Block releases Dessa, a full‑duplex speech model trained only on speech
Block’s AI research lab announced Dessa on October 23, 2026 as a research preview. The model is a full‑duplex speech‑to‑speech system that listens and talks simultaneously, trained solely on human conversation audio with no text or pretrained models. Block says it learned from real conversation cues such as backchannels, mid‑sentence laughs, and natural interruptions.
Key points
- Block’s Dessa is a full‑duplex speech‑to‑speech model trained only on human conversation audio, with no text or pretrained models.
- It was trained on a single NVIDIA GB200 NVL72 rack (72 GPUs) and runs faster than real time on a MacBook Pro M4 Max.
- No benchmark scores, weights, or API are released; access is via a waitlist and the model is a research preview.
Dessa was trained on a single NVIDIA GB200 NVL72 rack, which links 72 Blackwell GPUs and 36 Grace CPUs. The lab claims the model runs faster than real time on a MacBook Pro M4 Max or better. No benchmark scores, weights, paper, API or price are released, and access is currently via a waitlist.
The article notes that other full‑duplex models, such as Moshi and PersonaPlex, use different training pipelines and larger compute budgets. Block argues that a speech‑only approach preserves emotional and prosodic signals that are lost in text‑first pipelines.
The story so far
2 episodes →- Block releases Dessa, a full‑duplex speech model trained only on speechthis story
Block Unveils Dessa, a Full-Duplex Speech Model Trained Only on Speech
MarkTechPost · 29 October 2026
Loading the full article…
This text was published by MarkTechPost and written by Michal Sutter. It is reproduced here with attribution so you can read it in full; the rights remain with the publisher. Read it at the source ↗
The headline, key points and digest above were generated by Digest AI's editorial model from the linked sources. Automated summaries can contain errors: the sources are the record. Spotted a mistake? Tell us. Published by Martin K., who runs Digest AI and handles corrections.
More in Generative AI & Models
All →- Moonshot AI and DeepSeek reportedly used GPT-6 Astra outputs to build new models · 3 src
- DeepSeek V4.1 Flash 1017 rivals Claude Opus 5 in benchmark tests · 1 src
- Google launches Gemini 4 Carbon for cybersecurity defenders at $2/1M input and $10/1M output tokens · 7 src
- NVIDIA releases open-weight Nemotron 3.5 Transcribe model and Granary v2 dataset · 10 src
- Google announces Gemini 4 Carbon and Genie 4 models with staged rollout for cyber defenders · 4 src
Comments
via GitHub Discussions