Your Voice Agent is Just a Walkie Talkie — Neil Zeghidour, Gradium

English
AAI Engineer
Computing/SoftwareInternet Technology

Transcript

00:00:00Nelsi Ghidour: Hi everyone. I'm Nelsi Ghidour, co-founder and CEO of Gradium. Gradium is a
00:00:19startup based in Paris. Most of our background is from research. In particular, we have invented
00:00:26algorithms such as speech-to-speech models, codecs, and so on. Basically, we started from
00:00:36a research project called Qtai, a non-profit research lab that has been focusing on voice
00:00:41since day one. In particular, we released in 2024 the first full duplex speech-to-speech
00:00:46model called Moshi, the first real-time speech-to-speech translation system called Hibiki, and the
00:00:51first TTS models that can run locally on a smartphone. Basically, I will just say a few
00:00:58words about what we do, but we are a model company that trains models for building voice
00:01:03agents and voice applications, so we do TTS, API and on-device, speech-to-text, speech-to-speech
00:01:08translation, and much more to come. What we do is that we train foundation models for audio
00:01:13and then we can apply them for a lot of different tasks. So, I'll go quickly on voice agents because
00:01:19it's a fourth talk about the topic, but basically now we have these voice interfaces that we can
00:01:27use to do a lot of things across a variety of products and types of interactions with NPCs,
00:01:32with customer agents, language learners, coach, and so on and so forth. And in this talk, I try to go
00:01:39through the history of this technology and where I see it going in the next years. And maybe to start,
00:01:47I think we can take a look at the announcement of Siri back in 2011. And you'll see that it's actually,
00:01:55you know, I think it aged pretty well. What is the weather like today?
00:02:06Here's the forecast for today. It is that easy.
00:02:13Lots of things. We've integrated with the stocks, so you can ask it about the stock market. Something like,
00:02:18how is the NASDAQ doing today? NASDAQ composite is down right now at 2321.70.
00:02:29Again, you can ask this from the lock screen anywhere. Just press the button and ask. You can ask about,
00:02:32you know, the NASDAQ, the DAO. So, what you just saw is was kind of a voice agent. It was a bit
00:02:39constrained, but it was technically a voice agent. And the architecture behind it, so you have seen a
00:02:43thousand times today the STT, LLM, TTS. Back then, it was even worse, right? So, there was no LLM,
00:02:49obviously. So, there was what was called natural language understanding. So, you will go from the
00:02:54transcript and try to do basic classification of what is expected, what is the app that is supposed to
00:03:01be controlled, what is the action to trigger, and so on. So, it was a very complex pipeline and
00:03:08very constrained to very specific use cases. So, it was what was called close-ended dialogue back then.
00:03:18Then, the next generation of models, I could say, I think, was illustrated pretty well by the
00:03:25OpenAI voice mode, not the advanced voice mode, the original one. In which case, you just have
00:03:31a conversational experience that is open-ended. So, it's much less constrained than Siri. You can talk
00:03:35about pretty much anything. At the same time, you don't have the same agent capabilities, let's say,
00:03:42in the sense that it cannot really give you the NASDAQ or give you the weather.
00:03:50We want to hear a bedtime story. Tell us a story about the super-duper sunflower hedgehog named Larry.
00:03:57Start with telling us a little bit about him.
00:04:03Larry was a unique hedgehog unlike any other. He had bright sunflower petals instead of spines.
00:04:09Residing in Meadowville, Larry spread joy and color wherever he rolled.
00:04:15So, in that context, basically, that was, you know, kind of second generation of models where
00:04:19all the logic, the very complex logic was replaced by an LLM. And if we compare Siri and the voice mode from
00:04:28is open AI, not the advanced one, the original one, the big strength is open-ended dialogue but the
00:04:36advantage capabilities were inexistent. The naturalness was obviously much better than with Siri.
00:04:41The latency was still pretty bad. So, we are talking about several seconds and that was documented in the
00:04:46previous presentation. And so then, you know, we're in the third phase, which is a voice, what we call voice agents.
00:04:55And here, I'm going to show a quick demo of a drive-through bot vibe coded with our with our voice models.
00:05:02Hi there. What can I get started for you today?
00:05:05Hi, I'm actually in a hurry and I'd like to order a chicken sandwich quickly.
00:05:10We have the classic, spicy, deluxe or grilled chicken sandwich. Which one would you like?
