Can you translate a book with DeepL?

The simple answer, of course, is yes, you can, but can it also save you time and money? I tested it out.

Recently, I had the unexpected opportunity to actually test whether using DeepL (an online machine translation tool like Google Translate) can save an author or publisher time and money to translate a novel. The myth that all you have to do is run a book through DeepL and then have it edited by a native speaker has become naggingly persistent!

As it happened, in the spring of 2022, I found myself in a situation where a publisher hired me to work on three books from a series of crime novels by the same author. All three books had been, or were to be, translated from English into German.

• Book no. 1 was a translation by a computer. The translator had pulled a fast one on the publisher and simply submitted a machine translation of the book. When I was assigned to edit the translation, the publisher and I had no idea this was the case.
• Book no. 2 was a good translation done by an actual literary translator.
• Book no. 3 was to be translated by me.

I used this opportunity to keep track of the total time I spent working on these three books. And the results were quite eye-opening, even for me!

Disclaimer: This post is about the translation of novels, NOT operating manuals, business letters, or legal notices like data privacy statements. Those are kinds of text that DeepL can handle quite well, because they consist mainly of blocks of boilerplate text passages. Novels, on the other hand, are a completely different beast. This post is just about novels.

Editing the human translation

I began my test with book no. 2. The translation was the usual high quality I have come to expect from this publisher. But even a good translation of a novel needs editing, because two sets of eyes always see more than just one, and an editor can bring a more objective perspective to any piece. As a translator, at some point you become blind to your own work and no longer see your own mistakes. As an editor, I can read the translated book like a normal reader, and whenever something trips me up, I can smooth it out, which sometimes means I have to refer back to the original. Common things I have to watch out for are borrowed words, false friends, the correct use of tenses, idioms, puns, connotations, repetitive word choices, etc.

I had one month to edit this translation and, as always, I broke up the work into a manageable schedule. I set myself a daily workload of 50 pages a day, which equaled about three hours in total each day. Including a second read-through of my edits, this 480-page book took me around 40 hours of work.

Nightmare editing of the computer translation

And then I came to book no. 1. Here, as I already mentioned, the translator had simply turned in a computer translation. In the beginning he had made some attempt to cover up his use of machine translation by changing a word here and there, but as the story went on, it became clear he had stopped putting in even that little amount of effort. But I had agreed to do the editing, and thankfully the publisher tripled my pay, so I had little choice but to grind out this nightmare.

For this editing task, my daily workload was 20 pages – I simply could not stand to do any more than that. I started work around 8 a.m. every day and usually sat at my desk, slogging through this nightmare editing, until around 4 p.m. Not counting breaks, I managed only about three pages an hour – less than a fifth of the sixteen pages an hour while editing the good translation. So far, this sounds like I’m just comparing the editing of a good translation to that of a bad translation; of course editing a bad translation will take longer! But even more than the time, which I was paid for, this nightmare editing cost me in terms of my energy and sanity, which no one can ever compensate me for. The main things I learned were:

• Machine translation does not save you time in terms of typing. This is perhaps the biggest misconception that people have with computer translation: editing a computer translation and not translating it yourself will save you time because you won’t have to type as much. All the words are already there. You just have to edit them and almost nothing will need to be retyped at all. Well, I can assure you that that is quite a naive understanding of the situation, and only proves that anyone who believes that has never tried it themself.

For one thing, I’m a pretty fast typist. And I can also type without looking. When I translate, I read the section I’m working on in the original and at the same time type the German translation. Reading and writing is a simultaneous process for me. Sometimes I don’t even use a keyboard, I dictate. Then I’ll read the section in the original and at the same time dictate German to the computer. So I waste practically no time at all in typing, because typing and reading happen practically simultaneously, and I always have to read the original no matter what.

For another, when working with a DeepL translation, I have to read both versions: the original and the translation. First I read the translation to see how good it is, then the original to determine whether the translation captured the gist of the original message. This process has already taken up twice the amount of time as translating the original section from scratch. Even if the DeepL translation was perfect, which is almost never the case, I would spend more time working with a DeepL translation than without. And I still haven’t even gotten to the point of correcting or editing the translation.

