Block's Dessa is a voice model trained only on human speech that runs on a MacBook Pro. It is not public yet

Most voice agents turn the caller into text first. That step throws away how the caller sounds.

What happened

Block, the company behind Square and Cash App, unveiled Dessa, a speech to speech model it says was trained from scratch only on human conversation, with no text and no pretrained models. Block says it is the first AI you can have a realistic live conversation with that was trained only on speech.

Dessa is full-duplex, so it listens and talks at the same time, and Block says it knows when to cut in and when to backchannel with things like "yeah" or "got it". It is meant to pick up emotion, sarcasm, laughter and a frustrated sigh. Block says inference runs faster than real time, the model runs locally on a MacBook Pro M4 Max or better, and it was trained on a single NVIDIA GB200 NVL72.

There is no public access yet. Block is first asking AI security researchers to test it for unexpected risks. TechCrunch notes that Square already uses voice AI for ordering, and that it is not clear whether Dessa's emotion is steered by some kind of tags. The Decoder adds a fair caution: sounding emotional is not the same as reasoning well, and for advanced speech AI the quality of the answer matters more than how natural the voice sounds.

My take

You cannot build on Dessa today, so treat this as a signal, not a tool. The useful part is Block's critique of current pipelines: speech to text, then a language model, then text to speech. That chain loses tone.

For voice agents built for service businesses right now, that means:

1. Do not rely on the transcript alone to spot an upset caller. Add clear escalation triggers: repeated questions, the word "cancel", a request for a person.
2. Keep a fast human handoff path on every call flow, and log when it gets used.
3. Test interruptions. Callers talk over agents, and a bot that cannot handle it feels broken.
4. Keep the voice layer swappable, so a better speech model can drop in later without rebuilding the booking and CRM logic behind it.

The value in a voice agent is still the work it does after the call: the booking, the CRM note, the follow up.