Back to News
Block releases Dessa, a voice model trained only on human conversation

Block releases Dessa, a voice model trained only on human conversation

AIAI-assisted
Block's AI research lab has released Dessa, a full-duplex speech-to-speech model trained from scratch on human conversation audio alone, with no text data. The research preview includes demo recordings and a public waitlist, with early access going to a small group of AI-safety researchers; Block says it runs locally on a MacBook Pro M4 Max.
Read the full story on TechCrunch
Details

Why full duplex is the point

Most voice assistants work in turns: the user speaks, the system waits, then answers. That rhythm collapses in real conversation, where people interrupt, overlap and interject short acknowledgements. Block says Dessa listens and speaks at the same time, reproducing pauses, emotional shifts, sarcasm and backchannels such as "yeah" or "got it" rather than waiting for a turn to end.

The bet against transcripts

Speech systems are usually anchored to text — transcripts, captions or tokenised audio — and inherit the cadence of written language. Block argues that starting from conversational audio instead keeps the output closer to how people actually talk, rather than to how they write.

Access and the safety framing

Dessa is a research preview, not a shipping product. Demo recordings went out with the announcement, a public waitlist is open, and a small group of AI-safety researchers received early access first — a sequencing the company frames around studying risks in conversational voice before any wide release. The on-device demo, running on a MacBook Pro with an M4 Max chip, also signals a local-first path for anyone weighing on-device voice interfaces against cloud-hosted ones.

Share: