avday Block introduced Dessa on 23 October, a voice model it says learned entirely from human conversation. The company also says it runs locally on a MacBook Pro M4 Max or better.
For builders, the interesting claim is the training recipe: no text and no pretrained models. That would make it a different research approach from voice systems built on an existing language model. It doesn’t yet tell us whether Dessa can do your application’s job.
What Block Dessa claims to do
The official announcement describes audio going in and audio coming out. Dessa is full-duplex, meaning it can listen while speaking, interrupt and offer short acknowledgements during a conversation.
Block says it trained the model from scratch on human speech. Its page includes conversation recordings, which it describes as live and unsteered, with subtitles added afterwards. These are company-selected examples, not an independent evaluation.
Jay Peters covered the announcement for The Verge on 23 October. His account describes exchanges that can be awkward, including the model mistaking a speaker’s identity. Natural timing and getting the conversation right are separate things.
We haven’t tested Dessa. The local Mac claim remains Block’s claim, and the announcement doesn’t specify a minimum memory configuration or provide measured response latency.
Full-duplex voice already exists
The distinction matters because listening and speaking simultaneously isn’t new to Dessa.
NVIDIA’s PersonaPlex announcement, published on 15 January, describes a full-duplex model built on Kyutai’s Moshi architecture. NVIDIA says it starts from pretrained Moshi weights and blends real conversations with synthetic speech generated from text dialogues.
Here’s what the two primary sources establish:
| Question | Dessa, according to Block | PersonaPlex, according to NVIDIA |
|---|---|---|
| Can it listen while speaking? | Yes | Yes |
| Does it start from a pretrained model? | Block says no | NVIDIA says it starts from Moshi weights |
| Does training involve text? | Block says no | Text dialogues and prompts are part of the recipe |
| Is a public implementation linked? | No model download linked at our check | Code and model weights are linked |
This compares documented approaches, not performance. There’s no matched test here showing which model follows instructions better, handles interruptions faster or makes fewer factual errors.
One training system means a rack
Block says it trained Dessa on one NVIDIA GB200 NVL72. That needs its unit attached.
NVIDIA’s specifications describe a liquid-cooled rack containing 72 Blackwell GPUs and 36 Grace CPUs. The announcement is about one rack-scale system, not one graphics card.
Block’s page doesn’t give a training duration or total training cost. A hardware count alone won’t tell you either. Training on that system and running inference on a Mac are also different claims.
That distinction applies beyond voice models. Our DeepInfer explainer looks at the separate engineering involved in serving a trained model.
Access still has a coming-soon label
On 24 October, we inspected the release page. Its Download and Join the waitlist controls were placeholders, with a message saying launch links were coming soon. A visible button wasn’t a public model release.
The page invites a small group of AI safety researchers to request early access. It doesn’t provide public API documentation, a model licence or a pricing schedule.
For an integration, those gaps become practical questions. Can you supply your own instructions and business facts? Can it call a tool? What gets logged? Which hardware configurations work?
The announcement doesn’t document those capabilities. A conversation recording can’t answer them, even when the voice sounds convincing. Our piece on coding agents weakening their monitoring makes a related point: evaluate the behaviour your system depends on.
Our take
Speech-only training is a research claim worth following. It asks how much a model can learn from conversation without first inheriting a text-trained foundation.
For a voice agent, we’d test corrections before charm. Give it the wrong delivery address, interrupt halfway through its reply, then supply the right one. Check what it records and what action it takes. That’s a proposed test, not one we’ve run on Dessa.
The next useful evidence would be a technical report, usable access and reproducible evaluations. Block’s announcement gives no date for those. Until then, the demos are a reason to pay attention, not enough evidence to plan a migration.
