Block's new AI model Dessa was trained entirely on human speech to catch emotions text-based models miss

Block, the payments company behind Square and Cash App, has introduced a speech AI model called Dessa. Block describes it as the first model trained entirely on human speech, with no text data involved, and built from scratch. The announcement was reported by The Decoder.
The central design choice is that Dessa processes speech directly as audio rather than routing it through a transcription step first. According to Block, this direct approach lets the model pick up on emotions, sarcasm, laughter and pauses in real time. That matters because a conventional voice pipeline typically converts audio to text, hands the text to a language model, and then synthesizes a spoken reply; prosody, tone and timing are largely discarded along the way. By working on the audio signal itself, Dessa can in principle retain those cues.
Block also says the model supports full-duplex communication, meaning it can listen and speak at the same time. This is a meaningful departure from turn-based voice assistants, which wait for the user to finish before responding. Full duplex is what makes natural interruption and back-channeling possible in human conversation, and it is considerably harder to engineer than a request-response loop.
On infrastructure, Block says Dessa runs locally on a MacBook Pro M4 Max and was trained on a single NVIDIA GB200 NVL72. Those two facts sit at opposite ends of the compute spectrum and suggest the model is small enough to be served on a high-end consumer laptop after training on a data-center rack-scale system.
Block frames the motivation as a combination of curiosity, a wish for better human-AI communication, and safety concerns, arguing that current voice AI ignores emotional cues. The model is not publicly available. The team is first looking for AI security researchers to test it for unexpected risks, and interested parties can sign up via a form.
The emotional-voice space has seen several attempts. Hume AI launched EVI, its Empathic Voice Interface, in March 2024. Kyutai followed in September 2024 with Moshi. LAION's Open Empathic project has been working since late 2023 on training empathetic speech synthesis models using an open-source audio dataset. OpenAI overhauled ChatGPT's Advanced Voice Mode in June 2025 and has kept refining it, adding better intonation, pausing, and more realistic expressions of empathy and sarcasm.
A useful distinction here is between sounding emotional and understanding emotion. Much of the prior work focuses on synthesis: producing a voice that seems warm, amused or empathetic. Dessa's claim is about perception, extracting emotional and interactional signals from the incoming audio. In practice, the two are complementary, since a system that detects sarcasm in the user's tone can only act on it if the rest of the stack can respond appropriately.
That points to the harder problem. Recognizing a pause or a sarcastic inflection is a perception task; deciding what to say in response requires reasoning, and reasoning is typically the domain of large text-trained models. A model trained with no text data at all would not, by construction, inherit the language understanding of a text LLM. This suggests Dessa is more plausibly a front-end perceptual layer that could feed a reasoning model than a complete replacement for one. How Block intends to combine the two is not something the announcement settles.
Running locally on a MacBook Pro M4 Max is notable for privacy and latency. Audio never has to leave the device, and full-duplex interaction avoids network round trips. A likely trade-off is capability versus the largest hosted models, though the source gives no benchmark numbers to judge that.
Why it matters: developers building voice agents currently assemble audio capture, transcription, an LLM and speech synthesis, and lose emotional signal at the transcription boundary. A model that handles audio natively and supports full duplex could simplify that stack and enable more responsive interfaces, if it holds up. Because access is limited to security researchers for now, product teams should treat this as a signal of where voice AI is heading rather than something they can build on today; that reading is inference from the stated availability. The safety-first testing posture also hints that Block sees emotional perception as carrying risks worth examining before release.
Based on reporting from the original publisher. Visit the source for full context and later updates.
Publisher excerpt
Block, the company behind Square and Cash App, has introduced Dessa, a speech model that processes audio directly instead of converting it to text first. Dessa is designed to recognize emotions, sarcasm, and pauses in real time. The model runs locally on a MacBook Pro M4 Max but isn't publicly available yet. The article Block's new AI model Dessa was trained entirely on human speech to catch emotions text-based models miss appeared first on The Decoder .