00:05:16Uh, let's go with the classic chicken sandwich.
00:05:19Anything else for you?
00:05:20Uh, what else do you have in the menu?
00:05:23We have sandwiches, sides, drinks and desserts. What can I get for you?
00:05:29Can I get some mac and cheese maybe?
00:05:31You have a classic chicken sandwich and mac and cheese. Your total comes to eight.
00:05:37So, now, you know, unlike what we just saw, it's an actual agent. It's taking actions. It's
00:05:43keeping track of the order. It's then going to make you pay. So, it's an actual voice agent that can
00:05:48do real tasks. So, here, instead of having an LLM that is just a conversational interface,
00:05:54we have a real agent that is empowered with tool call, reasoning, planning and all this stuff. So,
00:06:02what we see now is we have gained back agenting capabilities and actually they are much more
00:06:09powerful and generic than before while keeping a very good level of naturalness and that's where
00:06:17speech-to-speech LLM came. In particular, what we could see here is the latency is better with cascaded
00:06:24system, but it's still higher than you will have the human conversation. And as also as explained before,
00:06:30the naturalness is fundamentally limited by the fact that you go through text. So, you lose a lot of
00:06:35information about what is said, the tone, the emotion of the user and so on and so forth. So, now that we
00:06:41have tackled intelligence and agenting capabilities, speech-to-speech seems like a natural next step for
00:06:46naturalness and latency. And so, here, it's an announcement from the OpenAI advanced voice mode.
00:06:53Hey, ChatGPT. I'm Mark. How are you?
00:06:56Oh, Mark. I'm doing great. Thanks for asking. How about you?
00:07:01Hey, so I'm on stage right now. I'm doing a live demo and frankly, I'm feeling a little bit nervous.
00:07:06Can you help me calm my nerves a little bit?
00:07:09Oh, you're doing a live demo right now? That's awesome.
00:07:14I think we all remember it was very impressive. Very impressive release.
00:07:18And in that context now, all the steps of STT, LLM and TTS have been absorbed into a single one.
00:07:26And so, now, intelligent, you know, like naturalness is
00:07:32still very good. Actually, it can be better because it can understand non-linguistic information.
00:07:37Latency is really, really nice. Honestly, it doesn't make sense to go better than that.
00:07:44Interestingly, and everyone who has used any speech-to-speech model can
00:07:48attest that the intelligence is still much more limited in that context than the cascaded counterpart.
00:07:57So, the speech-to-speech models are fundamentally still limited compared to the textual models.
00:08:03Another limitation is turn-taking. So,
00:08:07people tend to mix speech-to-speech and full duplex. But basically,
00:08:13when you do have a speech-to-speech model like GPT real-time, it's still based on fundamental
00:08:17turn-taking. In the sense that it's going to segment the conversation into either the model is
00:08:21speaking or the model is listening. And to give, to show you how this can make an interaction
00:08:27unnatural, I'm going to show a small demo with what is called backchanneling, which is these very human
00:08:32things that you do when someone talks to you is that you say, hmm, yeah, and so on.
00:08:37Hey, how's it going? I'd just like to brainstorm a bit about it with you.
00:08:40Oh, that's a great topic. Yeah, I'd love to help you brainstorm. Are you thinking more?
00:08:47Exactly. Yeah, I was saying. No, no, I didn't mean to interrupt. You know, I was just saying,
00:08:50yeah, like that. You can just keep going. You know, don't mind me. That's something I typically do.
00:08:55Ah, no worries at all. Yeah, I was just going to say we could break it down into a few aspects.
00:09:00Sure. Yeah, exactly. No, no, please stop interrupting. You know,
00:09:03it's called backchanneling. Humans do it all the time. It shows that you're just following the
00:09:07conversation. That shouldn't, you know, like interrupt you in your flow. Just,
00:09:10just going. Ah, got it. Thanks for letting me know. No problem. Oh, come on. Yeah, so you see,
00:09:17you know, it's still very annoying. You can have lightning speed latency. Fundamentally, this is
00:09:25an issue that can not be resolved when you're using turn-taking. So here, that's the walkie-talkie.
00:09:31Any real-time model today, I mean, now there is a bi-directional one that will come from OpenAI,
00:09:36but it's called half-duplex. So the model is listening or speaking. A human conversation
00:09:42has a constant flow between two people. People do backchanneling. People interrupt one another,
00:09:48talk on one another, and so on. If you're having a relative on the phone, there is up to 20% of the
