Narrly There 🏄

I’ve always been interested in seeking side projects that are interesting, and that solve real world problems.

Although like many engineers there’s many a side project on the shelf at any time. However, this article is a little different, and introduces a project I’ve been working on recently, and have become quite proud of.

Oh, and one that works well!

Where It Started #

This all started with a friend mine Ross Martin. Ross and I go back to the University of Huddersfield, and he’s the same good friend I built DartsNow with through the pandemic. So we’ve got form for this sort of thing and work well together.

Ross is a freelance editor and colourist, with a wide array of experience across the TV & Film space. A while back he was telling me about a chunk of his work that he was certain could be automated. He repeated this work on every project, every time.

Ross walked me through it properly, mostly because I was being nosey. But I’d also been whittering on to Ross about having a crack at building something new. With some of the new AI capabilities I was confident we could give something a punt, and it turns out the problem itself is straightforward. It’s just repetitive, and there’s a lot of it.

Guide, Master, and a Lot of Scrubbing #

For anyone who’s never been near an edit suite here’s the shape of it.

When you’re cutting a video, you usually put a rough voiceover down on the timeline first. A scratch track. It’s not the one that ends up in the finished thing, it just gives you something to cut the pictures against. That’s the Guide.

Later, the proper voiceover gets recorded in a session. One long recording, the script read from top to bottom, with retakes, coughs, false starts, “sorry, can I go again”, the lot. That’s the Master.

Now the fun bit. Someone has to go through that long Master recording, find the take that matches the first line on the timeline, drag it into position, and then do exactly the same for the second line. And the third. And every line after that, for as long as the script is.

It’s not particularly hard work. That’s what makes it so annoying. It’s just scrubbing, squinting at waveforms, and dragging, over and over, and any editor with a deadline is doing it when they’d rather be doing the actual editing.

So We Built Narrly #

Narrly does that matching bit for you.

You drop in your Guide and your Master, and it works out which takes in the Master belong to which lines in your Guide. You get back a multi-track WAV with your Guide on the first channel and the takes stacked above it, each one sat at the right point on the timeline. Import it like any other file, then mute, solo and cut as normal.

Alternatively you can export a single-track, which is Narrlys combination of the best matched takes. Interested to see how much this option gets used, but it’s there if required.

That’s it. It’s not trying to edit your video, and it isn’t going to write your script. It does the repetitive part and then gets out of the way.

Under the bonnet there’s a web app, a backend, and a chunk of audio processing doing the heavy lifting. I’ll leave it there for now, partly because the interesting bits deserve a post of their own, and partly because we’re still working on refining it all.

Shipping It #

It would be hard to not include that using LLM’s has enabled me to work faster, albeit in areas. It’s not all been smooth sailing though, it’s certainly created its fair share of mess, but I’ve taken a spec driven development approach to it so that I can help steer each small step. I also wanted this to be built in technologies that I could still step into, still be able to wrestle with when I eventually hit the usage limits.

There was lots to consider; Deciding what happens to those files (they’re processed on our own kit in the EU, they never go near a third-party AI platform, and they’re binned when you close the tab). Writing an FAQ in plain English rather than the version I’d write for another engineer, after all it isn’t a purely technical write up. Also how on earth to a make the interface simple enough to just do the task while being informative enough to give you confidence it succeeded.

There’s no sign-up, no login, and we don’t ask for your name or your email. It’s free while it’s in early access, mostly because it’s early and early feedback is valuable right now.

None of that is glamorous. All of it took longer than I expected too! Because I want to be confident in the outcome we ship. I want to confidently stand behind the stack, know it inside and out and have confidence that it still solves the problem at hand.

A Head Start, Not a Finished Edit #

Narrly won’t place every take perfectly, and it isn’t meant to make the final call. It shows you what it found before you download anything, so nothing quietly disappears on you, and you should still check the output before you rely on it for something that matters.

It’s meant to save you the repetitive part of the job. That’s the whole pitch in a nutshell.

Give It a Go #

It’s live at narrly.tech. Drop two files in and see what comes back.

If you edit video, and it’s rubbish for your workflow, I genuinely want to hear about it. You can reach me at hello@jamesmillner.dev.