The following article is an excerpt from an article published in Fortune Korea. It references a movie, Snowpiercer, a 2013 film directed by Bong Joon-ho, the director of Parasite. In the movie, a character, who speaks only in Korean, is able to communicate with the others via a two-way, voice-activated translation machine.
On January 9th, 2023, around 500 Korean entrepreneurs and IT engineers gathered in Silicon Valley for the "82 Startup Summit." This event, named after Korea's country code, was filled with innovators who are changing the game in their respective industries. One of these innovators is Tim Jung, CEO of XL8, a startup that is vying for dominance in the field of machine translation, surpassing Google in accuracy.
Bringing The Machine Translator in the movie into reality
There is a Korean entrepreneur who is working to break down language barriers, with the help of AI. This person, Tim Jung, the CEO of XL8, says that in addition to the current slate of products they offer, their live, handheld speech to speech translation service will be available for paid use within a few years. Before starting XL8, Tim was a tech lead manager at Google.
At CES 2023 in Las Vegas, there were more than 3000 companies in attendance. The more I heard fluent English conversation, the more I squeezed the smartphone in my pocket. Making sure that the mobile app was on, I headed to the scheduled business meeting. I had a chance to try out one of XL8's products, the "TravelCAT" app. The app, a machine translation tool, allowed me to communicate with ease in a different language. If you press the microphone button at the bottom of the screen and speak, it will be translated into the targeted language within seconds and provided in text and voice. There was no problem with communicating in a space with low background noise.
XL8, founded in 2019 by Tim Jung, is the only machine translation company that specializes in media and entertainment. Tim, who had worked at Google as a TLM (Tech Lead Manager) after working in the search team and event search saw the potential for machine translation in the field of media content. While Google's machine translation relied on data scraped from the web, XL8 was able to create high-quality datasets in real time by working with data from thousands of human translators. Tim says that the translation results are thoroughly reviewed by humans through multiple quality control steps, which leads to a hand-curated, high-quality dataset and improved translation performance.
XL8 is currently in the process of raising its Series A funding round, and has already partnered with Iyuno, a leading translation service provider known for their high-quality work. With their translation expertise, XL8 is looking to move into real-time subtitle generation for events such as sports and concerts.
When asked about the future of machine translation, Tim predicts that within a few years, we could see accurate, real-time translators like those portrayed in movies, become a reality. It is an exciting time for the development of machine translation technology.
What will be the diverging point where quality becomes more important than quantity?
Tim : It takes about 10 million sentences to train one language pair. Once you get that much data, it is worth creating a new language engine.
I think Google already knows the importance of data quality. (Why didn't they jump into this domain?)
Tim : I believe there has been no good way for them to get a super high-quality dataset. However, we have partnerships with leading LSPs. Our business model provides the LSP with tools to increase work efficiency, and we get an opportunity to improve the quality of our machine translation. It's a difficult thing for big companies to do. I'm 100% sure Google does better web translation. We specialize in media translation, which is centered on colloquial language.
Along with the quality of data, the performance of the translation engine is also important, I believe.
Tim : We have a technique specialized for media translation. Machine translation usually takes a sentence and thinks about how to translate it. If you understand the sentence by reflecting the context before and after, the accuracy of the translation will increase. This is called the 'Context Awareness (CA) technique'. It was more effective when translating colloquial words than when applying it to documents. There is emotion and nuance among people, and the next person can change the direction of the conversation.
Tim pointed out an example of applying the CA technique. To translate “I like this flower. Put it in the bag” into French, note the pronoun “it,” in the second sentence that refers to the flower in the previous sentence. Unlike English, flower is a feminine noun in French. Considering the context of the previous sentence, the feminine pronoun ‘la’ should be used instead of the masculine pronoun ‘le’. XL8's translation engine returns, 'J'aime cette fleur. Mets-la dans le sac’, but Google's translation engine translates it as 'Mettez-le dans le sac'.
What other technologies were applied?
