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

Read the talk

Your Voice Agent Is Just a Walkie-Talkie

Neil Zeghidour traces voice interfaces from rigid assistants to tool-using agents, then shows why fast speech-to-speech still feels unnatural—and how audio tokens, simultaneous streams, and a hybrid voice-plus-text architecture could fix it.

From a talk by Neil Zeghidour

At a glance

Ideas worth remembering

  • Open-ended dialogue, external action, expressive speech, low latency, and simultaneous conversation are distinct capabilities; improving one does not guarantee the others.

  • Speech-to-speech can remove the explicit text cascade, but a half-duplex system still alternates between listening and speaking. Backchannels therefore become unwanted turn changes.

  • Neural codecs and multi-stream modeling solve different problems: codecs compress long waveforms into audio tokens, while two streams represent simultaneous speech and silence for both participants.

  • Full duplex makes overlap resilient but does not guarantee good timing or strong reasoning. The early system could interrupt too often while remaining less capable than cascaded agents.

  • The hybrid proposal gives a small full-duplex model responsibility for conversational flow and delegates difficult reasoning and tools to a replaceable text backend. Its success depends on reliable delegation and asynchronous coordination.

Voice assistants first traded agency for conversational range

Gradium co-founder and CEO Neil Zeghidour frames voice-agent progress around capabilities that often get conflated: open-ended conversation, useful action, natural vocal expression, low latency, and the ability to listen while speaking. Gradium grew from voice research at Kyutai and trains audio foundation models for tasks including speech recognition, synthesis, translation, and spoken dialogue. The history matters because each generation improved a different part of the experience rather than advancing all of them together.

Recording frame at 98 seconds
Recording frame at 98 seconds

The 2011 Siri demonstration already contained real agency. A person could ask for the weather or the Nasdaq, and the assistant connected that utterance to an application action. Behind the voice, speech recognition produced a transcript; natural-language understanding classified the request, selected the application, and triggered a supported operation. This was a complex, closed-ended pipeline: useful inside its predefined territory, unable to roam far beyond it.

OpenAI’s original voice mode reversed that balance. An LLM replaced much of the hand-built dialogue logic, so a user could request a bedtime story about Larry, a hedgehog with sunflower petals instead of spines, and receive an improvised response. The system gained conversational breadth and more natural speech, but in Zeghidour’s comparison it lacked Siri’s ability to retrieve weather or market information and still took several seconds to answer. Open-ended dialogue and agency turned out to be separate features.

Compare the ideasThree capabilities that evolved separately

Classifies a transcript, selects an application, and triggers a predefined action.

The early systems did not improve along one simple axis: action, conversational range, and natural timing arrived through different architectures.

0:120:42
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

0:12 · section reference included

A drive-through order shows what tool use adds—and what text removes

The drive-through demonstration develops the next step through one concrete order. The customer asks for a chicken sandwich, selects the classic option, asks what else is available, and adds mac and cheese. The observable change is persistent task state: after several exchanges, the agent can report that the order contains both items. Zeghidour says payment would follow, although the shown exchange stops before completing it.

Recording frame at 340 seconds
Recording frame at 340 seconds

The causal chain is straightforward: the system interprets each utterance, consults the menu, records the chosen sandwich, answers a category question, adds the side, and computes the current order. An LLM is no longer merely generating a conversational response; it participates in an agent equipped with tool calling, reasoning, planning, and task state. That recovers agency in a more general form than the original closed assistant.

The remaining weakness sits in the cascade: speech-to-text converts the voice into words, a text model reasons over those words, and text-to-speech generates a new voice. This architecture can use a strong text model and dependable tools, but every component adds delay. More subtly, the transcript flattens information that was present in the input audio—tone, emotion, hesitation, and other nonlinguistic cues. Speech-to-speech promises to preserve that information while reducing latency.

4:455:03
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

4:45 · section reference included

Backchanneling exposes the walkie-talkie underneath

A single speech-to-speech model can absorb the explicit speech-recognition, text-model, and speech-synthesis stages. In the advanced voice example, a person says he is nervous during a live demonstration and receives an immediate spoken response. Direct audio modeling can retain nonlinguistic information and deliver latency that Zeghidour considers already good enough. The cost, in his qualitative assessment, is that these speech-to-speech systems remain less capable than their text-based, cascaded counterparts; the talk does not present a benchmark for the size of that gap.

Recording frame at 432 seconds
Recording frame at 432 seconds

Speech-to-speech still does not imply full duplex. A half-duplex system divides the exchange into two states: the model speaks, or the model listens. That works until the listener contributes a brief “mm-hmm,” “yeah,” or “sure.” Humans use these backchannels to signal attention without asking for the floor, but a turn-taking system can interpret any incoming speech as an interruption.

The demonstration makes the failure visible. The model starts brainstorming, the user says “mm-hmm,” and the response stops. After the user explains that the acknowledgment was not an interruption, the model resumes—only to stop again at the next “sure.” Even the explicit lesson about backchanneling does not fix the exchange, because the failure is structural rather than instructional. The system must switch turns whenever it detects user speech.

