Block debuts Dessa, a speech-only full-duplex conversation model

Block surfaces Dessa, a speech-only conversation model

Block, the fintech group whose brands include Square, Cash App, Afterpay and TIDAL, has put a new AI research project online. A page for Dessa appeared on the company's website, and on October 23, 2026 the link reached the front page of Hacker News, where the project was described as a conversation model trained exclusively on speech.

Block's own headline goes further than the Hacker News summary: the company characterises Dessa as the most lifelike full-duplex speech model anyone has built to date. That is a marketing claim rather than a benchmarked result, and nothing outside the company has verified it, but it indicates where Block's research effort is pointed.

What full-duplex means here

Nearly every voice assistant in production is turn-based. Audio is transcribed into text, a language model composes a written reply, and a synthesiser converts that text back into sound. The system is either listening or speaking, never both, and each exchange strictly alternates.

Full-duplex models try to break that constraint. They listen and generate audio simultaneously, which in principle lets a user interrupt mid-sentence, lets the model offer back-channel acknowledgements, and preserves natural pauses and timing instead of queueing every utterance behind a transcription step. It is a harder modelling problem, because the system must continuously decide when to speak, when to stay silent and how to react while the other party is still talking.

Training only on speech compounds both the difficulty and the potential payoff. Text is a lossy intermediate format for conversation: tone, emphasis, hesitation and emotion mostly vanish during transcription. A model that learns directly from audio can, in principle, carry those signals through, which is likely what Block means by realism, though the company has not yet published a technical explanation.

What is actually known

Public detail is thin. Beyond the superlative on the project page, the material retrievable from Block's domain at the time of writing is mostly a privacy notice for the Dessa website, effective October 21, 2026, issued by Block, Inc. It confirms the project sits inside Block's corporate structure but says nothing about the model itself.

There is no published description of Dessa's architecture, training data, scale or benchmarks, no release details, and no indication whether a paper, an API or open weights will follow. The record, in other words, currently amounts to a claim.

The name has history inside the company. Block, then operating as Square, acquired a Toronto machine-learning startup called Dessa in 2019, a team that had drawn attention with a widely reported AI voice-cloning demonstration. Whether today's project continues that line of work has not been stated.

Why it matters

Speech-native, full-duplex conversation is one of the conspicuous open problems in applied AI, and it has attracted serious competition: Kyutai's Moshi model and OpenAI's realtime speech offerings both target the same limitation. The motivation is straightforward. Text pipelines flatten everything that makes human speech expressive, and assistants built on them sound stilted no matter how capable the underlying language model is.

A payments company entering that race is notable in itself. Block is not among the major consumer-AI platforms, and Dessa suggests it is funding fundamental speech research rather than only wiring up other firms' models. What the capability is intended for, whether merchant support tooling, features inside Cash App, or pure research, is unstated.

For now, the honest reading is that a bold claim has been made and not substantiated. If Block follows with a paper or a public demo that survives scrutiny, Dessa becomes a serious entrant in the push toward voice interfaces that behave like genuine conversation partners. Until then, it is a headline to watch rather than a result to cite.