00:09:54time where you are both speaking at the same time. And that makes, you know, this very flexible dynamics
00:10:00in the conversation makes it much more comfortable for humans. And so to understand how we can make a
00:10:07model full duplex, I'll give a very short presentation of how we train such models.
00:10:14So the way you create a speech-to-speech model, half-duplex or full duplex, is the following one.
00:10:18So you start from a text LLM, which is probabilistic models over words. And instead of predicting the
00:10:24next word based on the past, what you want to do is rather predict the next audio based on the past
00:10:30audio. The issue now is that if you pass a raw audio to your model, which is, you know, a waveform,
00:10:36it's air pressure variations, basically you take this sentence, it's eight words. It takes around
00:10:45three seconds to pronounce it. And so at 24 kHz audio, instead of having eight words, the audio form
00:10:50is 72,000 time steps that you will need to feed to your LLM. Given that LLMs have quadratic complexity
00:10:57with sequence length, so the complexity is a square of the sequence length, a 10,000 times longer sequence
00:11:03is 100 million times more expensive to process. So there is no way you can train an LLM on a raw audio.
00:11:08So the way you addressed it is by creating neural codecs, or you can also call them audio tokenizers.
00:11:14And basically it's an encoder that takes an audio and compresses it in a very dense compressed
00:11:20representation, a bit similar to text. And then you have a decoder that can reconstruct high quality
00:11:25audio from it. So now you have gone from the audio domain into an abstract representation domain
00:11:31where you can train an LLM exactly like you will train it on text. And the speech-to-speech model
00:11:36from Japan AI that was shown before works in this fashion. So instead of having text tokens into your
00:11:42model, you have audio tokens that represent either the LLM or the user and you put them one after the
00:11:48other and the model predicts the audio tokens that should be saved by the model being given the context
00:11:54from both sides on the conversation. However, you can see that it's still a sequence between user
00:11:59and system which is still half duplex. So how did we made the first full duplex model ever?
00:12:06Very simple. We call it multi-stream language models. That's the technology now used also by thinking
00:12:11machines for their interaction model and most likely by the directional model of OpenAI is that instead of
00:12:18having a transformer that models one sequence of tokens, it models two of them so that both
00:12:24parties can be active at the same time, inactive at the same time, one active and one inactive.
00:12:29And I just show a very quick demo of how it sounds like but that's the release of Moshi in August
00:12:362024 where we did announcement live on stage talking to it for the first time.
00:12:41And you'll see that the model often guesses the end of the question and swerves over the speaker and both
00:12:48speaking at the same time is not breaking the flow like we saw with GPT. The whole thing is just
00:12:54extremely resilient to the most chaotic situations. So the planet is Sirius 22. Can you plot a trajectory
00:13:01course to it, please? Yes, sir. Okay. How long is it going to take us to get there? I've mapped it out.
00:13:06Okay. It's approximately five months to get there. Okay. That's not too bad. Do you think we have all
00:13:12we need on board the ship to start the mission? Yes, sir. We have everything we need. So back then
00:13:17it was even a bit irritating to people because it was interrupting you all the time. But the thing is
00:13:21that you can use, you could still use it in extremely noisy environments with a lot of noise, people
00:13:26coughing and so on and you know the flow is just constant. You don't get this very irritating break of the
00:13:32conversational flow. So these full duplex models, they are the highest level of naturalness you can
00:13:38expect. That's the same conversation we have a human. The thing is with our models, it was even more
00:13:45stupid than speech-to-speech models that were already less intelligent than cascaded systems. It's probably
00:13:50fine for some use cases if you just want to have a chit-chat, you know, the model doesn't need to be very
00:13:55intelligent. But if you want to make an actual full duplex voice agent, there is no way we can give up on
00:14:00intelligence just to gain speech-to-speech abilities. So how do we finally make models that tackle all
00:14:06these aspects jointly? And I think interestingly, if you look at the history I showed, there is a tension
00:14:13between naturalness and intelligence. So every time we improve naturalness or humanness of the models,