• Mistakes are overlooked. This is probably the most insidious thing about DeepL: the translation of each sentence reads grammatically correctly, and in most cases is even logical by itself. This means it’s very easy to overlook mistakes. During my nightmare editing, for example, several errors came to light only after I had read through my revisions for the third time. One example comes to mind in particular: the English word boot. According to the nightmare computer translation, a man took two flashlights out of his Stiefel (boot). Sure, I thought, why not? Sounds plausible! Why should I judge where other people keep their flashlights?! Only during a later read-through did I wonder why the man hadn’t put on his boots afterward. And that’s when it hit me: the German translation should be Kofferraum (vehicle boot in British English, or trunk in American), not Stiefel! The man had taken the flashlights out of the trunk of his car, not his footwear! I was only able to find such mistakes because I have a very good command of both the source and target languages. A native speaker of the target language, whose only qualification is their native language and who is merely supposed to polish the DeepL translation, would never discover a mistake like that.

• Translating a novel is not about simply translating each word or phrase from one language into another. A novel is not written and translated for informational purposes, but to entertain. To arouse feelings in the reader: suspense, love, fear… all depending on the genre and story, of course. And for the reader to feel those feelings, the author has to be able to convey them. And to be able to convey those feelings, the author has to feel them. he same is true for the translator. She reads the original, feels the emotion and action in the story, and crafts her translation so that the reader feels the same way. While I was editing the DeepL translation, I felt absolutely nothing – nothing except anger and frustration, which had more to do with the translation itself and very little to do with the original story. Admittedly, there were also quite a few times I burst out laughing, but verbal diarrhea like “Ein schrecklicher Gedanke traf sie so stark wie der Vogel, der vom Dach flog und ihr die Haare schnitt” (basically meaning: A terrible thought hit her as hard as a hairdresser-bird) is both maddening and maddeningly hilarious. My favorite machine-translated sentence was probably “Her temper boiled just below the surface of indignation” (“Ihr Gemüt kochte knapp unterhalb der Oberfläche der Empörung”), which was actually a pretty accurate description of how I felt while editing the translation. But suspense, love, fear? Worry that the hero’s sister had been abducted? I didn’t feel any of those things. And that’s probably why I wasn’t able to convey those feelings very well either.

• I was never in a state of flow. Authors know the feeling, when they’re writing their book and the rest of the world seems to fall away: they’re in a rhythm where their full concentration is focused on one task. They can see the action and plot of their story in front of them, and the words seem to type themselves into the computer. They are feeling what they are writing. And then, if someone interrupts them, they’re torn out of that state and often have to struggle to find their way back after the disruption goes away. I experience the same thing when translating novels. But not when editing DeepL’s output. I was so often torn away from the story that I couldn’t get into any sort of flow. There was an obstacle in every single sentence. I never got to read and experience an entire scene; I had to edit every single sentence. Try to imagine what it might feel like to be writing your own story and then to be interrupted during each and every sentence. If you can imagine that, you can perhaps begin to understand the frustration of editing a DeepL translation, and how poor the results are when you can’t work with your full concentration.

Let’s take a look at some of the completely practical weaknesses of DeepL & Co.:

• Quotation marks are often eliminated for some reason, but frustratingly not in any consistent way, which makes editing even more time-consuming. Often you have to not only change the style of the quotation marks and reinsert them, but you also have to go back to the original to see where they belong in the first place. That’s really annoying.
• Acronyms are usually not translated. The novel in question had constant mentions of SOCO and FLO, two acronyms that German readers would almost certainly not understand. So I had to research what those acronyms meant and then also research an appropriate German term to translate them. It was easy enough to find a German counterpart to SOCO (scene-of-crime officer), but there was no simple German equivalent to FLO – Family Liaison Officer. To be clear, finding a German equivalent is my job as the translator – but the time required to do so is the same when using a DeepL translation and when producing a new translation myself.
• The difference between the German familiar (“Du”) and polite (“Sie”) addressing of ‘you’ is always a problem. As a translator into German, I always have to decide whether to use the polite or familiar form of ‘you’ when someone is being addressed in dialogue – and then, of course, I have to make sure to be consistent, so that a character who uses the familiar form in one chapter does not suddenly switch to the polite form in a later chapter. DeepL does not pay attention to this problem at all. Sometimes it will use the familiar form and sometimes the polite, seemingly at random. And again, this problem must be corrected, which takes time. Time that DeepL was actually supposed to save you.
• DeepL does not pay any attention to the multiple meanings of words, and almost all words have multiple meanings. Sometimes the differences in meaning are slight, and sometimes merely stylistic – but sometimes the differences can be quite significant. Consider, as an example, the English sentence “Call in the station in the morning,” which was translated by DeepL into German as “Rufen Sie morgens beim Bahnhof an.” There are multiple mistakes in just this one sentence: 1. The German “anrufen” (call on the phone) is not correct; the person is supposed to stop by the station in person, not call on the phone (to callhas both meanings in English and must be translated according to context). 2. The German needs to convey not just “in the morning,” but “sometime before noon the next day” (this is implied in the colloquial use of English, but a German speaker would spell it out more explicitly. A similar situation is true for the English night, which has come to be used for both evening and nighttime in English and is often translated as the German Nacht, but often corresponds more to the German Abend, the time of a day around sunset). 3. Here, station is not a Bahnhof (train station), but a police station. Okay, and here’s a bonus fourth mistake: the person is supposed to stop by the station only once, the next day, and not every morning (a common use of morgens in German would suggest an action done every morning). Those are all problems that I am aware of as a professional linguist. A computer is not. And a native speaker of the target language, whose only qualification is being a native speaker, is not aware of the nuances in the source language. And, again, it took me longer to correct and rewrite this sentence than it would have taken me to produce a new translation from scratch.
• Other mistakes include: consistency (the same word is translated one way in one location and another way in a different location, which sometimes doesn’t matter, but sometimes means a connection is lost); wordplay; allusions (bunny-boiling bitch will be understood only if the reader has seen the movie Fatal Attraction); mismatched styles; local references; ambiguities; gender (is lawyer, cop, murdererreferring to a man or a woman? In English, we have to get that information from context, but in German, every noun has to explicitly state its gender); descriptive names; and many others. Those are the things that take time to translate, not typing. And DeepL doesn’t give you any help with those things.

Can you save money by using DeepL?

It’s a perfectly legitimate question, but in order to save money, you need to be able to save somewhere else too: time or qualifications, for example.

• DeepL does not save time. I have already mentioned this fact several times. I still have not talked about my new translation of book no. 3, which provided a wonderful opportunity to compare the time requirements for the different tasks. And why should I charge less for revising a DeepL translation than for a new translation, when the revision takes just as long, or even longer?
• DeepL still requires a qualified professional.An experienced translator who knows both languages intimately is still needed for revising a machine translation. An unqualified native speaker would miss many mistakes, which means that the translation would still be bad even after it was revised. And a qualified professional costs more than an amateur.
• DeepL is free. That is actually the biggest misconception: Why should I, as a translator, charge less money for editing a computer translation than for translating a book myself? DeepL is (practically) free! You can translate a whole book for just $9 a month! If using DeepL helped my translating work even in the slightest, I would happily pay that low monthly fee. But why on Earth should I agree to give up 30, 50, or even 70% of the cost of a new translation, representing hundreds of dollars, just because the author provides me with a computer translation that cost at most $9? It makes no sense.

Translating from scratch

So now we come to book no. 3, which I translated myself from beginning to end. As usual, I started around 8 a.m every morning and set myself a daily workload of 20 pages, the same number of pages as for the nightmare editing. And every day I was done by noon at the latest. Thus, I managed to work at a rate of 5 pages per hour. Therefore, producing a new translation was about twice as fast as editing the DeepL translation.

To be completely forthcoming, it should be noted that this was the time needed just to produce a first draft. My translation still required editing, of course, meaning it needed extra time similar to the amount of time for editing the good translation of book no. 2. But this work was needed both for editing the new translation and for the nightmare editing of the computer translation.

Time for the actual results:

All three books were around 480 pages.

For editing the good translation, my work amounted to:

• 16 pages per hour, for a total of 30 hours
• plus reviewing my edits and running the text through the Duden spell checker: 10 hours

Grand total: 40 hours of work

For the nightmare editing, my work amounted to:

• 3 pages per hour, for a total of 160 hours
• plus my own editing of the text, with time between revisions: 20 pages per hour, for a total of 24 hours

Grand total: 186 hours of work

For the new translation, my work amounted to:

• 5 pages per hour, for a total of 96 hours
• plus my own editing of the text, with time between revisions: 20 pages per hour, for a total of 24 hours
• plus editing by an editor: 16 pages per hour, for a total of 30 hours

Grand total: 150 hours of work

Conclusion: Using DeepL for translating novels is not worth it. It does not save you any time – and the results are worse. A clear case of a lose-lose situation, if you ask me!

This post was translated into English by William Knapton. The German original can be found here.

« Editing of translated novels How is the translator to be named in the book? »