Long Files That Don't Die

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

Here's the complaint we heard more than any other. Long files just die. You're forty minutes into a two-hour interview, and the tool quits, or the upload bounces, and the thing you actually needed never comes back as text.

So let's talk about why that happens, and how to get past it.

There are really two walls, and people mix them up. The first is a size wall. A lot of transcription runs on an API that caps a single upload at twenty-five megabytes. That's not about how long your recording is. It's about how many bytes you're sending. At a normal speech quality, twenty-five megabytes only holds about twenty-five minutes of audio. So a one-hour call already overflows it. A two-hour tape never stood a chance.

The second wall is a length cap. Some models flatly refuse a single request past about twenty-five minutes, no matter how small the file is. Size and length. Two different walls, and you can hit either one first.

Now here's the part that changes how you think about it. Splitting a long file feels like a hack. It isn't. Speech-to-text already reads audio in tiny windows. The underlying model processes sound in fixed thirty-second chunks, decoding one slice while remembering the last. Every clean transcript you've ever seen was already chunked under the hood. Cutting a two-hour file into twenty-minute pieces is the same idea, one level up.

So the fix has a name. Split and stitch. You cut the audio into pieces under both caps. You transcribe each piece. Then you stitch them back into one file.

But there's a trap in the stitching, and it's the whole game. Each chunk restarts its clock at zero. So piece three thinks it starts at zero, when really that audio lives at forty minutes into the recording. If you just glue the pieces together, every timestamp is wrong. Your two-hour transcript reads like six twenty-minute recordings stacked on each other, and every quote points at the wrong moment.

The fix is to re-offset. Before you join anything, you add each chunk's real start time back onto its timestamps. Chunk two's zero becomes twenty minutes. Do that, and the seams vanish. The transcript reads as one continuous timeline you can actually cite from.

Now, you can do all this by hand. Compress the file down to sixteen kilohertz mono to shrink it, split on the silences so you never cut through a word, re-offset the math across every seam. It works. It's also tedious and easy to get wrong on the fourth chunk at midnight.

Which is the whole argument for letting a tool do it. That's how Pepys handles a long file. No thirty-minute cap, no twenty-five-megabyte wall, because the chunking happens before anything reaches the model. You upload the two-hour recording. You get back one stitched, correctly-timed transcript.

Length was never the real problem. It was just a wall nobody told you was made of paper.

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.