The Hermes barge-in feature lets you interrupt the agent at any point in its voice turn — speak over it while the reply is playing and the microphone catches you, the text-to-speech playback cuts off instantly, and the model stops generating mid-sentence. No button, no wake word, no waiting politely for it to finish a paragraph you already know is wrong. You just talk, the same way you would cut across a colleague heading down the wrong track.
📺 Watch: Hermes Agent: New Voice Activation Update is INSANE!
🔥 Get the Agent OS as a free bonus: AI Profit Boardroom members get the full Agent OS zip, prompt libraries, daily tutorials and weekly live coaching calls. → Get inside
I have been living in voice mode since the Herald release dropped, and barge-in is the feature that turned it from a novelty into something I actually work with. Here is what it does, how it works under the hood, the one known issue worth understanding, and where it sits in my wider voice stack.
What Barge-In Actually Promises
The official Voice Mode documentation puts it in one line: "You can interrupt the agent at ANY point in its turn — the microphone stays live from the moment you finish speaking until the reply has fully played." That is the entire contract. From the second you stop talking until the second the reply finishes playing, the microphone is hot and listening for you.
In practice, that means when the agent launches into a ninety-second spoken answer and you realise ten seconds in that it has misread the question, you do not sit through the rest. You speak, playback stops, and the conversation picks up from your interruption. It is the voice behaviour every phone assistant has promised for a decade and almost none have delivered properly.
How the Hermes Barge-In Feature Works Under the Hood
Per the docs and the project's own technical write-ups, two things happen the moment you speak over a reply:
- Playback cuts off. The instant the microphone detects your voice, the text-to-speech audio stops. No fade-out, no trailing sentence.
- Generation stops too. Because voice runs in-process, the interruption triggers agent.interrupt(), which stops the model generating mid-sentence — not just mid-playback.
That second point is the one that matters. External voice wrappers typically just mute the audio while the underlying model keeps generating in the background — the wrapper goes quiet, but the model carries on talking to nobody, burning tokens on a reply you have already rejected. Hermes actually stops thinking when you interrupt it.
And because generation halts at the exact moment you cut in, the agent tracks precisely which part of its sentence you actually heard before the interruption. Its memory stays in sync with the live conversation: it knows what it said out loud versus what it never got to say, so a follow-up answer will not lean on an explanation you never heard. The write-ups describe adaptive turn-taking as part of the design, and that heard-versus-unheard tracking is what makes the voice experience feel adaptive rather than mechanical.
Building voice agents into a real business rather than a weekend demo? That is exactly what we do inside AI Profit Boardroom — get my Hermes voice agent workflows here. Or if you would rather map your setup one-to-one, book a free SEO strategy session and we will plan it live.
Why It Hears You Over Its Own Voice
The obvious objection: if the microphone stays live while the agent talks through your speakers, why does its own audio not drown out your interruption — or worse, trigger one? The documented answer is noise-floor calibration. The local voice activity detection calibrates against the room's noise floor at the start of a turn, so speaker bleed from the agent's own voice becomes part of the expected background rather than a blocker. Your interjection registers as a fresh signal on top of that floor, and barge-in fires. It is the piece of engineering that makes open-speaker voice conversation viable at all.
📺 Watch: Hermes3D Is INSANE: AI Agents Now Have a 3D Office
The Herald Release: Where Barge-In Shipped
Barge-in arrived in Hermes Agent v0.20.0 — the Herald release, August 2026 — alongside real-time conversational voice and on-device wake words. According to the release coverage and the project's own posts, the Herald name comes directly from the voice features: streaming text-to-speech with barge-in and wake words. Voice was not a footnote in that release; it was the headline act. The same version also introduced the Agent-to-Agent protocol, which is a story of its own — my v0.20 breakdown covers the full changelog, and my Herald release page goes deeper on why the update mattered.
The Known Issue: The TTS Feedback Loop
Honest coverage means covering the wart. There is a public GitHub issue reporting a regression where CLI continuous voice mode can transcribe Hermes' own text-to-speech output as a user utterance — the agent hears itself through the microphone, treats its own reply as your input, and drops into a TTS-to-microphone-to-STT feedback loop. If you run continuous voice on open speakers, that is worth knowing before it happens to you mid-demo.
The noise-floor calibration described above is the built-in defence. My own practical advice is simpler: wear headphones. That is common sense rather than anything from the documentation — with no speaker bleed there is nothing for the microphone to mis-transcribe, the loop cannot start, and interruption detection gets cleaner as a bonus.
📺 Watch: Hermes Bot Mode DESTROYS Grok Bot? (FREE!)
Barge-In vs Text Interruption: The /busy Parallel
If you use Hermes in text, you have already met this idea wearing different clothes. The /busy setting controls what happens when you message the agent mid-task — queue the message, interrupt the task, and so on. Barge-in is the voice-world equivalent of the interrupt option: your voice is the message, and it always wins. I have broken down the text side in my /busy command guide, and the two are worth reading together, because they are the same philosophy — you outrank the agent's current turn — applied to two different channels.
Community Builds: LiveKit, Pipecat and WebRTC
Beyond the native experience, community builds wire Hermes voice into LiveKit and Pipecat stacks, including browser-microphone setups running over WebRTC. I am deliberately not giving setup steps here because those integrations move quickly, but the pattern behind them is consistent: the barge-in behaviour is a big reason builders pick the native voice mode over generic wrappers when interruption quality decides whether the product feels human. If that is the road you are on, start with my voice agent guide.
Where Barge-In Fits in My Voice Stack
This page is one piece of my Hermes voice cluster. The voice mode overview gives you the full feature picture, the voice activation update covers how turns start (barge-in covers how they end early), and the Jarvis-style build shows the always-on assistant this all adds up to — barge-in being the thing that stops always-on becoming never-shuts-up. Voice is also one of the main surfaces I run through Agent OS, and if you are weighing up which model brain should sit behind a voice agent, my Goldie Bench testing is where I compare the candidates.
Barge-In Behaviour at a Glance
| Behaviour | What happens | Source |
|---|---|---|
| Interrupt at any point | Microphone stays live from the end of your speech until the reply has fully played | Official Voice Mode docs |
| Speak over playback | Text-to-speech cuts off the instant the microphone detects your voice | Docs and project write-ups |
| Generation stop | agent.interrupt() halts the model mid-sentence — no background generation continues | Project technical write-ups |
| Heard-vs-unheard tracking | The agent records which part of its reply you heard before interrupting | Project technical write-ups |
| Noise-floor calibration | Voice activity detection calibrates to the room at turn start, so speaker bleed does not block you | Official Voice Mode docs |
| Feedback-loop regression | Continuous CLI voice can transcribe its own TTS as user input on open speakers | Public GitHub issue |
| Release | Shipped in v0.20.0 (Herald), August 2026, with wake words and real-time voice | Release coverage and project posts |
Hermes Barge-In FAQ
What is Hermes barge-in?
It is the voice mode behaviour that lets you interrupt the agent by simply speaking over its reply. The microphone detects your voice, playback stops instantly, and agent.interrupt() stops the model generating mid-sentence — a genuine interruption, not a mute button.
Can I really interrupt at any point?
Yes. The official documentation says you can interrupt the agent at any point in its turn, with the microphone staying live "from the moment you finish speaking until the reply has fully played". There is no dead zone in the voice turn where you are locked out.
Does the model keep generating after I interrupt?
No. That is the difference between Hermes and external voice wrappers, which mute the audio while the model keeps generating underneath. In Hermes, the in-process architecture triggers agent.interrupt() and generation stops on the spot — which is also how it knows exactly what you heard.
Why does it hear me over its own voice?
The local voice activity detection calibrates against your room's noise floor at the start of each turn, so the agent's own speaker output does not mask your interruption. On open speakers that calibration is doing real work; on headphones the problem disappears entirely.
Which release added barge-in?
Hermes Agent v0.20.0 — the Herald release, August 2026 — shipped real-time conversational voice, on-device wake words and barge-in together, and the project has said the Herald name comes from those voice features.
My Verdict on Barge-In
Barge-in is the feature that decides whether a voice agent feels like a tool or a teammate, and Hermes has built it properly: playback stops, generation stops, and the agent knows exactly what you heard. Most of the market still mutes and pretends. Until the feedback-loop issue is fully closed, run headphones for continuous voice — beyond that, this is the most natural interruption behaviour I have used on any voice agent, and it is the reason voice mode stayed in my daily workflow instead of becoming a demo I show people once.
Want the exact voice-agent stack I run — barge-in, wake words and all — plus a community stress-testing it every week? Join me inside AI Profit Boardroom. And for a personal roadmap on AI and SEO growth, book a free SEO strategy session — it costs nothing and we will build the plan together.
