Block Dessa

Overview

Dessa is a speech model unveiled on October 23, 2026 by the small AI research lab inside Block, the payments company Jack Dorsey runs. It takes a different route from most voice assistants. Those transcribe your speech to text, reason over the text, then synthesise a reply. Dessa takes audio in and puts audio out with no text stage in between.

That choice drives everything else. Transcription throws away timing, pauses, where the laugh fell, the hesitation halfway through a sentence. Dessa keeps those in the signal, which is why it can pick up before you finish, stop when you cut in, and read emotion from tone. Full duplex here is an architectural property, not an interrupt button wired into a pipeline.

Training runs against the mainstream too. Block says Dessa was trained from scratch on recordings of human conversation alone, with no text corpus and no pretrained language model underneath, so language was learned by listening. Training used a single NVIDIA GB200 NVL72 rack, and Block says the model needed hundreds of times fewer words than a language model. Inference runs faster than real time on a MacBook Pro M4 Max or better.

There is no public API and no open weights yet. Early access goes to a small group of AI safety researchers, with a public waitlist on the site.

Key Features

• Audio in, audio out, no text stage: The model takes audio and returns audio with no intermediate text. Tone, pauses and laughter survive instead of being flattened in transcription.
• Full duplex, listens while it talks: Both audio streams run the whole time. An interruption is something that happens in the stream rather than an exception the system has to handle.
• Trained on conversation audio only: Trained from scratch on recordings of human conversation, with no text and no pretrained base. Block says it picks up language the way a child does by listening, needing hundreds of times fewer words.
• Rack-scale training, laptop inference: Training ran on one NVIDIA GB200 NVL72 rack; inference beats real time and runs locally on a MacBook Pro M4 Max or above, with no round trip to a server.
• Emotion read from the voice: No tags marking happy or sad in text. Laughter, a cracking voice or sarcastic intonation stay inside the audio signal. Block frames this as a safety requirement for voice agents acting on your behalf.
• Application-based early access: Early access prioritises AI safety researchers, with a public waitlist alongside. No public API and no weights released.

Use Cases

• Real-time voice assistants where interruptions have to feel natural
• On-device voice interaction without sending audio to a server
• Research on speech-native training and emotion recognition
• Companion or support conversations where tone matters as much as content

Pros

• Full duplex is built into the architecture rather than simulated with a state machine
• End-to-end audio keeps emotional cues out of the transcription loss
• Inference faster than real time on a high-end laptop
• Conversation-only training gives the speech-native route a concrete reference point

Pricing

Early access runs through an application form on the official site, prioritising AI safety researchers, with a public waitlist open alongside. There is no public API, no released weights and no announced pricing. Local inference requires a MacBook Pro M4 Max or better.

Summary

The interesting thing about Dessa is not whether you can use it today but which training route it picked. Mainstream assistants chain transcription, reasoning and synthesis; Dessa deletes the middle step and learns language from sound directly. If that route holds, latency and emotional awareness in voice interaction move together.

For now it is a research artifact: application-gated, no API, no weights, and demos Block selected itself. Judge it once weights or a public interface appear. For teams building voice agents, the durable value today is in how the full-duplex part is engineered, which is the hard part of this architecture.

Version History

• Block unveils Dessa, a full-duplex speech model (2026-10-23): Block research lab announces Dessa, a full-duplex speech-to-speech model trained from scratch on human conversation audio with no text stage and no pretrained base, on a single NVIDIA GB200 NVL72 rack. Inference runs faster than real time on a MacBook Pro M4 Max or above. Early access goes to AI safety researchers with a public waitlist; no weights, API or pricing announced.