SC-112 · Live · Wallifornia
The Real Story Behind Music and Streaming Algorithms at Deezer
Guest: Aurélien Hérault, Chief Innovation Officer at Deezer
Summary
Aurélien Hérault is Chief Innovation Officer at Deezer, where he has worked since the company launched 18 years ago, moving from technical and data work into leading research and development, and for the last five years broader innovation. He explains that recommendation started as simple editorial tagging by genre, but as Deezer's catalog and user base grew, the company built recommender systems around behavioral and acoustic data, since streaming listeners increasingly organize music by mood and situation rather than by genre alone.
Hérault says about 30% of listening comes directly from algorithmic recommendation, though most consumption still comes from a listener's own built-up library, with recommendation acting mainly as a way to feed that library through discovery. Deezer also studies local versus global listening, finding that factors like city size and national radio quotas shape how much local versus international music people hear, largely independent of the recommendation algorithm itself.
Deezer built its own detector for fully AI-generated tracks and found the share of newly delivered music that is 100% AI-generated rose from about 10% in January toward nearly 20%, driven mostly by fraud, bots streaming AI tracks to capture royalties rather than genuine listening. Deezer does not remove this music but chooses not to promote it through recommendation, and Hérault says detection keeps getting harder as generative models improve, so he believes the real fix has to come from changing the industry's economic model rather than only better detection tools.
As of the episode's release on 30 May 2025.
Key takeaways
- 01Deezer's own detector found that fully AI-generated tracks grew from about 10% of new deliveries in January to nearly 20% within months, a trend Hérault calls an industry-wide issue rather than a Deezer-specific one.
- 02Deezer does not remove AI-generated tracks from its catalog but chooses not to recommend them, since the company found this music is mostly used in fraud, bots generating and streaming tracks to capture royalty payments.
- 03About 30% of listening on Deezer comes directly from algorithmic recommendation, but most consumption still comes from a listener's own built library, which recommendation mainly feeds through discovery.
- 04Local factors like city size and national content quotas shape how much local versus international music people listen to, largely independent of what Deezer's recommendation algorithm actually suggests.
- 05Hérault says detecting AI-generated music keeps getting harder as models retrain to leave fewer audio artifacts, so Deezer built its detection system to be resistant to new models without constant retraining.
- 06Hérault believes solving AI-generated spam in music ultimately requires changing the industry's economic model rather than relying on technical detection alone, since detectors keep failing each time generative models are retrained.
Guest
- Aurélien Hérault, Chief Innovation Officer at Deezer
Questions this episode answers
Why does Deezer choose not to recommend fully AI-generated music?
Deezer built a detector for 100% AI-generated tracks and found their share of new deliveries rose from about 10% in January toward 20%. Rather than removing this music, Deezer decided not to recommend it, since the company found it is mostly used for fraud, bots streaming AI tracks to capture royalty payments rather than real listening.
How much of what people listen to on Deezer comes from its recommendation algorithm?
Aurélien Hérault says roughly 30% of listening on Deezer comes directly from algorithmic recommendation. Most consumption still comes from a listener's own built-up library, and recommendation mainly works as a way to feed new discoveries into that library over time rather than replacing it.
How does Deezer detect AI-generated music, and why is it getting harder?
Deezer's detector looks for mathematical artifacts that generative models leave behind in the audio signal. Hérault says detection keeps getting harder because each time a model is retrained to sound more realistic, its artifacts change, so Deezer built its system to stay resistant to new models without needing constant retraining.
What does Deezer think is the real fix for AI-generated spam in music?
Hérault says better detection alone will not solve the problem, since generative models keep improving and detection is always a step behind. He believes the real fix has to come from changing the music industry's economic model, so that flooding platforms with fraudulent AI tracks stops being profitable in the first place.
we believe the solution will not come from technical part, but more on evolution of our economic models
Episode notes
PARTNERSHIP WITH WALLIFORNIA 2025
We explore the evolution of music discovery alongside Aurélien Hérault, Chief Innovation Officer at Deezer.
From early editorial playlists to today’s AI-powered recommendation engines, Aurélien explains why Deezer labels—and withholds promotion of—100 % AI-generated tracks, how it detects fraudulent streaming activity, and the metrics that guide “lean-back” versus “discovery” modes.
You are about to listen to a clear, behind-the-scenes look at the technology and ethical choices shaping what you hear next.
