← stuffs

Tamlilתמליל

My search for what's said in YouTube videos, in Hebrew and English.

YouTube's search only looks at titles and descriptions, and I wanted to search what people actually SAY in the videos. You remember that someone explained something, but not in which video, or where in it. Tamlil keeps an eye on the few channels I pick, gets the text of every video, and then one search goes through all of it at once. Every result opens the video right at the moment it was said.

Sign in The code Docs My own Tamlil is just for me and the few people I invite, since it runs on my server. The code is free to use.
Searching שמים: four videos, each with its best moments, the word highlighted in each line, and a label that says whether it matched exactly, with a prefix, or in another spelling
Searching שמים: the word itself, with a prefix like ה, and spelled שמיים.

Captions, or it listens itself

Where a video has good captions from whoever uploaded it, Tamlil uses them, and for English it takes YouTube's automatic ones too. A lot of Hebrew videos don't have any, so for those it listens to the video itself with Whisper, using ivrit.ai's model that was trained on Hebrew. That runs on my server's own processor, which is slow (about 25 minutes for an hour of Hebrew), but it goes newest first, 10 videos at a time, so the newest ones are searchable within minutes and the rest fill in overnight. Adding a channel is pasting its link, from YouTube or from Invidious or Materialious if that's how you watch.

A lecture open at 1:04: the player, a tick under it for each of its 11 matches, and the transcript with each match highlighted
A lecture open at the moment, with a tick for every match and the whole transcript.

A search that knows Hebrew

Hebrew writes the same word in more than one way, and a plain search misses most of them. Tamlil finds שמים in השמים and ושמים, spelled שמיים, and with or without niqqud, and it never mixes it up with המים. Every line says which kind of match it is, so I can tell at a glance. English works too, and the two can sit side by side in the same library. When I know roughly where it was said, I narrow it to one channel, a few words from the title or a range of dates, and sort it newest first.

Searching רקיע on a phone
The Library: three sources with their language, range, how much is searchable, how many went through Whisper and their state, then Add and People
The Library: the channels, their dates, how far along each one is, and who I've invited.

Just for a few people

There are no passwords. I invite people by email, and they sign in with a code that comes by email, picking a username the first time. Everything that talks to YouTube goes through Tamlil's own VPN, and the videos play in a private player that runs next to it, so nothing touches my server's own address.

Basically, it's the search I always wished YouTube had, for the channels I actually care about. The pictures are from a demo with made-up channels and lectures.