Editorial illustration for Block trains Dessa on human speech to catch emotions text-based models miss
Block's Dessa AI Reads Emotions Directly From Speech
Block trains Dessa on human speech to catch emotions text-based models miss
Block, the company known for Square and Cash App, has built an AI speech model that skips text entirely. The model, called Dessa, was trained from scratch on raw human speech, no transcripts, no converting audio to words first. Block says that's a first for the industry.
Most voice AI today works by translating speech into text, processing that text, then converting the response back into audio. Dessa drops that middle step. It listens to audio directly, which means it can catch things text strips out: sarcasm, laughter, the pause before someone says something they're not sure about. It also runs full-duplex, so it can listen and respond at the same time, closer to how an actual conversation works.
The hardware side is notable too. Dessa was trained on a single NVIDIA GB200 NVL72 and runs locally on a MacBook Pro M4 Max, no server farm required. Block isn't releasing it to the public yet. Right now the company is recruiting AI security researchers to stress-test it for risks before anyone else gets access, with sign-ups open through a form.
Block says Dessa is the first AI model trained entirely on human speech, with no text data at all, built from scratch. Instead of converting audio to text first, Dessa processes speech directly as audio. That lets it pick up on emotions, sarcasm, laughter, and pauses in real time.
Why this matters
Hume AI's EVI and Kyutai's Moshi both promised emotionally fluent voice AI, and neither became the default way anyone talks to a machine. Block is now making the same bet with more distribution behind it: Cash App and Square processed payments for millions of people who already talk to customer support bots. If Dessa actually catches sarcasm and hesitation instead of transcribing dead text, that's a real product difference for anyone building voice agents, not just a demo trick.
But "trained entirely on speech, no text" is also a strong claim that invites scrutiny on benchmarks, latency, and how it handles languages or accents outside whatever data Block collected. For builders, the question isn't whether emotional nuance is nice to have, it's whether Dessa ships as something you can actually integrate, or stays a research showcase like its predecessors. We'd watch for an API, a paper with real eval numbers, and whether Block uses it inside its own products before anyone else gets access.
Until then, this is a claim, not yet a track record.
Related Reading
Business & Startups Nordic pilot adds Gemini for Education, NotebookLM to boost AI literacy
Business & Startups Kling launches Video O1, all-in-one model with MVL bridge using transformer
Business & Startups DeepSeek Seeks More Capital Weeks After USD 7B Funding Round
From Archives Coforge Revenue Rises 54% as AI Engineering Drives USD 1.04B Order Intake
From Archives Block's AI model Dessa, trained from scratch, talks like an awkward new friend
Common Questions Answered
How does Dessa differ from traditional voice AI models in its training approach?
Dessa was trained entirely on raw human speech without any text transcripts or conversion to text data, making it the first AI model to take this approach according to Block. Most existing voice AI models convert speech to text first, process that text, then convert responses back to audio, but Dessa skips this middle step entirely by processing audio directly.
What emotional nuances can Dessa detect that text-based models miss?
By processing speech directly as audio rather than converting to text first, Dessa can pick up on emotions, sarcasm, laughter, and pauses in real time. These vocal cues and emotional indicators are often lost when speech is transcribed into text, allowing Dessa to provide more emotionally fluent interactions.
What competitive advantage does Block have with Dessa compared to other emotional AI models?
While competitors like Hume AI's EVI and Kyutai's Moshi also promised emotionally fluent voice AI, Block has significant distribution advantages through its existing Cash App and Square platforms that already serve millions of users. This means Dessa can be integrated into customer support bots and voice agents that are already used by a large user base, giving it real product differentiation beyond just demo capabilities.
Why is processing speech directly as audio instead of converting to text important for voice agents?
Direct audio processing allows voice AI to capture vocal nuances like tone, hesitation, sarcasm, and emotional inflection that are inherently lost in text transcription. For voice agents and customer support bots, this means they can understand user intent and emotional state more accurately, leading to more natural and contextually appropriate responses.
Further Reading
- Dessa - the Conversation Model trained only on speech - Block
- Jack Dorsey's Block launches Dessa; AI voice model trained entirely on 'human conversation' - The Economic Times
- Jack Dorsey's new AI finally sounds human - TheStreet
- Check Out the Many Uncanny, Human-ish Voices From Block's New AI Model - Gizmodo
- Block unveils Dessa: a full-duplex speech model trained from scratch only on human conversation audio - AICoder
