Talk by Neil Zeghidour, co-founder and CEO of Gradium, arguing that current real-time voice agents are half duplex—they listen or speak, never both—unlike human conversation where speakers overlap. He traces the evolution from a 2011 transcript-to-action phone assistant, through open-ended voice chat without agency, to cascaded voice agents that call tools but inherit text-pipeline latency and flattened emotion, then to speech-to-speech models that fix latency but still fail at back-channeling, demonstrated live. The technical core covers why raw audio is costly for language models (long timesteps, quadratic attention) and how neural codecs compress audio into token-like representations, with full duplex requiring simultaneous modeling of two token streams. Zeghidour notes a tension between naturalness and intelligence in a fixed weight budget and outlines two paths forward: scaling the speech-to-speech model, or splitting the interface from the intelligence. # Video Identity Observations — "Your Voice Agent is Just a Walkie Talkie — Neil Zeghidour, Gradium" ## Speaker & Channel Self-Description - The speaker introduces himself as **"co-founder and CEO" of Gradium**, "a startup based in Paris." - He describes the company's research lineage: > "we started from a research project called QTI, a non-profit research lab that has been focusing on voice since day one." - Claimed technical credentials: > "we have invented algorithms such as audio LLMs, speech-to-speech models, neural codex, and so on." - Specific claimed releases: **Moshi** ("the first full duplex speech-to-speech model," 2024), **Hibiki** ("first real-time speech-to-speech translation system"), and "the first TTS model that can run locally on a smartphone." ## Claimed Scope - **Business scope**: > "we are a model company that trains models for building voice agents and voice applications. So we do TTS, API, and on-device speech-to-text, speech-to-speech translation, and much more to come." - **Research positioning**: > "we train foundation models for audio, and then we can apply them for a lot of different tasks." - The speaker claims authorship of a hybrid architecture approach: > "an approach that we introduced in one of our recent papers called Moushiraq" (multi-stream / split naturalness-intelligence models). ## Intended Audience (Demonstrated) - A **technical audience already familiar with voice AI** — the speaker notes: > "I go quickly on voice agents because it's the fourth talk about the topic" — indicating a conference/talk-format channel context with multiple speakers on the same theme. - Audience expected to know ML fundamentals: he references LLM quadratic complexity, audio tokenizers, and technical comparisons without basic explanation. - **Recruitment and product pitch included**: > "you can use our models on gradium.ai. You can apply to gradium. We are recruiting research scientists and engineers." — suggesting the channel reaches practitioners and potential customers/hires. ## Observable Presentation Patterns - **Historical-narrative structure**: walks through generations of voice agents (Siri 2011 → OpenAI original voice mode → cascaded voice agents → advanced voice mode → full duplex), with archived demo clips embedded. - **Live demos of own products**: a drive-thru ordering bot ("live-coded with our voice models"), a backchanneling demo, and a Moshi full-duplex demo ("the release of machine August 2024, where we did an announcement live on stage"). - **Opinionated technical commentary**: e.g., the talk title's thesis — voice agents are "the walkie-talkie" — expanded via: > "the model is listening or speaking. A human conversation has a constant flow between two people." - **Transparent acknowledgment of limitations**: he concedes their full-duplex model was "even more stupid than speech-to-speech models that were already less intelligent than cascaded systems." ## Claims vs. Demonstrated — Distinctions - **Creator claim (unverifiable from transcript)**: "first" statuses of Moshi, Hibiki, and the on-device TTS model. - **Creator claim about third parties**: multi-stream models are "now used also by Thinking Machines... and most likely by... OpenAI" — hedged ("most likely"), not confirmed. - **Demonstrated**: audio of working demos (drive-thru agent, backchannel failure case, full-duplex conversation) appears in the talk itself. - **Self-deprecating transparency**: > "with our models, it was even more stupid than speech-to-speech models" — creator voluntarily states a weakness of his own product. # Delivery Observations — Neil Zeghidour, Gradium ## Tone - **Conversational, low-key enthusiasm.** The talk is delivered in an informal register with frequent hedging and self-corrections: > "So, basically I will just say a few words about what we do" — repeated **"basically"** functions as a running discourse marker throughout (appears at least 10 times). - **Candid, direct judgments**, sometimes blunt: > "with our models, it was even more stupid than speech-to-speech models" and > "Honestly, it doesn't make sense to go better than that." - **Self-aware humor and admissions of ignorance**: > "I'm not at OpenAI, so I don't know because they don't release their model." ## Structure & Teaching Habits - **Historical, generation-by-generation framing.