Text-to-Speech Agent Barge-In 2026: LiveKit Turn Detector

Your voice agent finally hits sub-500ms TTFT, and then a caller says “wait—” mid-sentence and everything falls apart. The agent talks over them for another four seconds. Or the opposite: it clips itself every time someone says “mhm,” because a 700ms silence threshold cannot tell a backchannel from a completed thought. Meanwhile a TV in the background triggers a barge-in, the drive-thru mic picks up the car behind, and your echo path feeds the agent’s own TTS back into VAD as user speech. In 2026 the endpointing problem is no longer “detect silence” — it is deciding, in real time, whether a human is done, thinking, or genuinely interrupting, and the naive threshold that shipped your demo is the single biggest reason your production call quality scores are stuck.

This is for developers building production voice agents — you already ship Python or Node services, you understand streaming audio, WebRTC as a concept, and async event loops, and you have at least run a LiveKit Agents hello-world. You should be comfortable reading SDK source and reasoning about latency budgets. Out of scope: telephony carrier provisioning, training your own VAD or turn-detection models, TTS voice cloning, general prompt engineering for LLMs, and non-LiveKit runtimes except where covered as migration alternatives.

Honest framing: semantic turn detection is genuinely good at what silence thresholds cannot do — recognizing that “my account number is four seven two…” is unfinished even after a long pause, and that “yeah” is not a turn grab. It is unreliable with heavy accents outside its training distribution, code-switching mid-sentence, low-bitrate telephony codecs, and any environment where the noise floor moves. It will also confidently mis-endpoint on domain jargon and spelled-out identifiers. Human review is non-negotiable on two things: listening to real recorded calls from your own traffic before you tune a single parameter, and reviewing every cut-off and false-interrupt case your eval harness flags — the metrics tell you the rate, only your ears tell you whether the failures are tolerable for your use case.

What This Guide Covers

• Why interruption handling broke the previous generation of voice agents, and what changed in the 2026 stack that makes it solvable
• A precise mental model of the turn lifecycle — VAD, streaming STT, endpointing, and the interruption loop — so you can reason about which layer is actually failing
• The tradeoff between semantic end-of-utterance models and naive silence thresholds, including when the simpler approach is still the right call
• A working barge-in agent built on LiveKit Agents 1.x, from project scaffold to a call you can interrupt cleanly
• How to tune Silero VAD sensitivity and endpointing delay windows for your acoustic environment instead of copying defaults from a blog post
• Wiring modern streaming STT — Deepgram Flux and AssemblyAI Universal-Streaming — and what each gives you for turn detection specifically
• Practical defenses against false barge-in from television audio, bystander speech, and conversational backchannels
• Getting echo cancellation right so your agent stops interrupting itself through the acoustic feedback loop
• Preemptive generation for latency wins, plus clean rollback when the interruption turns out to be real
• An eval harness you can run in CI, measuring EOU latency, cut-off rate, and unnecessary interrupt rate against your own audio
• A full latency and cost budget across the stack, showing where your milliseconds and dollars actually go
• Twelve specific failure modes with the diagnostic signal for each and the fix that resolves it
• Case studies from three hostile environments — a support line, a drive-thru, and multilingual outbound calling
• Where the field is heading: Pipecat Smart-Turn, full-duplex models, and a migration checklist for evaluating alternatives without rewriting everything

Delivered as instant online access the moment checkout completes — no waiting on an email, no upsell, no subscription, no follow-on course pitch. You get the complete guide.