Voice AI’s Big Moment: Why Everything Is Changing Now (ft. Neil Zeghidour, Gradium AI)

Hey, thanks for having me.

I think it's both having a big moment and we're still early. It's having a big moment because there is progress all around AI models, and in voice, for example, the progress in latency, naturalness, and accuracy has been really, really huge in the past years, in particular in the last two years. And at the same time, text models have evolved into what we now call agents, which are not only text models, but they can actually take actions and manipulate data, access information, and so on and so forth.

And now, when you bring both together, you can have voice interfaces that at the same time are going to solve complex problems. And so I think there is a moment now because, for the first time, it actually can be enjoyable and even more convenient to talk to an AI on the phone than talking to a human, because you can call any time of the day or night, and the interaction is working pretty well, and it sounds really nice, and the latency is low, and so on and so forth.

So it's definitely having a moment because I think, in a way, now it can be used in much more use cases than it used to. But it's still early because it's still quite experimental. So anybody who is using even the most advanced voice agents and compares that to the movie Her from 12 years ago, it's obvious the gap that is still remaining. And there are so many topics that are completely unaddressed at the moment. In particular, every time you watch a voice agent demo, just realize that it's someone talking to a phone in a quiet room.

So the day where you will have someone shouting to a robot in the middle of a factory, and having the robot understand what's happening and who's talking to it, that will be when we'll be there, and we are not there at all.

I don't want to be mean to my people, the speech scientists, but historically, for some reason, voice did not attract the visionaries in machine learning, right? So even if you looked at the dynamics in conferences, if you proposed a new method, like a fundamental algorithm, and you wanted it to be accepted in a prestigious venue, you had to have an application either in computer vision, like image classification, or NLP. If you did it in speech, you would get rejected because it was, like, too speech.

And at the same time, the prestige of speech conferences used to be much lower than that of computer vision or NLP. So honestly, I don't really know why, because when you look at the details, the first big success of deep learning—everybody knows the AlexNet model in 2012, where for the first time, you had a deep learning model outperforming every single alternative on image classification. But actually, the really first big success in deep learning was speech recognition with work from Geoffrey Hinton himself.

It was in, I think, 2007 or 2008.

Yes. And so I think it was just not as prestigious, and so it wouldn't attract a lot of people that could have made significant contributions, I would say. And then what happened, and was very nice, is there was kind of a convergence of algorithms around transformers and LLMs, such that now, pretty much regardless of the task or modality you are looking at, you're always looking at the same technology. And there started to be much more progress, also thanks to now the similarity between different modalities.

Because in particular, what I contributed to with my team was to take a lot of inspiration from successes in vision and NLP and apply them directly to speech. But it's still interesting because, in a way, there are way fewer people who can train a competitive speech model than in text or in vision. So, I mean, it's a good position to be in because very few people have really gone into the depth of this topic. And it's one that is very challenging because it's bringing—ideally, if you want to solve the problems—you need to understand machine learning, signal processing. There is a much more completely different literature around telecommunications, audio compression, and so on, along with psychology—not cognitive psychology, but psychoacoustics.

How does human hearing work? How does speech production work in humans? So all of this. When you bring them together, you can make competitive models. But it requires a very wide scope of expertise in very different domains.

Between 10 and 100? No, I would say 50. I don't know. It's hard to say. Tiny. I think it's very few. And really meaningful contributions that have pushed the field forward have been made by very small groups of people. And I think that's also what's nice. So AI, I think, in general, is one field where individuals can have a disproportionate impact because the amount of things you can do by yourself—you have access to compute and datasets—is huge.

And in voice in particular, since the required compute is much lower, and that's the same for data, really a few individuals can make stuff that is completely changing applications at very large scales. Great.

So latency. Latency is already something that only makes sense if you are in a turn-based conversation, because then the definition of latency is how much time there is between two turns. One of the things we contributed to is getting rid of speaker turns completely with what we call full-duplex conversation. So the idea is that it's always listening, always speaking, and when it's not speaking, it's just that it's producing silence. But it's always on. And in that context, there is no real latency anymore.

Because the model can basically talk at any time, and it can talk over you, and you can talk over it. And that makes the conversation really natural. Then naturalness, it's not only these dynamics in terms of tempo, like when the model can jump in the conversation, when it should remain quiet. There is this dynamics question. And then there is emotion. And so there is emotion in what the AI expresses. Its emotion is natural, but also appropriate, that if you start feeling confident enough to start sharing about stuff that makes you unhappy or sad or feeling miserable, it's not saying, "Oh, I'm so sorry for you, let's talk about it."

And it should also understand when you're getting upset, when you're getting confused, and so on and so forth. This will make voice AI, in terms of interaction, extremely natural and as close as possible to human, which is basically one of the things we see in the Her movie, which I hate mentioning as a reference because it's so overused. It's annoying. But at the same time, everybody understands the gap between where we are right now and the movie. So I think it's still a relevant one.

And it's even interesting how relevant it still is despite the fact that there is so much work around voice. But then there will be other questions about how voice is integrated into our lives. So there are paradigms in voice AI such as wake word detection. So when you use Google Home or Alexa or whatever, you have a wake word that is going to turn the speech-to-text on. So now let's say you want to work with your assistant that is always listening to you.

So in a way, you would have something that is just running constantly without even necessarily having a wake word. So all of this, I think, is going to be both technical challenges and product challenges around where they sit, how we're interacting with them, the link to the hardware as well. So I think what is a good sign for voice as a field as well is that, in my perception, all the new hardware companies have voice at the heart of the product.

All the prototypes that we see, whether it's glasses or pendants or the new stuff that Jony Ive and Sam Altman are working on, voice is at the heart of the product and will be the main way of interacting. So all of these devices, they got rid of keyboards. They don't really have a screen or an interface, rather. And voice is going to be the main one.

So, for example, I used to think that one obvious application where voice was kind of irrelevant was coding, because fundamentally, you're not going to read code out loud, right? Yet now, since coding is going more and more toward vibe coding, which is natural language, it makes a lot of sense to do it by speaking. And now people are developing products that allow you to dispatch orders to coding agents in a way that is much more efficient than if you had to type in each different window to each of them.

Even prompting LLMs now by voice is much more convenient than typing. I still agree that there is one part which is more social about what the office environment will look like. I don't know, maybe we will also rethink the way we structure office environments. What is sure is now people have AI assistants that are almost colleagues, right? I mean, you talk to any software engineer, the anthropomorphization of Claude Code is—I find it extremely funny. Even the verb coding, I mean, it's going to be Claude-ing pretty soon.