That is the walkie-talkie problem. Faster turn changes can make half duplex feel responsive, but they cannot represent both sides speaking at once. Human calls include acknowledgments, interruptions, overlap, and moments when neither person speaks. Zeghidour says calls with a relative can contain simultaneous speech for up to 20% of the time; the talk supplies no measurement behind that figure, so it serves as an illustration of normal overlap rather than a quantified general rule.

Compare the ideasWhy low latency cannot repair half-duplex turn-taking

The model accepts user audio but does not speak.

The key difference is not response speed. It is whether both streams can remain active at the same time.

6:536:56
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

6:53 · section reference included

Raw waveforms are too long, so codecs turn audio into tokens

A text language model predicts the next token from earlier tokens. A speech model needs an analogous prediction target for audio, but a raw waveform is far longer than a sentence’s word sequence. Zeghidour’s example takes about three seconds to say and contains eight words. At 24 kHz, those three seconds contain 72,000 waveform samples: 3 × 24,000.

Recording frame at 656 seconds
Recording frame at 656 seconds

That length is prohibitive for ordinary transformer attention. Under the quadratic-cost relationship used in the talk, multiplying sequence length by 10,000 multiplies attention cost by 100 million. The 10,000 factor is a rounded illustration rather than the exact ratio between 72,000 samples and eight words, but it captures the scale problem: treating every audio sample like a language token makes the sequence astronomically expensive.

A neural codec, also called an audio tokenizer here, changes the representation. An encoder compresses the waveform into a dense sequence of learned audio tokens; the language model predicts over those tokens; a decoder reconstructs high-quality audio. Tokenization makes speech manageable for language-model training, but it does not solve duplex behavior. If user and system tokens are still placed one after another in a single sequence, the conversation remains turn-based.

How it fits togetherHow a waveform becomes modelable audio

Three seconds at 24 kHz contains 72,000 samples.

Compression solves sequence length; it does not by itself allow simultaneous conversation.

10:1610:46
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

10:16 · section reference included

Full duplex requires two timelines, not one alternating sequence

Full duplex adds a second representation change. Instead of one token stream alternating between user and system, a multi-stream model represents both participants independently. At any instant, both may speak, both may be silent, or either one may speak alone. Overlap is no longer an exceptional event that forces a turn transition; it is an ordinary state in the model’s input and output.

Recording frame at 789 seconds
Recording frame at 789 seconds

The Moshi demonstration makes that relationship audible through a fictional spacecraft exchange. A person asks for a route, travel time, and mission readiness. The model sometimes anticipates the end of a question and begins answering before the speaker finishes. Crucially, simultaneous speech does not terminate the interaction. The demonstration establishes overlap handling and conversational timing, not actual trajectory calculation or verification of the ship’s supplies.

Full duplex brings its own behavioral problem: the early model interrupted too often. Zeghidour describes that as irritating, even while reporting that the conversation remained continuous through noise and coughing. Two streams make overlap survivable; they do not teach politeness or determine when anticipating a speaker is appropriate.

Natural flow also did not make the model capable enough for serious agent work. Zeghidour bluntly describes the early full-duplex system as less intelligent than speech-to-speech models that were already behind cascaded agents. Casual conversation may tolerate that. An agent entrusted with tool calls, planning, or transactions cannot give up reasoning merely to sound more human.

His explanation is finite model capacity. A text model adapted to speech must understand language, interpret audio, and generate audio using a fixed collection of weights. Those extra modality responsibilities consume capacity that could otherwise support reasoning. This is a proposed explanation for the observed tradeoff, not a measured law quantifying how many reasoning capabilities each audio feature costs.

Compare the ideasOne alternating stream versus two concurrent streams

Contributions occupy one sequence, so the conversation alternates turns.

The multi-stream representation makes silence and overlap explicit for each participant.

11:4612:16
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

11:46 · section reference included

Two paths forward: scale everything, or separate conversation from intelligence

The first path is to keep the system end to end and scale it: use a larger speech-to-speech model, better pretraining, and stronger post-training until its intelligence becomes sufficient for more tasks. The operational advantage is simplicity—a single model handles the interaction. The development disadvantage is that adapting a frontier text model to understand and produce audio is complex, slow, and expensive. Zeghidour speculates that a major provider follows this path but explicitly says he does not know its internal architecture.

Recording frame at 1074 seconds
Recording frame at 1074 seconds

The second path separates the voice interface from the brain. A small, potentially on-device full-duplex model maintains timing, overlap, and natural speech. When the conversation requires reasoning, tools, planning, or external knowledge, it delegates asynchronously to a background text model. A shared backend could serve many small voice interfaces and return text for them to express conversationally.

This hybrid depends on a difficult handoff: the small interface must know when it does not know. The talk identifies that requirement without explaining how uncertainty is detected, how delegation is trained, or how a delayed answer is reconciled with a conversation that may have moved on. Those coordination details determine whether the split feels natural rather than like a cascade wearing a better voice.

