Jack Dorsey’s Block has unveiled Dessa, a voice model with a strange pedigree: it was trained entirely on human speech, with no text data at all. Where a conventional assistant hears you, transcribes you, thinks in tokens and reads its answer back aloud, Dessa processes audio directly — sound in, sound out — which is why its demo clips carry laughs, pauses, sighs and something close to genuine emotion rather than a well-read paragraph. TechCrunch reported the model is full-duplex — it talks and listens simultaneously, the way two people do — and that Block trained it from scratch on a single NVIDIA GB200 NVL72 rack.
The company’s argument, laid out in its own announcement, is that the text hop in today’s voice assistants throws away most of what a human voice carries. “Interacting with current voice AI can just plain suck,” Block wrote — its position is that turning a voice into text discards tone, hesitation and distress “text alone can’t carry,” and that a model which has never read can still hear how you feel. In the demos, a Dessa voice said it was tired from work and sounded tired; another clip had it audibly close to breaking down while a tester spoke to it. Dorsey’s own review on X: “Talking with it feels like talking with a person…full of emotion, nuance, pauses, and a natural cadence.”
🔍 THE BOTTOM LINE: Every serious lab is racing toward speech-native models because the text pipeline is a bottleneck and a latency tax; Block is the first to claim an entire model crossed the line with zero text training. The catch is that a voice with no verifiable paper trail cuts both ways — the same demo that awed an education founder reads to a fraud researcher as a scam call with the guards off. Block has not published benchmarks, and its answer to that tension is to gate access behind a wall of safety researchers first.
Why a speech-only model is a genuine technical claim
The idea is not new — Hume AI’s Empathic Voice Interface, Kyutai’s Moshi and LAION’s Open Empathic all chased emotional speech, and The Decoder’s history of those attempts notes none stuck — but the training recipe Block describes is more radical than “sounds nicer.” Most production voice stacks, including the ones with emotion tags bolted on, keep a language model somewhere in the loop. Dessa’s knowledge, per Block’s account to TechCrunch, comes from the conversational data itself. That is a bet that discourse — turn-taking, repair, emotional register — can be learned without a single written word, and it produced a model compact enough to run locally on a MacBook Pro with an M4 Max chip.
It slots into a shift this site has tracked all year: OpenAI’s GPT-Live made full-duplex voice mainstream, and the frontier question stopped being whether voice AI sounds human and started being where it runs. Dessa’s answer is “on your laptop, no cloud round-trip” — which matters for price, privacy and latency all at once, and which a payments company has obvious reasons to care about. Block’s lab traces back to Square’s 2020 acquisition of the Toronto startup Dessa, and it reports directly to Dorsey.
The reaction: awe in one column, fraud in the other
The public reaction split exactly along the fault line this technology always runs down. On one side, Binance founder Changpeng Zhao called it a natural fit for teaching children in his Giggle Academy project, where robotic AI voices had been “a challenge”; early listeners on X reached for sci-fi comparisons. On the other, the same humanlike qualities drew immediate warnings — filmmaker Justine Bateman asked why an AI voice needs to sound human at all, “so you can trick people into lowering their guard,” and journalist Shiv Aroor put it plainly: “Scam calls about to attain pandemic level.” As CoinDesk’s report framed it, the model drew reactions “ranging from amazement… to warnings that such technology could make phone scams harder to detect.”
Both reactions are reasonable, and the asymmetry between them is the whole story. Block has disclosed no model size, no benchmark results and no release date — the audio samples are company demonstrations, not independent tests, so even the “sounds human” claim is currently self-graded. What Block has done is invert the usual launch order: instead of shipping and apologising, it is offering early access to a small group of AI safety researchers before any public rollout, explicitly to probe the risks the company may have missed. That is a more responsible posture than most voice launches have managed — and it is also an admission of exactly how much damage this model class could do in the wrong hands.
The New Zealand angle: a voice scam economy is already here
This matters more in a country that has lived through the precursor wave. NZ’s own bank complaints data through 2025-26 showed cold-call and impersonation scams among the most costly fraud classes reported to banks, and when NZ researchers and journalists documented AI voice cloning being used on families here — a case where a synthetic voice was convincing enough to require spectrographic analysis to debunk — the question was never whether emotional synthetic speech would arrive, only whether it would arrive cheap, local and offline. Dessa is all three. A model that runs on a consumer laptop with no network dependency and no server logs is a scammer’s feature list, and it will not stay inside Block’s labs: speech-native architectures are being published across the field, and indigenous-language voice models have already shown the same techniques carrying mātauranga rather than fraud.
The honest read sits in the middle. A voice model that hears distress, sarcasm and hesitation could make phone support genuinely better — and Square already runs AI voice ordering in US food service, so Block has a distribution channel for it — while the same properties make “prove who you are by voice” permanently untrustworthy. The institutions that still use voice as an identity factor have a shrinking window to retire the practice. Block’s safety-first preview window, if it is real and not theatre, is the right sequencing; the scam-economy clock did not stop for it.
Sources: TechCrunch, The Decoder, CoinDesk, Block, Mashable
