Live Captions are supposed to track speech almost in real time — but yours trail the audio by several seconds, making videos and calls feel dubbed by a distracted translator. On a Copilot+ PC this lag is a performance problem, not a feature limitation: the caption pipeline is designed to run on the NPU at near-zero latency, so multi-second delays mean the fast path is broken.

The most common culprit is silent CPU fallback. When the NPU driver stalls, Windows keeps captions alive on the processor, which is dramatically slower at speech recognition. This guide shows you how to confirm the fallback, restore the NPU path, and trim the other sources of lag. If captions are also failing on microphone input, see our microphone captions fix — the two share audio plumbing.

Measure the delay against a news clip, apply the fixes in order, and watch the gap shrink.

Why this happens

  • NPU driver stall forces CPU fallback: Speech recognition on the CPU is several times slower than on the NPU.
  • Translation doubles the workload: Same-language captions are fast; adding translation stacks a second model in the chain.
  • System under heavy load: Gaming, compiling, or 50 browser tabs starve the caption process of CPU/NPU time.
  • Wrong language model: A mismatched speech pack makes recognition slower and less accurate.
  • Audio pipeline latency: Bluetooth audio and enhanced audio processing add delay before captions even start.
  • Outdated caption components: Older Windows builds had measurably slower on-device recognition.

How to fix it

Work through these in order — diagnose first, then act.

Fix 1: Restart to reset the NPU recognition path

The quickest test for CPU fallback is a clean restart. It resets the NPU driver, and if captions snap back into sync, you have your diagnosis.

  1. Close the Live Captions window completely.
  2. Restart the laptop (not sleep — a full restart).
  3. Reopen Live Captions and play the same test clip.
  4. If the delay shrinks dramatically, the NPU path was stalled — continue to Fix 5 to keep it healthy.

Fix 2: Turn off translation temporarily

Translation adds a whole second model to the chain. Testing without it isolates whether the lag is recognition or translation.

  1. In the Live Captions settings, set the caption language to match the audio (English to English).
  2. Replay the test clip and judge the delay.
  3. If same-language captions are snappy, translation was the bottleneck — use it only when needed.
  4. If delay persists without translation, the recognition path itself is slow (Fixes 3-5).

Fix 3: Reduce system load during captions

Speech recognition competes for NPU/CPU time. A busy system stretches caption latency noticeably.

  1. Close heavy apps (games, video editors, dozens of browser tabs) while testing.
  2. Open Task Manager and check overall CPU and NPU load during captions.
  3. Pause Recall snapshots and other background AI features during important captioned sessions.
  4. Re-test the delay with a quiet system — the difference is often dramatic.

Fix 4: Switch to wired audio output

Bluetooth adds its own latency before audio even reaches the caption engine. Removing it from the chain tightens sync.

  1. Disconnect Bluetooth headphones and switch to the laptop speakers or wired earbuds.
  2. Replay the test clip and compare the caption delay.
  3. If Bluetooth was the cause, keep wired audio for captioned content.
  4. Also disable audio enhancements (Settings > System > Sound > device properties), which add processing delay.

Fix 5: Update the NPU driver and Windows

The permanent fix for CPU fallback is a healthy NPU driver on a current Windows build. Caption performance improves measurably with updates.

  1. Download the latest NPU driver from your maker’s support page and install it.
  2. Install all pending Windows updates and restart.
  3. Verify the NPU appears in Task Manager > Performance with activity during captions.
  4. Confirm your hardware’s expected performance with our guide on checking NPU TOPS — if the silicon is fine, the driver was the issue. Related mic-side lag is covered in Live Captions microphone not working .

Fix 6: Reinstall the speech language pack

A corrupted language pack makes recognition slow and inaccurate. A fresh install restores full speed.

  1. Go to Settings > Time & language > Language & region, remove the speech language, and re-add it.
  2. Let the pack fully download before testing (check Settings > Time & language > Speech).
  3. Restart Live Captions and test the same clip.
  4. If delay persists across everything, report it via Feedback Hub — Microsoft tunes these models per build (general guidance on the Windows support hub ).

Closing

Caption delay on a Copilot+ PC is a processing-path problem: the NPU should deliver near-real-time captions, and anything slower means fallback, overload, or a bad model. Restart to reset the path, drop translation to isolate it, quiet the system, ditch Bluetooth for testing, update the NPU driver, and refresh the language pack. Get the fast path back and captions feel live again — because they are. If system-wide AI search also feels slow, see Windows Search AI indexing slow .

Frequently asked questions

Why are my Live Captions so far behind the audio?
Delay usually means the speech recognition is running on the CPU instead of the NPU — the CPU path is much slower. It can also be translation overhead (translating to another language adds processing), a busy system, or an overloaded audio pipeline. Restoring the NPU path is the biggest single fix.
How much delay is normal for Live Captions?
On a Copilot+ PC with a healthy NPU, expect well under a second of lag for same-language captions. Translated captions add roughly another second. If you are seeing three or more seconds consistently, something is wrong — usually the NPU fallback.
Does internet speed affect Live Captions delay?
No. Live Captions on Copilot+ PCs runs entirely on-device via the NPU — that is the privacy and speed advantage. If captions lag, your connection is not the cause; look at local processing instead.
Why is the delay worse when translating?
Translation adds a second AI model to the chain: speech-to-text, then text-to-text translation. Both should run on the NPU, but if either falls back to CPU, lag multiplies. Confirming your NPU’s capability with our guide on checking NPU TOPS tells you what the hardware should handle.