And so these people, if it's more convenient to interact with their main tool through voice, that will justify also rethinking office spaces, I guess. So, yeah, I think there will be workarounds, and we will naturally find them if voice becomes the main way. I mean, if it's more practical to interact with AI through voice.

So I studied mathematics, and I started my career with a short internship in quantitative finance.

Yes, in Paris. And I was born and raised there. And what was interesting, during my internship, I had access to a Bloomberg terminal. And so I would see the news, like the constant news below the screen. I was thinking, what if I could have an algorithm that just reads this news and takes positions on the market faster than anyone because it was just able to analyze the news live? And I was looking, but I had no keywords for that, right?

So I literally Googled, "How can I analyze text automatically?" or whatever, and I found machine learning. It was an epiphany. I decided to completely stop and start studying again. My goal was to go back to finance with AI and machine learning. But I got passionate about all the possibilities there were around. So back then there were no LLMs, and it was not really about generative AI. It was about medical imaging, text understanding, a lot of things around audio, speech recognition, obviously.

And I looked for an internship which was about unsupervised learning, which was already pretty cool, and I just wanted to do it. And so I pretended that I was passionate about language. And I got the internship. And then Yann LeCun opened Facebook Paris. I was able to interview. So, for the anecdote, I did my coding interview with Soumith Chintala, who then invented PyTorch. And now I think he's the CTO of Thinking Machines. And so I didn't know how to code because I had only studied mathematics.

And so he asked me to implement k-means, a basic algorithm. And I asked if I could do it in MATLAB because I didn't even know Python. I knew very basic Python. And he was kind enough to let me do my coding interview in MATLAB. And I got the job, which, when I think about it, is so cringe because, oh my God, thank you, Soumith. And yeah, I did my PhD there. It was very interesting. It was already around speech, and I was spending half my time at Facebook and half my time at École Normale Supérieure in Paris, in a lab that was studying language acquisition in babies.

And in particular, the main observation on which the lab was built was that humans learn language from mostly two speakers, their parents, with a few hundred to 1,000 hours overall in the first four years, with huge variance between social backgrounds and without annotation, right? Because you learn to speak before you learn how to read. And that still makes us already pretty okay for conversation when we are kids. And speech recognition back then was trained with already hundreds of thousands of hours of annotated data.

Now it's millions of hours of annotated data. So the topic was more around efficient learning, which is interesting because it was 10 years ago, but now it's still as relevant as—I think there was a new company that raised a large round recently to make learning more efficient. So it's still as relevant as it used to be. And then I joined Google. At that time, it was interesting because I joined working on speech in Google Brain, and there was almost nobody working on speech in Google Brain.

It was not considered a vibrant research topic. It was like a product topic. A lot of people were saying, "Oh, but it's solved. Oh no, it's solved. It just works." So already back then.

And what year was that? 2019. And so I found someone to work with me, and we did a lot of work around speech. And then I got excited about a specific topic around compression, audio compression. So it was just out of passion. I wanted to do a new compression format that would not be MP3.

Exactly. Absolutely. And I wanted to do it with neural networks. So the idea was that it would be computationally more expensive to compress and decompress the audio, but then you could compress it much more efficiently. And that's something we worked on for Google Meet, and it was called SoundStream. That was the first, what we call, neural audio codec. And I had no plan of doing generative modeling back then. I really didn't care about that. But I was very lonely in a way. I was trying to lure some people around me who were working in reinforcement learning.

I wanted to get them to work on speech with me. I started a project around diarization, which is a task where you're listening to a conversation and you have to tell who said what, which is probably the least sexy research topic out there. I'm sorry. I think it's fascinating, but you cannot get people excited.

This is the least sexy part, like the monk project: very lonely. I was not very successful at getting people to work with me, and I thought, okay, generative models. The nice thing is that if we generate speech, people will listen to it and they'll be, "Oh, that's cool. My method has generated speech." So honestly, it was very opportunistic. For me, I thought that it would be a good way to get people to work with me.

And so we started a project, and the idea was that we started to see success around language models. So it was 2021, way before ChatGPT. But internally at Google, there were already quite a few projects that were successful around language models. And the idea was that just after the work we had done on the neural codec, so now if instead of using your codec for real-time communication, you just use it to compress audio, now you have—

Yeah, a codec is just compression-decompression. So you have audio, right? And you want to send it over the network when you're having a Zoom meeting, and you're not going to send the uncompressed WAV file because it's too heavy. So you're going to compress it into a much lighter file that you will send over the network, and then the receiver can decompress it back into audio. And the secret is based on a lot of science and knowledge around human hearing.

We know what kind of information we can remove from audio so that it won't create a perceived degradation, basically. So there is a lot of science around what specific information you can remove from audio that will make it almost as good for a human as the original one, despite the fact that you removed a lot of information, which allows you to compress. And the main idea was that instead of using hardcoded rules to do that, we would learn from data what are the transformations that allow us to compress audio while making it as transparent as possible for human ears, basically.

And so now we had this way of compressing extremely efficiently, much more than MP3 or Opus on audio. And in a way, you could consider that it was so compressed that it was almost like text. And so the simple thing we did was just train an LLM to predict this compressed audio instead of predicting text. And then you could do the exact same thing you can do with text. You could prompt. So I take a few seconds of audio, compress it, I pass it to the LLM, and I let it predict the next compressed audio.

And we realized, in one week, we had invented instant voice cloning. So we could replicate any voice with a few seconds of audio. And yes, this became extremely successful because it had all the advantages of LLMs that it could benefit from. So LLMs are great at modeling long context. They scale very well. So if you want to have a large model, you just scale the small model. It sounds obvious like that, but for a lot of architectures, it's very difficult to go from a 100 million-parameter to a billion-parameter model.

With transformers and LLMs, it's obvious. And I could go into more details, but in a—

AudioLM, and then it gave MusicLM, and then it gave NotebookLM, that were the automated podcasts. And that became the standard framework for audio generation. There were two families that were kind of fighting one another for some time. It was diffusion models, which I think is what ElevenLabs was based on early on. And we were the audio language model family. I think today, virtually everything is audio language models because, since they are autoregressive, they run in a streaming fashion. They are naturally compatible with real-time inference, which is kind of the main topic around voice right now.

