What Is Dessa?
Dessa is a speech-to-speech AI model developed by a small research lab inside Block, the financial services company co-founded by Jack Dorsey. Unlike every major conversational AI on the market — ChatGPT, Claude, Gemini — Dessa was trained from scratch on nothing but human conversation audio. No text data. No pretrained large language model. Just recordings of people talking to each other.
Block says the model learned to hold real-time conversations simply by listening. It handles full-duplex interaction, meaning it can listen and speak at the same time rather than waiting for the user to finish talking. It picks up on laughter, sarcasm, frustration, pauses, and backchanneling cues like "yeah" or "got it."
The model was unveiled on October 23 as a research preview. Access is currently limited to AI safety researchers, with a broader waitlist planned.
How It's Different
Every major AI companion today — from Character.AI to Replika to Kindroid — runs on text-based large language models with a text-to-speech voice layer bolted on top. The LLM processes your words as text, generates a text response, and then a separate TTS system reads it aloud. This creates latency, loses vocal nuance, and makes natural turn-taking difficult.
Dessa skips the text step entirely. It processes speech directly as speech. Block claims this makes it the first AI you can have a realistic live conversation with that was trained exclusively on audio. Early demos show it handling interruptions, changing its mind mid-sentence, and responding to emotional tone — behaviors that text-based models struggle to replicate in voice mode.
There's a trade-off. The Verge notes that Dessa's conversations can be "a little awkward and uncanny." In one demo, the model initially misidentified a speaker as "Ronaldo." It's clearly early-stage technology.
What This Means for Companion Users
For the AI companion industry, Dessa points toward a future where voice interactions feel less like talking to a walkie-talkie and more like talking to a person. Full-duplex conversation — where both parties can speak and be heard simultaneously — is something companion users have wanted for years. Current voice modes on Character.AI, Replika, and others require the user to stop speaking and wait for the AI to finish generating a response before they can reply.
Dessa also runs locally. Multiple reports confirm it can run on a MacBook Pro with an M4 Max chip, potentially without an internet connection. For companion users who worry about their intimate conversations being stored on company servers, a model that can run entirely on-device is a meaningful privacy development.
Of course, Dessa is not a companion product. It's a research preview from a fintech company, and there's no indication Block plans to enter the companion space. But the architecture — speech-native, full-duplex, locally executable — is exactly what many companion developers have been working toward. If Dessa's approach proves scalable, it could influence how the next generation of companion apps handle voice.
The Bottom Line
Dessa is worth watching because it represents a genuine architectural departure from the text-first paradigm that dominates AI conversation today. For an industry that has spent years trying to make LLMs sound less robotic in voice mode, a model that was never trained on text in the first place is a provocative alternative. Whether it scales beyond research previews — and whether Block chooses to do anything with it — remains to be seen.