Explore Deezer here; www.deezer.com
About the 9th edition of the Wallifornia Music + Tech Summit in July 2025;
https://walliforniamusictech.com/
___________________________________________________________________________
Sponsored by Allfeat: Decentralized Blockchain Solutions for the Music Industry - https://www.allfeat.com/
Produced by Amplitude Ventures Consulting: Partners in Early-Stage Music Tech - https://amplitude.ventures
Topics
Transcript
Transcribed from the recording by the production team. Names and terms may be misspelled.
Read the full transcript
Jakob Wredstrøm
If you haven't tried it before, is that it streams the conversation, downloads it locally and uploads at the same time, which means at the end I'll get full quality, which also means that it might get a bit laggy once in a while or grainy. You don't need to worry about that. That's not gonna show at the end product. And then also we have some pretty good audio suppression software and audio enhancement. We actually just use Adobe's podcast, which is fine. And it means if there's background noise, you don't need to worry about it. It really just filters everything.
Aurélien
Mm-hmm.
Aurélien
Yeah, I saw that on your YouTube video.
Jakob Wredstrøm
Yeah, it's fun. I have a master's as a tone meister as my first master's. Like audio definitely was my thing. And I can't even be bothered to bring my studio mic for these podcast recordings because the difference in quality is minute. And you can almost argue that I'm pretty sure these models have been trained on this type of equipment. So like, I almost feel like it's better. So it's funny. And like, if people use their... AirPods, it's terrible. There's something like, yeah. So it's, it's funny. You just need to sort of understand how it works. Yeah. Okay. I get sort of ran through most of it. The, one of the most important things that really brings the best worth to listeners is if the conversation is dialogue based. So, I might be asking some questions that, you know, requires a lot of explanation and contextualization.
Aurélien
Yeah.
Jakob Wredstrøm
But if you are aware of trying to break your answers up to like natural breaks where I could maybe, you know, ask the follow up question rather than you answering the follow up question before I ask it, if that makes sense, that leads to the best conversations. Yeah. And sometimes actually it's funny. Sometimes I forget to mention it. And then people who usually very dialogue based goes into like a presentation mode, not in a bad way.
Aurélien
Yeah, it's like a conversation when we have a drink and... okay.
Jakob Wredstrøm
But it just means that the dialogue automatically sort of becomes less of a priority. So if we can just be aware of that, like typically that makes conversations so much better. And my angle is that I think I understand enough about technology to answer like technology questions. And sometimes I'll ask for...
Aurélien
Yeah? Okay. Perfect.
Jakob Wredstrøm
Explanation some certain things that I might know the answer to but that's for the audience sake. So I'll just I always try to take the position of the audience. So you don't need to if you just assume that sounds if you assume that I'm most likely informed enough to know what you're talking about. And if I ask questions that are more stupid it's to lead the audience into it. So you don't necessarily need to dumb down your points. I'll ask for clarification on certain things.
Aurélien
No, no. We talk to experts, we assume the audience is expert too and we don't need to be too generalist.
Jakob Wredstrøm
Well, you're the expert. Yeah, and the audience.
Jakob Wredstrøm
Yeah, most likely. Yeah. No, yeah, definitely experts in the music industry, whether or not they're technologists as well, that's sort of where I'll try to come in and create clarification. But again, you do you, and I'll make sure that's taken care of. Perfect. Do you have any questions before I do the last information? I know there's a lot of information, but yeah.
Aurélien
No, I think if we keep simple as a conversation and just having a conversation with a reaction, then I think, yeah, feel free to ask me any details and I will try to give you as much if it's possible. And especially I have all history in Deezer, so I can start. From where we came with this choice. And it's also for me, I think, important information because we have developed a lot of experience in distributing music in digital world.
Jakob Wredstrøm
Yeah.
Jakob Wredstrøm
Hmm.
Jakob Wredstrøm
Yeah. What's you guys policy on commenting on other companies like competitors? I don't think I'm going to go there, but I just want to be clear.
Aurélien
Usually we don't comment because we don't have details what they do and so for some part we have a comment topic especially for app store or stuff like that but for technical stuff, we don't have so many details so usually we say we don't know.
Jakob Wredstrøm
Yep. Great. Super. Okay, the last two things. So when you're done with the conversation, if you don't leave the meeting, because it just needs to finish uploading. I think with your internet, it's going to be finishing uploading in three seconds. But it's just to be sure that you don't do it before I say it's uploaded. And the last thing is just how we start the episode. We have a standard way of doing it, which is fairly simple. And that is, I'll do a very brief introduction. And when I say very, I mean, very brief introduction. And then I ask you, who are you and what do you do? I'll obviously say that you're Chief Innovation Officer from Deezer and I'll briefly mention sort of the topic today. But then you introduce yourself for 30 seconds, one minute or so. And then I'm gonna ask you about Deezer and then you can use a couple of minutes explaining Deezer. And of course not explaining what a streaming service is, but more like, you know, differentiation and market and that kind of stuff.
