Download

Windows and setup

Substrike is a Windows desktop app, launching soon. Check your PC and the download sizes here first.

What you get

• The Substrike Windows installer, about 190 MB in the 0.3.0 test build. Two small speaker-detection models (about 46 MB) are already inside it.
• Batch and watch-folder captioning with crash-safe resume, word re-timing, AI cleanup that flags unsure words, exports sized for Discord or vertical video, a reel builder for finished batches, and optional Discord posting.
• One licence, commercial use included (a monetised channel or paid client work is fine), for use on up to 2 of your own PCs. One purchase for every 1.x release. A later major version may be an optional paid upgrade, at half price for existing owners. You enter a licence key once, online, and Substrike then runs offline. The licence terms are a draft until launch.
• Captions are burned into the exported video. You can also save an editable SRT subtitle file next to it, with the same lines and timing, for Premiere, Resolve or CapCut, and the styled ASS file too.

Try it before the big download. First-run setup includes a 15-second preview on your own clip, using a small speech model already inside the installer, so you see captions on your own audio before downloading the larger models.

System requirements

• Windows 10 or 11, 64-bit, with the Microsoft Edge WebView2 Runtime (already present on current Windows). No macOS or Linux version.
• Disk space: about 190 MB for the installer, about 3.1 GB for the recommended first-run downloads below, and room for your exports and temporary files.
• Internet: for the first-run downloads, update checks, optional Discord posting and one licence activation. Captioning and export work offline once the models are downloaded. No footage is uploaded to Substrike or a speech service for processing; files leave your PC only for those connections.
• Speech recognition runs on the CPU. A recent NVIDIA, AMD or Intel GPU speeds up video export. Substrike test-encodes at start-up and only offers encoders that work on your PC.
• Speech recognition: English today. Spanish, Portuguese, French, German and Italian are in beta in the app, and more languages are coming soon. Their accuracy has not been measured, so check the captions before posting. Captions can also be translated into other languages after transcription, using the same local model.
• Nothing else to install. The Microsoft Visual C++ runtime Substrike needs ships inside the app.
• On one development PC (Ryzen 7 9800X3D, 16 logical CPUs, 31 GB RAM, CPU only), the recommended speech model processed clips at roughly their own length: a real-time factor of about 0.86, including loading the model each time. That machine was also running other heavy background jobs during the test, so treat this as a rough, one-machine data point, not a promise for your PC.
• Minimum RAM and typical processing time per clip on a clean PC will be published after clean-PC testing.

First-run downloads

The speech and timing models are downloaded once, from inside the app, from Hugging Face. Each file is pinned to an exact version, checked against a fixed SHA-256 checksum, and resumes if interrupted. Sizes are decimal megabytes as listed in the app.

First-run model downloads

Download	Size	Needed for
Whisper large-v3-turbo (recommended speech model)	1,625 MB	Transcription
wav2vec2-base-960h	378 MB	Precise word timing
Qwen2.5-1.5B-Instruct (Q4_K_M)	1,117 MB	Optional AI cleanup (flags unsure words)

Recommended setup total: about 3.1 GB for the three downloads above.

Smaller, faster speech models are available if you have less disk space or a slower PC: tiny.en (78 MB), base.en (148 MB), small.en (488 MB) and medium.en (1,534 MB). They are less accurate. There is also large-v3 (3,095 MB); on the clips we have scored it did not beat large-v3-turbo, so start with large-v3-turbo.

Without the timing model, captions still work but word timing is less precise. Without the cleanup model, Substrike cannot flag unsure words, so by default a batch holds each clip for review instead of exporting it as if it were checked.