Dessa learns conversation from speech instead of text

What this is about

Dessa, a research model from Block, is designed to hold conversations without first converting speech into text. Block introduced the model on October 23, 2026. It takes audio as input and produces audio as output, while the company says it can listen and speak at the same time.

That may sound like another voice assistant, but the technical difference matters. Many systems split a conversation into speech recognition, a text model and speech synthesis. Block says Dessa was trained from scratch on human conversation, without text data and without using a pretrained model as its starting point.

What Dessa actually does

Dessa is described as a full duplex model. Both sides can speak at the same time, as they do in ordinary conversation. The system can therefore process interruptions, brief acknowledgements and changes in speaking rhythm directly as audio. It does not always have to wait for a complete sentence to be recognized and passed to another model as text.

Block released audio samples and opened a waitlist. The company says the model runs faster than real time on an M4 Max MacBook Pro. It says six researchers trained Dessa on a single Nvidia GB200 NVL72 system. These are company claims and have not yet been confirmed through a full technical paper or independent measurements.

Why it matters

Spoken language contains more than words. Pauses, hesitation, emphasis, laughter and the timing of an interruption all carry meaning. A text pipeline can lose some of those signals or add them only later. A direct speech model could therefore respond more naturally while reducing the delay between a question and an answer.

That could matter for accessibility, language learning, customer service or devices without screens. The approach is particularly interesting for languages and speech patterns with limited written training data. However, it is not yet clear which languages Dessa supports, how it handles accents or what protections it includes against voice abuse.

In plain language

A conventional voice assistant works like a conversation through several intermediaries: one writes down every spoken word, another drafts an answer and a third reads it aloud. Dessa tries to work more like a direct conversation partner. It hears sound, rhythm and interruptions and answers with sound again, without needing a written note in between.

A practical example

A visually impaired user asks an assistant for the next train connection. Four seconds into the spoken answer, she says, “No, only after 6 p.m.” A rigid pipeline might stop its output, transcribe the new recording and then calculate again. A full duplex model can detect the interruption while speaking and respond immediately to the corrected condition.

If that removes only two one second pauses per exchange, a ten minute dialogue already feels noticeably smoother. This example illustrates the potential benefit of the architecture. Block has not yet published independent latency measurements for Dessa.

Scope and limits

First, Dessa is a research preview, not a generally available product. Second, there are no publicly documented benchmarks for intelligibility, latency, multilingual performance or robustness in noisy settings. Third, realistic voices can be abused for deception, unauthorized imitation or social pressure, and Block does not yet describe a complete safety framework on the product page.

A direct speech model still has to process facts reliably. Natural timing says nothing about whether an answer is correct. Without text as an intermediate representation, auditing, logging and moderation may also work differently from today's assistants. Dessa therefore demonstrates an interesting architecture, but not yet a proven advantage in daily use.

SEO & GEO keywords

Dessa, Block, speech model, speech to speech, full duplex, Voice AI, audio model, natural conversation, speech recognition, generative AI