And so everybody is using this technology today. And yeah, so it was very successful and extremely easy to apply to new tasks. So we did it on speech first. And so then we collected a dataset of piano performances, and then we had a model for piano, and then we did more general music. And then we could do pretty much anything. It's even used by a nonprofit lab working on animal vocalizations to try to decode the language of animals called the Earth Species Project.

So yeah, it's as flexible as an LLM for text, basically.

At the time when Gemini started at Google, that's when I left. So I wanted to create a small research environment that reminded me of the early days of FAIR or Google Brain. So, a very small team, elite, no distractions. Sorry to say that: no product managers, just research scientists, no emails, just locked in a room with the machines and focused on science. And in particular, the goal for me was really to keep working on fundamental research and keep pushing the field forward and training students and so on, because I felt very grateful to have been able to do research in such an open environment.

It was also obvious to me that, and for this I think I agree 100% with Yann LeCun, what made AI dynamic and get from ImageNet in 2012 to where we are today is open research. That's because it's kind of a worldwide collaboration, and everybody benefits from the progress of everyone. So, for me, it was important for the field itself to keep this going. And so we decided to create a nonprofit with the help of Eric Schmidt, Xavier Niel, and Rodolphe Saadé.

So, for the anecdote, the codename was Sphere because that's the name of the restaurant where we discussed the project. And we then understood that we could never trademark the name Sphere, obviously. So we just asked ChatGPT for "sphere" in a few languages. And sphere in Japanese is Kyutai. And there was AI in it. So we were like, okay, that's the name of the lab now. And so that's how we created Kyutai. The first person I reached out to was Alex Desfossés, who is also now co-founder of Gradium.

He's our chief science officer, because we had done our PhDs together at Facebook. And then when I joined Google, we kind of became rivals because we were working on the same stuff at the same time. And every time we'd meet one another, we'd just not talk about anything because, like, what are you working on? Oh, nothing. Okay. And, yeah, we had a small team, but with big expertise in speech. Again, it was an opportunistic decision that we made.

So I looked at the kind of stuff we could do around voice. And for our lab, okay, we had 1,000 GPUs. That seemed like a lot in 2023. It was still already at least one order of magnitude lower than bigger labs. So we wanted a project where we could make a difference despite the fact that we were four people. It shouldn't need too much compute, and it should be very innovative such that, just by being smart about what we did, we could make a difference.

And so we focused on conversational AI and real-time conversation because I had seen from the inside at Google that nobody was daring to touch this topic because it was so challenging, and it really seemed extremely far-fetched to be able to cast the task of dialogue into an LLM, right? People were just working on TTS and speech-to-text. The interactive stuff was really not there yet. So we thought, okay, we're going to work on it, and we're going to make it full duplex. Might as well do something really innovative.

What we didn't know at the time is that OpenAI had already been working on speech-to-speech conversation for a while. But in six months, with a core team of four people and then six, we were able to ship a model trained from scratch called Moshi that is still, to this day, the only full-duplex model. You can talk to it. It's a bit dumb because it's archaic in terms of intelligence, but the latency is still, one year and a half after, the best in the world by far.

And I think what was very interesting was that a very small team could make such a difference because then we shipped the first speech-to-speech translation system, and then streaming speech-to-text, and then streaming text-to-speech. And our models have been used across all industries. We are always proud to hear that a lot of big companies are using them. A lot of small companies are using them. The Maya and Miles demo from Sesame was built around our open-source models. I think what I really like about voice, and what makes Gradium a project that I deeply believe in, is that it's one of the modalities in AI where a very small team can make a difference.

I don't really see benefits from having extremely large organizations with a lot of people and resources because you don't need that many resources. You don't need 10,000 GPUs to train a speech model. You don't need 1,000 people. The ability to go fast, iterate fast with the right people, to me, is far superior to the advantage of having a big organization.

Yes. So, as I said, our open-source models were very successful. They are still downloaded millions of times a month. And we started having companies reach out, from all sizes, small and very large. They saw the potential. Sometimes it was a bit weird because I was talking to people leading extremely large teams, and I was explaining how I could train, by myself, streaming speech-to-text that was better than all alternatives. There was something we were doing very well, but at the same time, our open-source models remained limited.

They were fundamentally prototypes. For us, they were not even the actual contribution. For us, the contribution was the invention and the related research publication. For us, it was kind of like an artifact accompanying the paper, these models. But people wanted such models to be multilingual, higher quality, all the things that make an actual product. And so we considered kind of outsourcing this part, working with another team that would lead the product and so on. And honestly, after a few interactions, I realized nobody could carry such a project except us.

Nobody can believe in something if they have not developed it from scratch. We were in conversations with companies that wanted to create partnerships to improve our open-source models, and I would look at the specifications they wanted and realize it's something we could do in a few days, and they had been struggling for months. So, yeah, it seemed that we were also the best team to do that. And from a personal point of view, I was kind of addicted to academic prestige, having best paper awards, all this stuff.

It was always nice, but it was never enough. Every time I did a paper, I was happy for like two hours, and then I was already thinking about the next version. And at the same time, it eventually felt a bit weird that I could not imagine that, at the end of my career, I would have worked on something so applied and so close to actual real-life applications and not done them myself, not contributing to them directly. So I think in the end, even in terms of achievement, academic success is one thing, but nothing is near in terms of impact to the fact of having people using your models in the real world.

And in a way, I think it's something that is kind of generalized now in the industry, and it was also something that made me think. I looked at all the people I respected the most, the main one being, to me, a living legend, Aaron van den Oord from Google DeepMind. All these guys, they were making such nice contributions scientifically, and then they decided to focus on products. And I want to think it's because they also realized that academic prestige is one thing, but making a real impact is having your models being used in the real world.

And so, for me, it's the ultimate impact we can have. We still do science. In particular, Kyutai keeps doing open source and open science and so on. For me, the upside about it is mostly to be able to train the next generation of AI researchers and keep the field alive, as I said, because I think to have a healthy and vibrant AI field, you need to have scientific dissemination, so scientific exchange between institutions. And there also, the Chinese labs are doing remarkable work and are kind of forcing everyone to stay open to some extent, because otherwise it also hurts the ego, I think, of the people who are in the labs that don't publish.

So that was also something where I was very opportunistic about. So my strategy was like, if we publish in a world where the others don't publish, they will get so pissed off at us claiming all the inventions that it will make them join us eventually. And it's true that it's kind of hard because some people, they want to be in the place that is the cool place where the cool stuff happens. And it's not only about compensation and so on.

