Block’s Dessa has released a full-duplex speech model that reached the front page of Hacker News. The real question is whether it can move real-time voice from an impressive demo to a deployable Agent. Here is a reproducible framework for evaluating it.
Project page: block.xyz/dessa
Discussion: Hacker News, 92 points and 26 comments
What’s New About Dessa
Traditional voice flows are usually turn-based:
- The user starts speaking;
- The system decides whether the user has finished;
- After a pause, the model responds;
- The user waits for the next turn.
Dessa takes a full-duplex approach. It can continue understanding new input while producing speech. If the user says, “Wait, don’t refund it yet,” the model does not have to wait for the end of the sentence before deciding to stop.
The project describes it as “the most authentic full-duplex speech model ever built.” According to the page, it supports natural interruptions and incorporates the context created by an interruption into subsequent responses rather than simply canceling the previous playback.
This is not the same as keeping a microphone open while TTS audio plays. That approach can capture sound but cannot rewrite a response before it is played. A full-duplex model must handle three things simultaneously:
- Continuously understand user input;
- Determine whether that input adds information, challenges the response, or asks the model to stop;
- Adjust the current response to the new context.
This is also the main difference between Dessa and fixed-script voice bots. A scripted bot can insert a few reassuring phrases, but it has trouble handling a mid-request change such as, “Change the address to Shanghai.”
Still, the capabilities described on a product page do not automatically translate into production performance. Whether the model frequently misinterprets interruptions or reliably preserves information across multiple turns requires testing in the real world.
The Real Question Behind 92 Points
A score of 92 points and 26 comments show that developers genuinely care about real-time voice.
But discussion activity cannot substitute for evaluation. A customer support service might handle tens of thousands of calls every day. Every false interruption wastes another model generation, while every missed interruption forces the user to repeat themselves. In payment workflows, these errors turn directly into support tickets.
More importantly, developers tend to treat “natural” as a subjective impression.
In one demo, the model stopped within 300 milliseconds of the user speaking, which made it seem remarkably responsive. The conclusion could be completely different in a noisy environment, with long context, or over an unreliable network. A natural conversation has at least three layers:
- Responsiveness: sufficiently low latency;
- Interaction judgment: knowing when to stop and when to continue;
- Task performance: still completing the job after understanding a change.
No matter how well the first two layers perform, they are of little use if the task fails.
The most revealing comments on Hacker News were not remarks about how human-like the model seemed, but the metrics developers would use to challenge its behavior. Product teams should test failed interactions directly rather than clipping the best moments from a recording.
Full Duplex Does Not Equal Natural Conversation
“The most authentic full-duplex speech model ever built” is a strong claim, but it combines two concepts that need to be separated.
The first is full duplex. This is a communication mechanism in which the model continues receiving and understanding input while speaking. It addresses whether information can flow in both directions at the same time.
The second is naturalness. This describes the interaction outcome: users do not have to learn the machine’s phrasing, and interruptions do not require repeating a fixed wake word.
Full-duplex operation does not guarantee a natural result, and the reverse is also true. A team can pre-record many realistic responses and adjust pauses based on the user’s speaking speed, but that remains a single-turn script with better timing control.
Evaluation should at least record the following separately:
| Layer | Question to answer |
|---|---|
| Audio experience | Are the timbre, speaking rate, and pauses suitable for continuous conversation? |
| Interaction experience | Can the model distinguish interruptions, acknowledgments, and silence? |
| Task experience | Can it resume and complete the original task after an interruption? |
A demo in which the user changes an address halfway through a sentence does not cover routine cases such as the user coughing, someone speaking nearby, or the user saying “uh-huh” without making a correction.
What to Test in Practice
Start by fixing the model version, audio device, sample rate, and network conditions. Then record five groups of metrics.
1. Time to First Audio
Measure P50, P90, and P95 from the moment the user finishes expressing an intention to the moment perceptible audio begins playing.
Do not report only the average. In customer support, P95 better reflects the waiting time experienced by a large number of users. Keep network latency separate from model latency; combining them will slow down diagnosis.
2. Recovery Time After an Interruption
After the user interrupts, record two intervals:
- Stop latency: the time from receiving the interruption until playback stops;
- Recovery latency: the time from stopping playback until the model produces a useful response again.
Stopping quickly does not mean recovering quickly. Some implementations mute the audio immediately and then rerun the entire inference process, which still sounds awkward.
3. False Trigger Rate
Prepare non-command audio, including coughs, keyboard sounds, television audio from the environment, and simple acknowledgments.
Report at least the following:
- The rate of incorrectly stopping playback;
- The rate of incorrectly continuing playback;
- The rate of missing valid commands that contain no wake word.
In real conversations, false triggers are usually more damaging than missed triggers. When a model abruptly goes silent, users are likely to hang up.
4. Multi-Turn Context Retention
Run continuous tests of at least 10 turns, changing a key constraint midway through. For example:
“Book me a ticket to Shanghai tomorrow.”
“Change it to the day after tomorrow.”
“Keep the budget under 800 yuan.”
Observe whether the model continues using the old date, city, or budget.
These tests are harder than a single-turn test of whether the model can remember something. A full-duplex model may retain audio state, or it may have only a continuously growing text context. The two behave differently after an interruption.
5. Task Completion Rate
In customer support, the outcome must ultimately be a completed task rather than a chat-quality score.
Record at least the following:
- Intent recognition accuracy;
- Completion rate for collecting required information;
- Tool-call success rate;
- End-to-end task completion rate;
- Escalation-to-human rate;
- Average handling time.
Also retain transcripts of failures. Do not keep only successful recordings. Without a distribution of failures, teams cannot determine whether problems originate in ASR, the conversation policy, model reasoning, or a downstream API.
The Product Boundary for Voice Agents
Full-duplex interaction is best suited to reducing the user’s waiting time.
In customer support, booking, and after-sales consultations, users do not need to wait for a long answer to finish before saying, “That’s wrong; the order number is 1234.” This is especially useful for short tasks.
Companion products face more constraints. Users being willing to talk at length does not mean the system can store and use that content indefinitely. Recording retention, sensitive-information detection, use involving minors, and deletion mechanisms each require separate design.
Some tasks cannot be solved by sounding more natural:
- Identity verification;
- Payment authorization;
- Address and amount confirmation;
- Privacy consent;
- Refunds and order cancellations.
These steps need a deterministic toolchain as a backstop. For example, the voice model can collect intent, a state machine can confirm the current step, tools can submit actions, and critical fields can be confirmed through explicit repetition or a button press.
A safer division of responsibility is:
The model handles understanding and expression, business systems manage state, and rules govern authorization.
Natural interruption handling should not also make the model responsible for everything. Natural interaction and strongly constrained execution are not the same thing.
Editorial Perspective
The scarce part of full-duplex voice is not whether a model can interrupt at any time. That capability already has a clear technical definition, and Block’s Dessa has turned it into a released product feature.
What matters more is whether the system can recover the task after an interruption. When a user says, “Wait, use a different address,” the system cannot merely demonstrate that it heard the correction; it needs to know which parameters are now invalid and which steps must be repeated.
This differs from earlier voice bots because the competitive focus has moved from how natural the script sounds to whether real-time state can update continuously as the user speaks. That makes it closer to an AI Agent than to a voice interface.
Dessa’s false-trigger rate in real telephone environments, its long-context stability, and its end-to-end task completion rate remain uncertain. A score of 92 points and 26 comments demonstrate that demand exists, but not that the product metrics are sufficient. The next step is to examine public evaluations and failure cases rather than another polished demo.