Block unveils Dessa, a full-duplex, on-device speech-to-speech model trained from scratch on human conversation, with live demos and early access for AI safety researchers.

Dessa runs locally on a MacBook Pro M4 Max or better, enabling on-device operation and keeping raw voice data off servers, though exact memory, latency, and model size aren’t disclosed.

Unlike text-first systems, Dessa processes audio directly to preserve emotional cues, tone, and timing in conversations.

The report raises open questions and risks, including data provenance and consent for training data, impersonation risks, and data provenance concerns.

CZ Zhao highlights potential uses in education, such as GiggleAcademy, while noting safety considerations.

Analysts call for meaningful benchmarks to assess conversational behavior, turn-taking, interruptions, response quality, and safety, with emphasis on reproducible testing and transparent documentation.

There’s a cautious but intriguing possibility that training language from speech alone may need far less data, though definitive assessment awaits further data, papers, and independent testing.

There are concerns about deception and impersonation, as convincingly human voices could boost scam effectiveness.

The article references accompanying media and directs readers to Dessa’s site for more demonstrations and examples.

The evidence base consists of curated live clips with subtitles, not a comprehensive evaluation; independent benchmarks and held-out test sets are absent.

Block envisions benefits for customer service and accessibility, with a local-first, on-device voice model offering potential competitive advantages.

Key researchers include Ragavan Thurairatnam (lead) and teammates, with interns contributing; project origin traces to Toronto-based Dessa under Block.