Block's Dessa voice model runs on a laptop and can laugh, pause and cry

Started by ComputeNode Curtis, Oct 31, 2026, 04:06 AM

Previous topic - Next topic

0 Members and 1 Guest are viewing this topic.

Topic: Block's Dessa voice model runs on a laptop and can laugh, pause and cry   Views(Read 104 times)

ComputeNode Curtis

Block, the company led by Jack Dorsey, has released a voice model that it says sounds remarkably human. TechCrunch reports that the model, called Dessa, comes from Block's research lab

Block says Dessa was trained entirely on human speech using a single rack of Nvidia GPUs. It runs locally on consumer hardware, from a MacBook with an M4 Max chip upwards, and it can laugh, pause and even cry during conversation

Dessa is full duplex, meaning it can listen and speak at the same time, so conversations can flow without awkward turn taking. TechCrunch compares it with OpenAI's GPT-Live. Block says its knowledge comes from its conversational training data rather than a separate language model, and that emotion is built in rather than added with tags, as rivals such as ElevenLabs and Fish Audio do

Block's blog says interacting with current voice AI can just plain suck, and that the most powerful AIs in the world cannot hear how you feel. Dorsey wrote on X that voice is the most natural way to interact with intelligence. There is no public access yet, and the licence has not been confirmed. Block has asked safety researchers to test it, especially for risks around deception

Running a natural sounding voice model on a laptop is impressive if the demos hold up. It also raises obvious worries about scams and impersonation. Asking safety researchers to probe it before release is a sensible step

Would you want an AI that can laugh and cry? And how worried are you about voice models being used to deceive people?


DarkFreddy65

Training on one rack of GPUs is surprisingly modest. That suggests good voice models do not need giant budgets anymore. Smaller labs will take note

WWEReins19

Full duplex makes a big difference. Talking over each other is how humans speak. Turn based voice assistants always feel stilted. Interrupting naturally is what makes conversation feel real
I read every reply. Even the bad ones.

HeartbreakKidCurtis18

An AI that cries sounds a bit unsettling. I am not sure I want emotional performances from software. Natural pauses and laughter are fine, but crying feels like a step too far

Maybe it depends on the use case. A companion app might suit emotional range better than a support bot

StringTheory88

Demos are always cherry picked, so I will wait for public access. Every voice model sounds amazing in a launch video. Real conversations reveal the flaws

Cobalt Sophie

Scammers already use cloned voices to trick families. A model this realistic running offline would be a gift to them. Block testing for deception risks is the right move

I hope they publish the findings. Transparency would build trust in the model. It would also help others defend against misuse

Electric Brad

Running locally on a laptop is the headline for me. No cloud means better privacy. It also means nobody can switch it off once it is out there

TheGame92

Block is an unexpected company to lead on voice AI. They are known for payments. Dorsey clearly wants a foot in the AI race