Find a TikTok by the words you remember saying

Human memory does not store dates. It stores phrases.

You remember saying “we only had one shot at that light”. You do not remember that it was a Tuesday in March 2023, and you certainly do not remember what you wrote in the caption, because you wrote the caption in eleven seconds and never read it again.

So the thing you have is the exact key nobody built a lock for.

Why the caption is the wrong index

Captions are short, written in a hurry, and rarely repeat what you said in the video. They exist to hook someone scrolling, not to describe the content. Searching them is searching a marketing line, not a transcript.

The spoken track is where the actual content lives. A three-minute video is several hundred words, and every one of them is a possible search term. None of them is text until something transcribes it.

This is not a niche problem, by the way. Products like ReelRecall exist because transcript-based retrieval for short video is a real and recurring need. The distinction worth keeping is what is being indexed: FindMyTok indexes your own published TikToks, not a general inbox of videos other people sent you.

What the phone is doing

Apple’s Speech framework can run recognition locally, and Apple documents the flag for it: requiresOnDeviceRecognition, which forces the request to be handled on the device rather than sent to a server.

For text visible in the frame there is a second mechanism, Vision’s text recognition, which reads the words drawn on screen. Title cards, on-screen captions, anything you held up to the camera.

Two different sources of text, one index. Words you said and words you showed.

What a good result looks like

Not a thumbnail. A thumbnail is a guess and you will have to open five of them.

What you want back is the sentence with your phrase inside it, the post it belongs to, and the second at which it was said. Then you open TikTok at that video and skip straight to the moment, instead of watching two minutes of yourself to find eleven seconds.

That is the format FindMyTok returns, and it is the reason the app exists in the shape it does.

The honest limitations

Transcription is not perfect. Accents, music over speech, three people talking at once, a bad microphone in a windy street. Some words will come out wrong and some of those words will be exactly the one you searched for.

Search a longer phrase and you have more chances of hitting a stretch that transcribed cleanly. Search a unique word rather than a common one. This applies to every transcript search product that has ever existed, including the ones that promise otherwise.

And the video has to be one you published from the account you are searching. Something you watched once and never saved is a different problem: see watch history, likes, favourites, and your own posts.