Aurélien
Where we came.
Jakob Wredstrøm
That's primarily it, and then we're pretty much ready to go. Do you have water and everything you need?
Aurélien
No? Let's... Yeah, have everything. It's okay. Let's...
Jakob Wredstrøm
Perfect, I'm just gonna pull up the agenda again. And the agenda is, you know, a guiding post, but we'll see. We'll see how we cover
Jakob Wredstrøm
Great. I'm ready. So if you're ready, we'll start in five seconds.
Aurélien
I'm ready.
Jakob Wredstrøm
Great. How do I pronounce your name without butchering it?
Aurélien
Aurélien.
Jakob Wredstrøm
Or really, yeah? Or really, yeah, okay.
Aurélien
Yeah, it's not so easy. And yours is Jacob.
Jakob Wredstrøm
Over to you, yeah. I go by both, Jacob and Jakob. Thanks. Great, we'll start in five seconds.
Jakob Wredstrøm
Hey guys and welcome back to the Sound Connections podcast. Today we have Auréliean in the studio, Let's do it again. I also messed up the name. Orillian. Okay.
Aurélien
Now it's an effect onlying.
Jakob Wredstrøm
Starting in five seconds.
Jakob Wredstrøm
Hey guys and welcome back to the studio. Today we have already Anne in the studio from Deezer. Welcome.
Aurélien
Hello, Jakob.
Jakob Wredstrøm
I'm looking very much forward to a nerdy conversation today. Hopefully that's what I love. And we're going to talk about recommendation engines. We're going to talk about AI and Deezer in general and how you use technology and who better than the chief innovation officer from Deezer to talk about that. Welcome. For the people who don't know you, who are you and what do?
Aurélien
So I'm Aurélien, I'm Chief Innovation Officer for five years now, but I joined Deezer at the beginning, at the creation with the two funders. I was in charge mainly on technical aspect and data aspect and until 12 years now, I'm in charge of research and development mainly and for the last five years, it's more broader innovation in Deezer. That include research and development and data science topic mainly.
Jakob Wredstrøm
And for people who sort of just want to place Deezer in the landscape of streaming services, what does Deezer do and how do you differentiate from other alternatives?
Aurélien
So Deezer is a pioneer in the music streaming service. I mean, we have started at least 18 years ago. It was blog music at the beginning. Yeah, we have built a platform early stage on the beginning of the streaming. And we have invested a lot in research and development to develop personalization, recommendation, and try to understand as much as possible how users consume music and how to provide the best music services.
Jakob Wredstrøm
Hmm. And you're based out of France, right? Yeah.
Aurélien
Yeah, we have been in France, in Paris.
Jakob Wredstrøm
And there's only a few streaming services around the world that you would call large and Deezer will be one of them. And I've also read a lot about Deezer because you've been very public about statements that are sort of pro industry, pro artists, pro rights. And obviously that's one of the things that we're going to talk about today. One of the biggest topics, well, there's several very big topics when it comes to recommendation that's been through the years. One currently, of course now is AI, but AI has been around for a while. But the overarching topic that I've sort of been redefining a lot more is like what music is recommended, like what's the bias, what informs what people should hear because streaming services does have a huge social responsibility because music is effectful, it influences and the services that provides this consumption, it does have a say in how it gets shown and selected. Maybe we can set the scene with trying to understand, technically, how music is recommended and how Deezer approaches it.
Aurélien
Yeah, definitely. It's a really interesting question because for the last, I would like to say, 15 years, that evolved a lot. mean, at the beginning when I joined Deezer, it was pretty small catalogue. So, recommendation of music was pretty simple to do by editorial and people just choosing and animating the platform. But... with the democratization of streaming, the catalog is growing. We have stored a lot of data and behavior and to be able to recommend music, we have invested a lot in technical part because not only because we want, but because a platform and if we want to ensure... good understanding of user experience, it's important to learn how to build recommender system for that. And especially because in music, everyone have their own test. And we discovered that also. It was pretty easy 20 years ago, you have JOR, you have Rockfan, Reggaefan, Popfan. and not so much.