Now, I would say everybody working in AI and doing a good job in AI is going to get good economic outcomes. So then, glory is also very important, and it can be scientific or it can be just being proud that you are making the best products. But I think it's also an important part.

So one thing I mentioned was, if you have the right team, it can be extremely small and still make a significant impact. Another argument, I think, is focus. So, for example, if you look at large multimodal models, like generic models that understand images and can generate text and can produce code and so on, you have a limited budget, which is the number of parameters and data you're going to feed to your model when you want to add speech to them.

You're fighting with coding and image understanding and so on. So you are playing with a lot of trade-offs that are irrelevant to the task that you want to solve. And at the same time, these models are so large, they cannot run at scale because they will just make everyone lose money in the process. The only format that makes sense for speech models to run at scale is to be extremely compact, which also means that the training resources you need to train them are much smaller than what you need to train other kinds of models.

So the resources are not as challenging as for text models. I think also another aspect is, in a way, not trying to just make a conversational product. So really making building blocks so that people can build the product. So we could make the Gradium conversational assistant and think a lot about its capabilities and what it can do and what it cannot do and what the use cases for people will be, and so on, and hope that, like the voice mode of OpenAI, it's used for this task and this task and so on.

But it's impossible to cover everything with that. And so now, if we want to cover NPCs, fake sports commentators in video games, a language learning app, an annoying character in a cartoon, and the customer service agent, and so on, then we just make the building blocks. And this, again, is not really, I think, in the DNA of big companies to do this kind of very specific models that are targeted towards developers rather than trying to solve a lot of things at the same time.

I think what is very challenging right now is if you want to have the full intelligence on-device, like your full conversational AI on-device. Honestly, I would say at this point, if you want such a model to be useful, we are not there yet, right? You can have something that can chitchat a bit, and it will be decent. Or we also have shipped models on-device, but they are much more constrained in terms of applications. So, for example, we started a year ago with on-device speech-to-speech translation, which is something that makes a lot of sense because when you're traveling, maybe you don't have a data plan that is going in every country.

So it makes sense to have something that works on your phone if you want to order at a restaurant or something like that. I think it's a particularly adapted use case. But now we also released, two weeks ago, a model called Pocket TTS that not only is on-device but CPU-only. So there are already mods for AAA video games where the NPCs can be powered through these voice models. And now you unlock a completely new kind of applications because on-device models allow you to do very large-scale personalized content.

That would be economically not realistic with an API. So again, these kinds of things—if you want to make meaningful progress in that direction, making small models in voice is much more difficult than making large models in voice. So keeping the quality while reducing the size of the model, that's where the big challenge is. And our CPU model, in terms of algorithm, it's like the cutting edge of what we know. It's really the latest generation of everything we've been doing so far.

So if you read the paper, you will find our names on several pages. It's mostly inspired from the Moshi architecture, like pretty much every model right now. Even the Voxtral model that was released by Mistral two weeks ago is also based on our framework. I think that's really interesting because this proves that there are things that we do right because everybody is building on them. At the same time, I would say it's quite an advantage because there are two kinds of research papers.

There are research papers that are meant to be as explicit and reproducible as possible, which is what we try to do when we do one. And there are some that are more about, I would say, marketing, in the sense that they are mostly focused on the results and the performance rather than explaining the underlying mechanism and the data and so on and so forth. So the nice thing of people building around the frameworks we introduce is that even when they don't give details, we can infer the details.

So in a way, in a competitive landscape, I think it's quite an advantage because it would be more challenging if people were transitioning to something completely different from what we've been doing, because then we could not infer anything from reading their papers. Now all of this is very familiar when we read those. And at the same time, we have all the issues that people are facing. We have been facing them for a while, and so we have already workarounds and new versions and so on.

Open source at Kyutai, that's like the end goal of the lab. At Gradium, that's not the end goal of the lab. The end goal of the company is to make competitive products that outperform every alternative. But open source is, I think, a good way of allowing developers to prototype stuff, understanding what people expect, what they want. It's also a way to train talent. A few days ago, we released Hibiki Zero. So Hibiki was our speech-to-speech translation system.

Now it's a new algorithm that makes it even lower latency, better voice cloning, multilingual, and so on and so forth. And the PhD student who was working on that project at Kyutai is joining Gradium for a few months to do a visiting PhD, and then he will start his PhD again. So I think there are very healthy relations between open source and closed source that can be done. At the same time, I don't think it hurts any defensibility, quite the opposite, honestly.

And so I think it's also, for talent, very attractive because people, when they join us, they know that they can work on really competitive products, but at the same time they can keep sharing models and more exploratory research and do really frontier stuff. And I think if we want to stay at the cutting edge, we should be a product company, but also be a frontier lab. And being a frontier lab, you need to do fundamental research too.

So what is interesting in voice AI is now I think a high school student could make something that is decent. But then what people want is the last mile, and the last mile is extremely difficult. And the last mile is pronouncing well all the difficult cases, having a latency that not only is low but is robust, and almost zero downtime, and being able to clone voices regardless of accents and so on. So all these hard cases, for me, the only way to find the energy to solve them is because that's your business, because otherwise, if you look at benchmarks, like the benchmarks for TTS, it's LibriTTS.

So it's books, a lot of them from Dostoevsky. So it's more about whether you are going to miss an I or a Y in a Russian name, or it doesn't evaluate: can you pronounce phone numbers and email addresses and URLs and all this stuff? So when you optimize for these benchmarks, you lose a lot of the actual real-world cases. Yeah. So I think the main difference is the incentive to do things really with the finest details only makes sense when you want to be competitive in product.

And it's not only about the models, right? The infrastructure to run these models at scale is extremely important. I think particularly in an API business model, your margins mostly depend on the efficiency of your inference. And so, in particular, I'm very lucky to have in my team one of my co-founders, Laurent, who did several years in—he did all his career in quantitative finance, mostly at Jane Street. And now we have more people coming from quantitative finance. And these people, they are really passionate about efficient inference.

That's their bread and butter, because in finance, like in high-frequency trading, that's kind of the only way to exist. And the engineering challenges are really significant. Having a team in engineering as well, and not just training models, makes a huge difference. And now, if we're talking about on-device inference, that's even more complex, right? If you want a model to run on all Androids and iPhones and so on, that's a big engineering challenge.

