Twitter founder Jack Dorsey's voice model Dessa is not trained on text, forming a contrast with the Grok Voice 3.0 approach

Twitter's co-founder built an AI that "can't read" but is very good at talking

Ifanr Download the app | Exclusive early access October 28, 2026 09:58:00 From Beijing Updated October 28, 2026 10:41:39

Dessa, the speech model launched by the former Twitter co-founder, completely bypasses the bridge of text, attempting to prove that a machine's knowledge and conversational abilities can grow directly out of sound itself.

"Since human babies learn to speak before they learn to read, why can't machines take the same path?" With this conviction, Jack Dorsey set up an AI lab and did exactly that.

Jack Dorsey is one of the founders of Twitter and now runs his own company, Block, whose core business is financial payment products such as Square and Cash App. Dessa is a research team under the company that focuses solely on AI research and reports directly to company founder Jack Dorsey.

Text was removed from the training process because text naturally filters out a great deal of information, fixing a sentence into a string of static characters. Every utterance of sound, by contrast, is vastly different. The Block team says Dessa, raised fully immersed in sound, can not only listen and speak at the same time and interject naturally, offering acknowledgments like "mm" and "yes" at any moment, but can even run locally on a MacBook Pro equipped with an M4 Max chip. Autoplay

Limited by training details and model scale that have not yet been disclosed, it is still hard for outsiders to assess the ceiling of this technical approach, but it does point to a long-obscured "non-mainstream" path: speech-to-speech.

First it talks, then it answers the phone

If Dessa represents a pure exploration of the nature of speech in the lab, then during the same period, another pragmatic path for speech models is accelerating into everyday workflows.

Almost simultaneously, xAI also launched Grok Voice 3.0, which likewise focuses on real-time end-to-end speech-to-speech interaction.

Unlike Dessa's purist training that "never touches text," mainstream commercial models represented by Grok care more about how to combine the text reasoning foundation of large models with native speech capabilities, so that voice can truly be put to use.

Take Tesla as an example: owners can talk directly to Grok in the car's infotainment system without separately picking up their phone and opening an app.

Interestingly, the evolution direction of this batch of mainstream speech models all revolves around an extremely subtle span of time: half a second.

When typing on a keyboard, waiting an extra half second hardly changes the nature of the exchange; but on a phone call, half a second of silence is enough to push the conversation to the opposite extreme. This is also why, even if the AI's answer is flawless, users are instantly pulled out of the experience if it rigidly waits for the person to finish speaking before responding, laughs at inappropriate moments, or stubbornly reads out its original script after being interrupted. In evaluations of Grok Voice 3.0, half a second is its optimization target.

Differences measured in milliseconds directly affect real commercial applications. At present, in-car spaces and customer service hotlines constitute the scenarios where voice AI can best play its role. According to internal operating data disclosed by xAI, in addition to handling massive volumes of in-car Tesla conversations every day, this speech system also handles about 20,000 customer service calls per day for Starlink, of which about 70% of issues are resolved without transferring to a human agent.

Beyond speed, the model must be able to clearly grasp requests in calls mixed with accents, background noise and frequent interruptions, smoothly retrieve information in the background, invoke tools, handle faults, and finally politely hang up. This whole series of operations needs to be completed seamlessly, and combined with a natural voice, only then does it truly count as "on duty."

In mid-October, OpenAI updated its speech model GPT-Live, not only improving the naturalness of multilingual accents in the ChatGPT client, but also adding nine new voices to its developer-facing API in one go, including English with Japanese and Korean accents, and supporting voice customization directly through text descriptions.

Human imagination of what voice can achieve has a long history. As early as 1982's Blade Runner, the protagonist Deckard repeatedly issued verbal commands to an image analyzer, making the image pan and zoom until hidden clues were locked in.

By Her, the operating system Samantha went a step further, transforming from a machine that executes commands into a soulmate that accompanies you at all times. It can not only complete all daily tasks, but its voice is also extremely lively, which is why it can build a relationship with the male lead.

No matter how cutting-edge the machine, it must have a voice. MOSS in The Wandering Earth, with only a few lines and only a handful of utterances, is indeed the representative of "artificial intelligence" in the film, reflecting the overall level of technology.

Technological evolution in reality has indeed realized the earlier imagination. Early mobile phone voice assistants could only handle simple commands with fixed phrasing, such as setting alarms and checking the weather. After smart speakers entered living rooms, the most common tasks they did were nothing more than playing music and setting countdowns. Only when large language models were combined with voice agents did machines truly begin to understand free-flowing colloquial speech and start helping users send and receive text messages, book restaurants, and follow developers' verbal descriptions to invoke coding tools to modify code.

This is not because the models are not smart enough, but because of the physical properties of sound as an information medium.

Sound is linear. When it comes to orchestrating AI, it is especially suited to expressing a vague intention. For example, an idea in programming, or directing an agent to arrange tasks.

Claude Code's voice mode

But in work and production scenarios, it is not good at letting people review multiple parallel options at the same time. When writing an article, we need to see the paragraph structure at a glance and go back to revise at any time; when writing code, we must stare at the screen to compare code differences and run results.

Even just booking a flight, glancing at the prices, departure and arrival times, and change and refund rules of several flights laid out on the screen at once is far more efficient than listening to a voice assistant read them out one by one.

The physical environment and social etiquette also quietly draw the boundaries of voice. When driving, cooking, or holding a child with both hands, both eyes and limbs are occupied, and speaking up is undoubtedly the least effortful savior; but once the scene switches to a quiet open-plan office, one person continuously talking to a computer still seems too bizarre. If the whole office talks together, it is even more like a large customer call center.

These scenarios involve constraints on information bandwidth and privacy boundaries, and no matter how many more milliseconds engineers shave off model latency, they cannot be solved.

The sci-fi prophecy of "say a word to the computer and everything is done" is gradually coming true. It's just that the more machines learn to speak like humans, the more clearly we realize which moments allow speaking and which moments still require keeping quiet.

"Special statement: The above content (including videos, pictures and audio) is uploaded and published by users of "Dafeng Hao," a self-media platform under Phoenix Net. This platform only provides information storage space services.

Notice: The content above (including the videos, pictures and audios if any) is uploaded and posted by the user of Dafeng Hao, which is a social media platform and merely provides information storage space services."