Full-Duplex Voice: What It Is and How Dessa Does It

⏱️ Reading time: 13 min

Dessa can laugh in the middle of your sentence, before you even finish speaking. It doesn’t wait its turn, the way Siri or Alexa do. It listens and responds at the same time, just like a real person. Block’s AI learned to hold a full conversation by listening only to audio, without a single word of text in its training. That detail makes it the most concrete example that exists today of what a full-duplex voice model means, and the best way to explain how it differs from any traditional voice assistant.

📑 En este artículo

1. TL;DR
2. What Is a Full-Duplex Voice Model?
3. Why a Voice-to-Voice System Matters
4. How This Real-Time Voice Conversation Works
5. Dessa, Block’s Conversational Voice Model
   1. What the Simultaneous Speaking-and-Listening Demos Show
6. Comparing Conversational Voice Models
7. Use Cases for a Real-Time Voice Conversation
8. Limitations of a Voice-to-Voice System
9. Going Deeper: The Architecture Behind Simultaneous Speaking and Listening
10. Frequently Asked Questions
    1. What sets full-duplex voice apart from a normal voice assistant?
    2. Is Dessa a product you can use today?
    3. What hardware does Dessa’s conversational model run on?
    4. What data was Dessa trained on if it didn’t use text?
    5. Does a voice-to-voice model like Dessa understand emotions?
    6. What company is behind Dessa?
11. References

TL;DR

• Full-duplex means listening and speaking simultaneously, without taking turns the way a normal phone call does.
• Dessa, from Block, was trained only on audio from real conversations, with no text and no prior language models.
• The model runs faster than real time and works locally on a MacBook Pro M4 Max or higher.
• Interrupting, laughing mid-sentence, or raising its volume when asked are reactions nobody hand-programmed.
• The classic cascade (voice to text to text to voice) loses tone; a full-duplex system keeps it end to end.

What Is a Full-Duplex Voice Model?

A full-duplex voice model is an artificial intelligence system that processes incoming audio and generates outgoing audio continuously and simultaneously, without fixed turns of “one speaks, then the other.” It can listen while it talks and react to an interruption mid-sentence, in real time.

The term comes from telecommunications. A full-duplex channel can transmit and receive data at the same time, like a regular phone call, while a half-duplex channel forces turns, like a walkie-talkie. Applied to voice, the difference is the same. A half-duplex assistant records your whole sentence, processes it, and only then responds. A full-duplex one listens and generates audio at the same instant, which is why it can interrupt or be interrupted without the conversation getting stuck.

Why a Voice-to-Voice System Matters

Today’s voice assistants (Siri, Alexa, the classic voice mode of chatbots) convert your audio into text, that text goes through a language model, and the response is synthesized back into audio. Along the way, almost all the information that isn’t words gets lost: tone, irony, a tired sigh, a voice breaking. The result is a conversation that works, but feels stiff and always a step behind the human rhythm.

Block, the company behind Dessa, argues that this loss of information isn’t just a user experience problem, it’s also a safety problem. If an AI can’t tell whether the person talking to it is panicking or just in a hurry, it can hardly react appropriately. According to Block, an assistant that doesn’t hear the tone of the person speaking to it could, in an extreme case, keep giving instructions while that person is in real pain.

Building a voice-to-text-to-voice cascade is much cheaper and more predictable: each piece can be bought separately, audited separately, and replaced without touching the rest. A full-duplex model trained end to end loses that modularity. If it fails, it’s harder to know which part of the network caused the problem, and so far almost no lab outside Block has published a model of this kind trained entirely from scratch on audio.

How This Real-Time Voice Conversation Works

The most common architecture in voice AI today is a three-piece cascade: speech recognition (ASR), a language model, and a synthesizer (TTS). The ASR turns audio into text, the language model reasons over that text, and the synthesizer converts the response back into audio. Each piece processes in turns: first it transcribes everything you said, then it thinks, then it speaks. It’s a chain of sequential steps, which is why the assistant can’t listen while it generates its response.

