Dessa, Block's voice model that listens to emotions

Block presents Dessa, a speech model trained solely on human voice. The goal is to better capture emotions, sarcasm, laughter, and silences, before possible public release.

Block, the company behind Square and Cash App, has unveiled Dessa, a voice AI model presented as the first trained solely on human speech, without text. The announcement, dated October 24, 2026, according to Block, targets a use beyond simple transcription: understanding emotions, pauses, sarcasm, and laughter in real time. The model is not yet public, as the company is first seeking security researchers to test potential risks.

A model that treats voice as voice, not as text

Dessa works directly on audio, without prior text conversion. This technical choice changes how the system interprets a conversation, as it can capture cues that text-based models often miss. Block also states that the model supports full-duplex, meaning the ability to listen and speak simultaneously, a feature useful for smoother exchanges.

The group says it trained Dessa on a single NVIDIA GB200 NVL72 and specifies that it can run locally on a MacBook Pro M4 Max. This point matters because it suggests a model optimized for local execution rather than a simple cloud-hosted demonstrator. The promise is clear: reduce the gap between what a human feels and what a machine understands.

Why Block is interested in vocal emotions

Block explains it built Dessa out of curiosity, but also to improve human-machine communication and for security reasons. According to the company, current voice assistants neglect emotional signals that can nonetheless change the meaning of a sentence. A request said with nervous laughter, silence, or irony does not convey the same intention. And machines, so far, sometimes tend to take everything literally. Not very subtle, but quite typical.

This approach can matter for customer service, digital assistants, or payment interfaces, where tone influences the quality of the exchange. However, a more expressive voice does not guarantee better conversational intelligence. Block indirectly reminds us: for advanced voice systems, the ability to reason when responding often weighs more than a simple natural rendering.

A field already occupied, but without a clear winner

Dessa arrives in a segment already explored by several players. In March 2024, Hume AI launched EVI, presented as an empathetic voice interface. Kyutai followed in September 2024 with Moshi. On its side, OpenAI revised its “Advanced Voice Mode” in June 2025, then further refined it since, with more polished intonation, better-placed pauses, and more credible empathy according to public descriptions.

LAION has also been working since late 2023 on Open Empathic, an open-source project training empathetic voice models from an audio dataset. This shows that the industry is less focused on making machines talk than on teaching them to listen better. The challenge remains: producing a voice that seems sensitive without giving the illusion it understands everything. It’s subtle, and overly brilliant demos have already tripped up more than one product.

What restricted access says about what’s next

For now, Dessa is not available to the public. Block favors an evaluation phase by security researchers to spot unexpected behaviors or risks of misuse. This caution is logical for a voice model capable of detecting affective signals, as this type of AI can touch on trust, persuasion, and privacy.

The real question is therefore not only whether Dessa speaks well, but whether it can do so without manipulating the user or misinterpreting their emotional state. In this context, Block is moving on ground where technical performance matters as much as safeguards. And this is probably where the difference between a nice showcase and a truly usable tool will be decided.

By the numbers

  • October 24, 2026 — date of the announcement by Block.
  • 1 — single NVIDIA GB200 NVL72 used for training according to Block.
  • 0 — textual data used in training, according to the company.
  • 2024 — launch year of EVI by Hume AI and Moshi by Kyutai.
  • 2025 — update of OpenAI’s Advanced Voice Mode in June.

To read