RealSub
RealSub FAQ
English · 简体中文 · 繁體中文 · 日本語
App 1.0.1 / last updated 2026-10-04. Free tiers, model names and endpoints of third-party
services are the vendors’ to change; their own documentation is authoritative.
Q1. Does translation work from mainland China?
- The default “Local” translation runs on your own PC and never touches the network, so it works in mainland China as is (just install the free local translation model DLC).
- “Microsoft Translator” also works from mainland China (it is used automatically when the local translation model is not installed). We tested the full request flow from a mainland node on 2026-08-13 and got translations back in about 1.2 seconds. Nothing extra to configure.
- Do not pick “Google Translate” in mainland China — it times out there in our tests and will simply keep failing.
- Advanced users can pick “Local / LAN server”: enter the address of an OpenAI-compatible translation server running on your own PC or LAN (Ollama, LM Studio, llama-server and the like). No account or API key is involved. Only local and LAN addresses are accepted; public internet addresses are rejected — RealSub does not integrate any third-party service that requires registration or payment.
- If a translation source fails, RealSub switches to Microsoft Translator automatically; transcription is never blocked. Recognition is fully offline — subtitles keep coming even with no network at all.
Q2. Why are no subtitles appearing?
Check these in order of likelihood:
- What is playing is singing, pure music, or a voice with radio/walkie-talkie processing. Those do not produce subtitles at the moment.
- The wrong audio source is selected. If you picked a specific app under Settings → Transcription → Audio source, only that app is transcribed. Switch back to “Entire system (default)” to capture everything.
- The app you want is not in the list. Windows only creates an audio session for an app once it has actually made a sound — play something first, then refresh the list. Per-app capture also requires Windows 10 version 2004 or newer.
- System volume is at zero, the app is muted, or the volume is very low. There is no audio to capture while muted. Very quiet audio is boosted automatically (up to 20x) and a one-time notice appears; for the best recognition quality, raise the Windows volume or the playback app’s volume. You can also open the output device’s additional properties in Windows sound settings (Windows 11: “More sound settings” → the device → Properties) and enable “Loudness Equalization” on the Enhancements tab to even out the output level.
- Transcription is not running. Check whether the tray menu shows Start or Stop, or toggle it with
Ctrl+Alt+S.
- The subtitle window is hidden or off-screen. Toggle it with
Ctrl+Alt+H, or use “Reset window position” under Settings → Hotkeys & More → Window (the tray menu has the same entry).
- Only the translated line is missing. The free version does not include translation; the recognized text still appears normally. Translation is unlocked by the Full Version DLC.
If none of that helps, use “Export diagnostics…” in the tray menu and send the resulting zip with a short description to our support address.
Q3. Should I buy the Full Version DLC?
Use the free version first, then decide. That is a real recommendation, not a formality:
- The base app is free and gives you complete live transcription, with no usage counter. Installing it answers the three questions that actually matter: is recognition accurate enough for your content, does it run smoothly on your GPU, and do you like the overlay format.
- A 7-day full-featured trial (once per Steam account) also starts the first time you use RealSub, so translation and history export are available during that week.
- When the trial ends, the app switches to the free version automatically.
- Only once you are satisfied does buying the DLC make sense. That order removes almost every “bought it, then found out it was not for me” refund.
If you have already decided to refund, that is completely fine. We would just ask you to spend a few minutes with the free version first — several common surprises (no subtitles for singing, some GPUs may not get acceleration) are stated openly in the “what it does not do” section on the store page.
Q4. What GPU do I need? What if I don’t have an NVIDIA GPU?
- NVIDIA GTX 10-series or newer, from 2 GB of VRAM; 4 GB or more is more comfortable. The default recognition model needs a little over 2 GB, so a card with exactly 2 GB can get tight when other programs are also using VRAM. If the model fails to load on the GPU — not enough VRAM, driver trouble — RealSub switches to CPU mode automatically, so there is nothing to set up in advance.
- We strongly recommend an NVIDIA GPU — that is where RealSub performs at its best. Other GPUs go through Vulkan on a best-effort basis: the first run does a short performance test (tens of seconds) and only enables it if it passes. We have verified it on some GPU models, but cannot promise every card. If the test fails or Vulkan errors at runtime, RealSub falls back automatically, shows a tray notification, rewrites the device setting to what is actually running, and you can run “Re-test GPU performance” in Settings any time.
