What Happens to Your File After Transcription

How long your file actually exists

Time to first byte and total time differ by less than a millisecond on every run, which means the answer arrives in one piece — there is no second request and no redirect to a results page. A five-minute recording therefore exists in memory for roughly half a minute. Then the request ends, and so does the file.

• 30 s of audio / 90,512 B — 9.3 s, 5.2 s, 10.2 s
• 60 s / 180,476 B — 12.4 s, 12.2 s, 14.3 s
• 300 s / 900,512 B — 45.1 s, 44.7 s, 30.5 s

What comes back — and what never does

What is not in there matters more. No job id, no result URL, no retrieval token, no handle you could query later. There is nothing to come back for, and we checked: a plain GET to the same endpoint answers 404 Not found. The transcript is handed to your browser once, and after that it exists only wherever you put it.

• success — true
• transcript — the text, and nothing else
• language — reported as unknown
• engine — Cloudflare Workers AI (@cf/openai/whisper)
• fileSize — 61200, the exact byte count you sent
• fileName — the name of your file, echoed back to you

There is nowhere in our code to save it

A search of the deployed source for storage calls returns nothing at all — no KV, R2, D1, cache or local storage writes anywhere in the file. A worker with no storage binding can hold your bytes in memory for the duration of a request and then lose them. That is precisely what happens here, and it is a stronger guarantee than a deletion policy, because there is no code path that could keep the file even by mistake.

• bindings: one entry — name AI, type ai
• no key-value namespace
• no object storage bucket
• no database
• logpush: false
• tail consumers: empty

What we checked today, and how you could repeat it

• We uploaded a clip renamed with a random marker, then fetched the home page, the blog index, the sitemap, the privacy page and the about page. The marker appears zero times on all five.
• The transcription response sets no cookies.
• A GET to the transcription endpoint answers 404, so there is no stored result waiting to be fetched.
• The plain HTTP URL answers 301 to the HTTPS one; every page is served over HTTPS.

The part we will not claim

• Your audio is transcribed by Cloudflare Workers AI, not by hardware we own. That is a real third party in the path, and we are not going to hide it behind the word "encrypted".
• We do not run a training pipeline on uploads.
• We cannot show you what happens inside the isolate after the response is flushed, because we cannot inspect someone else's memory allocator. What we can show is that our code never writes the file anywhere and has nowhere to write it.
• If you need a guarantee that no third party ever touches the audio, the only version of that promise that holds is a tool running on your own machine. Ours is not that tool.

What this means in practice

• Do not upload audio you are not willing to send across a network.
• Copy the transcript out when you see it. It is delivered once.
• There is no account to delete and no history to clear, because there is no record for a history to live in.
• If the audio is legally sensitive — a recorded call, a medical or legal conversation — check your own obligations first. "We deleted it" is not a legal answer.