Dessa: The Voice AI That Skipped Text Entirely—Here's Why Creators Care

By Mirrorize ·

Dessa: The Voice AI That Skipped Text Entirely—Here's Why Creators Care

On October 23, Block—Jack Dorsey's fintech company—dropped something you can't get yet but can't stop watching: Dessa, a speech-to-speech AI trained only on recorded human conversation. No text layer. No pretrained language models. No transcription step between listening and speaking. This is the opposite of how voice AI normally works.

Most voice AI runs speech through a text pipeline: speech-to-text, then text processing, then text-to-speech. Each step strips away tone, timing, and emotion. Block argues that this pipeline throws away tone, hesitation and emotion. Dessa skips all that. Dessa is a speech-to-speech model: audio goes in, audio comes out. Block says Dessa was trained from scratch on human conversation audio alone. That makes it different from most speech-to-speech systems, which start from a text language model and add audio on top.

Block describes Dessa as full-duplex: it listens and talks at the same time, and can cut in or offer short acknowledgments such as "yeah!" while someone speaks. In one example, Block says, Dessa laughs mid-sentence and speaks louder when asked.

What people are actually hearing

Block released 21 demo clips with Dessa speaking. The recordings show range: a tired voice explaining a rough work day, a person playing a guessing game with Dessa, conversations bouncing between voices. The most convincing moments capture what normal conversation actually is—interruptions that feel natural, laughter that lands mid-sentence, hesitation before an uncertain question.

Some exchanges sound close to natural conversational flow, including a genuinely convincing laugh, while other segments, especially a male voice in several clips, sound noticeably synthetic or "plasticky." One listener said: "There's no way I would be able to detect if this was AI from the examples alone." Others noticed the bot's knowledge is thin—it's trained on speech, not facts—and sometimes speech feels slightly off when you listen closely.

Why podcasters and voice creators are paying attention

Dessa could be an unlock for podcast creators. Tone, timing, and emotion make a podcast or interview work, and text-based AI struggles with all three. Dessa could make possible: dubbing that matches human speaking patterns, interactive audio experiences where fans can "talk" to a show, and AI co-hosts or producers that can read the room in an edit session.

For years, voice cloning and speech synthesis felt robotic because they operated at the edge of text pipelines. Dessa is a live conversation model with no text anywhere in its pipeline; emotion emerges from the model itself. That's not a minor detail for anyone working in audio.

What actually works and what doesn't

Dessa is a research preview. As of November 2026, Dessa is a research preview. A public waitlist is open, and AI safety researchers can request early access with an institutional email. Block has not announced an API, pricing or model weights.

It's a speech-to-speech model—it can't do any tool use or you can't guide it much. The model learned language from speech alone, so it won't reason or fetch facts like an LLM. It's trained on speech, not vast amounts of knowledge. That matters if you want an AI co-host that can reference a guest's previous appearances or look up a song title.

The technical win that stopped people mid-scroll: "it runs locally on a MacBook Pro M4 Max or better. And we figured out how to train Dessa on only one NVIDIA GB200 NVL72." If it really can replicate that low latency plus that expressiveness on local devices, that's one heck of an accomplishment for "indie" AI.

How to follow it

You can't use Dessa yet. Block says the Dessa team has already granted "a small group of AI safety researchers" early access to the model, which can run locally on any MacBook Pro M4 Max "or better." There's a waitlist at block.xyz/dessa, and you can browse the demo clips to hear what a conversation trained only on speech actually sounds like.

The demos don't hide imperfection. Block says Dessa AI drops the text support entirely, with no text model underneath, no transcript and no pretrained weights. The mistakes stayed in. That honesty—keeping the stumbles in the demos—is part of what's making voice researchers and podcast creators lean in. This isn't a sales pitch yet. It's a genuinely different approach to how machines learn to talk.

Sources









More from the Learning Center

Make AI art · AI video · AI music on Mirrorize, free to start.