- Four device options: “Auto” / “GPU (NVIDIA)” / “AMD / Intel GPU (Vulkan)” / “CPU”. Auto prefers NVIDIA, then tries Vulkan, then CPU. Vulkan is best-effort, so some GPUs may not get acceleration.
- Without an NVIDIA GPU and without usable Vulkan acceleration — or if the GPU fails to initialize or the model fails to load — RealSub automatically switches to CPU mode (and tells you via a tray balloon and the subtitle status line). CPU mode uses a lighter, faster recognition engine (SenseVoice): subtitles keep up, and the in-progress preview line and sentence splitting still work, but recognition accuracy is lower than on an NVIDIA GPU. Please try it in the free version on your own PC before buying the DLC.
- The hard CPU requirement is a 4-core processor from the last decade with AVX2 support.
- You can pin the compute device to “Auto” / “GPU (NVIDIA)” / “AMD / Intel GPU (Vulkan)” / “CPU” under Settings → Transcription, and run the Vulkan check again with “Re-test GPU performance”.
Q5. Does it need the internet? Is my audio uploaded?
- Audio is processed entirely on your machine and never leaves it. The recognition model (Whisper) and the voice-activity model (Silero VAD) are bundled with the app; recognition needs no network at all.
- The default local translation stays offline too: translation happens on your PC and the subtitle text never leaves it. Only when an online translation service such as Microsoft Translator or Google Translate is in use is the recognized subtitle text (text only) sent to that service to get a translation back.
- No telemetry, no account, no usage statistics, no automatic crash reporting. A diagnostics zip is only created when you click Export yourself, and it stays local; API keys, your Windows user name and file paths inside it are masked automatically.
- The full details are in the privacy policy bundled with the app and linked from the store page.
Q6. Which languages are supported?
- Speech that can be recognized: Japanese, Chinese (Simplified), Chinese (Traditional), English. You can also choose Auto and let the app decide (detection is sticky, so it does not flip back and forth; Auto writes Chinese in Simplified characters — pick “Chinese (Traditional)” explicitly if you want Traditional).
- Translation targets: English, Simplified Chinese, Traditional Chinese, Japanese.
- Interface languages: English, Japanese, Simplified Chinese, Traditional Chinese, following your system language by default (Traditional for Taiwan / Hong Kong / Macau systems).
- The three settings are independent — you can run an English interface, recognize Japanese, and translate into English.
- The recognition and translation models could in theory cover more language combinations; to keep quality up, only the languages we have tested thoroughly are offered for now. The settings pages inside the app are the authoritative list.
Q7. What happens when the 7-day trial ends?
The app switches to the free version automatically:
- Kept: complete live transcription (recognition languages, bundled models, overlay appearance, hotkeys, per-app capture) plus read access to history you already recorded.
- Locked: translation (the translated line), saving new history and recordings, SRT / TXT export.
- Buying the Full Version DLC restores all of it immediately — no reinstall and no restart needed. If it does not apply right away, use “Purchased? Refresh” in the purchase dialog.
- The trial is tied to your Steam account: reinstalling the app, wiping its data, or switching to another PC does not grant a new trial.
Q8. Can I use the exported subtitles directly on a video?
Yes — you can export SRT subtitles or plain TXT (Full Version required). Note that:
- Timestamps are anchored to when you listened, not to the video file’s own timeline. If you played the file from the start without pausing or seeking, the two line up closely; otherwise you will need to shift them.
- Merging of short segments and de-duplication can shift individual cue boundaries by a second or two.
- The text is machine-recognized and machine-translated. Proofread before publishing anything.
Q9. Why does the translation always lag behind the original text?
With the default local translation, translations appear almost together with the original. The rest of this answer applies when an online translation service such as Microsoft Translator is in use.
That is how sentence-level translation works — it is not a network fault. The original text streams in while the speaker is still talking; the translation is only requested once a sentence boundary is found, and the result takes another second or two to come back.
- Normal dialogue: sentences end in natural pauses, so the translation usually trails by only 2–3 seconds.
- Near-continuous speech (commentary, lectures, live streams, fast talkers): when no pause can be found, the app waits up to about 10 seconds before force-cutting a sentence, so the translation lags noticeably.
The original (white/grey) line stays real-time and is not affected. Only if translations stop appearing entirely or go missing frequently is the translation service itself likely unreachable — try a different source under Settings → Translation.
