Voice AI Solved? First Look at Dessa on This MacBook

Block's small AI research team Dessa has released a speech model, also named Dessa. Jack Dorsey posted the demo on X, and when Akshay Pachaar reposted it, he wrote: 'voice AI is solved.' First, the facts: audio in, audio out, full duplex, and it can run locally on a MacBook Pro M4 Max or higher. It was trained only on human conversation audio, with no text and no pretrained model. The team claims this is the world's first AI that can hold real conversations in real time and is trained only on human speech.

Traditional voice AI goes from speech to text, then processes it, then synthesizes speech. Emotion, pauses, breathing, sarcasm, and laughter are lost at this step. Dessa's selling point is that it bypasses this path. The official site puts it bluntly: run speech through text and those things are gone, which is why most voice AI sounds stiff, slow, and, worst of all, annoying.

In the demo, it laughs right when something is funny, gets stuck mid-sentence, like a person. Other models usually wait until you finish and then laugh on cue. Tell it to speak louder, and it gets louder. The official site says no one taught it these things. In another clip, the other person can't remember the movie title, and Dessa finishes the sentence. In yet another clip, a surprise scene: it gasps, says 'oh my god' very softly, and then sounds like it is thanking someone through tears. The official site says these clips are real-time recordings of human conversations, with no one puppeteering, subtitles added afterward, and some clips flagged as containing profanity.

The technical details are worth a look too. Dessa is a true speech-to-speech model, full duplex, listening and speaking at the same time; it can interrupt, and it uses backchannels like 'yeah,' 'got it,' and 'ooh' to show it is listening. Inference is faster than real time, so there is no awkward wait. It was trained using only one NVIDIA GB200 NVL72. The team says it needs hundreds of times fewer words to learn language than LLMs, and jokes that LLMs also have to learn C++ and Angular. Speech data is harder: say the same sentence twice and the data is different; one spoken word has thousands of times more bytes than text.

It can also converse with GPT-Live-1. The official site notes that both voices in this clip are AI-generated; GPT-Live-1's developer is unrelated to Block and has not reviewed or endorsed Dessa.

Why build this? The team gives three reasons: curiosity, better mutual understanding between people and AI, and safety. AI is getting stronger, yet it is nearly deaf to emotion in voice. Imagine someone in severe pain while the voice AI has no idea and keeps talking. That is a safety problem. The Dessa team is giving a small group of AI safety researchers early access to find risks they did not anticipate.

The cold water came quickly too. Some users asked how much memory full duplex takes when running on a laptop; others asked whether heavy accents get through; still others asked how, with no text in the loop, you give system prompts, call tools, or read tickets. Akshay said voice AI is solved, and Sachin Mour replied: 'Voice AI is solved the way hiring a great voice actor solves your support queue.' Sounding human is one layer; being able to do the work is another. Whether this version of Dessa solves tool calling, the material gives no answer; whether Block will later pair it with a text model is also an open question.

One user brought up the Lindy rule, saying humans only talk to living things: other people, themselves, animals, objects with spiritual meaning. Now you can sit down and talk to a computer; the technology is here, and what remains is whether people want to.

Some people are already thinking about AI girlfriends.

Another demo was posted, showing a speech recognition interface with labels reading Excitement, Gasp, quiet OMG.

There is another demo with labels Laugh, Blown away, All the secrets, Speaking louder.

What is worth watching about Dessa is the combination of text-free training and local full duplex, not how human-like its laugh is. There are already plenty of human-like demos. What is worth watching is that it can interject in real time on a laptop without going through an ASR/TTS pipeline. Whether voice AI is solved depends on whether it can keep the same rhythm in real-world scenarios with accents, noise, ticket lookups, and tool calls. The demo is impressive, but the problems remain.

Dessa official site and waitlist: https://block.xyz/dessa
Akshay's original post: https://x.com/akshay_pachaar/status/2113989676482605106