Subtitling 62 hours of travel footage, and working efficiently with AI

For 62 hours of travel footage, I changed the subtitle workflow: Premiere's transcription finds where people speak, and only the highlights get polished with Gemini. With AI, the cost turned out to depend on conversation length and the number of back-and-forths.

This is the last of the work notes. It covers subtitles, cost, and the old footage I tidied up along the way.

Subtitles: Premiere transcription + Gemini

Gemini's SRT output was the most accurate for subtitles. But I couldn't feed all 62 hours into Gemini, so I changed the approach.

Step	Detail
1	Export each per-day sequence as MP3 with Media Encoder
2	Run Premiere's transcription first to locate the speech. The default language was set to English, so I switched it to Korean
3	Drop the unrecognized segments from the transcription output (JSON), which filters out safari radio noise
4	Build an SRT from the JSON and add it to the v2 sequences as a caption track through the MCP
5	Polish with Gemini only the parts that will actually be used in the edit, not everything. Highlight sections dense with conversation were picked by score and cut into 2-minute chunks. The Gemini app only transcribed about 2 minutes reliably to the end, so I came down from the 3-minute chunks I started with

C1 at the top of the v2 sequence is the caption track. Because these are captions rather than text clips, they're easy to re-export as SRT or edit later.

I queued the transcriptions day by day in Premiere and let them run on their own. At first the AI checked every 3 minutes and handled them one at a time, but a mid-way review showed that was inefficient, so it switched to processing finished days in batches every 30–40 minutes.

I also tried having it send a phone notification when a job finished. On the AI's side it said it had sent several, but not one arrived. I decided to just check in more often myself.

Cost and efficiency

This was the first time I really paid attention to usage. Every time it sends a command, the AI rereads the entire conversation up to that point. So the longer the conversation, the more the same task costs, several times over. Waiting 20 minutes costs almost nothing. Lots of back-and-forth is what's expensive.

Partway through, I asked whether the way we'd been working was efficient. The answer was: about half and half.

Task	Assessment
Recovering GoPro times	Fixed dozens of files at once with two or three scripts, and passed verification on the first try
Sorting out where the project is saved	Done in a few steps
Checking the Canon times	Inefficient. Tried to pull audio from the whole external drive and gave up because of the speed, received a 260,000-character error log in full, and trusted an untested assumption through two rounds of audio comparison. One photo would have answered it immediately

So at the end, it wrote a handover document into the working folder, and we agreed to move to a new conversation timed to when the usage limit resets. It even wrote the first message to paste into the new conversation.

Read the handover notes in the Tanzania edit folder on the home PC (drive I:) and let's continue the Tanzania edit.
Ask for every permission you need all at once at the start, and run the work in parallel.

The handover document also includes the tips I picked up this time.

Tip	Detail
Reading external drives	Slow. Read only the parts you need, not whole files
Output	Get summaries only. A flood of logs is all cost
Verification	Before running something long, test the hypothesis on one sample
Bulk copying	7-Zip File Manager is faster than copying from the command line
Audio sync	Line up two clips by comparing how their loudness changes

Old footage, same structure

It wasn't only the trip footage. I also brought 14 folders of personal footage shot before the trip into the same structure.

Item	Detail
Structure	Originals\Video, Originals\Photos, Other, Project
Rules	Video and audio go in Originals\Video, photos in Originals\Photos, processed material like edits and graphics in Other, project files in Project
Verification	Checked the file count matched what was there before
Left alone	Folders that already had a structure, projects being edited (in case links break), folders without dates
Next	Create a project for each folder, with sequences split into per-camera tracks

Remaining work

To do	Note
DAY 4 and 6	Brought the laptop material over on an external drive. Connect it and add it
DAY 9	Empty folder on every storage device
343 RAW photos	Export to JPG from Lightroom and add
Subtitles	Transcribe the remaining days, transcribe only the new clips separately, polish the highlights with Gemini
Sequences for old footage	Per project, after transcription is done
Cloud copies	Two OneDrive 1TB accounts, Google Drive 400GB

That's the end of the work notes. From the next post, I plan to sort out my thoughts on what this stockpile of footage could become.