Block Unveils Dessa, a Full-Duplex Voice Model Trained Only on Recorded Conversation

• Model: "Dessa." Block has not published a version number or parameter count (Voice AI Space).
• Demonstrated behaviours:
• Short listening cues while the user talks ("yeah," "got it"), known as backchanneling.
• Laughing mid-sentence and speaking louder when asked.
• Expressing sarcasm, sighs and a voice cracking with sadness (Block).
• Access: public waitlist only. Safety researchers can apply for early access with an institutional email. "Launch links" are marked as coming soon. Block has not said whether it will release weights or open-source the model (video, 4:36; 4:58).

• Architecture: audio goes in and audio comes out, with no text step in between. The model learned turn-taking from real conversation data (Block). Moshi, by contrast, predicts time-aligned text tokens before its audio tokens (its "inner monologue") (Moshi).
• Compute: Block trained Dessa on one GB200 NVL72, a rack of 72 Blackwell GPUs (Voice AI Space). The video calls this "surprisingly little hardware" (3:21), but without training duration that comparison has no basis.
• Not disclosed: audio hours, model size, context length, latency in milliseconds and the evaluation protocol. The work cannot currently be reproduced (OrcaRouter).

• "First" model to hold a realistic conversation after training only on speech, and a "GPT-3 moment for speech": both from lab lead Thurairatnam (tech-insider).
• "Hundreds of times fewer words" than an LLM: Block's claim. The actual word count is unpublished (Block; video, 1:56).
• Faster-than-real-time inference on an M4 Max: stated without supporting numbers (Block).
• "Most authentic" full-duplex model: no comparison against Moshi or NVIDIA's PersonaPlex has been published (Voice AI Space).

All of these are vendor claims. No independent evaluation was found.

• Lineage: the lab grew out of the Dessa startup that Square acquired in 2020. It reports to Jack Dorsey and has worked on generative speech for more than eight years (Unite.AI; Square).
• Correction to the commentary: the video calls Dessa a standalone "startup from Toronto." It has been part of Block since 2020 (video, 0:31).
• Open alternatives: Moshi and PersonaPlex-7B are released full-duplex models built on text-based backbones (Voice AI Space).

• Demo selection: the public demos are curated, so they show the best case rather than typical reliability (OrcaRouter). Subtitles were added to the recordings afterwards. One demo uses a third-party model, "GPT-Live-1," whose developer has not endorsed Dessa (Unite.AI).
• No text-native features: training on speech alone means no native structured output or tool use (OrcaRouter).
• Safety framing: Block argues that voice AI which is "deaf" to emotion is a safety risk (video, 3:52). Independent research cuts the other way on content: existing full-duplex models challenged false claims in only 14–15% of their replies and warned about hazards in 4–7% (arXiv 2609.19596). Natural timing does not mean the model understands what is being said.

• Nothing to deploy yet: there is no API, no pricing and no weights. Teams shipping now should keep evaluating Moshi- or PersonaPlex-class models, or pipelines that chain speech recognition, an LLM and text-to-speech (Voice AI Space).
• On-device appeal: if local inference on a laptop holds up, it would address privacy and round-trip latency concerns. That remains unverified.
• Lower-cost speech models: if training on one rack is reproducible, strong speech models would no longer require frontier-scale compute.
• Agent builders: the lack of tool use means Dessa-style models would likely need a text-capable model alongside them to take actions.
• Before any rollout: test interruption handling, factual pushback and emotional manipulation risk on your own data.