What we're really fighting is people scripting computers to generate thousands of songs and uploading them in batches to… — TechRadar

RESOLVING THE LEDGER

AV Insider

AV Insider is our new series of interviews with influential people inside the AV industry. From execs to the people behind the technology, every Friday we'll bring you a new perspective on world of TV and audio.

See the full list of AV Insider articles

Deezer has never been backwards about coming forwards regarding the scourge of AI slop infecting our music streaming platforms. Just days ago, the French streaming site announced that over half of all new daily uploads to its site are AI — up from 44% in April and just over 30% at the end of last year . So, things are only getting worse.

Luckily, the platform not only implements its own proprietary filter to find, label and in some cases purge AI tracks from its site; it also recently made the tech available to everyone , irrespective of your chosen streaming platform.

But how was Deezer's own in-house AI filter developed, why is Deezer so keen on tackling AI tracks, and what does the company think the future of music streaming will look like — you know, for actual humans still trying to make it their career?

I spoke with Manuel Moussallam, Deezer's Head of Research, and Jesper Wendel, the company's Head of Communications, to get their takes on all this and more.

Moussallam agreed to speak with me on the morning after the company's summer party, held at a music venue in Paris. As he was "Checking if the hangover is showing on my face", I couldn't help but think it's good to know that proper live gigs and parties still matter to a firm whose success hangs on our love of music.

(Image credit: Deezer / TechRadar)

Identifying an issue (back in 2021)

How long has Deezer's research arm been thinking "Okay, this AI thing is an issue, and we have to do something about it"? Moussallam explains that the site's comprehensive AI filter, now available to anyone who wants it, has been almost five years in the making.

"It all started in scientific conferences back in later 2021 or early 2022, because the generative models that were already working for images and videos were starting to make real progress — like real scientific progress — for music," he says. "That was probably a couple of years before the first commercial services were appearing.

"So, since we participate a lot in these scientific conferences, we saw people from Google and from Meta sharing these foundation models that were starting to work quite well for creating whole pieces of music. Not just snippets of 10 seconds like before, but really whole songs. So I guessed that eventually, if you're able to generate a lot of music, you can also distribute it quite easily — and down the road, it would end up in our catalogs."

‘We made some mathematical discoveries on how AI models work, and realized that we didn't actually need a big AI to detect AI songs’

Manuel Moussallam

"I would say we started working on the project in 2022, first with some complicated methods, because that's unfortunately a reflex for people like me — to try to do AI to capture AI — but then eventually we simplified a lot because we made some mathematical discoveries on how those models work.

"We realized that we didn't actually need a big AI to understand and to detect these pieces. Just by relying on some plain signal processing and mathematical tools, we were able to actually detect these artifacts for most of the models out there.

"Yeah, I think we deployed the first version of the detector in September 2023 — and we communicated the first numbers when we were able to scale it up to the catalog a few months later."

(Image credit: Deezer)

A big problem — and it's only getting worse

I wonder why, given that various rival music streaming sites have implemented 'Transparency Tags' (read: optional AI badges that place the onus on distributors and labels to tell listeners that what they're listening to was not written or performed by a human), Deezer chose to create a comprehensive filter — and then to release it to all? While admirable and commendable, that's a lot of R&D budget on AI, no? Moussallam laughs. "There are many aspects, but I guess firstly, we did it just because we wanted to see if we were able to do it. On my team, we do a lot of audio analysis. So, knowing was the first motivation."

‘I guess firstly, we did it just because we wanted to see if we were able to’

Manuel Moussallam

"But then, I think since Suno, Udio, and all these tools are trained on unlicensed catalogs, I was pretty convinced that eventually we would have to take down all this music, because it would be considered as plagiarism or copyright infringement in some sense — which may be the case if the legal battles end up that way.

"So at some point we thought that, okay, if people agree with us that it's kind of unlawful to train this model with all this music and they ask us to remove all the content from Suno in the catalog, we need to be prepared. We need to be able to do it. So, that was another big motivation.

"It didn't turn out like that — for now — but it may be that one day people tell us, 'OK, everything that was generated with that AI model prior to this date in the future, when people actually have agreements and compensate musicians for their work, all of this needs to be removed'. And we would have to be prepared for that."

‘One day, people may tell us, “OK, everything that was generated with that AI model prior to this date… all of this needs to be removed” — and we would have to be prepared for that.’

Manuel Mousallam

Putting information into the hands of people who need to make decisions

I mention that across all of my audio streaming subscriptions (not a flex, I simply review audio kit for a living), I tried the Deezer AI filter, and none of my playlists contained AI, bar one: my Tidal catalog listed 1% AI content — and to know what the pesky tracks are I'd need to switch to Deezer. I ask what the reception has been since the release of the software for all. "Mostly people are happy wit...

What we're really fighting is people scripting computers to generate thousands of songs and uploading them in batches to… — TechRadar | AIPROPX