PC の音声を、その場で字幕に。Everything playing on your PC, subtitled as it plays.

First-run setup

Launch the app, accept the terms, then in "Set up models" go through "Check this PC" → "Download models" (several GB, a few minutes to a quarter of an hour). Optionally run "Measure speed".

Start and stop

Ctrl+Shift+T (Cmd+Shift+T on macOS) starts and stops translation. You can change the hotkey in Settings.

The overlay is hidden right after launch. It appears when translation starts and hides automatically after translation stops, leaving the last subtitle on screen briefly. You can also show or hide it manually, even while not translating, from "Show subtitles" in the tray menu or from the switch above the Start button in Settings.

Audio input: "System audio" captures whatever your PC is playing. Windows uses WASAPI loopback; on macOS 13+ all apps' audio is captured together via ScreenCaptureKit (no virtual audio device such as BlackHole is needed). The first time, allow OverlaySubs under System Settings → Privacy & Security → "Screen & System Audio Recording" (macOS 15+) or "Screen Recording" (macOS 13–14), then restart the app. Microphone input works on both platforms.

Tray icon and quitting

While the app runs, the system tray / menu bar icon offers "Start / Stop", "Show subtitles", "Settings" and "Quit". Closing the settings window quits the app.

What you can configure

• Audio input: system audio / microphone, and which device. Changes take effect after you stop and start translation again
• Languages: the spoken language (auto-detect or fixed), the target language, and the language to use when the speech is already in the target language. Changes made while translating apply from the next utterance
• Translation tuning: Japanese politeness (polite / plain), glossary (one "source=translation" per line), topic (a hint for the translator), use the previous utterance as context. Changes made while translating apply from the next utterance
• Subtitles: show the source text; background color and opacity for the overlay and for the caption box; font, size, color, weight and line height for the translation; size and color for the original text; caption alignment, vertical position and number of lines; and how long subtitles stay after updates stop. When the lines are taller than the overlay, the scrollbar on the right or the mouse wheel scrolls through them (the wheel moves half a line per notch; with click-through off). While the subtitles keep updating, older lines stay until the line count pushes them out. While translation is stopped, opening this page shows the overlay with a sample caption so you can check your changes live
• Click-through: off by default. When on, mouse input passes to the window underneath (you can't interact with the overlay itself while it's on)
• History: save recent subtitles as text / SRT
• Interface language: English / Japanese / Chinese (Traditional and Simplified)
• Advanced: GPU layers for the translation model (0 = CPU), whether speech recognition uses the GPU, model folder

New version notifications

At startup the app checks once for a newer version and, if there is one, shows an "Update available" banner at the top of Settings with the latest and current version. Click "Open release page" to open the release page, or "Later" to dismiss it (it stays hidden until the next launch). Nothing is downloaded or installed automatically.

Audio, transcripts and translations are never sent anywhere. There is no telemetry. The only network access is the model download and one anonymous GitHub API request at startup to check for a newer release (it sends no account, token or personal data; if it fails, nothing happens). Once the models are downloaded, the app works offline.

Leftover data and full removal

This folder holds the models (models), settings (settings.json), your acceptance of the terms (terms-accepted) and the detected hardware (hardware.json). Keep it if you don't want to download the models again when reinstalling; delete the whole folder to remove everything.

Model download on first launch (several GB)

First-run setup downloads models that fit your PC from Hugging Face automatically. No Hugging Face account or token is needed. Interrupted downloads resume, and files are verified with SHA-256. Once the models are downloaded, the app works offline.

Sizes are the sum of the sizes declared in models.json (including Silero VAD, about 0.9 MB). Running "Check this PC" during first-run setup determines your tier from the hardware and picks the models. You can change the Whisper / PLaMo models individually in Settings. The result is cached in hardware.json in the config folder, so later launches show it without probing again; run "Check again" after changing the hardware.

Tier guide: High is 12 GB+ VRAM, Mid is 8 GB VRAM, Low is 4–6 GB VRAM, and CPU is for no GPU or under 4 GB VRAM (Apple Silicon is judged by memory size).