00:14:19they were less intelligent than the cascaded system. The cascaded agents, they are basically as smart as
00:14:24the best text models. So if you have a voice agent that is powered by the latest model from
00:14:29Entropic or OpenAI, it's going to be extremely smart. I have all the same reliability for tool
00:14:32call and so on. Speech-to-speech has this naturalness aspect. However, you give up intelligence to get
00:14:39that. And the reason why you give up intelligence is remember that LLM is a model that has a certain
00:14:45number of weights that we call the capacity. And if you take a text model and now not only has to
00:14:52handle text but it also needs to understand speech and produce speech, it's taking some of its capacity
00:14:57and this capacity now is taken from the intelligence. So fundamentally, there is a cost of adding a new
00:15:04modality to a text model that is going to be paid in intelligence. So where do we go from here? There are
00:15:10two paths that are in front of us and both are going to be explored at the same time. The first one is
00:15:15scaling the model. So making your speech-to-speech model bigger, better pre-trained, better post-trained and
00:15:22so on, will likely progressively increase its intelligence until it's good enough for a lot of use cases.
00:15:29The second one is splitting the model between naturalness and intelligence. The first one, I'm not at OpenAI,
00:15:37so I don't know because they don't publish their model. I guess OpenAI is the path one. So it's a
00:15:42frontier text LLM with a lot of science around post-training, instruct tuning to fine-tune it on audio
00:15:48and teaching it to be quite smart while using audio. The nice thing about that is you have a single model
00:15:55to orchestrate, so it's quite easy to deploy. And one big aspect, however, is that it's an extremely
00:16:06complex and costly process to go from the text model to this speech-to-speech model. The second path
00:16:12is to split it. It's an approach that we introduced in one of our recent papers called Mochirag and that
00:16:16has been reused by, in particular, the thinking machine interaction models, where basically the
00:16:21idea is that now you have two models. The first one is a small, maybe even on device, full duplex,
00:16:28extremely natural speech-to-speech interface. And its only role is to keep a very natural conversation
00:16:35and be able to delegate all the thinking, tool calling, reasoning, agenting capabilities to a
00:16:42background text model. And so the way to see it is you have a background text LLM that receives
00:16:48asynchronously queries from hundreds to thousands of small voice interfaces and just gives them their
00:16:53text, you know. And basically what we did is a very small full duplex model that just needs to know
00:17:00when it doesn't know so that it can delegate to the background model. And the reason why we believe
00:17:05mostly in this approach, and I go back to it later, it's our, let's say our culture is more the first
00:17:16one, the bitter lesson. So every time we've been pushing for end-to-end systems and so on. But now
00:17:21I think the hybrid approach has two main advantages. The first one is cost. So speech-to-speech models are
00:17:28notoriously quite expensive. And when you think about it, it's a loss of money to do chit chat with
00:17:35gigantic speech-to-speech models that can resolve differential equations and so on. So it doesn't
00:17:41really make sense economically to get all your workflow through this gigantic multimodal mixture of
00:17:47experts. At the same time, we see that people are very attached to their ability to control the
00:17:52backend to be able to switch SONET 5 was released a few minutes ago. People want to switch the backend
00:17:57and the intelligence and get a lot of optionality on that, right? When you're using a speech-to-speech model, your
00:18:05hands are a bit tied with this model provider. And to give you an idea of that, until recently, the
00:18:10advanced voice mode from VNI was powered by GPT 4.0, despite the fact that there have been several
00:18:15generations of the text model since then, because this process is so expensive and so long. For this
00:18:22reason, we rather bet on the hybrid approach, because that will give something that is not only very
00:18:28natural and very nice for demos and impressive, but also will be a viable alternative from an economic
00:18:34point of view and agentic capabilities point of view to the best cascaded systems that are still most of
00:18:41the market today in voice. So what now? You can use our models on Gradium.ai. You can apply to Gradium. We
00:18:49are recruiting research scientists and engineers, and thanks for your attention.
00:19:13Thank you.

