What is Dessa AI?
Block describes Dessa as a true speech-to-speech AI model: audio goes in and audio comes out. Unlike a conventional assistant that turns speech into text before reasoning and synthesizing a reply, the research direction keeps spoken interaction in its original medium.
That distinction matters because a transcript cannot fully represent cadence, breath, laughter, hesitation, sarcasm, or the emotional weight of a pause. Dessa voice AI research explores whether a model trained directly on human conversation can respond to those signals more naturally.
What Block has shown
Block's published demonstrations include overlapping conversation, short acknowledgements, interruptions, guessing games, and emotionally expressive exchanges. These examples are useful evidence of the intended behavior, but they remain demonstrations selected and published by the research team.
Dessa should therefore be treated as a research preview rather than a generally available product. Claims about quality, safety, latency, and local performance still need broader independent evaluation when access expands.
Can you use Dessa today?
Block has said that the preview can run on a MacBook Pro M4 Max or better and that a limited group of AI safety researchers has received early access. There is no public Dessa API connected to this website.
The playground on this site uses Qwen Realtime to provide a real, usable speech session while keeping the research explanation honest. It is a way to explore realtime voice interaction, not a substitute for testing Block's model.