Block released Dessa, a full‑duplex speech‑to‑speech model that can run locally on a MacBook M4 Max and produces human‑like expressions such as laughs, pauses and cries (TechCrunch, “Block’s new voice model sounds eerily human”, 24 October 2026).
How does Block’s Dessa voice model differ from existing voice AI?
Dessa processes raw audio directly instead of converting speech to text before generating a response, eliminating the text‑to‑speech pipeline that most commercial voice assistants use (TechCrunch, “Block’s new voice model sounds eerily human”, 24 October 2026). The model was trained exclusively on human conversation data, with no books, web pages or pre‑trained language models, mirroring a baby’s learning process (TheStreet, “Jack Dorsey’s new AI finally sounds human - Crypto - TheStreet”, 25 October 2026). Unlike ElevenLabs v4 and Fish Audio Drama 3, which rely on voice‑tag annotations to inject emotion, Dessa embeds emotional nuance such as laughter and sighs directly into its neural weights (TechCrunch, “Block’s new voice model sounds eerily human”, 24 October 2026).
What hardware can run Block’s Dessa model locally?
The company states that Dessa is compact enough to run on consumer‑grade hardware, specifically a MacBook Pro equipped with an M4 Max chip or any device with a comparable GPU (TechCrunch, “Block’s new voice model sounds eerily human”, 24 October 2026). Training was performed on a single NVIDIA GB200 NVL72 GPU rack, demonstrating that the model does not require massive data‑center clusters for inference (TechCrunch, “Block’s new voice model sounds eerily human”, 24 October 2026). Coindesk confirms that the research preview can be executed on a MacBook Pro M4 Max, reinforcing the claim of on‑device capability (Coindesk, “Block’s new voice AI sounds eerily human, sparking awe and scam fears”, 24 October 2026).
What are the reactions and potential risks associated with Block’s Dessa voice model?
Public reaction split between amazement at the model’s natural cadence and concern that its realism could be weaponized for phone‑based scams (Coindesk, “Block’s new voice AI sounds eerily human, sparking awe and scam fears”, 24 October 2026). Binance CEO Changpeng Zhao highlighted possible educational applications for children, while also warning about the ethical implications of such lifelike agents (Coindesk, “Block’s new voice AI sounds eerily human, sparking awe and scam fears”, 24 October 2026). Block itself framed the launch as a safety‑focused research preview, inviting AI safety researchers to test the system for misuse scenarios (TechCrunch, “Block’s new voice model sounds eerily human”, 24 October 2026).
When will Dessa be publicly available?
As of the announcement, Dessa remains a research preview with no public release date; access is limited to a small cohort of AI safety researchers for testing and feedback (Coindesk, “Block’s new voice AI sounds eerily human, sparking awe and scam fears”, 24 October 2026). Block has not disclosed model size, benchmark scores, or a timeline for broader availability (Coindesk, “Block’s new voice AI sounds eerily human, sparking awe and scam fears”, 24 October 2026).
Comparison of Dessa with other contemporary voice models
| Model | Training data | Emotion control method | Local‑run capability |
|---|---|---|---|
| Dessa (Block) | Solely human speech conversations | Built‑in emotional representation (no tags) | Runs on MacBook Pro M4 Max or better (TechCrunch, 24 October 2026) |
| ElevenLabs v4 | Large multi‑speaker speech dataset | Voice‑tag annotations for emotion (TechCrunch, 24 October 2026) | Requires cloud inference; not designed for on‑device use |
| Fish Audio Drama 3 | Diverse speech corpus | Voice‑tag annotations for emotion (TechCrunch, 24 October 2026) | Cloud‑based service; no local execution claim |
Sources
- TechCrunch — Block’s new voice model sounds eerily human — article detailing Dessa’s launch, architecture and hardware claims.
- CoinDesk — Block’s new voice AI sounds eerily human, sparking awe and scam fears — coverage of reactions, safety preview and hardware specifics.
- TheStreet — Jack Dorsey’s new AI finally sounds human - Crypto — description of speech‑to‑speech pipeline and training methodology.
- AIChatDaily — Block unveils Dessa, a full‑duplex voice model trained only on human … — summary confirming full‑duplex, speech‑only training.