We Tested Block’s Dessa. Here’s What Pindrop Pulse® Heard.

Every so often, a new voice model arrives that resets expectations for what human synthetic speech sounds like. Last week, Block’s AI research lab introduced Dessa, and it is one of those moments.

Naturally, we had to test it.

Dessa is a different kind of voice agent

Today, most voice AI is a pipeline. Speech is transcribed to text, a language model writes a reply, and a text-to-speech engine reads it back. 

But even the best pipelines inherit limits from this design. Prosody follows the text; the system waits for silence before speaking, and emotion is added after the words are chosen.

Dessa skips the text entirely. Block describes it as a full-duplex speech-to-speech model trained from scratch on nothing but human conversation, with no text and no pretrained models. Full-duplex means the model listens and talks at the same time, and the published demos show it handling timing, interruptions, laughter, and pauses with surprising ease. It also runs locally on a MacBook Pro M4 Max or higher at faster-than-real-time speed.

That matters to us as much as the naturalness. Conversational, emotionally responsive synthetic speech that runs on a laptop is a very different threat landscape than a cloud API with usage policies.

To Block’s credit, Dessa is not yet broadly available. The company has published recorded conversation demos and granted early access to a small group of AI safety researchers, which is the right call. But the direction of travel is clear, and defenders should plan for this risk now.

Models like Dessa are a real test for detection

There is a common assumption that deepfake detection works by catching what sounds “off”: robotic cadence, flat emotion, or awkward pauses. If that were how detection worked, Dessa would be a serious problem, because it is built to remove exactly those cues.

But that’s not how it works. Pindrop Pulse® uses machine learning to detect synthetic voice patterns in the audio signal itself, not in the impressions a human listener forms. Every generation process, however natural it sounds, leaves traces in the signal. A new architecture changes some of those traces, which is precisely what makes it the right test. 

The question is not “does Dessa sound human?” It clearly does. The question is whether detection generalizes to a generator it has never seen before.

What our research team did, and what we found

We ran Block’s publicly released Dessa conversation samples (21 samples in total) through Pindrop Pulse® [model version 26], with no retraining or tuning on Dessa audio.

Pindrop Pulse® flagged all Dessa audio samples as synthetic. 

Here’s a quick demo of Pindrop Pulse® analyzing Dessa: The video shows Pulse correctly detecting the AI-generated voice as synthetic (red) while identifying the human voice as real (green).

Demo uses synthetic audio for illustrative purposes only.

What I find most interesting is what this result implies. Dessa represents a different way of generating speech, one that never passes through text at all, and our models still identified signals of synthetic speech in the audio. That suggests the features our detection relies on are tied to something more fundamental than any one generation pipeline.

I want to be careful here because this is a research blog, not a victory lap. A single evaluation is a data point. Generalization is something we have to re-earn with every new architecture, and the samples Block released are curated demos, not the full range of what the model can produce. We will keep evaluating as more access and more audio become available.

Pindrop tests against the newest models, as they emerge 

By the time a generator appears in a real attack, the window for learning from it in a controlled setting has closed.

Our process is simple in principle and demanding in practice. When a new model is released, we gather whatever samples we can, test our current production models against them unchanged, and study where confidence holds and where it softens. Those findings feed directly back into training and evaluation. 

Our research draws on 1.5B real-world interactions analyzed annually, providing a baseline for what genuine human speech looks like across channels, devices, and conditions. That baseline is what makes new synthetic speech stand out, even when it is very good.

There is also a nice bit of history here. The team behind Dessa has long considered voice misuse. The same team made the RealTalk voice fake in 2019, which is part of why it asks safety researchers first. I think that instinct is exactly right. The people building realistic voices are increasingly engaging with safety, and detection is the other half of that conversation. 

What this means for everyone else

For contact centers, help desks, and anyone who verifies identity over a call or meeting, the takeaway is straightforward: asking “does it sound human?” is no longer a useful test. 

Models like Dessa are built to pass it, so trust in a voice interaction cannot depend on a listener’s ear. It has to depend on answering a more basic question first: is this a machine? That is the question Pindrop Pulse® is designed to help answer, and it is the question we will keep pressure-testing against every new model that comes along.

Huge thanks to the Pindrop Research team for moving quickly on this one. And to the Dessa team at Block: it is impressive work, and we will be watching closely.

If you are a researcher working on voice synthesis or detection and want to compare notes, reach out. We would love to hear from you.