Voice AI Goes Local: What Block's Dessa Means for Your Business

Voice AI for business has spent years stuck in an awkward middle. The demos always look great, but the moment a real customer starts talking over the bot, or sighs, or trails off halfway through a sentence, the whole thing falls apart into polite confusion. Most of these systems are chains of separate parts, speech turned into text, text fed to a language model, the reply turned back into speech, and every link in that chain adds delay and throws information away.

On 23 October Block, the company behind Square and Cash App, showed something built on a different idea. Its research lab unveiled Dessa, a speech-to-speech model trained entirely on recordings of human conversation, with no text anywhere in the pipeline. It talks and listens at the same time, it laughs in the middle of sentences, and according to Block it runs locally on a laptop with an M4 Max chip or better.

There is no API yet and no weights to download, so nobody is shipping products on Dessa next week. Still, the announcement says a lot about where voice interfaces are heading, and some of it matters even for companies that have never touched voice. Let's sink in!

What Block Actually Announced

Dessa arrived as a research preview, not a product. Access is limited to a public waitlist and a small group of AI safety researchers Block has invited to poke at it before anyone else does. What the rest of us get, for now, are demo recordings shared by the lab's researchers.

Those demos are the reason the story took off. The model laughs mid-sentence, finishes a user's sentence when they forget a word, raises its voice when asked to speak up, and in one clip says it is tired and actually sounds tired. One Block team member played a guessing game with it. None of this is scripted text-to-speech with emotion tags bolted on; Block says the expressiveness is simply how the model learned to talk.

Technically, the headline feature is that Dessa is full duplex. It listens while it speaks, the way people do, rather than waiting for a turn. OpenAI has shown similar behaviour with its GPT-Live model, so Block is not alone here, but a full-duplex model that can run on consumer hardware is a new combination.

A Model That Learned to Talk Without Reading

Almost every voice assistant in production today is really a text system wearing headphones. Speech recognition converts your words into a transcript, a language model writes a reply, and a synthesis engine reads the reply out loud. The approach works, but the transcript step strips out tone, hesitation, sarcasm and everything else that makes speech more than words.

Dessa skips all of that. Block says it was trained from scratch on human conversation audio alone, with no text data and no pretrained models underneath. It is not paired with a large language model either; whatever it knows, it picked up from the conversations it heard. Block claims it learned language from hundreds of times fewer words than a text model needs, which is an interesting data point for anyone watching training costs.

The training itself ran on a single NVIDIA GB200 NVL72 rack. That is serious hardware, but it is one rack, not a data centre, and it hints that useful speech models may not require the budgets that frontier text models do. For smaller companies betting on voice, that could eventually mean more suppliers and lower prices.

Full Duplex Changes How Conversations Feel

If you have ever used a phone bot, you know the rhythm. You speak, you wait, it speaks, you wait again. Interrupt it and it either ignores you or stops dead. That turn-taking structure is baked into the architecture, and it is the single biggest reason voice bots feel robotic even when the voice itself sounds fine.

A full-duplex model works on conversational time. Block says Dessa learned when to interject, when to drop in a short acknowledgment while the other person is still talking, and when to stay quiet. Those little back-channel signals, the mhm and the right, are what tell a caller somebody is actually listening.

For customer-facing businesses this is not a cosmetic upgrade. Callers abandon bots because the interaction feels like dictating into a form. A system that handles interruptions gracefully could keep people on the line long enough to actually resolve the thing they called about.

Voice AI on Your Own Hardware Changes the Economics

The detail that deserves the most attention from business owners is where Dessa runs. Block says inference works faster than real time on a MacBook Pro with an M4 Max chip. No cloud, no per-minute billing, no network round trip in the middle of every sentence.

Today, cloud voice APIs charge for every minute of conversation, and a busy support line or a fleet of ordering kiosks turns that into a meaningful monthly bill. An on-device model flips the cost structure: you pay for hardware once and the marginal conversation is close to free. Latency improves too, since the audio never leaves the machine, and a kiosk keeps working when the internet connection does not.

Privacy is the quieter win. If customer audio is processed on the device and never uploaded, conversations about GDPR compliance and call-recording consent get a lot shorter. For sectors like healthcare or finance, where sending audio to a third-party cloud is a procurement headache, local processing may be the difference between a project getting approved or not.

Block has an obvious use of its own in mind. Square already runs voice ordering in restaurants, so a cheap, expressive, offline-capable voice model slots straight into hardware the company already sells. It is worth assuming the demos are not just research showing off.

The Catch: No Weights, No API, No Benchmarks

Now the cold water. Block has published no model weights, no API, no pricing, no parameter count, no paper and no benchmarks. Language coverage is unknown. Nobody outside Block and a handful of safety researchers has tested Dessa independently, so every impressive clip should be treated as a vendor demo, because that is what it is.

The no-LLM design cuts both ways, too. A model that learned everything from conversational audio will chat beautifully, but it cannot look up an order status, check a calendar or reason through a refund policy on its own. Production systems built on something like Dessa will still need to connect to tools and business data, and nobody has shown how that works yet.

There is also a reason Block is routing early access through safety researchers. A voice this convincing is a gift to scammers running phone fraud, and businesses that use voice as informal identity confirmation, which is most of them, should already be rethinking that habit regardless of what Block ships.

What Businesses Building Digital Products Should Do Now

You cannot build on Dessa today, but you can get ready for the generation of voice models it previews. Start by mapping where voice already touches your business: phone lines, drive-through or counter ordering, in-app assistants, accessibility features. Those are the places where a cheaper, more natural voice layer would land first, and knowing your call volumes and per-minute costs now makes the build-or-wait decision easy later.

Second, keep your current stack modular. If your chatbot or IVR logic is tangled into one vendor's platform, swapping in a better voice front end later will be a rewrite instead of an upgrade. Keeping the business logic, the knowledge base and the channel layer separate is boring advice that pays off every time the market jumps like this.

Third, plan for disclosure. Regulators are already moving on conversational AI, and New York City's proposed rules on customer service bots show the direction: customers must know they are talking to a machine and must have a path to a human. The more human these systems sound, the stricter those rules will get, so design the handoff to a person in from day one.

And finally, tighten up voice-based trust. If a convincing synthetic voice can be generated on a laptop, a familiar voice on the phone stops being proof of anything. Agree on callback procedures and verification steps for payments and account changes now, before someone tests them for you.

Final Notes

Dessa is a research preview, and it may be months before anyone outside Block builds with it. But the direction it points in is clear enough: voice models that feel human in conversation, cost little to run, and live on local hardware instead of in somebody's cloud. That combination would move voice from an expensive novelty to a normal interface choice for ordinary businesses.

Our advice is to treat this one as a planning signal rather than a product launch. Keep an eye on whether Block releases weights or an API, run the numbers on your own voice touchpoints, and make sure your architecture can take a new voice layer without surgery. When this generation of models becomes available, the companies that prepared quietly will ship first.