Zeghidour favors the split for two practical reasons:

  • Cost: Ordinary small talk does not need a gigantic multimodal model capable of solving differential equations. A small interface can handle routine interaction and reserve expensive reasoning for requests that need it.
  • Backend choice: Developers can replace the text model as better reasoning systems appear. A unified speech-to-speech product ties conversational behavior and intelligence to one provider and to a costly audio-adaptation cycle.

The closing bet is deliberately more demanding than an impressive demo. A viable voice agent must match the natural overlap of full duplex while remaining economically competitive and capable of the tool use available in today’s cascaded systems. The hybrid architecture is presented as Gradium’s route toward that combination, not as proof that the handoff, economics, and agent reliability have already been solved.

How it fits togetherThe hybrid voice-agent architecture

Includes words, timing, acknowledgments, and overlap.

A lightweight full-duplex interface handles the live conversation and delegates difficult work to a replaceable text backend.

15:1615:45
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

15:16 · section reference included

Resources

From the talk

  • Gradium: Solving voiceArticle

    Background on Gradium’s audio-model approach, research lineage, supported speech tasks, and intended production applications.

Read the complete timestamped transcript
  1. 0:01

    [music]

  2. 0:12

    >> Okay, hi everyone.

  3. 0:14

    I'm Nel

  4. 0:15

    co-founder and CEO of of Gradio.

  5. 0:18

    So Gradio is a startup based in Paris.

  6. 0:22

    Most of our background is from research.

  7. 0:24

    In particular, we have invented

  8. 0:26

    algorithms such as audio LLMs, speech

  9. 0:28

    speech-to-speech models, neural codex,

  10. 0:31

    and so on and so forth.

  11. 0:32

    And

  12. 0:33

    basically

  13. 0:35

    we started from a research project

  14. 0:37

    called QTI, a non-profit research lab

  15. 0:39

    that has been focusing on voice since

  16. 0:41

    day one. So in particular, we released

  17. 0:43

    in 2024 the first

  18. 0:45

    full duplex speech-to-speech model

  19. 0:46

    called Moshi, the first real-time

  20. 0:48

    speech-to-speech translation system

  21. 0:50

    called Hibiki, and the first

  22. 0:52

    TTS model that can run locally on a on a

  23. 0:54

    smartphone.

  24. 0:56

    And basically I will just say a few

  25. 0:58

    words about what we do, but we are

  26. 1:00

    a model company that trains models for

  27. 1:02

    building voice agents and voice

  28. 1:04

    applications. So we do TTS, API, and

  29. 1:06

    on-device speech-to-text,

  30. 1:08

    speech-to-speech translation, and much

  31. 1:09

    more to come. What we do is that we

  32. 1:12

    train foundation models for audio, and

  33. 1:13

    then we can apply them for a lot of

  34. 1:15

    different tasks.

  35. 1:17

    So I go quickly on voice agents because

  36. 1:19

    it's the fourth talk

  37. 1:21

    about the topic, but basically now we

  38. 1:24

    have these

  39. 1:26

    voice interfaces that we can use to do a

  40. 1:28

    lot of things across a variety of

  41. 1:30

    products and types of interactions with

  42. 1:32

    NPCs, with customer agents, language

  43. 1:35

    learners, coach, and so on and so forth.

  44. 1:37

    And in this talk I tried to go through

  45. 1:40

    the history of this technology and

  46. 1:44

    where I see it going in the next years.

  47. 1:46

    And maybe to start, I think we can take

  48. 1:48

    a look at the announcement of Siri back

  49. 1:50

    in 2011.

  50. 1:52

    And you'll see that it's actually, you

  51. 1:55

    know, I think it it aged pretty well.

  52. 2:00

    >> What is the weather like today?

  53. 2:06

    >> Here's the forecast for today.

  54. 2:09

    >> It is THAT EASY.

  55. 2:11

    >> [cheering]

  56. 2:12

    [applause]

  57. 2:13

    >> LOTS OF THINGS. We've integrated with

  58. 2:14

    the stocks. So, you can ask it about the

  59. 2:16

    stock market. Something like

  60. 2:18

    How is the NASDAQ doing today?

  61. 2:23

    >> NASDAQ composite is down right now at

  62. 2:26

    2,321.70.

  63. 2:29

    >> Again, you can ask this from the lock

  64. 2:30

    screen anywhere. Just press the button

  65. 2:32

    and ask. You can ask about, you know,

  66. 2:33

    the NASDAQ, the Dow.

  67. 2:35

    >> So,

  68. 2:36

    what you just saw is what kind of a

  69. 2:38

    voice agent. It was a bit constrained,

  70. 2:39

    but it was technically a voice agent.

  71. 2:41

    And the architecture behind it, so you

  72. 2:43

    have seen a thousand times today the STT

  73. 2:45

    LLM TTS. Back then, it was even worse,

  74. 2:47

    right? So, there was no LLM, obviously.

