Is it possible to make smart captions without generative AI?

This is very definitely one of those problems where you really need to define, mostly for yourself, exactly what you mean by “without AI.”

What you’re describing is automation, and it is, by necessity, artificially intelligent automation. That’s because human beings don’t make the same sounds the same way every single time, and those sounds aren’t going to be made against the same background noise every single time. So there is no way that you can ever deterministically transcribe spoken word to text. It requires guesswork, and when you’re talking about computers and guesswork, you’re necessarily talking about technology that falls within the general purview of “artificial intelligence.” This has been true for as long as computer assisted transcription has been a thing.

So what exactly is it that you’re trying to avoid here? Is it specifically your data going to an AI company that concerns you (ie, would a local model be an acceptable solution)? Are you perhaps trying to avoid giving money to certain AI companies? These are the sort of issues that might be addressable. But just a broad “no AI” prescription will basically rule out any useful answers. I hope that makes sense.

I mean… I don’t consider ’old school’ ML (TTS, STT, OCR, photo tagging) as ‘AI’

You mean you want speech to text? Yeah you have to figure out what exactly you object to. The better STT stuff does use language models, but I wouldn’t call it “generative” in the chatbot sense.

You mean have a human do it for you?

Speech-to-text or speech recognition can be a generative (transformer) model architecture, but there are a lot of models that are not. The transformer ones seem to be better in quality.

Non-transformer examples:

Coqui STT https://stt.readthedocs.io/en/latest/

Kaldi https://kaldi-asr.org/models.html

Creating subtitles / captions requires some more programming to create the text file with the correct timings.