Transcribe audio & video to text, without it leaving your browser

Yes. Because processing happens on your device, it costs us nothing when you transcribe more: no trial minutes, no per-file caps, no watermark, no sign-up.

Drop the file here. The audio never leaves your browser: you can verify it by disconnecting from the internet after the model loads — transcription keeps working.

Yes. After transcribing, click SRT or VTT to download the file with per-sentence timings, ready for YouTube, Premiere, DaVinci or any player.

Yes, on desktop Chrome or Edge: the "Record another tab" button captures the audio the video-call tab plays and mixes it with your microphone, so it works even with headphones on. No bot joins the call. The same button works for a video, a live class or a radio stream playing in another tab.

Yes. The "Transcribe a video or stream from another tab" button captures the sound that tab is playing, without using your microphone. It works well for live streams, webinars, online radio and videos with no captions. It does not work with paid streaming services (Netflix, Disney+, Prime Video): browsers block capture of protected content. And transcribing for yourself is not the same as republishing someone else's work.

Download the voice note's audio and drop it here. WhatsApp's OGG and OPUS format is accepted as is, with no conversion, and the text comes back with per-sentence timestamps. Because recognition runs in your browser, the note is never uploaded to any server.

Speech recognition never tells you when it is unsure: when it cannot make something out, it writes a sentence with punctuation and cadence that looks right. Gate32 flags the lines with something anomalous about them — repeated text, more words than fit the time they occupy, seconds of audio turned into two words — and tells you how many there are. Click the timestamp to jump to that point in the audio. It does not claim they are wrong: it points at where to look.

It runs OpenAI's Whisper on your device, with three model sizes. You can pick from 47 languages or let it detect one. The Accurate model delivers strong quality in English, Spanish and the other major languages; in lower-resource ones (Vietnamese, Thai, Urdu…) the small models lose a lot and it is worth moving up a size — the app tells you when you pick one. Very noisy audio or overlapping voices reduce accuracy, as with any transcriber.

The first run includes downloading the model (~50–250 MB) and preparing it, which can take a few minutes; everything is cached and later transcriptions start almost instantly. Speed depends on your machine: with WebGPU (recent Chrome or Edge) it's usually faster than real time. Modest machines fall back to WASM — use the Fast model there.

Yes, it processes in chunks with visible progress. For very long sessions (90+ minutes) a machine with plenty of memory helps; if something fails, try the Fast model or split the file.

Yes, with one iframe. No SDK, no API key, no server to call, and because transcription happens in each user's browser it costs you nothing per minute and their audio never passes through your servers. The code, the parameters and the limits are on adding Gate32 to your platform.

Those tools process your audio on their servers and cap their free plans (minutes, file counts, features). Gate32 is local: verifiable privacy, no limits, no account. In exchange, server-side features like speaker identification aren't available yet (we're measuring interest).