Aurélien
Mixing jar. But with the streaming, have also discovered people is not organizing music as we did 20 years ago. It's more like situation, mood. So that changed completely the way how we recommend music, how we dig into the catalog. So in one part, you have a big catalog. In other part, you have a big user base. And between you have a lot of data and to be able to have good recommendation, need to build system using all this information. And as you see, you have some responsibility and ethics and bias. And it's important when we build this kind of system to be able to study what you recommend. And that's why...
Jakob Wredstrøm
Hmm.
Jakob Wredstrøm
Mm.
Aurélien
In early time in the research and development department, we have made partnerships with laboratory universities and provided our data to study also our bias to be able to correct.
Jakob Wredstrøm
Yeah. think especially in the times now where the world is turbulent, maybe less so with music streaming services, but there's really a lot of focus on suggestions. It's really about how people have fed information because it can be either directly, indirectly weaponized, culturally speaking, practically speaking, democracy, but also
Aurélien
Hmm.
Jakob Wredstrøm
One of the concepts that I keep hearing coming up and I haven't formed an opinion about it, but that's sort of localization. How local regions get more and more of their own content, which is great for many different scenarios, but also have a lot of downsides. What is your experience with this whole movement when it comes to exposure versus...
Aurélien
Mm-hmm.
Jakob Wredstrøm
Protecting local language music and like that whole dynamic, like what's these thoughts about globalization and also, especially when comes to cultural exposure.
Aurélien
Yeah, definitely it's really interesting question because that depends a lot also on the country. You have some country like France for many years, have Cota in radio protecting French music ecosystem and you have some other country really open to other country but what we saw in the data, user consume a lot of local music probably because it's a
Jakob Wredstrøm
Mm.
Jakob Wredstrøm
Mm.
Aurélien
Cultural aspect more than just be influenced by algorithm and also what we saw in recommendation and suggestion we have a lot of external factor. It's not only the platform. What we saw in some, for example, in some study is the size of the city change the volume of
Jakob Wredstrøm
Mm.
Aurélien
Discovery and the volume of artists you will consume in one year, for example. So it's completely external to the platform, but it's important to know that. So that's why also we work closely with our editors, because our editors is linked to the local scene, is linked to the local labels, and we can also help to animate with algorithm.
Jakob Wredstrøm
Hmm.
Aurélien
the platform. And you have also some country investing a lot in exportation. mean, K-pop, it's worldwide phenomenon, but it's also a politics engagement of this country to promote this kind of genre. So you have really different situation and we try to to underline, understand
Jakob Wredstrøm
Mm.
Aurélien
Almost a lot and build the features for users to be able to have access to the local but also to international.
Jakob Wredstrøm
Hmm. Because the recommendation obviously works in many different levels. There is algorithms, AI, these kinds of things that build on data and I guess some sort of guidelines from you, but then there's also editorial. And then of course, that's what people choose to listen to directly. You don't necessarily need to provide the numbers per se, but how big of a part of consumption are those?
Aurélien
Mm-hmm.
Jakob Wredstrøm
Generally speaking, three categories, like algorithm-based consumption, editorial presentation, and active choice. How much do they matter to each other when it comes to the size?
Aurélien
I think it's 30 % of the consummation come directly from algorithm. What we saw, it's mainly people listen from their own library. They build their own library. And finally, the recommendation, it's a way to collect. And it's a back and forth between exploration with editorial recommendation.
Jakob Wredstrøm
Hmm.
Aurélien
But at the end, the most important is their own library. At the end, if they discover something through your recommendation, that will finish in their own library and they build their own of music.
Jakob Wredstrøm
Make sense, okay. So,
Aurélien
And just, I would love to tackle a topic in term of recommendation. For many years, we have only used user data, but since 10 years, we use also acoustic data. I mean, we analyze also the composition of the music through algorithm also. And that change the way how we recommend music.
Jakob Wredstrøm
Hmm. Yeah, no, that makes a lot of sense. And I guess that's also where the... the interesting matches happen where like, okay, this is maybe not within your usual listening environment, but this has the same characteristics of what you're used to or what you might like, and therefore we can present in different ways. So I guess it gets a more nuanced recommendation system. And AI obviously... is a very big part of that. One of the things that you guys made out and made a statement on is obviously the influx of AI-generated music is very large. Four months ago, I believe that number to be 10,000 per day. I would assume it's higher now. And you guys have made a statement that if you track that this is AI-generated, I don't know to what amount that definition is, you choose not to recommend it through your recommendation engines. Could you explain to me a bit more what the reasoning is for that? And why is that important?
