SRT subtitles from video with a local Whisper model for bloggers, 2023–2024

In late 2023 and early 2024 I built a system called aif for travellers and bloggers who create content for YouTube, Instagram and other platforms. It took a video or a link to a recording, extracted the audio, recognised speech with a local Whisper model and produced a subtitle file with timestamps. It was a related project for the same company whose collaboration is described on the Software development for a London consultancy, 2023–2024 page.

The task

Paid transcription services usually work on a subscription, and with a lot of footage that becomes a noticeable expense. The goal was to let users process their content with free tools instead of paying a third-party service.

How it worked

The inputs were videos from Google Drive, podcast links and YouTube links, and all three were in actual use. The MP4 videos sat in a Google Drive folder, and from the links the system fetched the audio track.

For videos from the folder, the audio was extracted first, and long recordings were split into parts when needed. From there the path was the same: speech was recognised by a Whisper model running locally rather than through a cloud API. The result was saved as an SRT subtitle file: text split into segments, each with a start and end time. The file was uploaded back to Google Drive into a separate transcripts folder.

For Google Drive videos, the system recorded the status and progress of each file in Google Sheets. Temporary files were deleted after upload, and already processed links were tracked so the same recording would not be run twice.

Engineering choices

Jobs were accepted by a small Python service on Flask. It confirmed receipt straight away and ran recognition separately, without holding the request open while a long video was processed.

Recognition ran on a local model and free tools, so there was no per-minute charge. For a blogger who publishes a lot, that could mean huge savings compared with a paid subscription. The server itself still costs money, but there was no separate charge for the volume processed.

What was delivered and what stayed in the plan

Users received an SRT subtitle file with timestamps and worked with it further in their editing. The next step was meant to be automatic burn-in of subtitles onto the video, producing a finished clip. I did not get to it: the client stopped the project.

I still think it was a really great project, and it was a pleasure to work on.

Have a similar task?

If you need to transcribe audio and video, or run a model inside file processing on your own server, see AI workflows, Python data processing and Google Workspace automation. The other projects for this company are on the Software development for a London consultancy, 2023–2024 page.

Tell me what recordings you need to process and what the result should look like. Discuss your project