A full-duplex voice model collapses those three pieces into a single network that takes raw audio as input and produces raw audio as output, without passing through text at any intermediate point. To do this, the model needs to listen and generate audio in parallel: two streams running at the same time on the same network, not a conversation resolved in turns. That architecture is what lets it react in milliseconds to an interruption, instead of waiting for the microphone to detect silence to assume you finished speaking, which is the trick traditional assistants use.

Latency is the other critical factor. If the model takes more than a few hundred milliseconds to react, the conversation stops feeling natural, even if the content of the response is perfect. That’s why Block designed Dessa to run faster than real time: it needs to generate each fragment of response audio in less time than it takes to play back, so it doesn’t accumulate delay sentence after sentence.

sequenceDiagram
    participant U as User
    participant C as Cascade assistant
    participant F as Full-duplex model
    U->>C: finishes speaking
    C->>C: transcribes audio to text
    C->>C: generates response with the LLM
    C->>U: synthesizes and plays audio
    Note over U,C: the response arrives after 3 sequential steps
    U->>F: speaks
    F->>U: listens and responds in parallel
    Note over U,F: audio flows both ways at once

Dessa, Block’s Conversational Voice Model

Dessa is the model a small research lab inside Block built to test whether an AI can learn to talk without ever seeing a written word. The team trained it exclusively on recordings of real human conversations: no text, no pretrained language models as a starting point. According to Block, this is the first time an AI has been documented holding a realistic live conversation after learning exclusively from audio.

The lab behind Dessa isn’t new to this. Its founders come from the deep learning startup Dessa, based in Toronto, and have spent more than eight years building generative voice models. Today that team reports directly to Jack Dorsey, Block’s co-founder and chairman. What changes with this model isn’t the ambition (they’ve been in deep learning since 2012), but the decision to cut text dependency completely.

To train it, they used a single NVIDIA GB200 NVL72 unit, a relevant detail since most frontier language models are trained on thousands of those units in parallel. Dessa also runs faster than real time and works locally on a MacBook Pro M4 Max or higher, with no need for a data center to respond.

Block also opened a specific call for AI safety researchers, before expanding access to the general public. The reason the company gives is that a model that learned to talk without going through text is new territory, and they’d rather have outsiders look for risks the internal team hasn’t anticipated before opening the waitlist to more users.

What the Simultaneous Speaking-and-Listening Demos Show

Block published live recordings of people talking with Dessa, with no script at all. In one, someone tells a funny story and Dessa laughs mid-sentence, at the exact moment something becomes funny, not after hearing the whole sentence the way models that wait for a full turn do. In another, someone can’t remember a movie’s title and Dessa finishes the sentence for them, with the correct answer, even though everything it knows it learned from sound, never from text.

There’s a demo where someone surprises Dessa with an anniversary announcement and it responds with a whispered “oh my god,” like someone holding back emotion, and keeps talking in what sounds like genuine gratitude. In another, someone asks it to speak louder and Dessa raises its volume: nobody programmed that specific instruction, it inferred it from the request. Block also published a conversation between Dessa and GPT-Live-1, a third-party voice model: both voices are AI-generated, and the developer of GPT-Live-1 has no relationship with Block and didn’t endorse the experiment.

Comparing Conversational Voice Models

To place Dessa within the voice AI landscape, it helps to compare three different approaches based on how they handle audio and interruptions.

Use Cases for a Real-Time Voice Conversation

Block doesn’t sell Dessa as a product yet. For now it has only opened a waitlist and a call for AI safety researchers. But the stated goal shows where a model like this is headed. A voice conversation where the AI perceives real distress, not just alarm words, could change how automated help lines or mental health assistants respond. Technical support that detects growing frustration in someone’s tone, before the user says it in words, could escalate the case sooner.

For people with speech difficulties, a model that understands long pauses or repetitions as a natural part of the conversation, instead of cutting off listening after a few seconds of silence, could make a voice assistant usable for that group for the first time. Block hasn’t confirmed it’s testing this particular case, but the full-duplex architecture enables it by design.

There’s also a pure research angle. Block raises open questions about this type of architecture: how efficient it is at learning new languages compared to a text model, what kind of in-context learning a model that never saw a written word can do, and what behaviors emerge that nobody explicitly programmed, like laughing mid-sentence or adjusting volume.

