Block Incorporated, the company owned by Twitter co-founder and former chief executive Jack Dorsey, has built a speech model called Desa that it says can hold a natural spoken conversation.
On X, Dorsey described voice as the most natural medium of communication, and Desa as a full two-way conversational speech model built to talk with people normally. One of its selling points is size: it is small enough to run on a laptop.
What it does differently
The behaviours Block demonstrates are the ones that make voice assistants feel mechanical when they are missing. Desa listens and speaks at the same time, taking in the user’s voice and answering directly in its own, rather than waiting for a turn to end. Block says it processes faster than real time, so the gap between hearing and answering is very short.
On a demo conversation published on Desa’s website, the model decides for itself when to speak, what to say and how to say it. It laughs mid-sentence when something amusing comes up, uses sounds of surprise where the situation calls for one, and offers sympathy. When the user trails off mid-thought, it holds the thread rather than resetting.
The training claim is the real story
Desa was trained by being played audio of humans talking to each other. No written data was used, and it was not built on top of an existing pre-trained model. Block says Desa is the first AI model trained only on human voice.
Why that matters takes a moment to explain. Most voice assistants are a relay: speech is converted to text, a text model works out a reply, and the reply is converted back to speech. Everything that is in a voice but not in its words — hesitation, irritation, a rising note of doubt, the breath before bad news — is discarded at the first step and invented at the last.
Block’s argument is that the world’s most powerful models cannot hear how a user feels, so conversation with them does not flow, and they sometimes respond inappropriately because they have missed the feeling entirely. Training on audio end to end keeps that information in the pipeline.
The caveat is that “the first model trained only on human voice” is the company’s own claim, made about its own product, at the point of launch. It is plausible and it is not independently verified. Desa is in a test phase, available only to a limited number of interested users, so the demo is the only public evidence — and a demo is chosen by the people who built the thing.
Running on a laptop is the part to watch
If the size claim holds, it matters more here than the conversational polish does. A speech model that runs locally is one that does not send the contents of a conversation to a server, does not need a reliable connection, and does not bill per minute. For a clinic taking patient histories, a call centre, or a government office recording complaints, those three properties are the difference between a pilot and a deployment.
We have seen the server-side version of this already: an English ambulance service has been letting AI listen to patients through an iPad and write the notes. The privacy questions in that trial are mostly questions about where the audio goes.
The question nobody has answered
Which languages. A model trained end to end on conversational audio inherits whatever accents and languages are in that audio, and nothing in Block’s announcement says Bangla is among them. For a text model you can at least measure the gap; for a speech model trained without text, there is not yet an obvious way to test Bangla performance from outside.
That is the standing problem rather than a new one. Bangla is still treated as a low-resource language, and the reason is simply how little usable data exists. A voice-first architecture could in principle help — spoken Bangla is abundant even where written, cleanly transcribed Bangla is not. Whether anyone trains on it is a commercial decision, not a technical one.
Prothom Alo credits the Indian Express for the report.
