DEV Community

Cover image for how caption generation works in our windows screen recorder
Cilevi Recorder Studio
Cilevi Recorder Studio

Posted on Originally published at cilevi.com

how caption generation works in our windows screen recorder

When developers build a screen studio for windows, handling offline audio transcription locally requires careful thought about model sizes and download mechanics. Cilevi runs speech recognition entirely on your machine, keeping your data private. Inside the studio interface, the Captions tab exposes settings for the engine, model, language, and custom prompts. You can choose between the Windows operating system recogniser or a bundled Whisper implementation that supports 99 languages. Since speech models are large, they are not shipped directly inside the main binary. Instead, the application offers three distinct file sizes that download on demand into your local application data folder. The Standard model takes 264 MB, the Better quality option requires 574 MB, and the Best quality variant occupies 1.1 GB. Resumable downloads via HTTP range headers ensure that interrupted fetches pick up where they left off rather than restarting from zero. Before loading any weight file into memory, the software computes an SHA 256 hash to verify integrity against the compiled build. This verification step takes about one second for the medium model. Processing times scale with hardware and audio length. On a high end desktop with an i9 13900K processor, transcribing fifteen minutes of English audio takes about 1.4 minutes using the 264 MB model and roughly six minutes with the 574 MB option. Shorter clips face a fixed overhead cost for initialisation, which makes short samples appear relatively slower per minute of audio. For developers building product demo video software or figuring out how to make a product demo video, language choice heavily dictates model selection. While the small model handles English well, other languages benefit immensely from the 574 MB or 1.1 GB downloads to avoid persistent spelling errors. Adding specific domain terms or brand names into the prompt box helps guide the recogniser toward correct outputs. You can purchase this Windows tool for 50 dollars with no subscription required. To learn more about how it compares, check out this screen studio for windows page.

Top comments (0)