** The core pedagogical device is a timeline: 2011 Siri → original OpenAI voice mode → voice agents → speech-to-speech → full duplex. Each stage is introduced with a phasing phrase: > "Then, the next generation of models" ... > "we're in this third phase." - **Demo-driven teaching.** Nearly every concept is illustrated with an embedded video clip or live demo (Siri announcement, hedgehog bedtime story, drive-thru bot, backchanneling demo, Moshi stage demo), followed by an analytical debrief: > "So, what you just saw is what kind of a voice agent." - **Named trade-off framing.** A recurring thesis is explicitly stated as a tension: > "there is a tension between naturalness and intelligence." - **Contrast pairs** used repeatedly: half duplex vs. full duplex, cascaded vs. speech-to-speech, "naturalness" vs. "agentic capabilities." ## Vocabulary & Recurring Phrases - **Repeated terms**: "naturalness," "agentic capabilities," "full duplex," "half duplex," "cascaded," "walkie-talkie," "backchanneling," "so on and so forth" (used at least 4 times), "you know" (very frequent filler). - **Deliberate repetition for emphasis**: > "it's an actual agent. It's taking actions. It's keeping track of the order." and > "very impressive very impressive release." ## Technical Explanation Style - **Concrete quantification** to justify claims: > "at 24 kHz audio... the audio form is 72,000 time steps" and > "A 10,000 times longer sequence is 100 million times more expensive." - **Plain-language reduction of complexity**: describes codecs as > "a bit similar to text," and explains model "capacity" with a simple allocation analogy. ## Uncertainty Markers - The speaker explicitly flags speculation: > "I guess OpenAI is the path one" and > "most likely by the... directional model of OpenAI" — hedging claims about competitors rather than asserting them. - Occasional uncertainty in the transcript about reference accuracy (e.g., "GPT-3 time" at [491], possibly a misstatement), suggesting some imprecision in live recall of product names. ## Closing Pattern - Standard startup-pitch close: product URL, recruiting call, brief thanks — > "You can apply to gradium. We are recruiting research scientists and engineers. And thanks for your attention." # Observation Document: "Your Voice Agent is Just a Walkie Talkie — Neil Zeghidour, Gradium" ## Speaker - **Neil Zeghidour** (self-identified as "Nel," co-founder and CEO of Gradio/Gradium), startup based in Paris, research background. ## Background Claims - Gradium originated from a non-profit research lab (called "QTI" in transcript) focused on voice since day one. - Claimed releases: **Moshi** (2024, first full-duplex speech-to-speech model), **Hibiki** (first real-time speech-to-speech translation), first TTS model running locally on a smartphone. - Gradium trains foundation audio models: TTS, API, on-device speech-to-text, speech-to-speech translation. ## Historical Framing: Four Generations of Voice Agents ### 1. Siri (2011) - "Technically a voice agent," constrained, **closed-ended dialogue**. - Architecture: transcript → natural language understanding (classification of app/action), no LLM; complex pipeline limited to specific use cases. ### 2. Original OpenAI Voice Mode - Open-ended conversation, much less constrained; but **agentic capabilities "inexistent"** (couldn't give weather or NASDAQ quotes). - Naturalness better than Siri; latency still several seconds. ### 3. Cascaded Voice Agents (STT → LLM → TTS) - Demoed a drive-thru bot: tracks orders, triggers actions, payment — "an actual agent" with tool calls, reasoning, planning. - **Claim:** Agentic capabilities regained, more powerful and generic than before, with good naturalness. - **Limitations:** Latency better than cascaded predecessors but above human conversation; naturalness fundamentally limited by passing through text — tone and emotion are lost. ### 4. Speech-to-Speech Models (e.g., OpenAI Advanced Voice Mode) - STT/LLM/TTS absorbed into one model; naturalness can be better (understands non-linguistic information); latency "really, really nice." - **Key limitation stated:** Intelligence remains much more limited than cascaded counterparts. - **Turn-taking limitation:** Speech-to-speech ≠ full duplex; models segment conversation into speaking/listening. Demo of **backchanneling** failure — model treats "yeah," "mhm" as interruptions. ## Walkie-Talkie Thesis - Any real-time model today is **half duplex** (he notes a bidirectional model "will come from OpenAI"). - Human conversation has constant bidirectional flow: backchanneling, interruptions, overlap — "up to 20% of the time" both parties speak simultaneously on a phone call. - This flexible dynamics makes conversation comfortable; turn-taking cannot resolve the backchanneling problem regardless of latency. ## Technical Explanation: Training Speech-to-Speech Models - Text LLMs predict next word; speech models predict **next audio** given past audio. - Raw audio is computationally infeasible: an 8-word sentence ≈ 72,000 time steps at 24 kHz; with quadratic LLM complexity, ~10,000× longer sequence = ~100 million× cost. - Solution: **neural codecs / audio tokenizers** (encoder compresses audio, decoder reconstructs), letting LLMs train on audio tokens like text. - Standard speech-to-speech models interleave user/system tokens — still half duplex. - **Multi-stream language models** (their innovation, claimed also used by Thinking Machines and "most likely" OpenAI's bidirectional model): transformer models **two** token streams, allowing both parties active/inactive simultaneously. - Moshi demo: model interrupts, guesses question ends, resilient to chaotic situations — though "back then it was even a bit irritating to people because it was interrupting you all the time." ## Central Tension: Naturalness vs. Intelligence - **Claim:** There is a fundamental tension — every improvement in naturalness reduced intelligence. - Cascaded agents are "as smart as the best text models" (Anthropic/OpenAI latest models, reliable tool calls). - **Reasoning for the tradeoff:** an LLM has a fixed weight "capacity"; adding speech understanding/production takes capacity from intelligence. "Fundamentally, there is a cost of adding a new modality... paid in intelligence." - Their full-duplex Moshi was "even more stupid" than speech-to-speech, which was already less intelligent than cascaded systems — fine for chit-chat, but "there is no way we can give up on intelligence." ## Two Forward Paths ### Path 1: Scaling end-to-end - Bigger, better pre/post-trained speech-to-speech models, progressively increasing intelligence. - Zeghidour speculates OpenAI is on this path ("I'm not at OpenAI, so I don't know because they don't release their model") — frontier text LLM fine-tuned on audio. - Pros: single model to orchestrate, easy deployment. - Cons: extremely complex and costly text→speech conversion. ### Path 2: Hybrid split (their bet) - Introduced in their paper "Moushiraq" (as rendered in transcript); reused by Thinking Machines' interaction models. - Architecture: small, possibly on-device, full-duplex natural speech-to-speech interface that "needs to know when it doesn't know" and delegates reasoning/tool calls to a **background text LLM** serving hundreds to thousands of voice interfaces asynchronously. - Two stated advantages: - **Cost:** chit-chat through gigantic speech-to-speech models that "can resolve differential equations" is "a loss of money." - **Backend control/optionality:** customers want to switch backend intelligence (he cites a model release "5 was released a few minutes ago"); speech-to-speech ties hands to one provider. Evidence cited: OpenAI's advanced voice mode ran on GPT-4o for a long time despite newer text generations, because the conversion process is expensive and slow. - **Self-acknowledged tension:** their culture favors end-to-end systems ("the bitter lesson"), but they bet on hybrid anyway. - Goal: an economically viable and agentic alternative to cascaded systems, which "are still most of the market today in voice." ## Closing - Models available at gradium.ai; Gradium is recruiting research scientists and engineers. ## Notable Qualifications / Unresolved Tensions Preserved - Zeghidour explicitly flags speculation about OpenAI's approach ("I'm not at OpenAI... I guess"). - He admits his full-duplex model was annoyingly interruptive at release and less intelligent. - He notes his own preference for end-to-end ("bitter lesson" culture) conflicts with his hybrid bet. # Observation Record: Speaker Roles & Provenance ## Video Identification - **Title:** *Your Voice Agent is Just a Walkie Talkie — Neil Zeghidour, Gradium* - **Video ID:** `01a15ca7-63ba-7f74-bf21-684c9f50094d` - **Format:** Conference-style talk with embedded demo and archival video clips ## Speakers & Roles ### Primary Speaker — **Neil Zeghidour** (self-identified as "Nel"; transcript rendering of name) - **Role:** Creator-side presenter; **co-founder and CEO of Gradio/Gradium**, a Paris-based startup - Speaks throughout the presentation portions, delivers demos, and closes with recruitment and product promotion ### Voices in Embedded Clips (no speaker names given in transcript) - **Siri 2011 announcement clip** — demonstrator and Siri voice (Apple event audio) - **OpenAI original voice mode clip** — storyteller voices reading bedtime story request - **OpenAI Advanced Voice Mode announcement clip** — speaker "Mark" and ChatGPT voice - **Drive-thru bot demo** — customer and bot voices (live-coded with Gradio's models) - **Backchanneling demo** — human user and turn-taking voice model - **Moshi August 2024 demo** — presenter and full-duplex model in spaceship-roleplay dialogue ## Speaker-Attributed Claims & Provenance ### First-party claims (Gradio/Kyutai, per speaker) - Originated from **Kyutai ("QTI" in transcript)**, a non-profit research lab - Released **Moshi** (2024) — claimed first full-duplex