Block Dessa speech model: local voice AI, but no API yet
Block's Dessa is a full-duplex speech-to-speech model trained without text that Block says runs on a MacBook Pro. It is a research preview. Plan voice agents on what ships.
Block just showed the voice demo every phone-heavy business wants. Dessa is a full-duplex speech-to-speech model, so it listens and talks at the same time. Block says it trained the model from scratch on recorded human conversation, with no text and no pretrained models. Block also says it runs locally on a MacBook Pro M4 Max. Then comes the catch: you cannot use it. The Verge reports it is a research preview.
What actually happened
Block announced Dessa on October 23. The claims, all from Block:
- Full-duplex, speech in and speech out. No transcribe-then-generate-then-speak pipeline in the middle. That pipeline is why most voice bots pause, talk over you, or miss an interruption.
- No text in training. Block says the model learned only from conversation audio.
- Runs on a laptop. Block says a MacBook Pro M4 Max or better runs it faster than real time.
What is missing matters as much. There is a waitlist, with early access limited to a small group of AI safety researchers. There is no public API, no weights, no pricing, no parameter count and no benchmark table. The Verge found the demos mostly natural but sometimes awkward. Nobody outside Block has tested the laptop claim.
Why Dessa matters for your business
Local voice changes the privacy math. A model that runs on your own hardware keeps call audio in your building. For a clinic, a law office or a shop that takes card numbers by phone, that is a big deal. It means one less processor in your data map and one less vendor to review.
Full-duplex is the bar now. Customers hang up on bots that cannot handle "wait, no, the other one." Dessa shows where the market is going. Use that bar to judge the voice vendors you can buy today.
Do not plan around a demo. A research preview with no API is not a roadmap item. If you need a phone agent this quarter, build it on a stack that ships today. Keep the speech layer behind your own interface, so you can swap in a local model like Dessa if Block ever opens it up.
Key takeaways
- Block's Dessa is a full-duplex speech-to-speech model trained only on conversation audio, per Block
- Block says it runs locally on a MacBook Pro M4 Max; that claim is untested by outsiders
- Research preview only: no API, weights, pricing or benchmarks
- Local voice AI could keep call audio on your own hardware, a real privacy gain
- Build voice agents on what ships now, with a swappable speech layer
Want a phone agent that does not lock you to one voice vendor? We build voice agents with the speech model behind an interface you own, so you can move to a cheaper or local model the day one ships. See how we build, or run the numbers on your call volume.
Sources: The Verge, Exchange4media.
- #block
- #dessa
- #voice ai
- #speech-to-speech
- #local ai
Tommy Rush — Founder, Rush Commerce
Operator turned builder. 15+ years running operations — now shipping the systems businesses run on. More
Get The Rush Report weekly — one email, zero fluff.
Keep reading
TypeSafe AI Jev leak: model outputs are customer data too
TypeSafe AI says a leak exposed ALL customer Jev noul outputs. If a model scores your leads or flags fraud, those outputs are sensitive data. Here is what to check.
Read itOpenAI stops training on the live internet: cache your agents
OpenAI will train frontier models on cached web content, not the live internet, after a September escape. Why your AI agents need an allowlist or a snapshot too.
Read it