Block's AI research lab has released Dessa, a full-duplex speech model trained exclusively on human conversation, announced as a research preview on October 23, 2026.
Key points
- Block's Dessa is a full-duplex speech model trained only on human conversation.
- Dessa listens and talks simultaneously, bypassing text-based pre-training.
- The model learned conversational nuances directly from audio data.
- Inference runs faster than real-time on specified hardware.
- Training utilized one NVIDIA GB200 NVL72 rack.
Dessa functions as a speech-to-speech model, processing audio input and producing audio output, enabling it to listen and talk concurrently.
The model was trained without relying on text data or pre-existing models, diverging from the text-first approach common in current voice agents.
By the numbers
- Blackwell GPUs in the training rack
- 72
Block states that Dessa learned conversational nuances, such as when to interject, use backchannels like "yeah," and incorporate natural interruptions like laughter.
The training process for Dessa involved only human conversation audio, with no text or pre-trained models utilized. Block claims this method allows Dessa to learn language with significantly fewer words than required by LLMs.
Inference for Dessa runs faster than real-time on hardware like a MacBook Pro M4 Max or superior.
The model was trained on one NVIDIA GB200 NVL72 rack, which comprises 72 Blackwell GPUs and 36 Grace CPUs.
Block's approach aims to preserve emotional cues, sarcasm, laughter, and sighs that may be lost in traditional speech recognition, text modeling, and speech synthesis chains.
Sources and references
Original report
Block Unveils Dessa, a Full-Duplex Speech Model Trained Only on Speech
MarkTechPost
Read on MarkTechPost (opens in a new tab)Linked in the original report
- Linkblock.xyz/dessa (opens in a new tab)
- Linktechcrunch.com/2026/10/24/blocks-new-voice-model-sound… (opens in a new tab)
- Linknvidia.com/en-us/data-center/gb200-nvl72 (opens in a new tab)
- Paperarxiv.org/abs/2410.00037 (opens in a new tab)
- Modelhuggingface.co/kyutai/moshiko-pytorch-bf16 (opens in a new tab)
- Modelhuggingface.co/nvidia/personaplex-7b-v1 (opens in a new tab)
Ref 007U4RTP
Cite this
APA
Interestana AI Editorial. (2026, October 29). Block Releases Dessa Speech Model Trained Only on Audio. Interestana. https://interestana.com/articles/block-unveils-dessa-a-full-duplex-speech-model-trained-only-on-speech-007u4rtp
MLA
“Block Releases Dessa Speech Model Trained Only on Audio.” Interestana, 29 Oct. 2026, interestana.com/articles/block-unveils-dessa-a-full-duplex-speech-model-trained-only-on-speech-007u4rtp.
Chicago
Interestana AI Editorial. “Block Releases Dessa Speech Model Trained Only on Audio.” Interestana, October 29, 2026. https://interestana.com/articles/block-unveils-dessa-a-full-duplex-speech-model-trained-only-on-speech-007u4rtp.
BibTeX
@misc{interestana_007u4rtp,
author = {{Interestana AI Editorial}},
title = {{Block Releases Dessa Speech Model Trained Only on Audio}},
howpublished = {Interestana},
year = {2026},
month = oct,
url = {https://interestana.com/articles/block-unveils-dessa-a-full-duplex-speech-model-trained-only-on-speech-007u4rtp},
note = {Summary of reporting by MarkTechPost, https://www.marktechpost.com/2026/10/29/block-dessa-full-duplex-speech-model/}
}Credit the source
Original reporting: MarkTechPost, “Block Unveils Dessa, a Full-Duplex Speech Model Trained Only on Speech”, https://www.marktechpost.com/2026/10/29/block-dessa-full-duplex-speech-model/(opens in a new tab)
Follow this story
Get Artificial Intelligence in your inbox — new stories in 40 words or fewer, once a week.
Subscribe