What automation does well
Automated tools can produce a draft transcript or translation quickly and at low marginal cost. For clear, single-speaker audio in a well-supported language, the draft is often a useful starting point. That speed is the main reason automation has become common, and it is a real advantage when volume is high or budgets are tight.
The question for quality is not whether automation is useful, but where its output can safely go out without a person looking at it, and where it cannot. The honest answer is that it depends on the material, the audience and the cost of a mistake, and that no automated system is error-free.
Where a person is still needed
Meaning, not just words
Automatic systems work mostly by matching sound patterns to likely words. They can produce a smooth sentence that is wrong. A human reviewer reads for sense: does this sentence make sense here, does it contradict what came before, does it fit the topic? Sound-alike errors that survive automated spell-checking are typically caught this way.
Names and specialist vocabulary
Rare words are where automated transcription is weakest, because there is less context to guide the guess. Names, places, product names and technical terms are often misrecognized. A person with the script, glossary or subject knowledge can fix these quickly. Adding a custom vocabulary can help a tool, but the result still needs to be read.
Difficult audio
Overlapping speakers, background noise, strong accents, quiet speech and poor microphones all lower accuracy. The effects of audio quality and accent recognition are well known, and a human listener can often work out words from context and by replaying a section. A reviewer can also judge when a passage is simply unintelligible, and decide what to do about it rather than letting the system guess.
Speakers, punctuation and structure
Automated output may merge speakers, attribute lines to the wrong person or place sentence breaks in the wrong spot. Automatic punctuation is often plausible but misleading, and a misplaced full stop or question mark can change how a line reads. A person decides where a subtitle should break so the text reads naturally, which is a matter of style rather than raw recognition.
Condensing for reading
Spoken language is faster than most people read. Subtitles therefore usually condense speech, and doing that well means deciding what matters. Verbatim transcripts are often too long for comfortable reading speed. Deciding what can be dropped without losing the point is a judgment call that draws on understanding of the content and the audience.
Translation and culture
Machine output can be fluent and still miss intent, register or cultural meaning. Idioms, humor, formality and references need a reviewer who understands both the source and the target language. A literal rendering may be offensive, confusing or simply flat. Reviewing machine-translated subtitles with a bilingual reader is a common way to balance cost and quality.
Sensitive material
Where an error could cause harm, such as medical instructions, legal notices, safety warnings or emergency information, a qualified human reviewer should check the text before release. This is not a matter of distrust in any one tool; it is that the consequences of a single wrong word or number can be serious and a fluent error is hard to notice. The same applies to personal or distressing content, where tone and wording matter beyond literal accuracy. Our wider discussion of the limits of automatic subtitling for sensitive material covers this in more detail.
Non-speech information
Deciding that a sound is meaningful, such as a door slam that alerts a character or music that sets the mood, takes judgment. Systems may omit such cues or label them generically. For viewers who are deaf or hard of hearing, this information is part of the content, and a reviewer decides what to include.
Fitting review to the risk
Not every video needs the same level of checking. A sensible approach is to match the review to the stakes:
- Low stakes, short life: a quick read-through by one person, focused on names and obvious errors.
- Moderate stakes: a full pass for text, timing and consistency, using a checklist.
- High stakes: review by someone with subject knowledge, plus a second independent check.
Using confidence information from a tool can help allocate effort by showing which words to look at first, but it should be used as a guide rather than a guarantee, since confident errors do happen.
How people and tools work best together
The most effective setups treat the tool as a first drafter and the person as an editor. Tips from teams that do this well:
- Give the editor the source materials and glossary up front.
- Let the editor see the video and audio, not just the text.
- Measure how much correction the drafts need over time, for example using word error rate on a sample, to decide whether the tool is working for your kind of content. Remember that a single measure does not show everything.
- Use reviewer corrections to improve vocabulary lists and style rules, which can raise the quality of later drafts.
No review is perfect either
Humans make mistakes too, especially when tired or rushed, which is why many teams add a checklist and a second reviewer. The point is not that people are always right and machines always wrong. It is that they fail in different ways, and combining them covers more of the gaps than either does alone. Where the stakes are high, the sensible assumption is that automated subtitles are a draft until a qualified person has checked them.