I wanted subtitles for a few long lecture recordings without uploading hours of audio anywhere. A recent laptop turns out to be more than enough for that.
Chunking beats model size. Splitting audio on silences into segments of twenty to forty seconds gave cleaner punctuation and far fewer hallucinated sentences than feeding the model long stretches at once.
Keep the timestamps honest. Each segment keeps its own offset, so the final subtitle file can be rebuilt without drift even if one segment has to be re-run.
A glossary helps. Lectures repeat the same names and terms. A short list of expected spellings, checked after transcription, fixed most of the errors I cared about.
On battery the transcription ran at several times real time, and an hour of audio finished well before the kettle cooled. The fan barely noticed.
The slow part was never the model. It was me deciding what “good enough” punctuation looks like.