  75. 2:50

    So, there was what was called natural

  76. 2:52

    language understanding. So, you would go

  77. 2:53

    from the transcript and uh try to do

  78. 2:57

    basic classification of what is

  79. 2:58

    expected, uh what is the app that is

  80. 3:00

    supposed to be uh controlled, what is

  81. 3:03

    the action to trigger, and so on. So, it

  82. 3:05

    was a very complex pipeline and very

  83. 3:09

    constrained to very specific use cases.

  84. 3:11

    So, it was what was called close

  85. 3:14

    uh ended um

  86. 3:16

    uh dialogue back then.

  87. 3:18

    Then, the next generation of models, uh

  88. 3:21

    I could say, I think was

  89. 3:24

    is illustrated pretty well by the OpenAI

  90. 3:26

    voice mode, not the advanced voice mode,

  91. 3:28

    the original one. In which case, you

  92. 3:29

    just have um a conversational experience

  93. 3:32

    that is open-ended. So, it's much less

  94. 3:34

    constrained than Siri. You can talk

  95. 3:35

    about pretty much anything. At the same

  96. 3:37

    time, you don't have um

  97. 3:39

    as the same agentic capabilities, let's

  98. 3:42

    say, in the sense that it cannot really

  99. 3:43

    give you the NASDAQ or uh give you the

  100. 3:46

    weather.

  101. 3:51

    >> We want to hear a bedtime story.

  102. 3:53

    Tell us a story about the super-duper

  103. 3:54

    sunflower hedgehog named Larry.

  104. 3:57

    Start with telling us a little bit about

  105. 3:58

    him.

  106. 4:03

    Larry was a unique hedgehog unlike any

  107. 4:05

    other.

  108. 4:06

    He had bright sunflower petals instead

  109. 4:08

    of spines. Residing in Meadowville,

  110. 4:11

    Larry spread joy and color wherever he

  111. 4:13

    rolled.

  112. 4:15

    >> So, in that context, basically that was,

  113. 4:17

    you know, kind of second generation of

  114. 4:18

    models where all the logic, the very

  115. 4:20

    complex logic was replaced

  116. 4:22

    uh by an LLM.

  117. 4:24

    And

  118. 4:25

    if we compare Siri and the voice mode

  119. 4:27

    from OpenAI, uh

  120. 4:29

    not the Ad Ons one, Ad Ons one, the

  121. 4:31

    original one,

  122. 4:32

    uh

  123. 4:33

    the big strength is opening the

  124. 4:35

    dialogue, but the agentic capabilities

  125. 4:37

    were inexistent. The naturalness was,

  126. 4:39

    obviously, much better than with Siri.

  127. 4:41

    The latency was still pretty bad. So, we

  128. 4:43

    are talking about several seconds, and

  129. 4:44

    that was documented in the in the

  130. 4:46

    previous presentation.

  131. 4:48

    And so, then, you know, we're in this

  132. 4:51

    third phase, which is a voice, what we

  133. 4:54

    call voice agents. Uh and here I'm going

  134. 4:56

    to show a quick demo of uh of a

  135. 4:58

    drive-thru bot uh live-coded with our

  136. 5:01

    with our voice models.

  137. 5:03

    >> Hi there. What can I get started for you

  138. 5:05

    today?

  139. 5:06

    >> Hi, I'm actually in a hurry and I'd like

  140. 5:08

    to order a chicken sandwich quickly.

  141. 5:11

    >> We have the classic, spicy, deluxe, or

  142. 5:14

    grilled chicken sandwich. Which one

  143. 5:16

    would you like?

  144. 5:17

    >> Uh let's go with the classic chicken

  145. 5:18

    sandwich.

  146. 5:19

    >> Anything else for you?

  147. 5:21

    >> Uh what else do you have on the menu?

  148. 5:24

    >> We have sandwiches, sides, drinks, and

  149. 5:27

    desserts. What can I get for you?

  150. 5:30

    >> Can I get some mac and cheese, maybe?

  151. 5:32

    >> You have a classic chicken sandwich and

  152. 5:34

    mac and cheese.

  153. 5:36

    Your total comes to eight.

  154. 5:38

    >> So, now you know, unlike what we just

  155. 5:40

    saw, it's an actual agent. It's taking

  156. 5:42

    actions. It's keeping track of the

  157. 5:44

    order. It's then going to make you pay.

  158. 5:46

    So, it's it's an actual voice agent that

  159. 5:48

    can do uh real tasks. So, here instead

  160. 5:51

    of having an LLM that is just a

  161. 5:53

    conversational interface,

  162. 5:54

    we have a real agent that is empowered

  163. 5:56

    with tool call, reasoning, planning, and

  164. 5:58

    and all this stuff.

  165. 6:00

    So,

  166. 6:02

    what we see now is we have gained back

  167. 6:05

    agentic capabilities, and actually they

  168. 6:07

    are much more

  169. 6:09

    uh powerful and generic than before,

  170. 6:12

    while keeping a very good level of uh of

  171. 6:15

    naturalness.

