Block's Dessa Trains a Speech-Only AI, but Demo Audio Falls Short
Dessa's audio-only, full-duplex training is a genuine technical departure, but the supplied demos draw criticism for cadence and interruption behavior that undercut the human-realism claim.
Block's Dessa research lab has introduced Dessa, a full-duplex speech-to-speech model trained from scratch on human conversation audio alone, with no text or pretrained models. The company says it listens and talks simultaneously, backchannels naturally, and runs locally on a MacBook Pro M4 Max or better with faster-than-real-time inference.
Dessa was trained on a single NVIDIA GB200 NVL72 and learned language from hundreds of times fewer words than text-based LLMs, according to the lab. The release includes live recordings of conversations with GPT-Live-1, a third-party model whose developer is not affiliated with Block, and an early-access program for AI safety researchers.
Key takeaways
- Dessa is full-duplex: it listens and speaks at the same time, learning when to cut in and when to backchannel with cues such as “yeah” and “got it” from real conversation rather than taking turns.
- The model is audio-only and speech-to-speech, trained without text or pretrained models; the lab says it natively hears emotion, sarcasm, laughter, sighs and inflection that text pipelines discard.
- Training used one NVIDIA GB200 NVL72, and inference runs faster than real time locally on a MacBook Pro M4 Max or better.
- The published demos are live recordings with no steering, subtitles added afterward; Dessa chooses what to say, when to say it, and its own voice and personality.
- Block's Dessa lab, founded by former members of the Toronto deep learning startup Dessa, reports to Jack Dorsey and is offering early access to AI safety researchers.
What people think
Analysis details
4 comments analyzed. Analyzed .
The sample selects active discussion branches and includes available parent comments. It may omit parts of the full thread. Selected themes do not measure community opinion or how common a view is.
Demo audio realism questioned
The homepage sample is judged to have uncomfortable cadence, tonal range and structure, falling far short of real conversation. That reaction challenges the claim of authenticity and suggests the demo does not by itself establish human parity. The criticism is qualitative, but it targets the exact evidence Dessa offers for its central claim.
Interruption training harms turn-taking
Simultaneous speech is acknowledged as technically notable, but training the model to interrupt as much as possible is criticized as artificial. In the first example, a laugh after “winter” seems unmotivated, and the human speaker never finishes a thought before Dessa monologues. Full-duplex overlap alone does not produce natural conversation.
AI companion market raises unease
Dessa is seen as aiming at full human parity, and the broader full-duplex trend is called “Her”-ish. A suggestion that AI girlfriends are a large market is met with a joke about population collapse, reflecting concern about emotional attachment and social consequences rather than technical performance.
GPT-Live-1 comparison
GPT-Live-1 is praised as fantastic, while Dessa is characterized as pursuing full human parity. The comparison positions Dessa against an existing third-party voice model, though the article notes that the developer is unaffiliated with Block and has not endorsed Dessa.