Speech Model Learns to Interrupt, but Hasn't Learned Chinese Yet

Dessa's demo went viral in English-language X/Twitter circles. Block has released a pure speech-to-speech real-time conversation model, and Jack Dorsey tweeted about it himself. After listening to its live test conversation with GPT-Live, Hedy Zhang said she could no longer hear the machine in it: the tone of interruptions, the breath in speech, stress, hesitation, and fluidity all feel like a real person.

First, what it solves. Most voice AI products on the market take a detour: speech to text, process with a text model, then convert back to speech. Tone, pauses, breathing, and hesitation get flattened the moment they become text. Dessa's approach is more direct: sound in, sound out. Block says it was built from scratch and trained only on human voice conversations, with no text and no pretrained model. It is full-duplex, able to listen and speak at the same time, knowing when to interject and when to use short backchannels like "yeah" and "got it" to show it is listening. Inference is faster than real time, and the company says it can run locally on a MacBook Pro M4 Max or higher.

Some details in the demo are more convincing than the specs. It laughs at funny moments, in the middle of a sentence, like a person who cannot hold it in halfway through. Other models wait until you finish and then laugh on cue. If you ask it to speak louder, it raises its volume, even though no one specifically taught it to do that. When someone forgets the name of a movie, it fills in the second half. Hearing about an anniversary surprise, it gasps, quietly says "oh my god," and then continues answering. The official site labels clips like these as live recording: no one is controlling it, and what it says, when it says it, and in what tone are decided by the model itself.

The technical choice is a bit counterintuitive. LLMs learn language from massive amounts of text; Dessa relies only on voice. Block says it uses hundreds of times fewer words than an LLM to learn language, and training used only one NVIDIA GB200 NVL72. Voice data is harder to handle than text: say the same sentence twice and the data is different; the audio for one word is thousands of times the bytes of its text. Accent, pauses, breathing, sarcasm, humor, stress, emotion, and when to interject—each one changes the meaning. The official site gives an example: by moving the stress in "I didn’t say she stole my CUDA code," the same string of words can produce several meanings.

Community reactions are not all amazement. Some ask how to deploy it on a Mac, some ask whether it is open source, and some just say creepy. One commenter pointed out that the hardest part of voice is the rhythm of interrupting and turn-taking: when to stop, how long a hesitation can be before it feels unnatural—previously these were all patched together by text models. If it really is pure speech-to-speech, then latency and tone can line up. The remaining question is whether Chinese intonation and interruption habits can be equally natural. Block's official site lists "how to efficiently learn a new language" as an open research question, and there is no timetable for Chinese support yet. Hedy Zhang replied that it is currently in internal beta, English only, and not open source.

There is another issue that demos can easily cover up: content. After listening, one commenter said that once the question of whether it seems human is solved, the problem shifts to content—there is too much AI flavor. Hedy Zhang's reply was direct: this part really has no solution; you can only adapt and stay away from AI slop. Just because the voice is realistic enough does not mean what it says is worth listening to. This judgment is colder than "it seems like a real person," and closer to everyday use.

For developers, Dessa is worth watching for whether end-to-end speech training can transfer to other languages, whether full-duplex scheduling is stable in noisy environments, the real latency of local inference, and whether it will actually be open sourced. For ordinary users, the most intuitive change is that talking to machines may no longer require waiting for it to "finish thinking before speaking"; it can interject, pause, and catch you with its tone. The demo is impressive; next, see whether it can avoid interrupting at the wrong moment in Chinese.