Block unveils Dessa, a voice model trained on human conversation

24 October 2026 — Block announced Dessa on 23 October, a speech-to-speech AI model that the financial technology company says learned to converse entirely from human audio, without text or a pretrained model. Block’s internal research lab also released recordings, opened a public sign-up queue and offered a limited number of AI safety researchers an opportunity to try the model early, Unite.AI reported. For anyone building voice software, Block’s proposed change is to keep both learning and conversation in sound rather than making written words the intermediate step.

Dessa learns from conversation audio alone

Block says it trained Dessa from scratch using human conversations as sound, with neither written text nor another model as a starting point, The Economic Times reported. That is a claim about the material used to train the model, distinct from its behaviour once a conversation begins. Block also says Dessa processes incoming audio directly instead of first turning speech into a transcript.

A transcript can preserve the words in a sentence while dropping the way they were spoken. Block says direct audio processing lets Dessa retain cues such as laughter, sarcasm and sadness. The company describes speech as a harder training medium than text: even repeated words arrive as different sound, and each spoken word occupies far more data than its written counterpart. Dessa must also produce sound, including the timing and manner of its replies.

Block puts Dessa’s language-learning word count at hundreds of times below that of large language models, while acknowledging that the latter also learn material beyond conversation. The comparison is Block’s account of its training, rather than a published test of how well Dessa answers questions. The company lists the efficiency with which a speech-only model learns language and facts, learns new languages and learns from conversational context among the research questions raised by the work.

The research lab is small and, Block says, was founded by former members of Dessa, a Toronto deep-learning startup. It reports directly to Jack Dorsey, Block’s cofounder and chairman. The shared name carries into the new model, whose release places a training claim alongside recordings that let people hear some of its conversational behaviour.

Block’s recordings capture overlapping speech

Block calls Dessa a full-duplex model: it can receive speech while producing speech, instead of waiting for one side to finish before the other begins. Think of a telephone connection that carries both voices at once. Block says Dessa learned from conversations when to interrupt and when to give a brief acknowledgement without taking over, making the timing of a reply part of the task rather than a fixed handover between turns.

Asking a question aloud could become less stop-start: an interruption would be possible while an answer was still being spoken, and a laugh or a hesitant pause could remain part of the exchange rather than being reduced to words in a transcript.

Block describes its published clips as live recordings in which nobody steers Dessa, with subtitles added afterwards. Its examples include the model laughing during a sentence, speaking louder when asked and helping complete a thought about a film title. Those are selected conversations supplied by the company. Kingy AI reported that it had not tested Dessa itself and that Block had not given the proportion of attempted conversations represented by the clips.

Ragavan Thurairatnam, the lab’s lead, called Dessa “a GPT3 moment for speech” in a post on X. Block’s recordings concern particular exchanges, including how the model sounds and when it speaks. The company also says inference runs faster than real time, a claim about producing speech quickly enough for conversation, while the published clips give prospective developers examples to hear rather than repeatable measurements across conversations.

One Nvidia GB200 NVL72 and limited access

Block says the training run used a single Nvidia GB200 NVL72 system. Separately, it says the resulting model can run locally on a MacBook Pro M4 Max or better. The two hardware claims describe different jobs: training builds the model from conversation audio, while local inference uses the trained model to listen and respond. Local operation would matter to developers considering where to run a voice system, although access to Dessa remains restricted.

Block is offering early access to a small group of AI safety researchers to look for risks the team says it may have missed. The company argues that a voice system unable to register distress in someone’s speech can carry on without responding to it. Its research access therefore puts the speech-only approach before outside specialists while the wider public has recordings and a waitlist.

Block had not announced pricing or a general release date as of 24 October, Tech Insider reported. The early-access request form asks applicants for their name and an institutional email address, with an optional link to a Google Scholar profile or blog.