Limitations of a Voice-to-Voice System

Training without text has a cost. A language model learns from trillions of written words, a volume no labeled audio corpus comes close to matching. Block admits Dessa learned language with hundreds of times fewer words than a text LLM, which limits how much factual knowledge it can retain compared to a model trained on the entire text of the internet.

⚠️ Heads up: Block hasn’t published benchmarks comparable to the ones used to measure text language models, so there’s no objective figure for how accurate Dessa is compared to a traditional assistant on factual tasks.

Access is also limited for now: there’s no public API or commercial launch date, just a waitlist and a call aimed at safety researchers. For anyone who wants to try it today, the only path is to sign up and wait their turn.

Going Deeper: The Architecture Behind Simultaneous Speaking and Listening

For two audio streams (the one coming in and the one going out) to coexist on the same network without the model getting confused about which is which, the system needs an attention mechanism that constantly distinguishes which part of the sound the other person said and which part it’s generating itself, frame by frame of audio. It’s similar to what a jazz musician does when improvising while listening to the band: they don’t wait for someone else’s solo to end before starting to think about their response, they build it in parallel and adjust it on the fly.

That backchanneling (the “uh-huh,” “I see,” brief laughs that don’t interrupt but show someone is following the conversation) is another of the behaviors Block highlights as learned, not programmed. A text model has no way to pick up that signal: it doesn’t exist in a transcript, because an “mmh” usually doesn’t even get written down. It only shows up if the model listened to thousands of hours of real conversation and there was an audio signal tied to that moment.

A spoken word takes up thousands of times more bytes than the same written word, and it doesn’t even sound the same twice: saying the same sentence twice produces two different audio signals, while written text is identical character by character no matter who types it. That difference explains why training from scratch on sound alone is a far more data-expensive bet for every concept learned, even when the result captures nuances text could never convey.

flowchart TD
    A["Microphone: incoming audio"] --> B["Full-duplex network"]
    B --> C["Real-time outgoing audio"]
    B --> D["Internal conversation state"]
    D --> B
    subgraph "Continuous processing"
    B
    D
    end

Your next step: sign up for the Dessa waitlist and listen to the demos Block published, paying attention to the interruptions and mid-sentence laughs, the detail that most clearly sets a full-duplex model apart from a turn-based assistant.

Frequently Asked Questions

What sets full-duplex voice apart from a normal voice assistant?

A normal assistant transcribes your audio into text, thinks through the response in text, and only then synthesizes the outgoing audio, in turns. A full-duplex system listens and generates audio in parallel, without that intermediate text step, which lets it react the instant it detects an interruption.

Is Dessa a product you can use today?

No. Block has only opened a waitlist for early access and a call for AI safety researchers. There’s no public API or confirmed commercial launch date.

What hardware does Dessa’s conversational model run on?

According to Block, it runs faster than real time and works locally on a MacBook Pro M4 Max or higher, without relying on a data center to respond during the conversation.

What data was Dessa trained on if it didn’t use text?

It was trained exclusively on recordings of real human conversations: only sound, with no transcripts or pretrained language models as a starting point, using a single NVIDIA GB200 NVL72 unit.

Does a voice-to-voice model like Dessa understand emotions?

It detects and reproduces signals like laughter, pauses, sarcasm, or volume changes because it learned them directly from the training audio. Block hasn’t published a standard metric for emotional accuracy, so that capability is described in the demos, not in an independently verifiable benchmark.

What company is behind Dessa?

Block, the company co-founded by Jack Dorsey. The team that built the model comes from the deep learning startup Dessa, based in Toronto, and reports directly to Dorsey within Block.

References

• Block: official Dessa announcement, with audio demos and training details.
• Wikipedia: technical explanation of full-duplex and half-duplex systems in telecommunications.
• NVIDIA: specifications of the GB200 NVL72, the hardware used to train Dessa.
• Block: corporate site of the company, co-founded by Jack Dorsey.

📱 Like this content? Follow @programacion on Telegram for daily tech content in Spanish: quick summaries, fresh content every day.

Featured image: Foto de Matt Botsford en Unsplash