Block has introduced Dessa, a speech-to-speech model it says learned conversation entirely from human audio. Block says Dessa was trained without text or pretrained models. That makes the training approach the story: the model had to learn spoken language and conversational timing from the same material.
The Verge reported the announcement on October 23. Block's launch page offers recordings and a waitlist, with early access aimed at safety researchers. This assessment covers the announcement and its documented limits. We haven't run Dessa or independently checked its audio demonstrations.
Listening while speaking
Dessa is full-duplex: it listens and talks at the same time. Block describes a model that chooses when to interrupt or offer a brief acknowledgment while someone else is speaking. The team also says it hears and produces emotional cues, including laughter and changes in vocal delivery.
For a developer assessing that claim, a transcript leaves out part of the behavior being promised. A correct written answer won't establish whether an acknowledgment arrived at a useful moment or an interruption cut someone off. Block's description of simultaneous listening and speaking calls for an evaluation that preserves both speakers' audio and timing.
Local inference comes with unanswered questions
Block says Dessa runs locally on a MacBook Pro M4 Max or better. The launch page also says inference is faster than real time, but supplies no numerical latency result or memory requirement. Those omissions leave developers unable to estimate the hardware cost or conversational delay from this announcement alone.
The team says it trained Dessa using one NVIDIA GB200 NVL72. That is a training-hardware claim, separate from the laptop inference claim. Block provides no training duration or total cost on the page, so the hardware name alone cannot establish that the approach is inexpensive to reproduce.
Expression needs a separate safety test
Block is offering early access to a small group of AI safety researchers. Its stated motivation includes the risk that a voice system might fail to recognize distress. Yet the launch page provides no evaluation showing how accurately Dessa recognizes distress or how safely it responds. Expressive delivery gives researchers something to investigate; it doesn't supply that missing result.
The next evidence to watch is what those outside researchers find. Block's invitation explicitly asks them to uncover unanticipated risks. Tests of misunderstood emotion and badly timed interruptions would help establish where speech-only training works, and where a convincing voice still gets the conversation wrong.