speech-to-speech model - Released **Hibiki** — claimed first real-time speech-to-speech translation system - Released claimed first TTS model running locally on a smartphone - **Multi-stream language models** — their full-duplex technique, claimed also used by **Thinking Machines** and "most likely" OpenAI's bidirectional model (speculative, flagged as such by speaker) - Paper **"Moushiraq"** (transcript rendering) introducing the hybrid split-model approach ### Third-party claims referenced (attributed, not verified) - **OpenAI**: voice mode generations, Advanced Voice Mode powered by GPT-4o until recently, upcoming bidirectional model, presumed path-one scaling approach ("I'm not at OpenAI, so I don't know") - **ElevenLabs**: speech-to-speech model architecture described - **Anthropic/OpenAI**: referenced as backend text models for cascaded agents - **"So they 5"** (transcript artifact; likely a model release) mentioned as released "a few minutes ago" ### Quantitative/experiential claims - Up to **20% simultaneous speech** in phone conversations (unsourced assertion by speaker) - Latency of original OpenAI voice mode "documented in the previous presentation" (external talk, not in this transcript) ## Promotional Segments — **Flagged** - Opening company introduction of Gradio/Gradium products (TTS, API, on-device STT, speech-to-speech translation) - Embedded **drive-thru bot demo** of their own voice models - Embedded **Moshi demo** from their August 2024 announcement - Closing: **"You can use our models on gradium.ai"** and **recruiting pitch** for research scientists and engineers ## Metadata & Availability Limits - **No curated metadata supplied**; no comment availability, channel, or date information present - Transcript contains **transcription artifacts** ("Nel," "Gradio" vs. title's "Gradium," "QTI," "Moushiraq," "GPT-3 time," "AdSense/Ad Ons voice mode") — uncertainty remains - Speaker's claims about competitors (OpenAI internal approach, Thinking Machines usage) are **explicitly speculative or second-hand** in places \>\> Okay, hi everyone. I'm Nel co-founder and CEO of of Gradio. So Gradio is a startup based in Paris. Most of our background is from research. In particular, we have invented algorithms such as audio LLMs, speech speech-to-speech models, neural codex, and so on and so forth. And basically we started from a research project called QTI, a non-profit research lab that has been focusing on voice since day one. So in particular, we released in 2024 the first full duplex speech-to-speech model called Moshi, the first real-time speech-to-speech translation system called Hibiki, and the first TTS model that can run locally on a on a smartphone. And basically I will just say a few words about what we do, but we are a model company that trains models for building voice agents and voice applications. So we do TTS, API, and on-device speech-to-text, speech-to-speech translation, and much more to come. What we do is that we train foundation models for audio, and then we can apply them for a lot of different tasks. So I go quickly on voice agents because it's the fourth talk about the topic, but basically now we have these voice interfaces that we can use to do a lot of things across a variety of products and types of interactions with NPCs, with customer agents, language learners, coach, and so on and so forth. And in this talk I tried to go through the history of this technology and where I see it going in the next years. And maybe to start, I think we can take a look at the announcement of Siri back in 2011. And you'll see that it's actually, you know, I think it it aged pretty well. \>\> What is the weather like today? \>\> Here's the forecast for today. \>\> It is THAT EASY. \>\> \[cheering\] \>\> LOTS OF THINGS. We've integrated with the stocks. So, you can ask it about the stock market. Something like How is the NASDAQ doing today? \>\> NASDAQ composite is down right now at 2,321.70. \>\> Again, you can ask this from the lock screen anywhere. Just press the button and ask. You can ask about, you know, the NASDAQ, the Dow. \>\> So, what you just saw is what kind of a voice agent. It was a bit constrained, but it was technically a voice agent. And the architecture behind it, so you have seen a thousand times today the STT LLM TTS. Back then, it was even worse, right? So, there was no LLM, obviously. So, there was what was called natural language understanding. So, you would go from the transcript and uh try to do basic classification of what is expected, uh what is the app that is supposed to be uh controlled, what is the action to trigger, and so on. So, it was a very complex pipeline and very constrained to very specific use cases. So, it was what was called close uh ended um uh dialogue back then. Then, the next generation of models, uh I could say, I think was is illustrated pretty well by the OpenAI voice mode, not the advanced voice mode, the original one. In which case, you just have um a conversational experience that is open-ended. So, it's much less constrained than Siri. You can talk