  172. 6:16

    And

  173. 6:17

    that's where speech-to-speech LLM came.

  174. 6:19

    In particular, what we could see here is

  175. 6:22

    the latency, it's better with cascaded

  176. 6:24

    system, but it's still higher than you

  177. 6:27

    will have with human conversation. And

  178. 6:29

    as also was explained before, the

  179. 6:31

    naturalness is fundamentally limited by

  180. 6:32

    the fact that you go through text, so

  181. 6:34

    you lose a lot of information about what

  182. 6:37

    uh is said, the tone, the emotion of the

  183. 6:39

    user, and so on and so forth.

  184. 6:41

    So, now that we have tackled

  185. 6:42

    intelligence and agentic capabilities,

  186. 6:44

    speech-to-speech seems like a natural

  187. 6:46

    next step for naturalness and latency.

  188. 6:48

    And so here it's the announcement from

  189. 6:49

    the uh OpenAI advanced voice mode.

  190. 6:53

    >> [clears throat]

  191. 6:53

    >> Hey, ChatGPT. I'm Mark. How are you?

  192. 6:56

    >> Oh, Mark.

  193. 6:58

    I'm doing great. Thanks for asking. How

  194. 7:00

    about you?

  195. 7:02

    >> Hey, so I'm on stage right now. I'm

  196. 7:03

    doing a live demo, and frankly I'm

  197. 7:05

    feeling a little bit nervous. Can you

  198. 7:07

    help me calm my nerves a little bit?

  199. 7:09

    >> Oh, you're doing a live demo right now?

  200. 7:12

    That's awesome.

  201. 7:13

    Just

  202. 7:14

    >> I think we all remember it was very

  203. 7:16

    impressive very impressive release.

  204. 7:18

    And in that context now, all the steps

  205. 7:21

    of STT, LLM, and TTS have been absorbed

  206. 7:24

    into a a single one.

  207. 7:27

    And so now,

  208. 7:28

    intelligence, you know, like naturalness

  209. 7:30

    is

  210. 7:32

    still very good. Uh

  211. 7:34

    actually it can be better because it can

  212. 7:36

    understand non-linguistic information.

  213. 7:37

    Latency is really, really nice.

  214. 7:40

    Honestly, it doesn't make sense to go uh

  215. 7:42

    better than that.

  216. 7:44

    Interestingly and everyone was used any

  217. 7:46

    uh speech-to-speech model can uh attest

  218. 7:48

    that

  219. 7:50

    the intelligence

  220. 7:51

    is still much more limited in that

  221. 7:54

    context than uh the cascaded

  222. 7:56

    counterpart. So, the speech-to-speech

  223. 7:58

    models are fundamentally still limited

  224. 8:00

    compared to the textual models.

  225. 8:03

    Another limitation is turn-taking. So,

  226. 8:07

    people tend to

  227. 8:08

    mix speech-to-speech and full duplex.

  228. 8:11

    But basically,

  229. 8:13

    when you do have a speech-to-speech

  230. 8:14

    model like GPT-3 time, it's still based

  231. 8:16

    on fundamental turn-taking. In the sense

  232. 8:18

    that it's going to segment the

  233. 8:19

    conversation into as long as the model

  234. 8:21

    is speaking or the model is listening.

  235. 8:23

    And to give to show you how this can

  236. 8:26

    make an interaction unnatural, I'm going

  237. 8:28

    to show a a small demo with what is

  238. 8:30

    called backchanneling, which is this

  239. 8:32

    very human thing that you do when

  240. 8:33

    someone talks to you is that you say,

  241. 8:34

    "Mhm, yeah." and so on.

  242. 8:37

    >> Hey, how's it going? Just like to

  243. 8:39

    brainstorm a bit about it with you.

  244. 8:42

    >> Oh, that's a great topic. Yeah, I'd love

  245. 8:44

    to help you brainstorm. Are you thinking

  246. 8:45

    what

  247. 8:47

    Exactly. Yeah, I was thinking

  248. 8:48

    >> No, no, I didn't mean to interrupt, you

  249. 8:50

    know, I was just saying, "Yeah." like

  250. 8:51

    that. You can just keep going. I know,

  251. 8:53

    don't mind me. It's just something I

  252. 8:54

    typically do.

  253. 8:56

    >> Uh no worries at all. Yeah, I was just

  254. 8:58

    going to say we could break it down into

  255. 8:59

    a few aspects. Like

  256. 9:01

    Yeah, exactly.

  257. 9:02

    >> please stop stop interrupting. You know,

  258. 9:03

    it's called backchanneling. Humans do it

  259. 9:05

    all the time. It shows that you're just

  260. 9:06

    following the conversation. That you

  261. 9:08

    don't,

  262. 9:09

    you know, like interrupt you in your

  263. 9:10

    flow. Just just going.

  264. 9:12

    >> Ah, got it. Thanks for letting me know.

