Nothing leaves your computer. Each tool is a small piece of code your browser downloads once, and after that your file is read and rewritten on your own machine. There is no upload to opt out of, because there is nowhere to upload to.
You can check this yourself. Open the network panel in your browser, convert a file, and watch that no request is made. Speech to text is the one exception, and it is explained further down: it fetches its model once, and never your audio.
Because it costs us almost nothing. Your computer does the work, so there is no per-file cost to pass on to you. That is the whole reason there is no cap, no account and no paid tier.
As long as your browser needs to read the file, which for a subtitle file is a few milliseconds. There is no queue because there is nothing to queue for.
Subtitle files in the wild are often slightly broken: stray formatting tags, a byte order mark, timings that run past each other. Where a tool repairs something it tells you what it changed and how many times. It will not quietly rewrite your file and say nothing.
Speech to text is the exception worth explaining. Turning audio into words needs a model, and the first time you transcribe anything your browser downloads one, about 76 MB. That is a real network request and we would rather point at it than let you find it. It happens once, your browser caches it, and it is the only thing that is ever fetched. Your recording itself is never sent, on that first run or any run afterwards.
It depends on your machine, which is the honest answer when the work happens on your machine. On the computer this was built on, eight seconds of audio took about eleven seconds with the model already loaded, using the graphics card. A device without WebGPU falls back to the processor and takes longer. Long video is the current weak point: the whole file is decoded in memory, so a two hour recording can exhaust what a browser allows.