One thing that is clear is that you can only trust human judgment. People have tried to make objective proxies of human judgment, like a neural network that listens to an audio and gives it a grade. It sucks. So many people try, and it works on their constrained setting, and on real audio it doesn't work at all and completely breaks. So we don't trust anything but our ears. So we do a lot of blind tests internally.

We do a lot of blind tests externally. So we are working with human judgment constantly. Every single decision we make is based on human listening. We don't trust metrics at all. So the quality of audio is fundamentally subjective experience. But there are some things that are going to be widely shared, like the prosody. Is the tone and the rhythm natural or not? A lot of people would agree on that. Is the voice nice or not?

Nobody agrees on that. And so then the only way for me to claim to have the best solution is to have the largest catalog and most diverse set of voices that people can pick. Because what voice people are going to like really depends between people. I had faced that in the past when we did MusicLM at Google. So for the first time, we made text-to-music. So you could type, like, death metal with a marimba or whatever, and you get death metal with marimba.

And so we put a website online where people could do this stuff. And at the same time, we were already planning to do, for the first time, RLHF, so reinforcement learning from human feedback on music. And so what we had designed is people could give—there would be two generations every time, and people could give a trophy. I had planned that because I wanted, for the first time, to integrate humans in the loop of music generation, because I think scientifically that would be really huge.

And what was interesting is that we made a paper about that called MusicRL, and it was quite nice, but the results were not extremely convincing. And then what we did is that we just did judgment ourselves. So with people, with my colleagues, we took, like, I don't know, 10 or 20 pairs of audio, and each time we chose our preferred one, and there was zero agreement between people. So there is no way an algorithm is going to learn human preferences when there is so much subjectivity.

So the only thing you can do is make it more steerable for each user to be able to customize it, because there is nothing that is going to please every user consistently.

All the models being the same, I think then people talk about TTS already, right? Which is much more constrained than voice AI. So in text-to-speech, there are factual metrics about accuracy and latency, and then there are more subjective things around expressivity and so on. But you can, again, really make a difference by making it more controllable, more customizable, and so on and so forth. So I don't think that it's clear that the best TTS, the most controllable, the most robust, the smartest in terms of expression, is in front of us.

Nothing is close to it yet. And then there is everything that is not TTS. You look at transcription now, again, what if a lot of people are speaking? So I was talking about diarization as the least sexy problem ever. At the same time, it's an extremely useful problem. And you look at the error rates, and they are very bad. I mean, it's still just not working. In difficult cases where you have a podcast with a lot of people talking at the same time, it just completely breaks.

Full duplex: we did Moshi a year and a half ago. Still nobody has made it into a product. There are so many things that are just not existing today that I think commoditization, maybe it will happen someday, but we are very, very far from it, honestly. And the gap in quality and abilities to make that as powerful as human production of speech and understanding of speech in complex environments and so on. And then even if you reach this point, there will always be challenges about getting the same performance with smaller models because then again, like a full-duplex smart voice AI that can run on a gamer GPU or on an iPhone, good luck.

So, a cascaded system, which is the old way, is actually pretty new. The old way is without LLMs. So when we talk about the old way, it's like Alexa and Siri and so on.

The old way is natural language understanding with parsing, with very limited vocabularies and so on. But the cascaded idea is you basically just take an LLM and wrap it in a speech input and speech output. So you have streaming transcription and streaming text-to-speech. The nice thing about that is you can plug in an LLM and customize its behavior through its prompt, like you would with a text model. A main limitation is, first, latency of the whole thing, because you're running three models in a cascade and each of them is going to add to the overall latency of the interaction.

And by going through the bottleneck of text, you lose what we call paralinguistic information, which is all the information we convey when we speak on top of what we say. So that's emotional states, irony, lying is one. A lot of information is conveyed that is not in what we say. Particularly, I encourage people to look at the—every few years, Interspeech, one of the main speech conferences, organizes a paralinguistic challenge with challenges that are more random every time.

So lying detection is one, but then you have trying to recognize the origin of the parent of someone from them speaking. Or there was one about people speaking while eating, and you had to recognize what they are eating. So anyway, when we speak, there is a lot of information that comes out about us, and this is lost through text. So what you're eating obviously is maybe not the most relevant one, but understanding in the customer care context that you're getting lost or annoyed and so on.

It's extremely important to be able to recover the conversation into a good state. Sometimes it's obvious from the text, like if the AI says, "Go to hell," yeah, probably they're upset. But sometimes it's not that obvious, and that kind of cue can be helpful. And so one way to address all of that altogether is speech-to-speech. And full duplex, on top of that, addresses what I think today is the worst part of voice AI. I hate it from the bottom of my heart.

It's turn-taking. In any introductory course to convolutional networks or image classification, there is always this analogy that is made that you can make rules to tell whether an image is shot in the morning or at night just based on colors, right? You can just look at color values and make a handmade rule. But now, if you want to recognize whether it's a cat or a dog, you cannot make rules that recognize all the shapes of cats and dogs and all the angles and so on and so forth.

So that's why we do machine learning, right? You just learn from the data when you cannot make the rules. With turn-taking, we are back at the archaic era of handmade rules, which is ridiculous. It's like you have an algorithm called the voice activity detection algorithm that just says whether it's silent or not. And if it's silent for more than X hundred milliseconds, and then this rule and this rule and this rule, then it's an interruption. But if this happens, it's not an interruption.

And so we have rules on top of rules on top of rules to decide whether the model should talk or not. And that makes it such that everybody who's talking with cascaded systems right now needs discipline when they talk to AI. You need to adapt to its flow. Otherwise, it gets lost and gets confused and interrupts and so on. And this is extremely annoying. Moshi, again, people can talk to it today. It's not really powerful, but the quality of the interaction is unmatched in the sense that you can really take five seconds to think about what you're going to say and then talk, and the model won't be lost.

It will just be extremely flexible. And what we did was an extremely simple thing. So we took the audio language model. Instead of having it model one stream of tokens, we called it multi-stream: just two streams of tokens, one for the user, one for the AI, and both can be active at the same time. And there is no turn-taking anymore. And you just train it on stereo data, people talking on the phone, and you have one person on the left channel and one on the right, and your model models both at the same time.