  265. 9:14

    >> No problem.

  266. 9:16

    Oh, come on.

  267. 9:17

    >> Yeah, so you see, you know, it's it's

  268. 9:18

    still very annoying. Uh you can have

  269. 9:20

    lightning speed latency. Fundamentally,

  270. 9:23

    this is

  271. 9:24

    uh an issue that can not be resolved

  272. 9:27

    when you're using turn-taking. So, here

  273. 9:29

    that's the walkie-talkie.

  274. 9:30

    Um any real-time model today, I mean,

  275. 9:33

    now there is a bidirectional one that

  276. 9:35

    will come from OpenAI, but it's called

  277. 9:37

    half duplex. So, the model is listening

  278. 9:39

    or speaking. A human conversation

  279. 9:42

    has a constant flow between two people.

  280. 9:45

    People do back channeling. People

  281. 9:47

    interrupt one another, talk on one

  282. 9:48

    another, and so on.

  283. 9:50

    If you have If you're having a relative

  284. 9:52

    on the phone, there is up to 20% of the

  285. 9:54

    time where you are both speaking at the

  286. 9:56

    same time.

  287. 9:57

    And that makes, you know, this very

  288. 10:00

    flexible dynamics in the conversation

  289. 10:02

    makes it much more comfortable for

  290. 10:03

    humans.

  291. 10:04

    And so, to understand how

  292. 10:07

    we can make a model full duplex, I'll

  293. 10:09

    give a very short

  294. 10:12

    presentation of how we train such

  295. 10:13

    models. So, the way you create a

  296. 10:15

    speech-to-speech model half duplex or

  297. 10:16

    full duplex is the following one. So,

  298. 10:18

    you you start from a text LLM, which is

  299. 10:20

    a probabilistic models over over words.

  300. 10:23

    And instead of predicting the next word

  301. 10:24

    based on the past,

  302. 10:26

    what you want to do is rather predict

  303. 10:28

    the next audio based based on the past

  304. 10:30

    audio.

  305. 10:31

    The issue now is that if you pass a raw

  306. 10:33

    audio to your model, which is, you know,

  307. 10:35

    a waveform, it's

  308. 10:37

    air pressure variations.

  309. 10:39

    Uh

  310. 10:40

    basically, you take this sentence, it's

  311. 10:42

    eight words.

  312. 10:44

    It takes around 3 seconds to pronounce

  313. 10:46

    it. And so, at 24 kHz audio, instead of

  314. 10:49

    having eight words, the audio form is

  315. 10:51

    72,000 time steps that you would need to

  316. 10:53

    feed to your LLM. Given that LLMs have

  317. 10:56

    quadratic complexity with sequence

  318. 10:58

    length, so the complexity is the square

  319. 11:00

    of the sequence length. A 10,000 times

  320. 11:02

    longer sequence is 100 million times

  321. 11:04

    more expensive to to process. So, there

  322. 11:06

    is no way you can train an LLM on raw

  323. 11:08

    audio. So, the way you address it is by

  324. 11:10

    creating neural codecs, or you can also

  325. 11:12

    call them audio tokenizers. And

  326. 11:14

    basically, it's an encoder that takes an

  327. 11:16

    audio and compresses it in a very dense

  328. 11:19

    compressed representation, a bit similar

  329. 11:21

    to text. And then you have a decoder

  330. 11:23

    that can reconstruct high-quality audio

  331. 11:25

    from it. So, now

  332. 11:27

    you have gone from the audio domain into

  333. 11:29

    a abstract representation domain, where

  334. 11:31

    you can train an LLM exactly like you

  335. 11:33

    would train it on text.

  336. 11:35

    And the speech-to-speech model from

  337. 11:37

    ElevenLabs, as I was showing before,

  338. 11:38

    works in this fashion. So, instead of

  339. 11:40

    having text tokens into your model, you

  340. 11:43

    have audio tokens that represent either

  341. 11:45

    the LLM or the user, and you put them

  342. 11:47

    one after the other, and the model

  343. 11:49

    predicts the audio tokens uh that should

  344. 11:52

    be said by the model, being given the

  345. 11:53

    context from both sides of the

  346. 11:55

    conversation.

  347. 11:56

    However, you can see that it's still a

  348. 11:58

    sequence between user and system, which

  349. 12:01

    is still half duplex. So, how did we

  350. 12:04

    make the first full duplex model ever?

  351. 12:06

    Very simple. We call it multi-stream

  352. 12:08

    language models. That's the technology

  353. 12:10

    now used also by Thinking Machines for

  354. 12:12

    their interaction model, and most likely

  355. 12:14

    by the for the by the directional model

  356. 12:16

    of OpenAI. Is that instead of having a

  357. 12:18

    transformer that models one sequence of

  358. 12:20

    tokens, it models two of them, so that

  359. 12:23

    both parties can be active at the same

  360. 12:25

    time, inactive at the same time, one

  361. 12:27

    active and one inactive. And

  362. 12:30

    I just show a very quick demo of uh of

  363. 12:32

    how it sounds [snorts] like, but that's

  364. 12:34

    the release of machine August 2024,

  365. 12:37

    uh where we did an announcement live on

  366. 12:39

    stage talking to it for the first time.

