Your recording is encrypted before it leaves this tab and stays encrypted the whole way — including while the AI is transcribing it. Nobody along the chain can read it: not the AI provider, not the cloud it runs on. Not the audio, not the transcript.
Ordinary encryption protects your data in transit and at rest. Confidential computing keeps it unreadable while it is being used. That is the whole reason this page can exist.
Sign up at portal.privatemode.ai and copy the key it gives you. You pay Privatemode by the audio minute — about €0.84 for an hour with Whisper, €0.24 with Voxtral. hushscribe estimates the cost of each file before the transcript comes back. The key stays in your browser; this page has no server to send it to.
Your browser then checks, before anything is sent, that it is really talking to a sealed machine and not an impostor. A green check means the check passed.
Audio or video. It is scrambled here on your device, unscrambled only inside that sealed machine, and the text comes back to you. Pick the spoken language first if you want timestamps.
Nothing you drop here is uploaded to us — there is no “us” to upload to. Transcripts are saved in this browser only, and you can delete them at any time.
Sign up at portal.privatemode.ai and copy the key from your account. It works like a password that also pays for transcription, so treat it like one. It is stored in this browser and sent only to Privatemode, never to anyone else.
Whisper, unless you are in a hurry. It copes better with accents, background noise, and poor recordings. Voxtral is quicker but wants clean audio — if the result looks wrong, run it again with Whisper.
You pay Privatemode per audio minute, not per file: €0.014/min for Whisper, €0.004/min for Voxtral. An hour-long recording is about €0.84 or €0.24. Prices as of September 2026, plus VAT where applicable — check privatemode.ai/pricing for the current figures. hushscribe shows an estimate for each file once your browser has read its length.
Pick the language actually spoken. Given the wrong one, Whisper does not fail — it translates into the language you chose, and the result reads like a plausible transcript that is not what was said.
Wrong language → Whisper translates instead of transcribing.
Two reasons. Naming it gets you a better, faster transcript than letting the model guess. And it is the only way to get timestamps — the clickable line-by-line transcript and downloadable subtitles. On automatic detection you still get the full text, just as one block with no times attached.
Naming the wrong language is worse than naming none. Whisper treats a mismatch as an instruction to translate, so German audio labelled English comes back as fluent English prose — readable, confident, and not a transcript of what anyone said. If you are unsure, leave it on automatic detection.
0 characters
Anything the model is likely to get wrong: people’s names, company or product names, medical or legal terms, acronyms, unusual spellings. Write them the way you want them to appear, in the same language as the recording. Keep it short — it is a hint, not an instruction, and only the first couple of hundred characters are read.
flac · mp3 · mp4 · mpeg · mpga · m4a · ogg · wav · webm
50 MB and 1 hour per file. Nothing is uploaded until it is encrypted.
Text and timestamps are kept in this browser. The media files never are — they only ever existed in the tab you dropped them into, so history has no player.
Results still appear above while this tab is open, but nothing is written to disk. Transcripts saved before you ticked this stay until you clear them.
Ticking “include API key” turns the export into a credential file. It travels further than this browser does.