Let's Talk

A transcription product whose two halves couldn't share a page

Audiotextify

Diagram of Audiotextify: audio runs through a custom FFmpeg pipeline that cuts it into pieces, the pieces go to the API and come back as one transcript, with API and infrastructure costs kept low, and Stripe handles the paymentsFFMPEGAPISTRIPEAUDIOCUSTOM FFMPEG PIPELINEAPI AND INFRASTRUCTURECOSTS KEPT LOWTRANSCRIPTPAYMENTS
FIG_001

The second kind

Audiotextify started as a client's simple idea, a transcription product with a GPT-3 integration, built around the time ChatGPT first came out. Some projects are big from the first call, and some only show you how hard they are once you're inside them, and this was the second kind.

Running FFmpeg in the browser needs the page isolated with a pair of security headers, and those same headers stop Stripe's checkout from loading, so every time I got one working, the other broke. AI couldn't help yet, and my question on Stack Overflow went unanswered, so I went to Stripe's own developers, who confirmed the two can't run under the same headers in one project. I split the product into separate deployments, each served with the headers it needed, and both worked.

The client got a product that takes audio and video files of up to 2 GB, of any length, slices them and transcribes them with very high accuracy. My team later rebranded it as AudioFlow, and it still runs and still earns.