You need to agree to share your Hugging Face information to access this model
This repository is publicly accessible, but you have to accept the conditions to access its files and content.
Access to Echo Omni Duplex is provided through Zero Runtime. Visit https://zeroruntime.ai/?utm_source=huggingface&utm_medium=referral&utm_campaign=huggingface to get started.
Log in or Sign Up to review the conditions and access this model content.
A full-duplex turn detector. It reads the user's microphone and the agent's output as two aligned channels and estimates, at 50 Hz, which side holds the conversational floor.
A 500 ms silence looks the same whether the speaker has finished, is thinking, or took a breath. Detectors that wait it out have to guess, and they either cut people off or leave dead air.
Echo Omni Duplex does not wait. It listens to the user and the agent continuously, on separate channels, and predicts who is about to speak next. That is how it separates a finished sentence from a breath, and catches a barge-in while the words are still arriving. At 5.8 million parameters it ranks 3rd of 26 on end-of-turn and 2nd of 26 on interruption on the TurnBench test leaderboard.
Model at a glance
| Task | Conversational turn-taking: end-of-turn and interruption detection |
| Parameters | 5.8M |
| Input | Two continuous audio streams. Channel 1 the user's microphone, channel 2 the agent's output. Never mixed down. |
| Output | Floor probability per side, committed END_OF_TURN and INTERRUPTION events, confidence |
| Frame rate | 50 Hz (20 ms) |
| Context | Bounded window of recent past, constant cost per frame |
| Transcript | Not used. Audio only, no speech recognition, no language model. |
| Compute | 5 to 10 ms per inference on a single NVIDIA L4 |
| Languages | English (en) |
| Access | Served through Zero Runtime. Weights are not distributed. |
Why two channels
Two failure modes motivate the design, and a detector listening to the caller alone cannot separate them: a short sound that means "keep going" and a short sound that means "stop".
The claim is measurable before any benchmark. On public recordings of two-person conversations, two seconds after a speaker goes quiet, a two-channel estimate of who holds the floor separates a real turn end from a mid-turn pause at an AUC of 0.943. The best signal available from the speaker's own channel reaches 0.837.
Why channel 2 matters. The strongest cue for "has this person finished?" sits on the other side of the conversation, not the speaker's own. In a voice agent that side is the agent, whose audio is known exactly, so the model receives it as a clean input instead of inferring it.
What it decides
| Situation | Event | The agent should |
|---|---|---|
| The user has finished their turn | END_OF_TURN |
Answer |
| The user takes the turn while the agent speaks | INTERRUPTION |
Stop speaking |
| The user pauses mid-turn | none | Keep listening |
| The user acknowledges while the agent speaks | none | Keep speaking |
Each event has a look-alike that calls for the opposite response. A mid-turn pause looks like an ending, so silence-based detectors fire on it. An acknowledgement looks like an interruption, so they stop mid-sentence. Telling these pairs apart is what the second channel is for.
The floor estimate across those four moments. The commit policy reads this trace to decide when to fire. Values are illustrative.
How it is built
Echo Omni Duplex is one network in two stages. A pretrained front end, frozen during training, encodes each channel separately, so the user's audio and the agent's are never summed into one signal. A causal model then combines the two encodings, reading only a bounded window of recent audio and never anything ahead of the current moment. Both channels share the same weights, because the question is the same whichever side is speaking.
It predicts who is about to be speaking rather than classifying the present, which is what lets it commit before a silence has run long enough to be conclusive on its own. Cost per frame stays constant however long the call runs. A deterministic commit policy, separate from the network and tunable without retraining, turns the estimate into end-of-turn and interruption events. The whole system is small enough to run alongside speech-to-text, an LLM and text-to-speech on a single GPU.
It was trained on an internally curated corpus of two-channel, full-duplex English conversations. The composition and the hours of data used are internal and not disclosed.
Results
Echo Omni Duplex ranks 3rd of 26 on end-of-turn and 2nd of 26 on interruption on the TurnBench test leaderboard. Predictions were submitted on 19 October 2026 with the operating point fixed beforehand.
| Task | Rank | Recall β | False-positive rate β | Latency p50 β |
|---|---|---|---|---|
| End-of-turn | 3rd of 26 | 0.970 | 0.072 | 768 ms |
| Interruption | 2nd of 26 | 0.984 | 0.106 | 604 ms |
On interruption it ties the first-placed system on recall and commits 143 ms sooner. It beats Voice Activity Projection, the strongest published baseline, on recall, false-positive rate and latency together, by roughly 300 ms on latency. (One published system commits interruptions sooner still, at materially lower recall.)