Tim : Our engines employ multimodal techniques to consider additional forms of information to aid in the translation process. For example, translating a description of an image is more accurate when incorporating the image itself. The engine takes into account information relevant to each language. When there are multiple speakers in a video, factors such as age, gender, and relationships are also reflected in the translation to better convey the situation or relationship between individuals. Developing a technology specialized for colloquial speech has been effective.
How did you get chosen to provide MT for the largest media LSP?
Tim : The process of reaching an agreement was not that difficult. In particular, when they took over a company in the US in 2021, the acquired company was using a machine translation solution. They tested 20 language pairs, and we performed well for every language pair. That's why they chose us the day after the results came out.
I have heard that collaboration offers are now coming in from several LSPs.
Tim : There is a strategic aspect to working with XL8. LSPs have no choice but to think about the future. Translation is a business that people do by hand, and there will be limits to using an external AI engine. However, self-development is not easy. If so, it is worth thinking about ways to grow together by investing in equity in our company. If we receive support from multiple LSPs in a neutral way, it will be easier for us to grow our businesses and make them more profitable.
There is a real-time translator portrayed in the movie ‘Snowpiercer’. While using TravelCAT, I remembered it.
Tim : Eventually, a device like that, where you press it to your throat and a translated voice is generated, will become a reality. It's colloquial conversation, so our engine would be the best fit. If more than 9 out of 10 sentences are reliably translated by a device, it could be commercialized. We come close to that point now with our MT. In 1-2 years, shouldn’t we be able to translate daily conversations without problems?
At the level of everyday conversation, what is your percentage of accuracy?
Tim : Approximately 95%. In business of course, you have to be more precise. However, accuracy is important, but in real-time, I mean how quickly the translation is done is also important. Now, it is translated only after the words are finished. In the Netflix original ‘Space Sweepers’, people can speak to each other at the same time. Accuracy and real-time are a trade-off, so you need to strike a good balance.
How many seconds does your engine currently take?
Tim : There is a 7 second version and there is a 40 second version, the 40 second version is more accurate. But in the future, I think there will be an engine that can be used in real time with only a difference of 4 to 5 seconds.
Are you also inspired by sci-fi works?
Tim : Certainly. There is a novel written by Andy Weir, the original author of the movie ‘The Martian’ called ‘Project Hail Mary’ (2021). There is a scene where people try to translate the language of the aliens. Aliens speak in harmony with a few tones. In terms of recognizing chords and researching how to handle them, the historical evidence was well proved as the author was an engineer. While reading the work, I thought that people always suffer from language problems, and that technology can solve it.
It is attractive that your business is not limited to one country.
Tim : Translation is a technology that cannot be dependent on one country. Humanity has been thinking about it for thousands of years, and now useful tools are starting to come out.
What do you think of director Bong Joon-ho's words, "The Oscars are not an international film festival. They’re very local.” There are times when the saying, “Korean startups must enter Silicon Valley” is like a slogan in the era of industrialization.
Tim : Although we are doing well as a startup, we need to continue to survive and thrive. Being global, inclusive and discerning especially applies to recruitment. There's no rule that we should hire Koreans because I'm Korean. A successful company should develop a team that can do their best work together. Looking at our results, I am pleased to say that our current team has done a very good job.
Even though you have been busy, you became the President of Bay Area K Group(BAKG).
Tim : I did BAKG activities for a long time. Soon After moving to California in 2015, I took a role as an executive member at BAKG. After working with a small group management team for about 6 years, I became the President. Initially, I was a bit concerned about taking a role because XL8 was still in its infancy and required a lot of time, but it was nice to be able to meet people and make important connections. I am most proud of the coding school, where people ended up getting good jobs because of their experience. Whether revitalizing their skills , or starting from scratch, many ended up as engineers in Silicon Valley.
Afterwards, I met CEO Jung again at the ‘82 Startup Summit’ event on January 9th. Standing on the podium, he introduced XL8's business model in front of hundreds of engineers and international students. Like the essence of translation, which encourages the dissemination of knowledge, he was also busy disseminating his experiences and insights to his colleagues and juniors.