  367. 12:41

    And you'll see that the model often

  368. 12:43

    guesses the end of the question, answers

  369. 12:46

    over the speaker,

  370. 12:48

    and both speaking at the same time is

  371. 12:50

    not breaking the flow like we saw with

  372. 12:52

    GPT. The whole thing is just extremely

  373. 12:54

    resilient to the most chaotic uh

  374. 12:56

    situations.

  375. 12:57

    >> So, the planet is serious 22. Can you

  376. 13:00

    plot a trajectory course to it, please?

  377. 13:02

    >> Yes, sir.

  378. 13:03

    >> Okay. How long is it going to take us to

  379. 13:05

    get there?

  380. 13:06

    >> it out. It's approximately 5 months to

  381. 13:08

    get there.

  382. 13:09

    >> Okay, that's that's not too bad. Uh do

  383. 13:11

    you think we have all we need on board

  384. 13:13

    the ship to start the mission?

  385. 13:14

    >> We have everything we need.

  386. 13:16

    >> So, back then it was even a bit

  387. 13:18

    irritating to people because it was

  388. 13:19

    interrupting you all the time. But the

  389. 13:21

    thing is that you can use you could

  390. 13:23

    still use it in extremely noisy

  391. 13:24

    environments with a lot of noise, people

  392. 13:26

    coughing, and so on. And you know, the

  393. 13:28

    flow is just constant. You don't get

  394. 13:30

    this very irritating break of the

  395. 13:32

    conversational flow. So,

  396. 13:35

    these full duplex models, they are the

  397. 13:37

    highest level of naturalness you can

  398. 13:38

    expect. That's the same conversation

  399. 13:40

    with a human.

  400. 13:42

    The thing is, in with our models, it was

  401. 13:44

    even more stupid than

  402. 13:46

    speech-to-speech models that were

  403. 13:47

    already less intelligent than cascaded

  404. 13:49

    systems.

  405. 13:50

    It's probably fine for some use cases if

  406. 13:52

    you just want to have a chit-chat. You

  407. 13:53

    know, the model doesn't need to be very

  408. 13:55

    intelligent. But make an actual full

  409. 13:57

    duplex voice agent,

  410. 13:58

    there is no way we can give up on on

  411. 14:00

    intelligence just to gain uh

  412. 14:01

    speech-to-speech abilities.

  413. 14:04

    So, how do we finally make models that

  414. 14:06

    tackle all these aspects jointly?

  415. 14:08

    And I think interestingly, if you if you

  416. 14:11

    look at the history I showed, there is a

  417. 14:12

    tension between naturalness and

  418. 14:14

    intelligence. So, every time we improve

  419. 14:16

    naturalness or humanness of the of the

  420. 14:18

    models, they were less intelligent than

  421. 14:20

    the cascaded system. The cascaded

  422. 14:22

    agents, they are basically as smart as

  423. 14:24

    the best text models. So, if you have a

  424. 14:26

    voice agent that is powered by the

  425. 14:28

    latest model from Anthropic or OpenAI,

  426. 14:29

    it's going to be extremely smart, have

  427. 14:31

    all the same reliability for tool call,

  428. 14:33

    and so on. Speech-to-speech has this

  429. 14:35

    naturalness

  430. 14:36

    aspect. However, you give up

  431. 14:38

    intelligence to get that. And the reason

  432. 14:40

    why you give up intelligence is remember

  433. 14:43

    that the LLM is a model that has a

  434. 14:45

    certain number of weights that we call

  435. 14:47

    the capacity.

  436. 14:48

    And if you take a text model and now it

  437. 14:51

    not only has to handle text, but it also

  438. 14:53

    needs to understand speech and produce

  439. 14:54

    speech,

  440. 14:55

    it's taking some of its capacity, and

  441. 14:58

    this capacity now

  442. 15:00

    is taken from the intelligence. So,

  443. 15:01

    fundamentally, there is a cost of adding

  444. 15:04

    a new modality to a text model that is

  445. 15:06

    going to be paid in in intelligence.

  446. 15:08

    So, where do we go from here? There are

  447. 15:10

    two paths that are in front of us, and

  448. 15:12

    both are going to be explored at the

  449. 15:14

    same time. The first one is scaling the

  450. 15:16

    model. So,

  451. 15:18

    making your speech-to-speech model

  452. 15:19

    bigger, better pre-trained, better

  453. 15:22

    post-trained, and so on. We likely

  454. 15:24

    progressively increase its intelligence

  455. 15:26

    until it it's good enough for a lot of

  456. 15:28

    use cases.

  457. 15:29

    The second one is splitting the model

  458. 15:32

    between naturalness and intelligence.

