Free AI Caption Generators Can Subtitle Your Film, but at What Cost?
Updated: 4 days ago
Not long ago I got a call from a prospective client who needed subtitles in English and Spanish for her film, spoken in Portuguese. She said something like, "Do you work with AI? With the wonders it can do nowadays, I was wondering if you use it to subtitle."
And I get this all the time. Some clients provide a version of the subtitles and reach out to me to polish them. How did they do them? Some will (flat-out lie and) say a person was responsible. To the trained eye, though, it is obvious that this was not the case. How to train the eye, you might ask? I have made a visual comparison to help you see exactly what you are getting when you use free AI tools to generate and translate subtitles.
Contents
Free Subtitles = Cost-Effective?
Yes, the words "free," "fast," or "cost-effective" are almost irresistible. So, it is more than understandable that when we are offered a tool that promises exactly what we need, at a lower price and at a faster pace, we will be inclined to adopt it. Especially in an expensive field like the film industry. And especially when it comes to something that, more often than not, is treated as an afterthought: the subtitles.
A film will most likely need captions (the transcription of all audio in its original language) for accessibility and subtitles (the translation of the spoken dialogue into other languages) for foreign viewers—mind you, this includes investors (key during the WIP phase), studios, festival programmers, and, of course, filmgoers. They usually comes at the end of the filmmaking process (although it not always should). By then, budgets are tight, deadlines are looming, and everyone is understandably looking for ways to save time and money.
What most people miss when it comes to the process is that subtitles need to be translated, yes, but also adapted, segmented, timed, proofread and read well on screen. They have to work within very specific constraints: characters per line (CPL), characters per second (CPS), shot changes, reading speed, rhythm, tone and context. And sometimes, the biggest challenge is figuring out what is actually being said in the first place. It takes skill, practice, and common sense.
AI-Generated Captions vs. Handmade Subtitles: The Experiment
For this AI-vs-handmade experiment, I used a 1-minute clip from the Argentine film Espejos rotos (2020), directed by Christian Mouroux and Edgardo Bechara El Khouri. The documentary follows the journey of two tango-singing immigrants in Buenos Aires: a Swiss man (El Tigre Tanguero) and a German woman (Agnes). Although they speak Spanish quite fluently, they still have an accent and interference from their native language. However, to any Argentine, it is perfectly easy to understand what they mean.
In the selected scene, Agnes is back in Hamburg, Germany, recalling a time when she came to Buenos Aires, and shots of her speaking to the camera alternate with those of a man reciting poetry in German.
The AI-Generated Subtitles
I worked on these subtitles back in the day, and now I needed to see what AI would do with them. So, I:
Uploaded the clip to YouTube (unlisted!).
Selected to create automatic captions.
Downloaded the resulting srt file.
Opened the file with Notepad and copied the contents.
Pasted the text on Deepl.
Voilà, I had my English version:
💭 What do you think?
The Human-Made Subtitles
To make them, I worked with industry-standard guidelines. To many in the film production pipeline, this might be unfamiliar territory, so allow me to give you a brief refresher.
Good subtitles are those that are easy to read and capture the meaning. They also do so within a limited number of characters per line (CPL) and per second (CPS). These limitations are not arbitrary; they are the result of numerous studies that have determined the optimal settings for people to enjoy a movie without feeling like them are interfering with their experience.
This should go without saying, but they are also well translated, in other words, the translation is accurate and sounds natural in the target language.
I was dealing with the standard challenges (how am I going to translate this? How am I going to do it respecting the CPL and CPS constraints?). But I had an added layer of difficulty: somebody on screen was speaking—actually, make that "reciting a poem that doesn't have an official translation"— German, which I wasn't familiar with. So, I did the obvious thing: I called a friend. I recruited the help of a German translator for those sections and left the rest up to experience and good judgment.
Here is the result:
Want to see the film? 📺 Grab a coffee and hit play here.
How Did the AI-Generated and Translated Subtitles Fare?
You may have already caught some of the obvious shortcomings of the AI-generated subtitles, but just in case, here they are, explained:
Poor segmentation. The sentences are cut at awkward points, making them harder to read and disrupting the natural flow of the dialogue.
AI failed to pick up the other language. AI did not correctly identify or understand the passages in German, so they were left untranslated.
Incorrect CPS. Some segments contain too many characters for the amount of time they remain on screen, exceeding comfortable reading speeds.
Incorrect CPL. Some lines are simply too long, taking up so much of the screen that the audience actually needs to move their head to read them.
Changes in meaning 😲. A few examples:
“A flyer for” → “advertising”. A seemingly small lexical choice, but one that changes what the person is actually referring to.
“Wrote a play” → “Made this piece”. The AI selected a word that was wrong in context.
“Let’s do it that way”. The Spanish sentence may allow the intended meaning to be inferred from context, but the English subtitle needs to make that meaning explicit. A literal rendering does not necessarily work.
“Book adaptation” → “cover”. The AI misinterprets what the speaker is referring to, producing a completely different meaning.
Grammatical errors in English. At one point, the subtitle reads, “(...) who when he went…” The problem is that 'who' already works as the subject, making the resulting construction ungrammatical.
Missing capitalization of a formal designation. “Guest of Honor” was not capitalized correctly when referring to the formal designation of the country during the Book Fair.
What Does a Bad Subtitle Actually Cost?
If you think subtitles like these are only used for demonstration purposes, think again. I have seen this quality in films submitted to international film festivals or to film grants. That’s when this stops being a harmless experiment with a new tool. These are often what the film that programmers, distributors and international audiences encounter first.
So, what does a bad subtitle actually do?
At the very least, it gives the audience a bad experience. People may not notice a subtle change in meaning, but they notice a mistake in their language or a couple of seconds with no lines. Not only are these things distracting (they pull the audience's attention away from the film), they can be, frankly, quite annoying.
If a programmer catches the problems and asks for changes, they will most likely ask a professional to retranslate, resegment, retime and QC the entire film. The money they thought they had simply moved the cost further down the line, and potentially added another round of work on top of everything. And that’s before we even consider the less tangible cost: a missed opportunity, a delayed delivery, or a poor first impression.
When so much has already been invested in making a movie, is saving the cost of proper subtitles really where to cut corners?
Are Your Subtitles Ready?
After years of working as a subtitler and as a Translation & Subtitling Manager for international film festivals, I’ve seen firsthand how much difference a good set of subtitles can make. I’ve also seen what happens when they don’t quite make the cut.
So, if you’re working on a film and wondering whether yours are really ready to go, I’m always happy to take a look. Sometimes a second pair of trained eyes is all it takes to catch what a machine, or a very busy filmmaker, might miss.
Comments