No signup and no upload. The model downloads once and then everything happens on your own machine — which is the only honest way to promise that a recording of your meeting stays private. Recordings up to 30 minutes; the app handles full-length meetings.
Drop a recording of up to 30 minutes and get a clean transcript. Works with mp3, m4a, wav, mp4, mov and more.
Open tool →עברית · HebrewTranscribe Hebrew recordings — including calls that switch between Hebrew and English mid-sentence.
Open tool →SubtitlesTurn a video into timed .srt or .vtt captions, ready to drop into YouTube, Premiere or a player.
Open tool →Your file is read by the page and transcribed in the tab. It is never sent to us or to anyone else, so there is no server that could keep a copy.
The first run fetches a speech model — around 80 MB — and your browser caches it. After that the tools keep working with no connection at all.
A browser-sized model is good, not perfect — it trips on accents, crosstalk and rare names, and it does not separate speakers. We measured the gap against Kika's own transcription: 13.7% of words wrong here, 3.9% there.
Every tool on this page sees a single file, with no idea who was speaking or what any of it referred to, and forgets it when you close the tab. That is fine — often a transcript is all you wanted.
But the questions people actually have about their meetings are never about one meeting. Why did we choose ten users instead of fifty? What did I promise this client? What has been sitting on their side for a month? What did someone ask me that I never answered? None of that lives in a transcript. It lives in the relationships between forty of them.
That is what the Kika app is — and we wrote up how it is built, including the parts that are still hard.