RealSub

Turn audio you cannot follow into subtitles you can read

RealSub is a live subtitle tool for Windows. It captures whatever your PC is playing — or the sound of one specific app you pick — recognizes the speech locally, optionally translates it, and shows the result in a floating subtitle bar on top of everything else. Japanese video with no subtitles, foreign-language streams, lectures, podcasts, interviews, game dialogue: if the sound comes out of this PC, it can be subtitled. Including the places a player plugin can never reach.

Hardware first, because the bar is lower than you expect

Starts at 2 GB of VRAM on a GTX 10-series card (4 GB or more is more comfortable. If the GPU fails to initialize or the model fails to load, RealSub switches to CPU mode by itself — nothing for you to configure)

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.

No NVIDIA GPU? It still works — RealSub tries Vulkan first; if Vulkan is unavailable, does not pass the test, or the model fails to load, RealSub switches to CPU mode with a smaller recognition model. Accuracy is lower than on a GPU and latency grows noticeably, but you get subtitles.

Models ship inside the app — nothing to download, no account, no API key to get started

What the free app does, and what the DLC unlocks

Free base app : complete live transcription. Recognition languages, the bundled models, overlay appearance, global hotkeys, per-app audio capture — none of it is locked, and there is no usage counter.

Full Version DLC : unlocks translation (bilingual subtitles) and session history + SRT / TXT export .

A 7-day full-featured trial starts the first time you use RealSub (once per Steam account; reinstalling does not reset it). When the trial ends, the app switches to the free version described above automatically.

Features

Subtitle one app, not your whole desktop — per-process audio capture (Windows 10 2004+). Pick your browser or media player and only its audio is transcribed; system notification sounds and other windows stay out. Whole-system capture is there too.

Recognizes Japanese, Chinese (Simplified or Traditional) and English , or detects the language for you (with sticky switching, so it does not flip back and forth)

Translates into English, Simplified Chinese, Traditional Chinese or Japanese ( Full version ) — local translation by default: add the free Local Translation Model DLC and translation runs right on your PC, so translations appear almost the moment the words do, even offline; without it, Microsoft Translator is used automatically, also with zero setup. You can also switch to Google Translate (not reachable on mainland-China networks) or a translation server running on your own PC or LAN (no account or key needed)

An overlay built for watching — one line of source text, one line for the in-progress guess, one line of translation. Clicks pass straight through so it never blocks you; hover to reveal a frame you can drag and resize. Font, size, color, opacity and line counts are all configurable.

History and export ( Full version ) — transcripts are saved per session and can be reopened any time, then exported as SRT subtitles or plain TXT. Optionally record what you heard as an MP3 (about 20 MB per hour). The free version can still browse history you already have; it just stops saving new sessions and cannot export.

English / Japanese / Simplified Chinese / Traditional Chinese interface , following your system language by default

Global hotkeys and a tray icon — show/hide subtitles and start/stop transcription with one key; the tray menu shows live RAM and VRAM usage

Two-step first-run wizard — checks your hardware, picks your languages, and your first subtitle line appears soon after

What it does not do (better to know before you install)

Singing, pure music and radio-effect voices do not produce subtitles at the moment. Anime openings, songs, and dialogue deliberately processed to sound like a radio broadcast will not be transcribed.

Four device options: "Auto" / "GPU (NVIDIA)" / "AMD / Intel GPU (Vulkan)" / "CPU" . Auto prefers NVIDIA, then tries Vulkan, then CPU. Vulkan is best-effort: some GPUs get no acceleration and fall back to CPU mode, where accuracy is lower than on a GPU and latency clearly higher. Usable, not comfortable.

Recognition is fully offline, and so is the default local translation (the Local Translation Model DLC is not offered in the EU, the UK or South Korea). Online translation services such as Microsoft Translator or Google Translate need an internet connection; while one is in use, the recognized subtitle text is sent to the service you chose. Audio is never sent anywhere.

With online translation services such as Microsoft Translator, translations run one beat behind the original. The original text appears while the speaker is still talking, but translation works sentence by sentence and only shows up once a sentence is complete. In normal dialogue that is a delay of a few seconds; with near-continuous speech (commentary, lectures, live streams) the app may wait up to about 10 seconds before force-cutting a sentence, so the translation trails noticeably. That is how sentence-level translation works, not a network or performance problem.

It captures what your PC plays, not your microphone. Apps that have not made a sound yet do not appear in the audio-source list.

Subtitles are machine-recognized and machine-translated, so they contain errors. They are an aid to understanding, not a record of what was said.

Only Japanese, Chinese (Simplified or Traditional) and English speech can be recognized right now , and translation is limited to English, Chinese (Simplified and Traditional) and Japanese. 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.

Windows only (10 version 2004 or newer, 64-bit). There is no macOS or Linux build.

Privacy

Audio is processed entirely on your machine and never uploaded or stored on any server. No telemetry, no account, no usage statistics. The default local translation also runs on your machine. Only the subtitle text is ever sent out, only while an online translation service such as Microsoft Translator is in use, and only to the service you selected. The full privacy policy is linked from this store page.

Footage shown in screenshots and video is licensed material.