And then you play the role of the left channel and the AI plays the role of the right channel, and the model will learn how to handle a conversation like humans on the phone. And so when we did the Moshi announcement in particular, there was a fun early demo that we did. At some point, we trained the model on the Fisher dataset. So it's phone conversations that were recorded in the US in the late '90s. And so you could literally give a phone call to the late '90s, and it would mention Saddam Hussein and Jacques Chirac and talk about a lot of political stuff.

It was very weird because you're talking to a guy who says, "Hey, I'm Bob from Arizona." And every time, it's a new guy. And you can talk about whatever they tell you about their job and so on. So it's a kind of paranormal experience. Very fun. I think eventually it's impossible to think that we will stick with turn-taking in the future. So obviously for us, the end goal is full duplex. That was always our motivation. At the moment, what we do is cascaded systems because that's where the market is right now.

People are still iterating a lot on the underlying text models that they want to use, on the tool use and so on. And so there is so much progress on the text side. One drawback of speech-to-speech models is that since everything is integrated, when you go from a text model to the speech-to-speech model, you need to fine-tune it on speech data. So now the cost to switch the underlying text model is extremely high because you will need to fine-tune everything from scratch.

People want something that is modular, plug-and-play. What will solve everything is providing the same flexibility as cascaded systems so that you can change the backend on demand, basically. And you get the same customizability and customization that you have with text models, but with full duplex. This, I think, is going to be a convincing solution to all the current limitations.

One of the frontiers, I would say. Yeah. Again, one where I would say a frontier is: I could bet every single speech team in the world that they don't solve it in the next year or so. It's a robot in a modern factory, and there is a lot of noise from machines, and you have a lot of people talking to the robot, and the robot has to figure out what the hell is going on. This, I can sign, is extremely challenging.

If I had to pick the worst topic, I mean, the one that will be the most challenging, I think that will be this one, even more than full duplex. Much more difficult, I think.

One thing that is really important is to have several microphones. So we do. We have two ears, and it's extremely useful because that allows us, for example, to localize an audio source. So the reason why we can localize where sounds come from is because we have a small time difference between the time when it arrives in this ear and this ear. So if it arrives here before, my brain will understand it will come from this, and you get, like, an angle. And then the acceleration of the phase gives you the other dimension.

And so that's how you can locate in 3D. So having the ability of doing this spatialization, understanding where the sources are coming from and which one is saying what, and so on. Again, it's both a hardware and software issue. Creating training data for that is extremely challenging. Honestly, the level of robustness that we have as humans on that is extremely high. Also, we are all lip-reading. We don't know that we are doing it, but everybody is lip-reading all the time.

That helps understanding a lot, in noisy environments in particular. We do it unconsciously, but we all do it. I think interactions on the phone are quite okay because the mouth is close to the phone, and the phone can do a lot of work to enhance the quality. But now people want to have a robot that can be even a static robot that is just in a room, and you want to shout to it from your living room. It's called far-field speech recognition.

It's really broken. Every company who has made these home assistants knows how difficult it is to have them work in environments where there are several people. The reason I say I think it's one of the biggest challenges is, in TTS, we see a lot of progress. For this kind of hard understanding problem, I hear people saying the exact same stuff as they did 10 years ago. Like, there was zero progress.

So if you do the math, basically, training on a few trillion tokens, which is what you will do for a basic text model, that will amount to hundreds of millions of hours of speech or something like that, which is the kind of amount that is very hard to get. I think it's a very interesting question that comes up in a lot of discussions, and everybody has their theories. In particular, one attribute of speech data was that if you train a conversational model on speech data, it's going to be much less intelligent than a model trained on text.

And I think it's because when you listen to speech data, the density of information is much lower than you will have in text. So you don't have Wikipedia in speech data, you don't have Stack Overflow, Reddit, and so on. Getting your model to learn about the world from speech, I think it's a terrible idea. I think you should start—I mean, we have text models. So what we did for Moshi is we started from a text model. And then we took this text model and trained it on speech while trying to prevent as much as possible a loss of intelligence.

So all the time, we will recompute the text metrics, and they will degrade. But we are trying to keep it a bit contained. But indeed, the quality of speech data is extremely variable. Honestly, I think we are training on way too much data. So for Moshi, we trained on 7 million hours of speech. It's ridiculous. We could probably do that with 10,000 hours if we had the right method. We did not care because we had the data. So, yeah, screw it, we just trained on that, which is great to capture all the very specific idiosyncrasies in voice, the French accent, like the things that make a voice unique.

So for the diversity of voice, the amount of data is extremely useful. But at the same time, you could think, okay, if my model is as intelligent as a text model, it could learn dynamics of conversation from a few thousand or dozens of thousands of hours of data and not millions of hours of data. That's one source of data, the largest volume. It's just unstructured conversation, like publicly available audiobooks and this kind of stuff. And then you get towards more specific data that you make for your applications.

For example, for TTS, you want to have expressive data of high quality. You don't want to have something that is recorded with arbitrary conditions. You want really studio recordings, very low levels of noise, professional or semi-professional actors. And then there is what kind of text do you give them to read? And so it's a very interesting problem because if you want to generate hundreds of thousands of hours of scripts for voice actors, it's not that easy. I mean, you cannot ask Claude or ChatGPT, "Write 100,000 hours of scripts."

And make them as diverse as possible. So it doesn't work. It's just going to be in a loop and collapse on a few topics. Even if you ask for phone numbers, they are not really random number generators, LLMs. They will eventually always produce the same phone numbers. So Alex and my team spent a lot of time on making very complex machines for script generation with a taxonomy of all possible topics and subtopics and sub-subtopics and sub-sub-subtopics, and sampling constantly. And every time we generate a phone number, we generate it with an actual random number generator so that we actually cover the whole scope of phone numbers.

So I think it's a very interesting problem. It's painful, it's annoying, it's really about details. And I love that because I think that's always where we can differentiate, just because we always, as gamers say, we tank the painful stuff. And this pays off a lot because, as I said, in terms of computational resources, we don't need as much as to train large reasoning models or video models. But for speech, it's not really about the volume of data. But having high-quality data is extremely useful and very important.

And the accuracy, for example, of annotations, it's paramount. And it's a lot of human labor to be able to annotate precisely.

Yes, there's a labeling aspect. It's interesting because automatic annotation works very well. And where humans are useful is really for the few mistakes that the best automatic annotations will do right now. And so the only way to make it worth the time of humans is if they have perfect annotation. If they have slightly imperfect annotation, not really worth it.