Aurélien
I think it's important to come back just before AI. What we saw for five, six years, it's really fast growing catalogs. Fast growing catalog, it's a lot of computing in term of ingesting new content, recommendation, stream manipulation. So a lot of topics is linked to the catalog because the catalog is a Earth of Deezer. It's really the center. We build features around our catalog. So it's really important to take care of what we have in our catalog, what we know about our catalog, to provide the best experience. And we have an anticipated AI because we saw and because we have a research and development team, we monitor a lot scientific paper and we saw the... how fast that go in terms of music generation. But our first main problem is we don't know how to identify that. We can talk about AI music, but how it's possible to say it's a problem or not. So that's why we have focused for one year and a half building tools to be able to detect 100 % Gen AI. music. We have released that in December last year and we were shocked to see the volume of delivery of music AI generated and we have started to communicate about this number because it's not a dessert issue. We have all the same catalog so it's an industry issue. And before to say it's an issue, we prefer to say, okay, there is a number, there is a fact. We have 10 % in January. We have almost and close to 20%. Now it's a little bit less. Yeah, every day. It's a little bit less, but it's 18 % today. So what we saw, it's few months.
Jakob Wredstrøm
20 % of all songs uploaded. That's insane. Yeah.
Aurélien
The number of delivery in terms of AI generative music is growing. So where we go? What we do with that? So first things, because we saw it's not a lot of consummation, we have decided to do not promote this music. That not mean we have removed from the platform. It's completely accessible. We are not here to do the police. We just say... Deezer don't want to recommend this kind of music. And we have started also to study how this music is used. And what we discover it's mainly used in stream manipulation. So in fraud behavior. So finally, you have some bots generating music and listen by other bots just to capture. And for us, we estimate it is an issue for the industry.
Jakob Wredstrøm
I can imagine, that's been a big talking point, like these sort of fraudulent streams. And I had a conversation with Andrew Beatty and Conraso a few months ago, and they presented some numbers that I haven't been able to find an academic confirmation of, but I also did talk with some professors that said that's probably not far off. And that is, I think the number he said was 10 % of all streams can be marked as fraudulent. Globally according to you know billions of songs they've analyzed so you know and you look at how much money there is in streaming in general that's an insane amount of money used for criminal activity used for you know so many things that shouldn't be there and it is an issue and obviously you guys addressing the I generated things that also I would say most likely also helps you see whether or not something maybe is from that work when you don't have the recommendation associated to it. Helps you also tackle those aspects, I guess.
Aurélien
It's important to take care of it because when you have recommendation on JANG and you don't take care of your catalogs, your recommendation on JANG can promote and accelerate and so you see what I mean. It can be dangerous and really accelerator of bad content or fraudulent content. That's why we have in one port
Jakob Wredstrøm
Hmm.
Aurélien
A strong system in terms of fraud detection and that's why we have some really specific system to analyze on the audio what kind of content we have. Yeah, technically speaking and I think I anticipate your question about AI detection, no?
Jakob Wredstrøm
Hmm. Technically speaking, sorry, sorry.
Jakob Wredstrøm
Yeah, go for it. Yeah, it was.
Aurélien
Gen.AI is using statistics. We call AI is machine learning. So it's discipline, but at the end of the day, it's neural network, you have different kind of model to generate, but it's statistics, it's mathematic. And what we have observed in the signal, all these kind of system leave some artifact in the signal. Artifact in the signal and that helped us to build models to detect this kind of artifact and it was a pretty good success at the beginning and We saw some limitation I mean when you retrain a model to generate music the artifacts change and our model is failing so we have focused our energy to have a system resistant to this modification. So that means if new models come on the platform, we are able to detect it without retraining our own system. Sometimes we need to do because the modification is too large and the modifications to really big change.
Jakob Wredstrøm
Hmm.
Aurélien
But for small modification or medium modification, we are able to detect it. And I would love to precise also we have published paper to talk about limitation of detector of AI. And why we do that? It's because we believe the solution will not come from technical part, but more on evolution of our economic models. It's another topic, but...
Jakob Wredstrøm
Mm.
Aurélien
That's why also we have this transparency about our own system because the industry needs to know we are able to capture a large amount of this kind of content but not all. It's really conservative.
Jakob Wredstrøm
Mm.
