Jack Dorsey’s company Block has announced Dessa, a new speech AI built by its small in-house research lab that was trained from scratch on nothing but human conversation audio — no text, no pretrained models. Block calls it “the most authentic full-duplex speech model ever built”, according to The Verge’s reporting on 23 October 2026. Full-duplex means Dessa listens and talks at the same time, like a person: it laughs mid-sentence, interrupts, backchannels (“yeah!”, “got it!”, “ooh”), and hears emotion, sarcasm and a cracking voice. It learned language from “hundreds of times fewer words” than text-based models see, trained on a single NVIDIA GB200 NVL72 rack, and runs faster than real time — even locally on a MacBook Pro with an M4 Max chip or better. Early access goes to a small group of AI safety researchers, with a public waitlist for everyone else.
Trained on speech alone
Dessa’s design breaks with how today’s voice assistants work. Most voice AI converts speech into text, runs it through a text model, and converts the answer back into speech — a pipeline that flattens tone, timing and emotion along the way. Block’s team, made up of former members of Dessa, the Toronto deep-learning startup, led by Ragavan Thurairatnam, took the opposite route: the model learned directly from conversation audio. The lab reports directly to Jack Dorsey, Block’s cofounder and chairman.
The result is a model that keeps everything text pipelines throw away. In live recordings of unscripted conversations the team published, Dessa mistakes its interlocutor for “Ronaldo” at first; another conversation veers from Christopher Nolan films into Bryan Singer’s Superman films. One clip even shows Dessa talking with OpenAI’s GPT-Live-1 voice model, a third-party system not affiliated with Block.
Small data, small compute
Two claims make the announcement stand out technically. First, Block says Dessa learned language from “hundreds of times fewer words” than the text data text LLMs train on — a genuine speech-native sample-efficiency result, if it holds up. Second, the whole model trained on a single NVIDIA GB200 NVL72 rack and runs locally on a high-end laptop. That points toward a very different compute path than the data-centre-scale training runs behind today’s frontier models — a direction with implications for products like IBM’s Granite Speech 5.0 open models and MiniMax’s M3.1 Flash open weights, where local or efficient inference matters.
Safety as the stated motive
Block says it built Dessa out of curiosity, for better human-AI understanding, and for safety. The argument: today’s voice AI is “nearly deaf to the emotion in our voices”, and that is a safety problem when people in distress are misheard. Rather than shipping wide, the lab is releasing early access to a small group of AI safety researchers and opening a public waitlist — treating a convincingly human voice as a risk to study before scaling it.
Why It Matters
Dessa matters because speech-native training is a genuinely new direction. Text pipelines preserve what was said and discard how it was said; a model that learns from conversation audio keeps timing, tone and emotion as first-class information. Local inference on a laptop hints at a compute path that could make capable speech AI far cheaper to deploy. And the limited, safety-first release is notable in itself: Block is treating a voice that laughs, interrupts and reads your mood as something to understand before it scales, not after.
Background
Block is Dorsey’s payments company, known for Square and Cash App, and it previously owned a Toronto deep-learning startup also called Dessa — the name now lives on in this new lab. Today’s voice assistants mostly shuttle speech through text, which is why they miss sarcasm, distress and the backchannels that make conversation feel human; Dessa is the lab’s attempt to skip the text step entirely. The release follows a busy stretch for speech models, from open releases like IBM’s and MiniMax’s to rankings such as Kimi K3.1 Pro’s fourth-place AI Index ranking.
Source: Block; UA.NEWS (The Verge reporting)
0 Comments