How does it work? What is interesting is that we see transfer between languages, in particular languages from the same family. For example, with Hibiki Zero that we released last week. So it was trained on 50,000 hours or maybe 100,000 hours for, for example, Portuguese and Spanish. And then to do Italian translation to English, 1,000 hours was enough. I would say for languages that belong to a completely different family, that's more challenging. What is nice is that fundamentally the whole pipeline is the same.

Right now we have been focusing on a few languages to find the recipe that works, and then it can be applied to most languages. What is very difficult is unwritten languages, so languages that don't have an official writing system. That's a lot of dialects where people are mixing languages and putting, like, a bit of French and then a bit of Creole and then a bit of something else and so on. This is very challenging because getting data is difficult. Getting annotated data is even more difficult because you have much fewer annotators that can do that.

But these ones are the most challenging, I would say. But it has been a challenge for a long time, and there are a lot of projects around collecting data in such languages. Interestingly, there is one. So, for example, the content that you can find in the most languages is the Bible. And so the main source for the most widely available content in the world is the Bible in all languages, because a lot of people spend time translating it and so on.

So that's a very valuable source of data. Obviously, you won't have mentions of modern vocabulary in it.

If you want to have voice models run at scale and make sense in terms of economics, you need to have these models compact anyway. If you think about having NPCs in a game where people are going to pay, I don't know, $70 or $90 or whatever for a game, and then they want to talk to it three hours a day because it's a very good game, so you spend a lot of time on it. If you have a large model, it's just impossible.

Not only does it not fit on the GPU, but even through APIs, it just would not make sense economically. So fundamentally, I think these models need to be small, but sometimes they require access to large models because they're solving a complex task. I'm quite, in a way, stating the obvious, but selective and adaptive compute usage based on the context and the difficulty of the task at hand. The task being as precise as the next few words. Can I just answer like that?

Because somebody says, "Hey," and I just say, "Hey." They ask for the next flight to SF, and I have to look up on the internet to find the time. So being very selective about when to use compute, I think that's the only way for all of it to make sense economically.

Yeah. So I think for this one, for us, the product is the underlying models, right? Because what we see is that people are building voice agents. So voice agents, it sounds, I think, a lot like a business agent, like a customer care agent or something like that. But an NPC, in a way, is an agent, right? Because they have a voice interface, they have an underlying text model. Eventually, in video games, they will be able to do function calling.

To launch a quest, to decide that you solved something, or to control an action, and so on and so forth. What we focus on is providing the best technology to the people who want to build the agents. So we don't build agents ourselves. We want to focus on the quality of the model because that's our specialty. That's where our talent is. And at the same time, I think it would be a bit unrealistic to think that we can address all the needs of the market with our models.

Because again, I think what is very exciting is when we look at our customers, every time there is a new one, it's completely unpredictable what it will be about. And it's about learning and customer care and video games and media and personalized press. And it's so wide that we'd rather just make it easy for people to build voice agents and provide them with reliable models and infrastructure. And that also works pretty well because, in terms of staff, we are less than 15 employees today at Gradium.

When we partner with companies that deploy voice agents in banks or hospitals or different kinds of businesses, it requires a lot of what we could call today a forward-deployment engineer. So there is a lot of staff, human labor, that is involved for these deployments into various complex systems and infrastructures. And we can just focus on the model. So that also allows us to remain pretty small and really focus on the science and the engineering.

Actually, at Gradium, we have the best in the industry. And the best means not only replicating the specific characteristics of someone, but also the accent, some unusual recording conditions. And so, in a lot of contexts, for example, if you want to create a vintage-sounding character with old radio effects, then it's something that we do pretty well. If you want to have a robot voice, we can do that pretty well as well. If you want to have any kind of accent or speaking style, like posh or more laid-back or more urban, any kind of social or individual aspects of voice, it's something that we replicate pretty well.

I think what is interesting is that voice cloning in itself, where I see the most potential, is about creating interactive experiences around licenses. I tried to pitch, for example, Who Wants to Be a Millionaire? There is a video game, and the questions can be generated on the fly. You would like the host to—the voice of the host—to pronounce them all the time. So this one makes a lot of sense with cloning a specific voice because you want to replicate the voice of a character, of a person, of an athlete.

Of a K-pop star or whatever, I don't know, any kind of experience where people want to engage with a voice that they know, but they want to go for an interactive experience. Then, in a lot of cases, for example, in customer care, people don't want a specific voice. They want a voice with specific characteristics, and that is nice for the use case. And so typically they clone the voice of a colleague that has a good voice. For these customers, I think the solution that makes the most sense is not cloning; it's voice design.

So being able to create voices from a natural language description and having your voice generation being really faithful to the prompt so that you can—it's really worth iterating on it, because there are a few solutions that exist today around voice design. But as far as I know, they are not very popular, and people still stick to the existing voice catalog because they cannot steer it precisely enough. But when this works, this will allow people to design the voices that they want for their use case, which I think is very interesting, even though sometimes I'm a bit surprised because we have some customers that just take a random stock voice from our catalog, which is one of ours, like me or my colleague, and they don't care.

I try to push them, like, let's design a voice together. Let's make a voice that represents your brand. But some people are like, no, it's fine.

So, first thing I want to say: watermarking is a scam. I'm sorry, I have to say it. It just doesn't work. I worked on it. We have an appendix in the Moshi paper around how we could break so easily any watermarking that was supposed to be state of the art. So people should not rely on that. Also, watermarking: who verifies the watermark, right? You will need a platform to do the verification and remove the audio. But if someone— explain what watermarking is in the context of audio.

Yeah, in audio, that's the idea of either finding a way to— so when you generate audio, you put a hidden stamp in it like you would do in an image, but now it's in audio, so you cannot hear it, but it makes it very recognizable that not only is it fake, but it's fake from this specific model. And also, there is deepfake detection, which is just recognizing that audio is synthetic, even if there is no watermark in it, just telling true from fake.

These are very difficult to do. Unfortunately, if you get a phone call from a scammer, unless you have automatic detection inside your phone, it's useless, right? You cannot upload it to a website that says it's true or fake. Honestly, I would say at this point, I don't know who has their grandma who lost their credit card and is at the airport and needs $1,000 today, because it's always a story I hear. It's probably not the case if someone gives you a phone call.

So I would say, just in that context, honestly, I think there is nothing as safe as just asking personal questions that only the person they pretend to be could answer.

