[Interview] Fast, Lightweight and On-Device AI: How Samsung Research Built AI Features That Translate in Real Time

Samsung Electronics’ Galaxy AI-enabled mobile devices allow users to enjoy seamless, barrier-free communication in even more countries. Now supporting Arabic, Indonesian and Russian, Galaxy AI’s Interpreter and Live Translate features have expanded from 13 to 16 available languages.

 

Samsung Research combined data and cutting-edge technology. Together with the Mobile eXperience (MX) Business R&D Office, they further honed this technology to develop the translation features powered by on-device AI — which can be used for real-time translation during calls and across various applications. Samsung Newsroom met with Yoonjung Choi and Yonghyun Ryu from Samsung Research’s Global AI Center to learn more about these ambitious features.

 

 

 

Smooth Communication and Strong Security With On-Device AI

On-device AI is the key differentiator to Galaxy AI’s Interpreter and Live Translate features — introduced to users through the release of the Galaxy S24 series earlier this year. Leveraging the advanced computing resources built into the devices themselves, smartphones with on-device AI can provide services without relying on servers or the cloud. Users can be assured that their data will remain private and secure since information is not shared with external sources.

 

Samsung Research’s Global AI Center contributed to integrating proprietary technology into its AI translation model to commercialize these features for widespread use.

 

Envisioning a wide range of applications, the Samsung Research team and MX R&D Office obtained an expansive amount of data. “We collected colloquial data for real-time translation during calls in Live Translate and travel-related data for Interpreter,” explained Yoonjung Choi, who led the project. “To provide the most accurate translation, we studied and incorporated casual language used in chatrooms and HTML tags used in web browsers.”

 

 

 

How Samsung Research Trained Its AI Model

The Samsung Research team’s AI translation model is based on deep learning technology that learns from its own data. Yonghyun Ryu, who is in charge of AI research and development, likened this process to raising a child. “Similar to how a child needs excellent educational resources and caregivers to grow and thrive, good language data and talented researchers are required when developing a high-performance AI translation model,” he described.

 

Samsung Research has both — since 2013, the company has been providing in-house translation services, conducting R&D related to AI translation and accumulating high-quality data.

 

Samsung Research’s team of deep learning experts played an important role in training the AI model. “If incorrect translations occur during the research and development process, it is necessary to identify the problem and make improvements. However, this can be challenging and time-consuming for researchers without sufficient capabilities and experience,” he explained. “Our researchers used their expertise and know-how to quickly analyze the cause of the issue and come up with a solution to enhance the AI translation model.”

 

 

To assess the performance of Galaxy AI’s translations, the Samsung Research team used quantitative metrics based on test sets as well as qualitative evaluations by human translators and the MX R&D Office.

 

In addition, the team gained credibility by competing in global machine translation competitions. “Although participation requires time and effort, good performance in competitions provides momentum for research and development,” Ryu emphasized. “We were able to achieve strong results because our team members could freely discuss new ideas and put them to the test.”

 

 

Politeness and Punctuation: Understanding the Quirks of Each Language

Each language has characteristics that are unique to its culture — these can include honorifics, tonal inflections and distinct punctuation symbols. To make translations as accurate as possible, the AI translation model considers all these linguistic idiosyncrasies. For example, in Korean and Japanese, honorifics are translated to maintain a respectful tone.

 

Samsung collaborated with regional R&D centers to fully understand languages in their cultural context. “By working closely with researchers and linguists in other countries, we were able to offer a more accurate and complete translation,” said Choi.

 

At the same time, dealing with different languages often involves trial and error. “Vietnamese, for instance, is a tonal language. However, we realized during the research process that Vietnamese users often omit tones in casual conversations when chatting,” she explained. “We needed additional data to help the features translate sentences without tones.” For Thai, Samsung Research developed a special sentence separator because the language does not use punctuation.

 

 

Why Lightweight Technology Is the Key to Effective On-Device AI Models

 

 

Samsung Research began developing its on-device AI translation model in 2019. “Unlike server-based AI models, on-device AI models must be driven only using users’ devices,” stated Ryu. “Developing lightweight technology that uses minimal resources is key.” To make the model lighter, the team used “knowledge distillation” and “quantization” technology.

 

Knowledge distillation is a method that extracts knowledge from a large, high-performing teacher model and delivers it to a smaller student model. This is similar to how a teacher summarizes a topic so that a student can digest the information more easily and efficiently.

 

Quantization simplifies AI algorithms to reduce model size and streamlines the process to increase response speed.

 

 

Ryu compared quantization to drawing strawberries. “You need a wide range of colors to paint lifelike strawberries — but strawberries can also be made with just red and green,” he explained. “Quantization is the process of minimizing the number of colors needed to draw strawberries while trying to make it as close as possible to the real thing.”

 

 

In the AI field, knowledge distillation and quantization are well-known approaches to making models lighter. Yet, implementing them on a commercial scale is not easy due to differences in each researcher’s detailed experimental methods and factors. Samsung Research developed proprietary technology by discovering an efficient quantization technique and creating a accelerated algorithm based on it. “Through constant experimentation, we found an optimal way to make the model lighter,” commented Choi.

 

By combining the high-quality AI translation model with algorithms that make models lighter and speed up response time, lightweight and fast on-device AI features were born.

 

 

The Culture Behind the Language: What Would the Perfect AI Translation Model Look Like?

The researchers at Samsung Research’s Global AI Center have bright goals as they lead the field of on-device AI. “My ultimate goal is to help users communicate smoothly and conveniently with people who speak other languages,” said Choi. Ryu revealed his vision to build the perfect translator. “One day, we want to create a translator that truly understands the cultures behind the languages it is translating, equipped with an extensive pool of knowledge,” he expressed. “I want to challenge myself to develop a translator the world has never seen before.”

 

 

Samsung’s on-device AI translation features allow anyone with an enabled mobile device to communicate freely — without worrying about internet connection or information leakage. Going forward, the Samsung Research team will continue to spearhead innovations in the rapidly evolving field of AI and bring new levels of convenience to users.