about pretty much anything. At the same time, you don't have um as the same agentic capabilities, let's say, in the sense that it cannot really give you the NASDAQ or uh give you the weather. \>\> We want to hear a bedtime story. Tell us a story about the super-duper sunflower hedgehog named Larry. Start with telling us a little bit about him. Larry was a unique hedgehog unlike any other. He had bright sunflower petals instead of spines. Residing in Meadowville, Larry spread joy and color wherever he rolled. \>\> So, in that context, basically that was, you know, kind of second generation of models where all the logic, the very complex logic was replaced uh by an LLM. And if we compare Siri and the voice mode from OpenAI, uh not the Ad Ons one, Ad Ons one, the original one, uh the big strength is opening the dialogue, but the agentic capabilities were inexistent. The naturalness was, obviously, much better than with Siri. The latency was still pretty bad. So, we are talking about several seconds, and that was documented in the in the previous presentation. And so, then, you know, we're in this third phase, which is a voice, what we call voice agents. Uh and here I'm going to show a quick demo of uh of a drive-thru bot uh live-coded with our with our voice models. \>\> Hi there. What can I get started for you today? \>\> Hi, I'm actually in a hurry and I'd like to order a chicken sandwich quickly. \>\> We have the classic, spicy, deluxe, or grilled chicken sandwich. Which one would you like? \>\> Uh let's go with the classic chicken sandwich. \>\> Anything else for you? \>\> Uh what else do you have on the menu? \>\> We have sandwiches, sides, drinks, and desserts. What can I get for you? \>\> Can I get some mac and cheese, maybe? \>\> You have a classic chicken sandwich and mac and cheese. Your total comes to eight. \>\> So, now you know, unlike what we just saw, it's an actual agent. It's taking actions. It's keeping track of the order. It's then going to make you pay. So, it's it's an actual voice agent that can do uh real tasks. So, here instead of having an LLM that is just a conversational interface, we have a real agent that is empowered with tool call, reasoning, planning, and and all this stuff. So, what we see now is we have gained back agentic capabilities, and actually they are much more uh powerful and generic than before, while keeping a very good level of uh of naturalness. And that's where speech-to-speech LLM came. In particular, what we could see here is the latency, it's better with cascaded system, but it's still higher than you will have with human conversation. And as also was explained before, the naturalness is fundamentally limited by the fact that you go through text, so you lose a lot of information about what uh is said, the tone, the emotion of the user, and so on and so forth. So, now that we have tackled intelligence and agentic capabilities, speech-to-speech seems like a natural next step for naturalness and latency. And so here it's the announcement from the uh OpenAI advanced voice mode. \>\> \[clears throat\] \>\> Hey, ChatGPT. I'm Mark. How are you? \>\> Oh, Mark. I'm doing great. Thanks for asking. How about you? \>\> Hey, so I'm on stage right now. I'm doing a live demo, and frankly I'm feeling a little bit nervous. Can you help me calm my nerves a little bit? \>\> Oh, you're doing a live demo right now? That's awesome. Just \>\> I think we all remember it was very impressive very impressive release. And in that context now, all the steps of STT, LLM, and TTS have been absorbed into a a single one. And so now, intelligence, you know, like naturalness is still very good. Uh actually it can be better because it can understand non-linguistic information. Latency is really, really nice. Honestly, it doesn't make sense to go uh better than that. Interestingly and everyone was used any uh speech-to-speech model can uh attest that the intelligence is still much more limited in that context than uh the cascaded counterpart. So, the speech-to-speech models are fundamentally still limited compared to the textual models. Another limitation is turn-taking. So, people tend to mix speech-to-speech and full duplex. But basically, when you do have a speech-to-speech model like GPT-3 time, it's still based on fundamental turn-taking. In the sense that it's going to segment the conversation into as long as the model is speaking or the model is listening. And to give to show you how this can make an interaction unnatural, I'm going to show a a small demo with what is called backchanneling, which is this very human thing that you do when someone talks to you is that you say, "Mhm, yeah." and so on. \>\> Hey, how's it going? Just like to brainstorm a bit about it with you. \>\> Oh, that's a great topic. Yeah, I'd love to help you brainstorm. Are you thinking what Exactly. Yeah, I was thinking \>\> No, no, I didn't mean to interrupt, you know, I was just saying, "Yeah." like that. You can just keep going. I know, don't mind me. It's just something I typically do. \>\> Uh no worries at