Audiotextify turns any audio or video into a transcript, then into summaries, translations, posts and articles. A solo founder came to us with an idea and a logo. We built it on Whisper and GPT, solved a clash between two pieces of browser technology that nobody online could answer, and it launched with paying subscribers.
Background
The founder came to us with an idea, a logo and two documents: one describing the product, one setting out how it would make money, free to start with paid plans above that. This was before ChatGPT was much use for writing code, and before coding agents existed at all. We built the brand around his logo, designed every screen, and got to work.
He was paying for the build himself, so it had to cost as little as possible to run. The easy route was a ready-made transcription service, but the good ones were expensive and the cheap ones were unreliable. So we built directly on the underlying models: Whisper to transcribe, GPT to write everything made from the transcript.
The problem: FFmpeg against Stripe
Partway through, we ran into a problem that took more effort than anything else in the build. Running FFmpeg in the browser needs shared memory, and browsers only allow that on a page locked down with two specific security headers. Turn the headers on and FFmpeg worked, but Stripe's checkout would not load. Turn them off and Stripe worked, but FFmpeg did not.
The hardest part was figuring out what the actual problem was. It played out like a riddle. Whenever Stripe broke, the trail led back into Stripe's integration. Whenever FFmpeg broke, it led back into the FFmpeg code. Saif audited Stripe's settings line by line and came up empty, which meant something outside both of them was getting in the way. After narrowing down every header combination, the cause turned out to be one specific header responsible for controlling which outside resources a page is allowed to load. ChatGPT cycled through suggestions that had already been tried, Stack Overflow had no answer, and it took a direct question to Stripe's own developers on their Discord to confirm it. Stripe's checkout did not support that header, with no date for when it would.
What we built
The fix was to split the product in two. FFmpeg moved to its own deployment, served with the headers it needed, while the main site kept Stripe and never sent them. When someone drops a file into the extension, a small window opens on the FFmpeg deployment and compresses the file there, on the user's own computer, before anything is uploaded.
That fix also slowed transcription down, since the compression now ran in a separate place. Moving it to a bigger server made it faster but too costly to run while the product was still in testing, so the real fix was in the compression itself. Stripping out video, mixing audio down to one channel and compressing it for speech cut an hour of audio to about 5 MB, small enough for Whisper to take in a single request. On a test of our own, a 2 GB, 48-minute film came back as a transcript in under two and a half minutes.
How it works
The extension is a shell
The extension itself holds almost nothing. Sign-in, transcripts and payments all live in the web app behind it, where they are safe, and the extension only opens windows into that app when it needs to.
Compressing before anything uploads
Each file is converted to a mono, speech-tuned audio format at 12 kilobits a second, which brings an hour of audio to around 5 MB, well under Whisper's 25 MB limit. That means a file almost never needs to be split into pieces before it is transcribed.
One request to Whisper
The compressed file goes to Whisper in a single request, and the response comes back broken into timed segments. Those timestamps are what let a summary cover just one stretch of a long recording instead of the whole thing.
Writing in whichever language the user picked
Summaries, posts, threads and articles are all written by GPT and streamed back as they are generated. The founder wanted his output in French, so every kind of post and article works in whichever language the user picks.
The result
Audiotextify launched on the Chrome Web Store with its web app live and Stripe processing payments, and it had paying subscribers from the start.
It's what I want. Perfect Saif
Jessy Mondesir, founder of Audiotextify
Saif listens, responds quickly and gives everything he has to his mission.
Jessy Mondesir, founder of Audiotextify
The compression pipeline built for this project became the base for NiggleAI's audio and video processing later on.
Who did what
Burhan Pasha was the client lead and owned delivery on this project. Saif Pasha designed the brand and every screen, built the product on Whisper and GPT, and found the fix for the Stripe conflict. The Silverthread Labs team built and shipped it.
If your product needs two pieces of technology to work together that were never designed to, book a call.