Two very different inputs
A transcript with timestamps already knows when each passage was spoken. Exports from transcription tools, YouTube transcripts copied from the page and many meeting notes look like this: a time, then the text. For these the conversion is reliable. Each line starts at its own timestamp and stays on screen until the next one begins, for seven seconds at most, and a long passage is divided into several captions spread across that span.
Plain text, such as a script or an article, contains no timing. No converter can know when a sentence is spoken without listening to the audio, and a tool that claims otherwise is guessing. What this page does is state the guess openly: captions follow each other without gaps, each lasting as long as it takes to read at the speed you choose, between about one and seven seconds. The result is a valid SRT that will drift away from the speech unless the speaker happens to keep that pace.