all. Yeah, I was just going to say we could break it down into a few aspects. Like Yeah, exactly. \>\> please stop stop interrupting. You know, it's called backchanneling. Humans do it all the time. It shows that you're just following the conversation. That you don't, you know, like interrupt you in your flow. Just just going. \>\> Ah, got it. Thanks for letting me know. \>\> No problem. Oh, come on. \>\> Yeah, so you see, you know, it's it's still very annoying. Uh you can have lightning speed latency. Fundamentally, this is uh an issue that can not be resolved when you're using turn-taking. So, here that's the walkie-talkie. Um any real-time model today, I mean, now there is a bidirectional one that will come from OpenAI, but it's called half duplex. So, the model is listening or speaking. A human conversation has a constant flow between two people. People do back channeling. People interrupt one another, talk on one another, and so on. If you have If you're having a relative on the phone, there is up to 20% of the time where you are both speaking at the same time. And that makes, you know, this very flexible dynamics in the conversation makes it much more comfortable for humans. And so, to understand how we can make a model full duplex, I'll give a very short presentation of how we train such models. So, the way you create a speech-to-speech model half duplex or full duplex is the following one. So, you you start from a text LLM, which is a probabilistic models over over words. And instead of predicting the next word based on the past, what you want to do is rather predict the next audio based based on the past audio. The issue now is that if you pass a raw audio to your model, which is, you know, a waveform, it's air pressure variations. Uh basically, you take this sentence, it's eight words. It takes around 3 seconds to pronounce it. And so, at 24 kHz audio, instead of having eight words, the audio form is 72,000 time steps that you would need to feed to your LLM. Given that LLMs have quadratic complexity with sequence length, so the complexity is the square of the sequence length. A 10,000 times longer sequence is 100 million times more expensive to to process. So, there is no way you can train an LLM on raw audio. So, the way you address it is by creating neural codecs, or you can also call them audio tokenizers. And basically, it's an encoder that takes an audio and compresses it in a very dense compressed representation, a bit similar to text. And then you have a decoder that can reconstruct high-quality audio from it. So, now you have gone from the audio domain into a abstract representation domain, where you can train an LLM exactly like you would train it on text. And the speech-to-speech model from ElevenLabs, as I was showing before, works in this fashion. So, instead of having text tokens into your model, you have audio tokens that represent either the LLM or the user, and you put them one after the other, and the model predicts the audio tokens uh that should be said by the model, being given the context from both sides of the conversation. However, you can see that it's still a sequence between user and system, which is still half duplex. So, how did we make the first full duplex model ever? Very simple. We call it multi-stream language models. That's the technology now used also by Thinking Machines for their interaction model, and most likely by the for the by the directional model of OpenAI. Is that instead of having a transformer that models one sequence of tokens, it models two of them, so that both parties can be active at the same time, inactive at the same time, one active and one inactive. And I just show a very quick demo of uh of how it sounds \[snorts\] like, but that's the release of machine August 2024, uh where we did an announcement live on stage talking to it for the first time. And you'll see that the model often guesses the end of the question, answers over the speaker, and both speaking at the same time is not breaking the flow like we saw with GPT. The whole thing is just extremely resilient to the most chaotic uh situations. \>\> So, the planet is serious 22. Can you plot a trajectory course to it, please? \>\> Yes, sir. \>\> Okay. How long is it going to take us to get there? \>\> it out. It's approximately 5 months to get there. \>\> Okay, that's that's not too bad. Uh do you think we have all we need on board the ship to start the mission? \>\> We have everything we need. \>\> So, back then it was even a bit irritating to people because it was interrupting you all the time. But the thing is that you can use you could still use it in extremely noisy environments with a lot of noise, people coughing, and so on. And you know, the flow is just constant. You don't get this very irritating break of the conversational flow. So, these full duplex models, they are the highest level of naturalness you can expect. That's the same conversation with a human. The thing is, in with our models, it was even more stupid than speech-to-speech models that were already less intelligent than