Yeah. I mean, it was already the case with emails, right? And people just need to be much more vigilant and probably will find ways to have authentication, for example, on the other phone side, that it's the actual person that is calling, and so on and so forth.

Only you can use it. You need to own it, but nobody will be able to use it. And so it's only for your own usage. And I think what eventually we could also do, which was done before, is allow people, if they want to opt in, to share their voice with the community such that they can also get financial compensation if their voice is used by other users. So I think it has a lot of value because it's a good way of sourcing a lot of voices.

Eventually, again, if we omit the case of replicating familiar voices from licenses or existing people who give their voice knowingly to create specific content, I think voice design is going to just remove this issue because then, again, people typically are going to clone the voice of someone, but what they wanted is someone from a specific gender, specific demographics, age, accent, and so on. And so they could just fill this information and get a few propositions for voices that will fit their need, and that will remove the need for voice cloning.

Yeah.

There are several things in particular. So if you want to do speech understanding, having access to the video can help a lot. Again, for diarization, for example, if you're filming, like if you're listening to presidential debates with several candidates, it's awful to understand. If you watch, you can see who is speaking at each time and so on. So audiovisual understanding is much stronger than audio understanding in itself. I think now also what is interesting is, so we see that in video generation, like with Veo 3 from Google, now there is native audio that is included.

And that's why I think for video generation, the most natural thing is integration of video and audio. I don't think it really makes sense to do video-to-audio generation separately because typically the data exists as a multimodal signal, right? When you have a video, you have the audio track as well. So you might as well just exploit both to train your models. However, for example, we released for Valentine's Day a small app called Bridger Clone.

Now, if I put my face in a photo and I put it to video and say, "Make a video," it's going to make up a voice that it thinks sounds like me based on my appearance, my age, and so on.

App.

Okay, exactly.

And so what it allows you to do is now you can clone your voice, you can record a small, short love message, and now you get a video of you in your voice. And this, again, I think that's why it's useful to have a voice that is treated as kind of a separate component. Because now you can have much more control on the actual voice of the virtual character rather than just having a likely voice that sounds like it could be yours.

And what is interesting is, at Kyutai, in two years, we were able to do conversation translation, TTS, speech-to-text, always competing with much bigger and much more mature teams. It's still the same at Gradium. And one of the big strengths is that we have this fundamental framework for audio generation around audio language models, and it's extremely flexible. So I think one of our strengths as well is our ability to do a new task. And when people come up with new needs, whether it's about annotating various stuff or generating various stuff, all of this can be cast pretty easily in our framework.

So then, if we see that there is huge demand for speech separation, it's pretty easy for us to do it, for voice transformation, for accent transformation, whatever. That's also why we are always interacting with the developers to understand what they want. We're also now giving access to alpha models that can do stuff that are still experimental but are world premieres. And I think that's something I'm pretty excited about because we can be much more creative than just speech-to-text and TTS. Obviously, these are kind of the master tasks where we want to be the best in the world because that's where most of the opportunities are.

But at the same time, we can do a lot of fun stuff that is completely orthogonal to that, in particular around transformation, audio effects, and so on. I think there are a lot of things that can be very cool.

A lot of things to say about that. I was born and raised in Paris. I did all my career there. Facebook arrived when I started my PhD, and then Google Brain moved to Paris, Google DeepMind afterwards. I'm what we can call terminally online, in the sense that I love the very mean memes against Europe. So there is this guy, I don't remember his name, it's like a fake Swedish name, and he keeps posting about how he has, after only 20 meetings, contributed a €10,000 check for a sitcom, and it's a compliance-first company and everything else.

But I love it. It's very mean. And honestly, I love it because it's very mean. I love when people have so much time to spend just to be mean. I mean, I think it's quite—I respect that. But at the same time, it's so far from reality. So French AI and European AI—European AI is mostly French AI, to be fair. There is also Germany, but a lot of it is in France.

And before French companies, there was French talent in American companies. So again, a lot of the current generative models of Google, a global company obviously, were developed between Paris and Zurich. Actually, most of it. LLaMA was started in Paris. DINO, which is the most groundbreaking vision work from Facebook, was developed in Paris. A lot of things have been developed in Paris that are not seen as Parisian because they were made in American companies. And now, I think the field in Paris is—the talent is so dense, and the people are extremely strong and extremely committed.

And the best signal that proves it is: we used to have Facebook and Google in Paris. Now there is OpenAI, there is Anthropic, there is Cohere. Pretty much everybody is opening an AI lab in Paris. And the reason why they do is for the talent. So I think we have everything in France to develop global companies. In particular, in AI, which is an economy really built around talent, that's a perfect deal for it. Also, as a French guy, I really don't want us to screw it up in France, because, in a way, when you see the most successful French companies, it's luxury or that kind of stuff.

And so, yeah, I really want AI to be a field where France can make a big difference, and that can become one of the biggest drivers in the European economy. If it doesn't work, then Europe will really have to look itself in the mirror and be like, how could we screw it up with so many strong people? Because the people are here, and capital—we can get capital in Europe as well. I think all the conditions are there for it to be competitive. In a way, the more people are mocking Europe, the more it can make the people who mock overconfident about themselves.

And everyone who is overconfident eventually gets displaced by the underdog. So people should always—when I watch the 996 LARPing on Twitter, it's ridiculous. Like, people coding at the gym, like what the hell? Code and then go to the gym. You're not doing a good gym and you're not doing good code. It's just pretending. So I think we don't have a culture of pretending, and that's whole Europe, from the most western to the most eastern part.

We tend to be a bit more pessimistic, maybe, and a bit more down to earth about our impact and the challenges and so on. We don't try to sugarcoat things. We rarely look enthusiastic or happy about our work, but it's a good discipline. So I think the results kind of speak for themselves. What I think—so Mistral, for example, was sometimes mocked because it's not considered the frontier lab like OpenAI, Anthropic, whatever. Okay.

But look at the staff and the resources and so on. I know the best people from Mistral. They can be compared to the top of the top of the biggest labs. So then there's a question of scale and scale of resources, indeed. So the unfortunate tweet you mentioned about the €30 million is one of them. But yeah, other than that, there are a lot of very strong people.

And they have also done a lot of great things for American companies. I mean, Yann LeCun is one of them, obviously, but it's not only him. Samy Bengio, who used to lead Google Brain, is now leading Apple ML Research. The amount of French people in the leadership of big tech AI research is very large.

Well, I love that. I get carried away a bit when I speak about this.

Thanks.