Jakob Wredstrøm
Yeah, because it is worrying. know, I've been very much
Jakob Wredstrøm
understanding AI song generation and I've sort of stopped practicing a lot the last six months but you know in sooner I've made at least 10,000 generations like you know I've really tried to understand what's going on and people might have opinions about that or not but that's not so important for me it's more to understand sort of what is happening and obviously video has moved very fast maybe so also faster than audio generation lately, probably because of how well it's financed. But as Google Vivo 3 came out a few weeks ago, and I start, I'm a big TikTok guy, like I get a lot of educational content there. that's, you we actually learned most of the stuff I do. And I've now totally changed the way that I consume content there because like, I know that I cannot be sure anymore that these sort of... videos that I see where people are speaking about a topic is real and even that the people it's supposed to be is that person. And I think there's a lot of foresight to it when it also comes to music. you know, again, we don't need to go into specifics, like, Suno 4.5, which is the newest technology they have out, like if you've prompted really well and you prompt, let's say, 200 songs, there's one or two there where you need to be very well versed. in music to understand that, hear that this is a generator. And I've also experimented with techniques myself where you actually take, you know, one version and you remaster it many, many times and you edit it together. So like all these, at least audible, visible artifacts disappear. And you can actually now, right now, create music where I have, you know, five years conservatory degree in music, cannot distinguish it. And sometimes when I hear songs on the radio right now when I'm driving my car and I hear like a human produced song, I feel like this sounds more AI generated, this song, than some of the songs that I know is AI generated. So we've right now come to a point in music where the human ear, even from an expert perspective, you're not guaranteed you can hear the difference. And that is a change in consumption that we've never experienced before.
Aurélien
Yep.
Aurélien
But it's how this model is evolving. mean, if you take models two years ago, it was really easy to detect, okay, it's AI-18 music. But more they train and retrain and their target is to be the most realistic as possible. I'm pretty sure now if you make a blind test and listen different music, it's really hard to make difference. And it's more and more difficult. And I had exactly the same approach on TikTok. You have some video and say, okay, it's true, it's not true, it's difficult and more and more difficult. And that's why at Deezer we bet on transparency saying,
Jakob Wredstrøm
Mm.
Aurélien
We build system, we change and we try to evolve the economic model and we give transparency to users. It's AI generated or not. We want to focus on 100 % generated music because we are not against AI. I mean, if an artist want to use it as a part of a composition and you have artistic approach. Why not? I mean, it's a tool. Our main concert is more the spam, saturating catalogs, trying to get money from real artists. And it's our main concert.
Jakob Wredstrøm
Yeah, makes sense and it's important. Music is a difficult place to thrive in the first place. And if you need to start competing by people with a totally different agenda than the fan interaction, it gets really complicated on top of that. But again, you want to look a bit about the initiatives, again, back to both Deezer and TikTok, there's the AI-generated sort of... that people can see our creator label as a. Yep.
Aurélien
Yeah. But we have a difference. It's not our own catalog. I mean, we receive catalog from others. So we don't provide the tools to create music. So we don't have the guarantee it's AI or not. That's the main difference with other platforms who provide the way to film and where they have more way to check if it's AI or not.
Jakob Wredstrøm
Yeah.
Jakob Wredstrøm
Mm.
Jakob Wredstrøm
Mm.
Jakob Wredstrøm
That's true. Absolutely, I do see you, Wing. And it's complicated, but my general comment is that I, as a consumer, I do experience that there's some sort of protection coming in place at large, but I'm also very much in doubt on whether or not that actually is true at scale. What's your sense of this? Is this being implemented to protect the consumers?
Aurélien
You see what I mean?
Jakob Wredstrøm
streaming service like these, you focusing very much on it? Like how successful are you and how do you believe the general industry is approaching this?
Aurélien
We try to provide to the users some protection by transparency, but also for creators. I mean, it's really important for creators to feel and know we are building safe place for them. And for users, for example, we plan to labels, so really have a notification on our product when it's something coming from AI.
Jakob Wredstrøm
Mm.
Aurélien
And for artists is more knowing on our platform we put their efforts to protect their creation and for industry sharing our numbers and our study. That answer your question.
Jakob Wredstrøm
Yeah, it does, it does, it definitely does.
Jakob Wredstrøm
What do you think is the responsibility Stream Platforms has of all this? Like AI is changing the world, not just in consumption, but also in energy, in resources. It's a, everything is tearing on what we already five, 10 years ago were very heavy proclaiming to be, you know, costly for the world. What's your thoughts on this?