Broken out by what the agent should have done, against our previous model. The two "keep" rows are one minus the false-positive rate; the other two are recall.
| Situation | Correct action | Echo Omni Duplex | Echo Omni |
|---|---|---|---|
| The user pauses mid-sentence | Keep listening | 92.8% | 92.9% |
| The user finishes speaking | Respond | 97.0% | 83.9% |
| The user says "mm-hm" while the agent is speaking | Keep speaking | 89.4% | 87.1% |
| The user interrupts | Stop speaking | 98.4% | 90.4% |
Latency is the benchmark's measure, not compute time. It is how long after the real event the decision was committed, and every reported timestamp carries a fixed 0.5 s offset so that it never precedes the moment of decision. The rest of the window is evidence the model would otherwise be deciding without. Compute is 5 to 10 ms per inference on an L4.
One operating point. It was chosen on the public dev split only, as the highest recall inside a 0.10 false-positive rate, and applied unchanged to the test set, which was never scored beforehand. On dev it scored 0.965 / 0.103 on end-of-turn and 0.988 / 0.098 on interruption. Splitting the 38 dev conversations in half at random, choosing the setting on one half and scoring it on the other, cost a median 0.0052 recall, so the setting is not specific to the conversations that produced it.
A held-out check on other data. Krisp turn-taking-test-v1: 2,730 labelled turns (976 shift, 1,754 hold) across 55 episodes and 30 speakers, roughly 4 hours of English audio. It is single-channel, so the second channel is empty throughout, and the model still reaches 0.884 AUC and 0.816 balanced accuracy.
Ranks are positions on a live leaderboard and can move as new systems are submitted. Recall and false-positive rates are fixed properties of the submitted predictions.
Causality
The output at any moment uses audio up to that moment and nothing after it. The model reads only a bounded window of the recent past, the voice-activity gate is causal, and the input gain uses only audio already heard.
Construction arguments are easy to get wrong, so we tested it directly. We took 38 real two-channel conversations, cut each at random points, re-ran the whole system from scratch on the shortened audio and compared the result against the run on the full recording. If a system were using audio from later in the file, removing that audio would change its earlier decisions. It does not.
| Check | Result | What it means |
|---|---|---|
| Random cuts | 84, across 18 conversations | how many times the pipeline was re-run on shortened audio |
| Events compared | 19,828 | every event early enough to be checked, summed over all cuts |
| Event lists identical | 330 of 336 | each cut produces 4 lists, one per speaker per task |
| Voice activity before each cut | identical | zero frames differed |
The Echo family
Echo Omni Duplex is the full-duplex member. The other three decide one turn at a time, which suits a half-duplex pipeline. Duplex decides continuously, which is what an agent needs when it is listening and speaking at once.
| Model | Modality | Decides | Best for |
|---|---|---|---|
echo-small |
Semantic | Per turn | The lowest latency, when responsiveness matters most |
echo-large |
Semantic | Per turn | Higher accuracy, when it matters more than raw speed |
echo-omni |
Audio and semantic | Per turn | Acoustic understanding on top of the semantics |
echo-omni-duplex |
Two-channel audio | Continuously | Full-duplex agents that listen while they speak |
Get access
Echo Omni Duplex is served by Zero Runtime. There is nothing to download or host.
# Set ZERORUNTIME_AUTH_TOKEN in your environment.
from zeroruntime.inference import TurnDetector
# Real-time end of turn + interruptions, English
turn_detector = TurnDetector(model="echo-omni-duplex", responsiveness="balanced")
responsiveness sets how quickly it commits:
| Value | Behaviour |
|---|---|
"relaxed" |
Waits for the clearest end of turn. |
"balanced" |
The default. Suits most conversations. |
"quick" |
Replies and yields soonest, for fast back-and-forth. |
Stream your agent's output as the second channel, keep your existing STT, LLM and TTS, and compare its end-of-turn and interruption events against your current endpointing on your own calls.
Echo Omni Duplex by Zero Runtime
Collection including zero-runtime/echo-omni-duplex
Evaluation results
- Recall on TurnBenchtest set TurnBench leaderboard0.970
- False-positive rate on TurnBenchtest set TurnBench leaderboard0.072
- Recall on TurnBenchtest set TurnBench leaderboard0.984
- False-positive rate on TurnBenchtest set TurnBench leaderboard0.106







