Skip to content
Subscribe

Block unveils Dessa, a voice AI that learned to talk only by listening

Jack Dorsey's Block has shown Dessa, an AI model trained only on recorded speech that listens and talks at the same time and runs on a laptop. It is not public yet. Here is what Block says it does, what is still unknown, and why it matters for how we will use AI.

Ammar Khan

Ammar Khan

4 min read
A microphone on a stand, lit against a dark, empty stage.
Photo by Matthias Wagner on Unsplash

Key takeaways

  • Block, the company behind Square, Cash App and the Bitkey bitcoin wallet, announced Dessa on 23 October. It is a speech-to-speech AI model: audio goes in, audio comes out, with no text in between.
  • Block says Dessa was trained only on human conversation, can listen and speak at the same time, and runs locally on a MacBook Pro with an M4 Max chip or better.
  • It is a research preview. There is no public access, and Block has not published the model's size, benchmark results or a release date.
  • Block is giving early access to a small group of AI safety researchers to look for risks.

What Block announced

Block's AI research lab unveiled Dessa on 23 October. "We believe voice is the most natural way to interact with intelligence," Jack Dorsey, Block's co-founder and chairman, wrote on X.

Most voice assistants today work in steps. They turn your speech into text, pass the text to a language model, then turn the reply back into speech. Block argues that this loses what text cannot carry, such as tone, emotion, sarcasm or a sigh. "The most powerful AIs in the world can't hear how you feel," the company wrote.

Dessa skips the text step. According to Block:

  • It learned only from speech. It was trained "from scratch on nothing but human conversation", with "no text, no pretrained models". Everything it knows, Block says, came from raw sound.
  • It is full-duplex. Like a person on a phone call, it listens while it speaks. It can be interrupted, can laugh mid-sentence, and can make the small sounds people use to show they are following, such as "yeah" or "got it".
  • It runs on a laptop. Block says it runs locally on a MacBook Pro M4 Max or better, faster than real time. It was trained on a single NVIDIA GB200 NVL72 system.

Block published several recorded conversations between people and Dessa, including one in which it talks with OpenAI's GPT-Live voice model. Block says nobody steered the model's replies. These are company demonstrations, not independent tests, CoinDesk noted.

Who built it

The lab is a small team inside Block that reports directly to Dorsey. It was founded by former members of Dessa, a Toronto deep-learning startup that Square, now Block, acquired in 2020. Block says the team has built generative speech models for more than eight years. The project is led by Ragavan Thurairatnam.

Block gives three reasons for the work: curiosity, better understanding between people and AI, and safety. On safety, it argues that an AI that cannot hear distress in a voice may miss when someone is in pain.

What is not known yet

  • No release date. There is no public access, and Block has not said when or whether there will be.
  • No published numbers. Block has not disclosed the model's size or benchmark results, CoinDesk reported.
  • Languages. The demos are in English. Block lists how efficiently such models learn new languages as an open research question.
  • How emotion is produced. Block says emotion is built into the model. TechCrunch reported that it is not clear whether it is steered in some other way.

How people reacted

Reactions online split. Some users called the demos the most human-sounding AI they had heard. Binance founder Changpeng Zhao suggested uses in teaching children. Others warned that voices this natural could make scam calls harder to spot, CoinDesk reported. Block's call for outside safety researchers is aimed at finding risks it has not anticipated.

Why it matters for Pakistan's tech and crypto community

Block is best known in crypto for bitcoin products, but Dessa is not a crypto product. It is part of a wider shift in AI, from typing into apps towards talking to them. OpenAI's GPT-Live, released earlier this year, works the same full-duplex way, according to TechCrunch.

For developers, the notable claim is size: a conversational voice model that runs on a laptop rather than only in a data centre. TechCrunch noted that Square already uses voice AI to take orders, and that an on-device model could lower the cost of such services.

For everyone else, the practical point is that AI voices are getting harder to tell from human ones. A voice on the phone is less proof of who is calling than it used to be. Our guide on how to report a crypto scam in Pakistan sets out where to go if a call turns out to be a fraud.

What do you think?

Block's bet is that the future of AI is less screen and more conversation: fewer menus and text boxes, and more talking to AI the way you would talk to a person.

Is that where AI, and the more powerful agents being built on it, is heading? Would you rather talk to an AI than type to it? And would you trust a voice you could not tell from a human? Tell us in the comments or on our social channels.

What we're watching

  • Whether Block opens Dessa to developers, and on what terms.
  • Results from the safety researchers given early access.
  • Whether speech-only models like Dessa learn languages such as Urdu.

Sources

Share WhatsApp X LinkedIn