What everyone hates about transcription

This is Talk Type, from the team at Pepys, where we turn talk into text.

Here's something we did that probably took longer than we'd like to admit. We read about a hundred Reddit threads about transcription. Real threads, from real people, venting about tools that failed them in the middle of a project.

We went to the obvious places. The PhD subreddit. Podcasting. Journalism. Places where people record hours of audio and actually depend on getting it back as text. And what struck us wasn't the variety of the complaints. It was the repetition. The same four problems, over and over, no matter who was posting.

The first one, and the loudest, was simple. Long files just die. People don't test a tool on a two hour file before they trust it, so they find the wall the hard way, halfway through the thing that matters. One person put it flatly. Everything I've tried kills it at ten or thirty minutes. Another watched an hour and a half meeting transcribe fine, then hit a wall at two hours, with no warning. That one was a lawyer's meeting. The fix here isn't clever. Don't cap by length. Split the audio, transcribe the pieces, stitch the timestamps back together. There's no real reason to quit at thirty minutes.

The second complaint was about money, but not the way you'd think. Transcription is bursty. You get a batch of interviews, a season of episodes, one long deposition, and then quiet. Monthly subscriptions are the opposite shape. So people described the workaround everyone lands on. Sign up for one month, batch everything, cancel. What they actually want is to pay for what they use, and keep what they don't.

The third one surprised us. A lot of tools transcribe your audio and then won't give you the file. They'll answer questions about it. They'll show it to you on a screen. But hand it over as a document you own? No. For anyone whose next step is a captions track, or a document, or a quote in an article, that is a dead end.

And the fourth. Accents and crosstalk quietly wreck accuracy. The threads were honest about where AI still struggles. Background noise, strong accents, people talking over each other. Most of that fight is upstream, before you ever hit transcribe. Record clean, place the mic well, and a good model does the rest.

Underneath all four ran a fifth worry, and it got louder the more we read. Where does my audio go, and does it train someone's model? Once a company holds your recordings, that's what an ethics board, or a source, or a nervous client actually cares about.

So we stacked the four complaints, and the wishlist basically wrote itself. Any length. Pay for what you use. Give me the real file. And don't train on my audio. We didn't invent those priorities. We just read the room.

That's this episode of Talk Type. The full write up, with the links and sources, is in the show notes. Pepys transcribes any file or link, any length, pay once, and we never train on your audio. Your first sixty minutes are free at pepys dot co. Thanks for listening, and we'll see you next time.