cascaded systems. It's probably fine for some use cases if you just want to have a chit-chat. You know, the model doesn't need to be very intelligent. But make an actual full duplex voice agent, there is no way we can give up on on intelligence just to gain uh speech-to-speech abilities. So, how do we finally make models that tackle all these aspects jointly? And I think interestingly, if you if you look at the history I showed, there is a tension between naturalness and intelligence. So, every time we improve naturalness or humanness of the of the models, they were less intelligent than the cascaded system. The cascaded agents, they are basically as smart as the best text models. So, if you have a voice agent that is powered by the latest model from Anthropic or OpenAI, it's going to be extremely smart, have all the same reliability for tool call, and so on. Speech-to-speech has this naturalness aspect. However, you give up intelligence to get that. And the reason why you give up intelligence is remember that the LLM is a model that has a certain number of weights that we call the capacity. And if you take a text model and now it not only has to handle text, but it also needs to understand speech and produce speech, it's taking some of its capacity, and this capacity now is taken from the intelligence. So, fundamentally, there is a cost of adding a new modality to a text model that is going to be paid in in intelligence. So, where do we go from here? There are two paths that are in front of us, and both are going to be explored at the same time. The first one is scaling the model. So, making your speech-to-speech model bigger, better pre-trained, better post-trained, and so on. We likely progressively increase its intelligence until it it's good enough for a lot of use cases. The second one is splitting the model between naturalness and intelligence. The first one, I'm not at OpenAI, so I don't know because they don't release their model. I guess OpenAI is the path one, so it's a frontier text LLM with a lot of science around post-training, in- instruct tuning to fine-tune it on audio, and teaching it to be quite smart while using audio. The nice thing about that is you have a single model to orchestrate, so it's quite easy to deploy. Um and one big aspect, however, is that it's a extremely complex and costly process to go from the text model to the speech-to-speech model. The second path is to split it. It's an approach that we introduced in one of our recent papers called Moushiraq, and that has been reused by, in particular, the Thinking Machine Interaction models. Where, basically, the idea is that now you have two models. The first one is a small, maybe even on-device, full-duplex, extremely natural speech-to-speech interface. And its only role is to keep a very natural conversation and be able to delegate all the thinking, tool calling, reasoning, agentic capabilities to a background text model. And so, the way to see it is you have a background text LLM that receives asynchronously queries from hundreds to thousands of small voice interfaces and just give them their text, you know? And, basically, what we did is um very small full-duplex model that just needs to know when it doesn't know, so that it can delegate to the background model. And the reason why we believe mostly in this approach, um and I go back to to it later, it's a our uh let's say our culture is more of first one, the bitter lesson. So, every time we've been pushing for end-to-end systems and so on. But now I think the hybrid approach has two main advantages. The The first one is cost. So, speech-to-speech models are notoriously quite expensive. And when you think about it, it's a loss of money to do chit-chat with gigantic speech-to-speech models that can resolve differential equations and so on. So, it doesn't really make sense economically to get all your workflow through this gigantic multimodal mixture of experts. At the same time, we see that people are very attached to their ability to control the backend, to be able to switch So they So they 5 was released a few minutes ago. People want to switch the backend and the intelligence and get a lot of optionality on that, right? When you're using a speech-to-speech model, your your hands are a bit tied with this model provider. And to give you an idea of that, until recently the AdSense voice mode from OpenAI was powered by GPT-4o, despite the fact that there have been several generations of the text model since then, because this process is so expensive and so long. For this reason, we rather bet on the hybrid approach because that will give something that is not only very natural and very nice for demos and impressive, but also will be a viable alternative from a economic point of view and agentic capabilities point of view to the best cascaded systems that are still most of the market today in voice. So, what now? Uh you can use our models on gradium.ai. You can apply to gradium. We are recruiting research scientists and engineers. And thanks for your attention. \>\> \[applause\]