Aurélien
I think we have really strong responsibility because we are the crossroad between industry creators and users. So we unique view on catalogs and what's happening globally in the music industry, how it's consumed and stuff like that. And the particularity of music, we have the same catalog on each platform, almost. We have almost. When we come back to AI and we saw we have almost 20 % of AI generating music every day, not consumed, we store that. We know our competitor having the same problem. And in term of responsibility, we have first one question, do we need that? I mean, it's not listening, it's not a good value for users that... cost us a lot in terms of storage, terms of investment, in terms of exposing us to stream manipulation. So at the end, I think we have responsibility to take action, but not alone with all industries. As I said just before, it's not a teaser issue. It's global issue. What do we want in music catalog for our users, for our creators and right holders. And we raise the question because it's really important for you to know almost all researcher or developer are musician at Deezer. We really care and we will love music and we see AI is starting to change the game for creators. Sometimes in the good way because it's a really strong tools to help on creation and stuff like that. But in other hand, it's so easy to industrialize fraud or generating music, but didn't bring any value to the users. We should raise this question and answer altogether as industry.
Jakob Wredstrøm
big question that's sort of tied to that, and I touched upon it earlier in the interview, is... it's quite political of nature or you know It's what music is recommended and how big a thought process is there in what is recommended. Obviously the way that I understand it, which might be incorrect, is that recommendations are built on user behavior. But there's also, again, the lack of better terms because I don't know the technology that well, that there must be a system prompt behind it. There must be some sort of like human made logical preference in how these work. And back to my point of music having incredible power. Those decisions matter. How does Deezer think about that?
Aurélien
So you're right, but we need to have a history of recommendation. Few years ago, it was only tags and really broad description of music and more we have technology but also computing power, more we can go to the details in term of describing music. I mean, at the beginning it was rock music and now you have the label works and what we call embedding. It's really a vector describing music mixing different information, not only user behavior, but also acoustic models, also location, time of the day. You have more and more complexity of describing music and we try to go deep and deeper in terms of granularity.
Jakob Wredstrøm
Mm.
Aurélien
The first thing we do is building a space like a universe with galaxy of tracks and we have clusters and stuff like that based on all this data. And the second thing is to put the users at the right time in the right moment in this space to recommend music. It's how we build our system and come back to the responsibility. it's important to monitor the space. For example, for a new artist, how he came into this space. If it's in the right place, what is the main risk when you don't have huge popularity? Because what we observe, for example, for a new artist, it changes a lot of place in this space because it's not stabilized in particular galaxy.
Jakob Wredstrøm
Hmm.
Aurélien
And it's really important because what we observe that can change the way how it's recommended and sometimes it's good if it's in the right place and sometimes it's bad if it's not in the right place because people will skip it that will reinforce the system to say okay it's not the right track to recommend but at the end the problem was not the track it's just the place on the space.
Jakob Wredstrøm
Hmm. Yeah, that makes sense. So if I can understand it correctly, is the logic more that it's music placed in galaxies that the users are, or is it users that are placed in galaxies? Like, what's the best way to look at it?
Aurélien
On both sides, that depends. But if you put the users in the bad place in the space, recommendation will be not good for you. But that's happened also on the artist side. I mean, that's why we care to see how this recommendation space evolves. And that's why also we build different.
Jakob Wredstrøm
Okay.
Aurélien
recommend a space based on the objective of the recommendation. I mean, if you are in lean back experience, you just want to listen music you like without so much discovery, we have a specific algorithm. If you want to discover, we change parameter and we reinforce discovery. And finally, we don't have only one engine. but different engine based on different objectives you have when you came to the platform.
Jakob Wredstrøm
And what are the parameters that you look at when you say this is a good place to be? Is that engagement, retention, these kind of things?
Aurélien
you have different and also that it's already depends on the which kind of algorithm you use. If for discovery, we will must look at the collection rate. I mean, on the discovery mode, people don't listen completely the tracks, they just want to collect and... you will look at this matrix for lean back experience is more listening time. How much time you spend, do you reconnect? So skip rate also. If we have a good recommendation on lean back experience, normally you don't skip. I mean, if you start to skip in this kind of experience, it's because the recommendation is not good. So based on the objective of the algorithm, we will look at different metrics to ensure it's the right way to do. It's the same for search engine.
Jakob Wredstrøm
Yeah, so is that data, these sort of data points associated to each track is like, okay, someone skipped, so now that data is associated to that track. Is that like how it works that, you know, those tracks that's pillars and then there's lot of data built around them? What's sort of the technical approach to destroying that?
