Block's new AI model Dessa was trained exclusively on human speech to detect emotional cues that text-based models overlook.

Meta description: Block unveils Dessa, an AI speech model trained on human audio with no text data, supports full-duplex communication, and runs on MacBook Pro M4 Max.

Tags: Block, Dessa, AI speech, full-duplex, voice AI

Featured image: abstract waveform visualization with neural network-style nodes and flowing audio lines, no people, no logos, no text.

Block's Dessa AI speech model is the latest attempt by Square and Cash App parent Block to build voice systems that respond to spoken cues rather than only transcribed words. The company says Dessa is the first AI model trained entirely on human speech, with no text data at all, and that it was built from scratch. The key point for readers is not simply that the model sounds human, but that it is designed to process speech as audio, which could matter for timing, emotion, and interruption handling in voice interfaces.

Background: Why Block moved away from text-first voice systems

Many voice assistants and chatbots convert spoken audio into text before a model generates a response, then send the answer back through a speech output path. Block says Dessa avoids that audio-to-text step and processes speech directly as audio. In that setup, the system can react to emotional cues, sarcasm, laughter, and pauses in real time. The company links the project to curiosity, a desire for better human-AI communication, and safety concerns, arguing that current voice AI ignores emotional cues.

The model also supports full-duplex communication. In practice, that means it can listen and talk at the same time, instead of waiting for a user to finish before responding. Full-duplex speech can make conversation feel more natural because people interrupt, backchannel, and adjust tone while speaking. Block positions Dessa as an attempt to move beyond turn-taking voice assistants toward systems that can handle overlapping audio.

Development: Local inference and a single training machine

Block reports that Dessa runs locally on a MacBook Pro M4 Max. The training setup is also specific: the model was trained on a single NVIDIA GB200 NVL72. Those details matter because they show the project is tied to a local MacBook Pro M4 Max and a single NVIDIA GB200 NVL72. Block has not said whether Dessa is available to users, and the company describes it as not publicly available yet.

The training approach is also distinct. Block says the model was trained entirely on human speech, with no text data at all. That means the system is not merely a text model with a voice wrapper. Instead, the audio signal itself is the main training material. Block says the model was built from scratch, which implies it was not adapted from a text-first architecture.

Earlier attempts: Emotional voice AI has arrived in pieces

Dessa is not the first project trying to make speech systems sound emotionally aware. In March 2024, startup Hume AI launched EVI, or Empathic Voice Interface, described as a voice AI with "emotional intelligence." In September 2024, Kyutai followed 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 continued refining it since then, adding better intonation, pausing, and more realistic expressions of empathy and sarcasm.

These projects show a broader trend: voice AI is moving from simple speech recognition and response toward richer audio behavior. But Block's framing adds a technical distinction. Sounding emotional is not the same as being emotional or intelligent, even if it creates that impression. For advanced speech AI, capabilities such as reasoning while responding can matter more than how natural the voice sounds, because they affect answer quality.

Latest progress: Security researchers are the first audience

The current stage of Dessa is pre-public testing. Block says the team is first looking for AI security researchers to test the model for unexpected risks. Anyone interested can sign up via a form. That means the immediate next phase is not public rollout, but safety testing.

For readers, the timeline points to a model that is still not publicly available. Block has given a concrete technical direction: speech-first training, no text data, full-duplex operation, local inference on a MacBook Pro M4 Max, and training on a single NVIDIA GB200 NVL72. The unresolved question is whether a speech-native model can handle reasoning and safety as well as it handles tone. The next signal will come when security researchers report unexpected risks, and when Block makes the model more widely available.