Key Takeaway

Multi-stream language models and hybrid architectures solve the latency, turn-taking, and intelligence trade-offs of voice agents by decoupling full duplex naturalness from background reasoning engines.

Highlights

  • Gradium released Moshi in August 2024 as the first full duplex speech-to-speech model.

  • Raw audio processing requires feeding 72000 time steps per three seconds at 24 kHz, creating quadratic sequence complexity in language models.

  • Conversations feature up to twenty percent simultaneous speech overlap where both participants talk at the same time.

  • Multi-stream language models use a transformer that models two token sequences concurrently to enable full duplex interaction.

  • Gradium introduced the Mochirag hybrid architecture to split naturalness and intelligence between a small on-device speech model and a background text large language model.

Timeline

Evolution of Voice Interfaces from Siri to Early OpenAI Voice Mode

  • Siri utilized a complex pipeline of speech-to-text, natural language understanding, and classification constrained to specific close-ended tasks.
  • OpenAI's original voice mode replaced rigid pipelines with an open-ended conversational large language model lacking agent capabilities.
  • Early voice modes suffered from multi-second latency and zero action-taking functionality.

Gradium operates as a Paris-based startup building foundation models for voice applications, including speech-to-speech models, codecs, and local smartphone text-to-speech engines. An examination of Siri from 2011 demonstrates early constrained voice control using natural language understanding rather than language models. Subsequent iterations like the original OpenAI voice mode introduced open-ended conversation and naturalness at the expense of external tool execution and low latency.

Agentic Capabilities and the Limitations of Half-Duplex Turn-Taking

  • Drive-through demonstrations illustrate modern voice agents empowered with tool calls, reasoning, and order tracking.
  • Advanced speech-to-speech models integrate speech-to-text, large language models, and text-to-speech into a single component with reduced latency.
  • Half-duplex models force strict turn-taking that breaks natural communication flow during backchanneling and simultaneous speech.

Voice agents advanced to execute complex real-world tasks such as order tracking and payment processing. While advanced speech-to-speech models achieve rapid response times, they remain fundamentally constrained by turn-taking dynamics. This limitation prevents natural conversational behaviors like backchanneling, where humans utter supportive sounds while another person speaks.

Technical Architecture of Full-Duplex Multi-Stream Models

  • Processing raw audio waveforms directly creates quadratic sequence complexity that makes direct large language model training impossible.
  • Neural codecs compress raw audio into dense token representations that function identically to text tokens.
  • Multi-stream language models model two token sequences concurrently so both parties remain active or inactive at the same time.

Training speech-to-speech models requires converting raw audio waveforms into dense compressed representations using neural codecs or audio tokenizers. Gradium released Moshi in August 2024 as the first full-duplex model using multi-stream language models. This architecture eliminates turn-taking restrictions, allowing the system to handle noisy environments and simultaneous speech without breaking conversational momentum.

Balancing Intelligence and Naturalness Through Hybrid Architectures

  • Expanding a text model to handle audio processing consumes model capacity and reduces reasoning intelligence.
  • Scaling monolithic speech-to-speech models represents one frontier path for improving intelligence over time.
  • The hybrid Mochirag approach delegates reasoning and tool calls to a background text model while a small on-device model handles full-duplex naturalness.

A fundamental tension exists between naturalness and intelligence because multimodal processing consumes model capacity originally dedicated to reasoning. Monolithic frontier models attempt to solve this via massive scaling and post-training, but remain economically costly and inflexible. Alternatively, hybrid architectures combine small on-device full-duplex interfaces with asynchronous background large language models to optimize costs and maintain backend optionality.

Community Posts

No posts yet. Be the first to write about this video!

Write about this video