Installing dependencies
AI Subtitle Studio does not bundle every tool it uses. It downloads and manages a set of well-known tools on demand, into its own storage folder, independent of anything already on your computer. You install and update them from Settings → Dependencies.
The tools
| FFmpeg (with FFprobe) | Reads videos and extracts and encodes audio and streams. Required to scan videos, extract subtitles and transcribe. FFprobe installs with it. |
|---|---|
| Whisper.cpp | The speech-recognition engine behind Transcribe. Needs at least one model, downloaded on Speech Engines. |
| FFsubsync | Lines subtitles up with a video’s audio or a subtitle already in sync. Powers Tools → Sync. Optional. |
| mkvextract | Part of MKVToolNix. Used only for DVD subtitles in MKV files that the faster built-in reader cannot serve. Optional. |
| yt-dlp | Fetches captions and audio from YouTube. Used by the YouTube Transcript screen and nothing else. About 35 MB, installed automatically on first use. Optional. |
| Qwen3-ASR | An alternative to Whisper: about 0.9 GB, CPU only, Windows only. Optional. |
| Parakeet | An alternative to Whisper for English and 24 other European languages: about 0.5 GB, CPU only, Windows only. Optional. |
| Speaker detection | Finds who speaks when, for Dubbing. About 55 MB. Optional. |
| Dubbing runtime | A private Python environment with PyTorch and four voice models, for Dubbing. About 6.4 GB with an NVIDIA card, about 9 GB on disk. Windows only. Optional. |
A tool that has no build for your system shows as unavailable instead of offering an install that would fail. OCR for image subtitles is built into the app and needs no external tool; its recognition model is downloaded the first time Convert uses it.
Installing and removing
Each tool has its own card with an Installed or Not installed badge and its version. Click Download & Install to install; once installed the same button becomes Update, beside Show files to reveal where it lives and a delete button. Downloads show progress and can be cancelled.
Installs are staged: the new version is built beside the old one, run once to check it works, and only then swapped in. A failed, cancelled or interrupted install never touches a working one, and closing the app mid-update is safe — the next launch finishes or undoes the swap. A tool cannot be installed or removed while a job is using it.
Choosing the Whisper build
Whisper comes in builds for different hardware: CPU, CUDA (NVIDIA graphics cards on Windows), Vulkan (AMD, Intel and NVIDIA graphics on Windows, needing the graphics driver’s Vulkan support) and Metal (Apple Silicon). Choose one under Build to install before installing. Once installed, the card tags the active build, and that installed build decides whether transcription may use the graphics card. If a GPU build cannot start its graphics support it quietly runs on the CPU. With several graphics devices, a Vulkan run uses the first dedicated card rather than integrated graphics.
Updates
The Updates area on the Dependencies screen has Check now and Update all, and shows a count on Settings in the sidebar when something is out of date. A tool is called out of date only when the app could read and compare both versions. Update automatically is off by default: when on, the app installs new versions in the background, only while it is open and nothing else is running, one tool at a time, and leaves a failed update at the current version. Very large downloads such as the dubbing runtime are offered but never installed automatically.
What happens if something is missing
If a core tool is missing you see an amber Setup required banner on the Dependencies screen listing what is missing, and a Required tools missing notice with a link to Dependencies on the screen you tried to use. FFmpeg is fetched from several sources in turn, so a single dead download link does not stop an install. The Storage location on the screen shows where everything is kept.