Aurélien
So you have different, this kind of metrics will be attached to your profile, but also to, will unpack also, as you say, a data point on track level, on album level, on artist level. I mean, you don't have only the tracks. We need also on the album, on the playlist. So all items will be impacted by user behavior. and often we resume by popularity because it's a simple matrix, it's a ranking. But in the competition of popularity, have sub-metrics will impact this data point.
Jakob Wredstrøm
Mm.
Jakob Wredstrøm
So it is incredibly, incredibly complex. And I think that's also... The difficulty in all of this is if you guys want to, again, I'm just assuming, if you guys want to implement the policy such as we won't recommend this, it's not just like doing a simple change. It's not just like writing a rule and then that happens. Like there's so many factors you need to take into consideration when you're of playing around with steering certain things, I guess.
Aurélien
It's a multiple dimension. I mean, I will take a simple metrics ranking. Ranking, you have ranking for tracks, album, playlist, artist. And after that you have ranking by country because you don't want to mix and you want to keep local popularity. So I mean. Just with one data point, you see the number of dimension and we have thousand of data point. So it's really a complex system and that's why it's important to do research because we know we can have some basis and we can see it and it's important to understand how we recommend music and that's why we have researcher dedicated to explainability. How we can explain why we recommend this track to these users. It's really important to do this research.
Jakob Wredstrøm
One of the things that people have complained a lot about in general with streaming, some very knowledgeable, some not, is the difficulty in breaking through as an artist with streaming. And well, that's always been the case. But now we're also seeing a lot of fan engagement, at least initiatives happening in, of course, in Asia, this has happened for a long time. But how is this affecting streaming behavior? Like, can you see that people are getting more engaged or are you not seeing that in streaming consumption?
Aurélien
Here you see some communities really engage around this artist and that's why we look deeper in fan engagement because with this, it's not the case for all artists but some artists have really hardcore fan and when you have a new release you see specific focus and the weight.
Jakob Wredstrøm
Mm.
Aurélien
midnight to stream and you see a pics. yeah, fan engagement is really important. But it's not for all kind of music or artists. But it's also I think linked to the tools available on internet now with TikTok, Instagram. A lot of artists are also influencers. They know, they do videos, they engage their community, they invest in their community in other ways than 10 years before. So, yeah, you have different kind of typology of artists.
Jakob Wredstrøm
This has been incredibly interesting and as a partnership episode together with Wallifornia, you'll be there as well. And I look very much forward to conversation with you in person. But what I value the most in this conversation is the positioning you guys have on behalf of the industry. I think I know for a fact that streaming is really, really difficult and it needs to be built at scale. So having a company that takes a positioning and also having a ethical standpoint is very important because it is a obviously a for-profit venture but it is so incredibly cutthroat competitive and that is so evident. So I think it's important and what I really hope is that technology at large will be very public about these things. It is very much missing and it might also just be drowned out. in all the announcements of technology. But again, I follow AI so closely and we have a lot of developers. We also have several PhDs on a team that has a focus on this. And it's for the first time in my life where I'm actually maybe scared is a big word, but worried. this is so different from what everyone knows. And I was listening to an interview with... some of my favorite authors and thinkers a few days ago and they just kept on coming back to whoever says they know what the future holds. They just don't know. Like, it is so un-
Aurélien
But I'm worried too, but I see also an opportunity for some company like Deezer, but it's available for and true for some other sectors to say, okay, we can do differently. We have a choice. We can invest also on technology. We can have a different approach and we can propose to the users and creators. a different way to handle AI. I mean, we are not against AI. We just want to use carefully and with an ethical approach. And for me, I see that as an opportunity and also to bring value in this world of AI.
Jakob Wredstrøm
Hmm.
Jakob Wredstrøm
Mm.
Jakob Wredstrøm
Well, thank you for the talk. It's been a pleasure and we'll see each other at Bluffonia.
Aurélien
Thank you. For sure. Thank you.
Jakob Wredstrøm
Before we stop the recording, actually, because I just need to redo something from very beginning. You actually don't need to be involved. I forgot to mention at very beginning that there was a partnership with Wellifonia. So if you just stay on the camera, I won't show you. I'll just say it at very beginning. So again, I don't want to put your name. Arlen? No. Yeah. Great. OK, so I'll just do that. So if you just stay on the camera, I'll give you two seconds.
Aurélien
Aurelien. It's...
Jakob Wredstrøm
Hey guys and welcome to the SoundCarnation Podcast. Today we have a partnership episode with Wellifornia coming in around a month and we have Orlan in the studio, welcome. That was it, I just need to get that on camera. So I had it instead of the one I did. Perfect, well I'll stop the recording.
Aurélien
to be.
