Block introduces Dessa, a voice AI it says learned from speech alone

Block has introduced Dessa, a voice AI it says learned to converse entirely from human speech, without text or pretrained models. The promise is conversation that preserves emotion, interruptions and timing. For now, the public evidence is Block’s recorded demonstrations; early access is reserved for a small group of AI safety researchers, with a waitlist for others.

Learning the conversation, not just the words

Dessa takes audio in and produces audio out. Block describes it as full-duplex, meaning it can listen and speak at the same time rather than wait for a rigid handoff. The company says it learned when to interrupt and when to offer brief acknowledgments such as “yeah” or “got it.”

The training approach is the central distinction. Block says it built Dessa from scratch using only human conversation audio. Its argument is that converting speech to text discards information carried by delivery: a sarcastic tone, a frustrated sigh, laughter or a voice cracking with sadness.

Block also claims Dessa learned language from hundreds of times fewer words than large language models. It qualifies that comparison by noting that those models must learn other material, including programming languages. That makes the claim narrower than a like-for-like demonstration of better learning efficiency across the same tasks.

Recordings put timing and expression on display

The introduction includes recordings that Block describes as live conversations between a person and Dessa, without anyone steering the model. Subtitles were added afterward. The examples emphasize small conversational behaviors, not just whether Dessa can produce a correct answer to a spoken question:

• In a conversation about the release, Block highlights Dessa laughing mid-sentence and raising its volume when asked to speak up.
• In a movie discussion, the company presents Dessa completing a sentence when the person cannot recall a title.
• An anniversary-surprise recording highlights a gasp and a response that Block describes as sounding like teary gratitude.

Another recording pairs Dessa with GPT-Live-1. Block explicitly labels both voices as AI-generated and says the third-party model’s developer is not affiliated with Block and has neither reviewed nor endorsed Dessa. That clip is a demonstration of interaction, not a partner endorsement.

Dessa picks what to say, when to say it and how, down to its own voice and personality.

Block, in its Dessa introduction

Outside testing starts with safety researchers

Block says Dessa runs faster than real time and can operate locally on a MacBook Pro M4 Max or better. It says training used one NVIDIA GB200 NVL72. Those are company-supplied hardware and speed claims, distinct from the conversational behaviors presented in the recordings.

The safety rationale goes beyond making exchanges less awkward. Block argues that an AI unable to recognize emotion in a voice could miss signs of serious pain and continue responding inappropriately. It presents Dessa as a step toward addressing that problem, rather than evidence that the problem has been solved.

Dessa lead Ragavan Thurairatnam invites a small group of AI safety researchers to use early access to look for risks the team has not anticipated. Other interested users can join the waitlist. The invitation makes outside risk-finding an explicit purpose of this initial access, not simply a chance to try the voice.

Block also leaves several research questions open: how efficiently speech-only models learn language and facts compared with text-trained systems, how they acquire new languages, and what they can learn from examples supplied during a conversation. Those are proposed directions for investigation, not capabilities established by the introduction’s expressive voice samples.