  459. 15:35

    The first one,

  460. 15:36

    I'm not at OpenAI, so I don't know

  461. 15:37

    because they don't release their model.

  462. 15:39

    I guess OpenAI is the path one, so it's

  463. 15:42

    a frontier text LLM with a lot of

  464. 15:44

    science around post-training, in-

  465. 15:46

    instruct tuning to fine-tune it on

  466. 15:48

    audio, and teaching it to be quite smart

  467. 15:51

    while using audio.

  468. 15:54

    The nice thing about that is you have a

  469. 15:55

    single model to orchestrate, so it's

  470. 15:57

    quite easy to deploy.

  471. 15:59

    Um

  472. 16:00

    and

  473. 16:02

    one

  474. 16:03

    big aspect, however, is that it's a

  475. 16:06

    extremely complex and costly process to

  476. 16:08

    go from the text model to the

  477. 16:09

    speech-to-speech model.

  478. 16:11

    The second path is to split it. It's an

  479. 16:13

    approach that we introduced in one of

  480. 16:14

    our recent papers called Moushiraq, and

  481. 16:16

    that has been reused by, in particular,

  482. 16:18

    the Thinking Machine Interaction models.

  483. 16:21

    Where, basically, the idea is that now

  484. 16:23

    you have two models. The first one is a

  485. 16:25

    small, maybe even on-device,

  486. 16:27

    full-duplex, extremely natural

  487. 16:29

    speech-to-speech interface. And its only

  488. 16:31

    role

  489. 16:32

    is to keep a very natural

  490. 16:34

    conversation and be able to delegate

  491. 16:37

    all the thinking, tool calling,

  492. 16:40

    reasoning, agentic capabilities to a

  493. 16:42

    background text model. And so, the way

  494. 16:44

    to see it is you have a background text

  495. 16:46

    LLM

  496. 16:47

    that receives asynchronously queries

  497. 16:49

    from hundreds to thousands of small

  498. 16:51

    voice interfaces and just give them

  499. 16:53

    their text, you know? And, basically,

  500. 16:55

    what we did is um

  501. 16:57

    very small full-duplex model that just

  502. 16:59

    needs to know when it doesn't know, so

  503. 17:01

    that it can delegate to the background

  504. 17:03

    model.

  505. 17:04

    And the reason why we believe mostly in

  506. 17:06

    this approach,

  507. 17:07

    um

  508. 17:08

    and I go back to to it later, it's a

  509. 17:12

    our

  510. 17:13

    uh

  511. 17:14

    let's say our culture is more of first

  512. 17:16

    one, the bitter lesson. So, every time

  513. 17:18

    we've been pushing for end-to-end

  514. 17:19

    systems and so on. But now I think the

  515. 17:22

    hybrid approach has two main

  516. 17:24

    advantages. The The first one is cost.

  517. 17:26

    So, speech-to-speech models are

  518. 17:28

    notoriously quite expensive.

  519. 17:30

    And when you think about it,

  520. 17:32

    it's a loss of money to do chit-chat

  521. 17:35

    with gigantic speech-to-speech models

  522. 17:37

    that can resolve differential equations

  523. 17:39

    and so on. So, it doesn't really

  524. 17:42

    make sense economically to get all your

  525. 17:44

    workflow through this gigantic

  526. 17:46

    multimodal mixture of experts.

  527. 17:48

    At the same time, we see that people are

  528. 17:50

    very attached to their ability to

  529. 17:51

    control the backend, to be able to

  530. 17:53

    switch So they So they 5 was released a

  531. 17:56

    few minutes ago. People want to switch

  532. 17:58

    the backend and the intelligence and get

  533. 17:59

    a lot of optionality on that, right?

  534. 18:02

    When you're using a speech-to-speech

  535. 18:03

    model, your your hands are a bit tied

  536. 18:06

    with this model provider. And to give

  537. 18:08

    you an idea of that,

  538. 18:09

    until recently the AdSense voice mode

  539. 18:11

    from OpenAI was powered by GPT-4o,

  540. 18:13

    despite the fact that there have been

  541. 18:15

    several generations of the text model

  542. 18:17

    since then, because this process is so

  543. 18:19

    expensive and so long.

  544. 18:21

    For this reason, we rather bet on the

  545. 18:24

    hybrid approach because that will give

  546. 18:26

    something that is not only very natural

  547. 18:28

    and very nice for demos and impressive,

  548. 18:31

    but also will be a viable alternative

  549. 18:33

    from a economic point of view and

  550. 18:35

    agentic capabilities point of view

  551. 18:38

    to the best cascaded systems that are

  552. 18:40

    still most of the market today in voice.

  553. 18:43

    So,

  554. 18:44

    what now?

  555. 18:45

    Uh you can use our models on gradium.ai.

  556. 18:48

    You can apply to gradium. We are

  557. 18:50

    recruiting research scientists and

  558. 18:51

    engineers. And thanks for your

  559. 18:53

    attention.

  560. 18:56

    >> [applause]

  561. 19:10

    [music]