Auto-Transcription and Subtitles: What Is Good Enough in 2026
Answers best automatic transcription for video subtitles
Two years ago the question was whether automatic transcription was usable. It is. The question now is where it stops being usable, and that line is specific enough to be worth knowing.
What it is genuinely good at
Getting you to eighty per cent in minutes. On clear speech from a decent microphone, modern models produce a transcript you can work from rather than correct. On an interview, that is the difference between starting the edit on the day and starting it the next morning.
Timing. The word-level timings that come out of these tools are usually more accurate than hand-timed subtitles, and they are what makes burned-in captions look right.
Searching footage. A transcript turns an hour of interview into something you can search. On long-form work this changes how the edit is done, not just how fast.
Accessibility at low cost. Captions that once needed a transcription service now need a pass and a check.
Where it still fails, reliably
Names and jargon. Product names, place names, industry words. Always wrong, always the same words, which means it is fixable with a glossary.
Numbers. "Fifteen" and "fifty", "two" and "to". Particularly in the accents these models handle worst.
Crosstalk. Two people talking over each other produces a plausible sentence that nobody said. Dangerous in an interview, because it reads as fact.
Heavy accents and quiet rooms. Still a real gap, and worth testing on your own material rather than trusting a demo.
Punctuation across a pause. Placing a full stop where the speaker was thinking changes the meaning of the sentence.
The practical workflow
First pass by machine. Accept the eighty per cent.
Glossary before, not after. Feed the tool the names it will meet. It is five minutes and it removes most of the correction.
Correct against the audio, on the sound. Not on the words on screen. A sentence that reads correctly and was said differently is the error that goes unnoticed until it is published.
Write the captions for reading, not for listening. Shorten, split at meaning, drop the filler. A verbatim transcript is a poor caption.
One person checks the whole thing. Splitting the work half-and-half produces two standards.
The tools, in the order worth trying
Whisper, the open model from OpenAI, is the one most things are built on and it runs locally, which matters if the footage is under an agreement. github.com/openai/whisper
Subsai wraps Whisper with a interface and subtitle output, which is what you actually want if you are not going to live in a terminal. github.com/absadiki/subsai
Built-in transcription in the editing application, which is convenient and usually a generation behind the open models.
The paid services, which are faster and hosted. Worth it on a deadline, less so on a job where the footage should not leave the machine.
What this means for a job
Subtitling is a service and it stopped being expensive. If a quote for captions looks large, ask what is in it: the machine pass costs minutes, and the hours are in the checking and the reading-edits, which is where the quality actually comes from.
If your project needs captions in more than one language, that is worth raising early, because it changes the timeline rather than the export.