Q10. The subtitles disappear as soon as the player goes fullscreen?
Once a second the subtitle window checks whether another window is covering it, and only then re-asserts its always-on-top status. With ordinary fullscreen playback (PotPlayer, VLC, mpv, browser fullscreen, apps like Netflix) it comes back on top within a second.
If it still does not show, the player is most likely using exclusive fullscreen (PotPlayer’s “Direct3D exclusive mode”, madVR’s fullscreen exclusive mode, mpv’s --d3d11-exclusive-fs, etc.). In exclusive fullscreen nothing can be drawn over the video — not even Windows’ own volume pop-up. If you change the volume and see no system volume bar, that is the case. Turn off exclusive mode in the player’s settings, or use borderless windowed fullscreen instead.
There is one more case, and it only happens on Windows 10: if a Store app such as the built-in Movies & TV is already fullscreen the moment it starts, Windows puts that fullscreen view on a higher display layer than ordinary always-on-top windows, and the subtitle window cannot reach it. Leave fullscreen and enter it once more — from then on the subtitles stay on top.
Q11. How do I fill in “Local / LAN server”?
This option is for people who already run an OpenAI-compatible translation server on their own PC or LAN. No account or key is involved. Only local and LAN addresses are accepted (localhost, 127.x, 10.x, 192.168.x, 172.16-31.x); public internet addresses are rejected.
| Server |
API URL (up to /v1) |
Model name |
Notes |
| Ollama |
http://127.0.0.1:11434/v1 |
a model you have pulled, e.g. qwen2.5:7b |
required; a wrong name returns “model not found” |
| LM Studio |
http://127.0.0.1:1234/v1 |
the identifier shown in LM Studio, or leave empty |
empty = the currently loaded model; start the local server in LM Studio first |
| llama-server (llama.cpp) |
http://127.0.0.1:8080/v1 |
anything, or empty |
serves one model; the name is ignored |
| another machine on your LAN |
http://192.168.1.20:11434/v1 |
as above |
replace 127.0.0.1 with that machine’s LAN IP and make sure the server listens on 0.0.0.0 |
To verify: after entering the address and model name, play a video with speech; the translated line appearing means it works. If the server is unreachable or the model name is wrong, the subtitle bar reports a translation-source failure and falls back to Microsoft Translator automatically. Quality depends on the model you pick; general 7B+ models are usually usable for Japanese → Chinese, smaller ones are noticeably stiff.
Q12. It sits on “Loading…” for a long time after launch. Is it frozen?
No. To keep transcription fully offline, RealSub ships the speech model and the GPU runtime libraries locally, and every cold start (the first launch after booting) reads about 4 GB from disk:
- Internal SSD: usually under half a minute
- Internal hard disk drive: can take around a minute
- USB external drive / flash drive: can take several minutes (over 2 minutes measured on one external drive), not recommended
While loading, the subtitle bar shows the current stage and the seconds waited; as long as the number keeps ticking, everything is fine. Launching again in the same Windows session is much faster because the files are already cached in memory.
To speed it up, keep your Steam library on an internal SSD. Steam’s “Settings → Storage” can move installed content without downloading it again.
Q13. After updating, the first launch was slow / it re-tested my GPU / why is my GPU not being used?
Updating to 1.0.1 changes three things:
- The first launch re-runs the GPU performance test, once. The test changed in this version, so the result saved by the old version no longer applies, and that one launch takes roughly ten seconds to under a minute longer. Later launches are back to normal. The new test is also stricter: weaker AMD / Intel graphics and integrated GPUs that only just passed before may now be placed in CPU mode, with a “Switched to CPU mode” notification and the device setting changed to “CPU”. That is intended, not a fault — on those PCs the CPU engine keeps up better than the GPU did (see Q4). NVIDIA GPUs are not affected.
- GPUs that an older version stopped using after a single error get another chance automatically. Older versions could give up on a GPU for good after one error at runtime; those records no longer apply. (From 1.0.1 on, errors at runtime only switch to CPU for the current session; restarting RealSub goes back to the GPU.)
- A device setting of “CPU” is not changed back for you. If an older version switched your device to “CPU” (you would have seen a “Switched to CPU mode” notification at the time), the update leaves it alone — it is part of your settings. To give your GPU another chance, open Settings → Transcription and set Device back to “Auto”. If the GPU passes the test it is used; if not, RealSub simply returns to CPU mode and tells you.