dataset covers 578 songs and about 354 hours of material.
The task moved quickly. MSRBench, released in October 2025, evaluates restoration across twelve degradations, including analogue effects and both traditional and neural codecs.
The reframing matters. A stem pulled from a finished mix carries the mix with it: the compression, the bus reverb, the codec smear. Restoring the source means modelling what was done to it and undoing it, which is a much stronger claim about what the sound was.
Separation by description
The other line of progress is in how a sound is named. In December 2025 Meta released SAM Audio, a single model that isolates a sound from a mixture using a text description, a click on the object in a video, or a marked span of time where the target sound occurs, alone or combined. Its paper reports that mixed prompts outperform single ones, and Meta released a benchmark and a reference-free judging model with it.
Together these results describe where audio has arrived. Isolating a part is becoming a request in plain words. Restoring it is becoming a modelled inverse of production. What remains hard is the part that touches artists directly: making the change inside the session where the music is made, and making it accountable.
Looper's instruments
Looper is Livgens's collaborator for sound and music. It works inside Ableton Live 12, and beyond it on the stage. Its instruments, as the site lists them, are all available now:
- Denoise: noise taken out of a recording, with what it removed shown.
- Restore: declipping and denoising in real time.
- Balance: loudness balanced across channels, into a look-ahead limiter.
- Equalize: a fifteen-band equaliser that matches its curve to a reference track.
- Master: equalisation, compression and a final limit on loudness, fitted to a reference.
- Voice: pitch correction now; voice conversion next.
- Generate: a part, a variation, a stem or a removal, from a brief and references.
- Time: a tempo-synced delay, and a reverb of Livgens's own design.
Alongside them sit three capabilities that also run now. Listen hears the tempo, key, chords and drum pattern of what is playing, with the room taken out of the way. Extract places drums, bass, voice or any other part of a recording on its own track. Generate makes a part that follows your own recording in seconds, or a whole song.
Show what was removed
The detail that matters most in that list is small. Denoise shows what it removed.
Every denoiser makes a judgement about which energy is noise. Some of what it takes will be breath, string squeak, room tone or the tail of a cymbal that an artist wanted. A tool that hands back only the cleaned result asks to be trusted. A tool that hands back the removed part as well makes its judgement inspectable: the engineer can listen to the residue and hear whether anything musical went with it.
That is the same principle that runs through the separation research. SAM Audio outputs a target and a residual. Music Source Restoration is scored against the unprocessed stems. In each case, the claim about what the sound was is set beside the evidence.
Match to a reference
Equalize, Balance and Master share a second principle: they work against a stated target. Equalize matches its curve to a reference track. Master fits equalisation, compression and a final loudness limit to a reference. Balance sets loudness across channels into a look-ahead limiter.
A reference turns a matter of taste into a measurement. The artist chooses the record they want to sit beside, and the instrument works towards it. What it produced can then be compared with what was asked.
Done means it looked
Looper's working rules are written on the site, and they are the reason the instruments can be used in a real session.
After every change, Looper reads Live again and compares. If the take is not where it said, it says so. Every turn is one undo step, so a whole run of effects or a whole scene comes back with one press in Live. Tempo changes, deletions and playback each need the artist's word, and nothing fires a clip that was not asked for. For proposals from the instruments, the artist sets a policy once, ask, auto or off, and changes it any time.
On stage, Looper reaches lighting, desks, QLab, Eos, TouchDesigner and Companion through the artist's own machine. Every send asks, and a show lock refuses them all.
Restoration is easy to overdo and hard to notice when it has been overdone. A system that checks its own work against the session, keeps every turn undoable, and asks before anything irreversible is the right shape for a tool that edits recordings people care about.
Where Livgens fits
Livgens represents the thing exactly, in sound, light, matter and motion, and Living Worlds runs it forward through its lifecycle. Looper brings that to recordings inside Ableton Live 12 and on the stage, in early access. Twin, through Looper and Render, carries the same idea into rooms: how a place will sound, zone by zone, heard before it is built, with every reading carrying its range. Living Worlds Lab has registered studies on that, including one in which a room's sound was predicted beyond the range of configurations it learned from, winning 6,846 of 7,869 held-out measurements, and one in a second room where neither registered condition was met. Both are published as they came out.
