Audio AI in 2026: licensed, labelled, and measured

Audio is the medium where generative AI met an organised industry first and hardest. In eighteen months the conversation moved from lawsuits to licences, from silence about machine-made tracks to labels on them, and from novelty to volume. What is left is technical, and it is the same question in three forms: where did this sound come from, what exactly is in it, and does the tool that changed it know what it did?

Volume

Deezer has published the clearest numbers. In April 2026 it reported receiving almost 75,000 fully AI-generated tracks a day, about 44% of daily uploads. By June, in its July 2026 update, the share passed half of all new music uploaded, at nearly 90,000 tracks a day. Deezer says it detected and tagged more than 13.4 million AI-generated tracks in 2025.

Listening tells a different story. Fully generated music accounts for between 1% and 3% of streams on the service, and Deezer reports that up to 85% of streams on those tracks in 2025 were fraudulent and were excluded from royalties. It now removes detected tracks from algorithmic recommendations and editorial playlists, takes down generated tracks used for fraudulent streams, and removes those that have not been played for six months.

Spotify announced its own measures in September 2025, saying it had removed more than 75 million spam tracks in the previous twelve months, adding a policy against unauthorised vocal impersonation and backing a DDEX standard for disclosing AI use in music credits, covering vocals, instrumentation and post-production.

The pattern is clear: generation is cheap enough that the supply of tracks is now limited by distribution, and most of the surplus is never heard.

From lawsuits to licences

The major labels changed course in late 2025. On 30 October, Universal Music Group settled its copyright suit against Udio and agreed to build a subscription service with it, powered by models trained on licensed music, with Universal's artists paid through the settlement and new licences. On 25 November, Warner Music Group settled with Suno; Suno said it would launch licensed models in 2026 and retire its existing ones, and that downloads would require a paid account.

Licensing settles who is paid for training. It leaves a practical problem for everyone downstream: telling, from the audio alone, what a track is and where it came from.

Provenance is harder than it looks

Regulation now requires an answer. Under Article 50 of the EU AI Act, transparency obligations apply from 2 August 2026. The Commission's Code of Practice on marking and labelling sets out how providers mark and detect generated content and how deployers label deepfakes, and the Commission and the AI Board treat it as an adequate voluntary way to show compliance.

The obvious tool is a watermark: an inaudible signal embedded at generation that a detector can find later. The evidence on robustness is sobering. A March 2025 systematisation of knowledge reproduced nine audio watermarking schemes and tested them against 22 removal attacks in 109 configurations, including eight new attacks that were highly effective against all of them, and concluded that none of the schemes withstands every attack. A mark has to survive everything a track meets on its way to a listener, and an attacker only needs one transformation that erases it.

Watermarks still have a place, as one signal among several. But provenance that has to survive the real world works best when it is recorded at the moment of making, by the tool that did the making: what was asked, what was produced, and what changed. Detecting it afterwards from the waveform is a contest that the remover tends to win.

Measuring sound the way people hear it

The same gap appears in tools that edit rather than generate. Source separation, pulling drums, bass or a voice out of a finished mix, is now a standard feature of music software. How it is scored is less settled. A 2025 listening study compared the standard objective metrics with human ratings and found that signal-to-distortion ratio remained the best predictor for vocals, but for drums and bass a different metric, scale-invariant signal-to-artefacts ratio, tracked what listeners heard better. A single number on a benchmark can rank systems in an order listeners would not.

Rooms raise the stakes further. Most recorded sound carries the space it was made in, and much of what people want from audio AI, a voice that sounds as if it were in a hall, a mix that translates to a venue, a stem taken out of a live recording, depends on handling that space correctly. A room's acoustic behaviour can be measured as impulse responses and summarised in quantities such as reverberation time and clarity, each with a known threshold below which a listener cannot tell the difference. That gives a measurement to score against, rather than an opinion.

Exact representation as the common thread

Put together, these threads describe what trustworthy audio AI needs. It needs to know what is in a sound: tempo, key, the parts, the room. It needs to measure what it produced against what was asked, in units that track hearing. And it needs to keep the record of each change so that provenance is a log, not a guess.

Where Livgens fits

Livgens represents the thing exactly, in sound, light, matter and motion, and Living Worlds runs it forward through its lifecycle; for sound, that means knowing exactly what a recording or a room is. Livgens Looper is the artist's collaborator for sound and music, working inside Ableton Live 12 and on stage. It listens for tempo, key, chords and the drum pattern of what is playing, with the room taken out of the way; extracts drums, bass, voice or any other part onto its own track; and generates a part that follows the artist's own recording in seconds, or a whole song. Its instruments include denoising with what it removed shown, and real-time declipping and denoising. Every turn is one undo step, and done means it looked: after every change Looper reads Live again and compares, and if the take is not where it said, it says so. Tempo, deleting and playback each need the artist's word.

For spaces, Livgens Twin predicts how a place will sound zone by zone, as impulse responses you can listen to, and every reading carries its range. The method is scored against measurement, and where it is not yet good enough, it says so. In one registered study, Living Worlds Lab predicted four public rooms it had never seen from twelve measured positions each, scored against the smallest difference a listener can hear, and met its bar in all four, while recording that the average of the five nearest measurements alone still scored higher. Looper and Twin are in early access.